sensors

Article
Design and Analysis of an Upper Limb Rehabilitation Robot
Based on Multimodal Control
Hang Ren 1 , Tongyou Liu 2 and Jinwu Wang 1,2, *

1 School of Health Science and Engineering, University of Shanghai for Science and Technology,
Shanghai 200000, China; [email protected]
2 School of Biomedical Engineering, Shanghai Jiao Tong University, Shanghai 201100, China;
[email protected]
* Correspondence: [email protected]

Abstract: To address the rehabilitation needs of upper limb hemiplegic patients in various stages
of recovery, streamline the workload of rehabilitation professionals, and provide data visualization,
our research team designed a six-degree-of-freedom upper limb exoskeleton rehabilitation robot
inspired by the human upper limb’s structure. We also developed an eight-channel synchronized
signal acquisition system for capturing surface electromyography (sEMG) signals and elbow joint
angle data. Utilizing Solidworks, we modeled the robot with a focus on modularity, and conducted
structural and kinematic analyses. To predict the elbow joint angles, we employed a back propagation
neural network (BPNN). We introduced three training modes: a PID control, bilateral control, and
active control, each tailored to different phases of the rehabilitation process. Our experimental
results demonstrated a strong linear regression relationship between the predicted reference values
and the actual elbow joint angles, with an R-squared value of 94.41% and an average error of four
degrees. Furthermore, these results validated the increased stability of our model and addressed
issues related to the size and single-mode limitations of upper limb rehabilitation robots. This work
lays the theoretical foundation for future model enhancements and further research in the field
of rehabilitation.

Keywords: upper limb rehabilitation robot; sEMG; kinematics analysis; joint angle

Citation: Ren, H.; Liu, T.; Wang, J.

Design and Analysis of an Upper
Limb Rehabilitation Robot Based on 1. Introduction
Multimodal Control. Sensors 2023, 23, The goal of rehabilitation is to improve deficits in motor control in patients through
8801. https://doi.org/10.3390/ strengthening exercises for the affected limb. Traditional rehabilitation therapy relies on
s23218801 one-on-one training with a rehabilitator focusing on the relearning of lost or weakened
Academic Editor: Enrico Meli functional movements [1], which can increase staff costs and is susceptible to human
factors. However, the integration of robotic technology and artificial intelligence has led
Received: 4 September 2023
to a new paradigm in rehabilitation training. In particular, upper limb rehabilitation
Revised: 24 October 2023
robots, combined with industrial robots and AI, have become an important means of
Accepted: 27 October 2023
addressing these challenges. Patients can leverage rehabilitation robots to efficiently finish
Published: 29 October 2023
repeated exercises while collecting real-time data on themselves [2]. Since their introduction
to the medical field in the 1990s, various types of structural and functional upper limb
rehabilitation robots have been developed [3]. These robots can be divided into two types
Copyright: © 2023 by the authors. based on their traction contact: exoskeleton and end-effector robots.
Licensee MDPI, Basel, Switzerland. The end-effector type rehabilitation training robot predominantly employs a series of
This article is an open access article mechanical structures or a linkage mechanism as its core framework. This system is securely
distributed under the terms and affixed to the human arm through the end-effector. Furthermore, it exerts traction force on
conditions of the Creative Commons the extremities of the patient’s limbs, facilitating upper limb rehabilitation training [4–8].
Attribution (CC BY) license (https:// Exoskeleton rehabilitation robots are wearable devices with mechanical structures that
creativecommons.org/licenses/by/ correspond to human joint movements. Unlike end-effector type upper limb rehabilitation
4.0/).

Sensors 2023, 23, 8801. https://doi.org/10.3390/s23218801 https://www.mdpi.com/journal/sensors

Sensors 2023, 23, 8801 2 of 32

robots [9], which have no fixed matching relationship with human joints, exoskeleton reha-
bilitation robots can effectively restrict the compensatory movements of the limb [10–12].
These robots achieve rehabilitation by providing traction and support to the upper limb
hand, wrist, and shoulder joints. Exoskeleton solutions designed for actively guided mo-
tions usually consist of tandem links driven by powered joints. These exoskeletons and
the user are connected through one or more interaction ports [13]. However, they have
coaxial joints and limb joint axes, and their mechanical structure has more contact points
with human joints, increasing the difficulty of human–computer interaction and control
strategies and reducing the diversity of the exoskeleton mechanical structures [14–18].
Despite these challenges, exoskeleton rehabilitation robots are still a focus of research.
The parameters of traditional controllers are fixed and cannot be adapted to changes in
the environment or tailored to the patient’s condition. In their study, Islam et al. introduced
a seven-degree-of-freedom rigid upper limb exoskeleton robot that effectively stabilized
the patient’s rehabilitation movements [19]. However, their control strategy remained
passive, lacking the active involvement of the patients during the intermediate and late
stages of rehabilitation. In a different approach, Paolucci et al. developed the MOTOR
rehabilitation system [20], which allowed patients to perform translational movements
with their hands in a two-dimensional horizontal plane. They incorporated trajectory
planning and gamified training to enhance patient engagement and visualize the training
process. Nevertheless, the system’s limited degrees of freedom restricted its applicability
to horizontal plane movements, and it could not effectively address rehabilitation in the
sagittal plane. Furthermore, Zhang [21] designed a novel four-degree-of-freedom upper
limb rehabilitation robot, enabling training movements for the shoulder and elbow joints.
They also created a virtual reality rehabilitation training system using position sensors
and Unity3D(2021.3.22f1c1). However, the robot’s steel body material hindered mobility
and added to the cost, making it less suitable for daily use. Additionally, the control unit’s
complexity created operational inconveniences and lacked wrist joint training. In summary,
none of these upper limb robots fully met the rehabilitation requirements of the patients at
different stages.
Given the variations in the structure, control mechanisms, and specific treatment
approaches employed in upper limb rehabilitation robots, the accurate provision of torque
tailored to individual patient movements has emerged as a focal point in contemporary
rehabilitation robotics research. The response of users to additional torque necessitates
the precise implementation of algorithms. Achieving precise torque delivery requires
the incorporation of key parameters, such as joint positions, velocities, accelerations, and
kinetic modeling of the impaired limb, into the control system. Presently, upper limb
exoskeleton rehabilitation robots are directed towards addressing the following challenges.
1. Enhancing patient comfort and safety: Traditional gear-driven structures tend to be
bulky. The implementation of wearable rehabilitation robots employing flexible and
elastic materials can reduce the weight of upper limb devices, thereby improving
patient comfort and safety.
2. Diversifying rehabilitation training: Conventional systems often offer a limited range
of motion, primarily focused on anterior movements, which might provide inadequate
stimulation for central nervous system rehabilitation. A dynamic robot controller,
adaptable to the patient’s specific rehabilitation needs and progress, along with the
design of effective training modes and their corresponding hardware implementations,
are critical components of robot-assisted rehabilitation.
3. Addressing joint movement uncertainty: Controlling the movement of specific mus-
cles during the joint movement process of the affected limb can be challenging. When
patients experience reduced grip strength, it’s essential to select a structural model
that is better suited for the arm and hand.
4. Enhancing portability and versatility: Many exoskeleton rehabilitation robots are
hindered by complex structures and limited portability, primarily due to the reliance
on motor technology with inherent power constraints.
Sensors 2023, 23, 8801 3 of 32

To address this, portable multifunctional rehabilitation robots should aim to combine

the advantages of both traditional and portable systems, all while ensuring a proper fit
with the human body’s joint axes. An example of innovation in this direction is showcased
in Harvard University’s work [22], which introduced a lightweight, fully portable, textile-
based, soft inflatable wearable robot. This technology, transparent when unpowered,
not only offers mechanical flexibility but also enables the quantitative assessment of a
user’s unrestricted movement. Such advancements may pave the way for future prosthetic
applications benefiting individuals with disabilities [23]. To enhance the grip strength of
patients’ hands, soft robotic gloves hold the promise of augmenting user autonomy and
freedom. This is achieved through their design featuring a portable belt pouch and an open
palm [24].
The upper limb rehabilitation robot in this study effectively addressed several key
limitations, as follows.
1. Enhancing patient safety by positioning the motor without direct contact with the
body, eliminating the need for a traditional gear-driven structure and thereby reducing
the weight of the upper limb device.
2. Implementing multimodal control to better cater to the rehabilitation requirements of
hemiplegic patients at various stages, ultimately reducing the workload of rehabilita-
tors while enabling easier data acquisition and visualization.
3. Mitigating joint movement uncertainty by implementing multiple safety measures
throughout the entire system, ensuring the patient’s safety from software to hard-
ware levels.
4. Embracing a modular design for simplified disassembly and maintenance, with
3D printed components to enhance portability and provide a foundation for future
personalized design enhancements.
The rehabilitation process can be categorized into three primary phases, which depend
on the time elapsed since the traumatic event. These phases include the acute, subacute,
and chronic phases. During each of these phases, rehabilitation therapy should facilitate
the gradual recovery of the range of motion and muscle strength in the injured limb. The
control strategy employed by the robotic system should be adapted to the specific needs of
each phase. In the initial phase, for instance, when the patient has lost a significant portion
of arm function, the robot should assist the patient in following a predetermined trajectory
to enhance limb mobility and mitigate muscle atrophy or tendon retraction.
Traditional rehabilitation strategies have typically prioritized proximal limb joint
rehabilitation, as nerve recovery often progresses from the proximal limb segment to the
distal limb segment. However, this approach may limit opportunities for distal limb control
using synapses. Bilateral coordinated movement of the upper limb is an effective neural
remodeling method for stroke patients who generally suffer from unilateral hemiparesis.
This method involves using the healthy limb to guide the rehabilitation training of the
affected limb. When training hemiplegic patients, it is important to increase human–
computer interaction as much as possible. This can enhance the motivation for training
and visualize the training process.
Unlike passive robots that constantly direct the movement of the affected limb, assis-
tive robots provide assistance only when the patient intends to initiate the movement. Both
passive and assistive modes are designed for the initial stages of stroke rehabilitation, when
the patient may lack the necessary strength to independently move the affected limb. On
the other hand, active and resistive modes are employed in the later phases of rehabilitation
as they necessitate active movement from the patient.
Facilitating a seamless human–robot interaction (HRI) that feels natural presents a
significant challenge in the development of robotic rehabilitation systems. The chosen
approach for detecting user intentions plays a pivotal role for ensuring a transparent and
user-friendly HRI. Certain robotic rehabilitation platforms adopt bio-signals as a source
for recognizing intentions [25–29]. Among these, electromyographic (EMG) signals are
widely utilized due to their strong correlation with human motion. These signals capture
Sensors 2023, 23, 8801 4 of 32

the electrical activity produced by skeletal muscles responsible for executing the intended
actions [30]. There is another branch of robot-assisted rehabilitation that incorporates
various therapeutic approaches alongside upper extremity exoskeletons. For instance,
functional electrical stimulation (FES) has demonstrated its ability in facilitating recovery
in the paralyzed arms of stroke survivors. Combining functional electrical stimulation with
the patient’s remaining voluntary effort has proven to enhance cortical plasticity [13].
Despite the development of numerous prototypes, research on robotic devices for
upper extremity rehabilitation remains a growing field and requires new approaches to
address key limitations in hardware design. In this study, we have designed an upper
extremity exoskeleton robot to tackle some of the key limitations in existing exoskeletons.
We have also developed a myoelectric angle synchronization acquisition system to enable
more accurate myoelectric control of joint movements. The main contributions of this study
are as follows.
1. The aim of this study was to conceptualize, analyze, construct, and validate a six-
degree-of-freedom upper limb rehabilitation robot that seamlessly integrates EMG
signal monitoring, PID control, and Unity3D simulation games. This comprehensive
approach set a notable precedent for similar inquiries in the field of rehabilitation research.
2. This paper devised an eight-channel sEMG acquisition system to facilitate passive
control of the affected side using an elbow rehabilitator. The feasibility of sEMG-
based elbow angle prediction was validated using a back propagation neural net-
work (BPNN).

2. Overall Design Solution of the Robot Mechanical Structure

2.1. Human Upper Limb Structure and Kinematics
In this work the interrelationships between bones, joints, muscles and other tissues
of the human upper limb were analyzed. Based on this analysis and relevant clinical
requirements, this paper proposed an overall design scheme for the mechanical structure
of an upper limb hemiplegic rehabilitation training robot.
The human upper extremity includes the shoulder, upper arm, forearm, and hand.
The main joints of the upper limbs are the sternoclavicular joint, the shoulder joint, the
elbow joint, and the hand joint. The palm of the hand has a complex structure and flexible
movements and is often analyzed and designed separately for rehabilitation training
systems. However, the research in this paper focused on the shoulder, elbow, and wrist
joints of the human upper extremity. The seven degrees of freedom of movement of
the human upper limb, including the shoulder, elbow, and wrist joints, involve different
muscle groups and joint mobility [31]. The relevant information for each degree of freedom
is presented in Table 1. In the absence of muscular non-rigidity in the upper limb, the
musculoskeletal relationships can be effectively described using the Hill model. This model
predicts moments within the range of 0–30 N·m for the shoulder and upper arm muscles,
while moments within the range of 0–7 N·m are expected for the forearm muscles [32].

Table 1. Upper limb joint muscles and mobility.

Joint Joint No. DOF Main Muscles Rom

Flexion Coracobrachialis 70~90◦
J2
Extension Latissimus dorsi 40~90◦
Shoulder
Internal rotation Latissimus dorsi 70~90◦
J1
External rotation Infraspinatus 40~50◦
Sensors 2023, 23, x FOR PEER REVIEW 5 of 36
Sensors 2023, 23, 8801 5 of 32

Internal rotation Latissimus dorsi 70~90°

J1
Table 1. Cont. External rotation Infraspinatus 40~50°
Flexion Biceps 135~150°
Joint
Elbow Joint
J4 No. DOF Main Muscles Rom
Extension Triceps 0~10° ◦
Flexion Biceps 135~150
Elbow J4 External rotation Pronator teres 80~90°
Forearm J5/J3/J2′ Extension Triceps 0~10◦
Internal rotation Supinator muscle 80~90°
0 External rotation
Flexion FlexorPronator teres
carpi ulnaris 80~90◦
50~60°
Forearm J5/J3/J2
J6
Internal rotation
Extension Supinator muscle
Extensor carpi ulnaris 80~90◦
30~60°
Wrist Adduction
Flexion Flexor carpi
Flexor carpiradialis
ulnaris 50~60◦
30~40°
J6
J6′ Extension Flexor/Extensor carpi
Extensor carpi ulnaris 30~60 ◦
Wrist Abduction 25~30°
Adduction
ulnaris
Flexor carpi radialis 30~40◦
J60
Abduction Flexor/Extensor carpi ulnaris 25~30◦
2.2. Design Analysis of the Upper Limb Rehabilitation Robot
It has been noted that the brain plasticity theory highlights the significance of stimu-
2.2. Design Analysis of the Upper Limb Rehabilitation Robot
lating compound movements of multiple joints and muscle groups for neural recovery.
It has been
Additionally, thenoted that the
correction brain plasticity
of abnormal theory
movement highlights
patterns after the significance
hemiplegia of stim-
necessitates
ulating compound movements of multiple joints and muscle groups for neural
the coordination of movements among joints. As a result, single-joint or few-joint motor recovery.
Additionally,
training alonethe
maycorrection of abnormal
not be sufficient movement patterns
for rehabilitation. after the
Therefore, hemiplegia necessitates
robot design aimed
the coordination of movements among joints. As a result, single-joint or few-joint
to facilitate the participation of multiple joints of the upper limb in the training simulta- motor
training alone may not be sufficient for rehabilitation. Therefore,
neously. By doing so, better rehabilitation outcomes could be achieved. the robot design aimed to
facilitate the participation of multiple joints of the upper limb in the training simultaneously.
The scheme of the exoskeleton was designed by first making certain simplifications
By doing so, better rehabilitation outcomes could be achieved.
to the models of the joints.
The scheme of the exoskeleton was designed by first making certain simplifications to
The shoulder joint was simplified as a ball-and-socket joint with a fixed center of ro-
the models of the joints.
tation.
The shoulder joint was simplified as a ball-and-socket joint with a fixed center of rotation.
The elbow joint was simplified as a hinge joint rotating around a coronal axis, provid-
The elbow joint was simplified as a hinge joint rotating around a coronal axis, provid-
ing one degree of freedom for flexion/extension.
ing one degree of freedom for flexion/extension.
The forearm joint was simplified as a carpel joint with one degree of freedom, which
The forearm joint was simplified as a carpel joint with one degree of freedom, which
allows for inward/outward rotation around the central axis of the forearm.
allows for inward/outward rotation around the central axis of the forearm.
Similarly, the wrist joint was simplified as an elliptical joint with two degrees of free-
Similarly, the wrist joint was simplified as an elliptical joint with two degrees of
dom: flexion/extension and retraction in the frontal and sagittal planes, respectively.
freedom: flexion/extension and retraction in the frontal and sagittal planes, respectively.
The degrees of freedom corresponding to each joint of the human upper limb are
The degrees of freedom corresponding to each joint of the human upper limb are
shown
shown in in Figure
Figure 1.
1.

Figure 1. Sketch of the degrees of freedom of the exoskeleton robot.

Flexion/extensionof
Flexion/extension ofthe
theshoulder
shoulderjoint
joint(J2);
(J2); internal/external
internal/externalrotation
rotation(J1).
(J1).
Flexion/extension of the elbow joint
Flexion/extension of the elbow joint (J4). (J4).
Internal/externalof
Internal/external ofthe
theforearm (J5/J3/J20 ).
forearm(J5/J3/J2′).
Flexion/extensionof
Flexion/extension ofthe
thewrist
wristjoint
joint(J6);
(J6);adduction/abduction
adduction/abduction(J6′). (J60 ).
In addition,
In addition, the
the sizes
sizesof
ofthe
thespecific
specificparts
partsofofthe
themechanism
mechanism areare
required
requiredto be
to adjustable
be adjust-
to accommodate different sizes of the affected limb.
able to accommodate different sizes of the affected limb.
Sensors 2023, 23, 8801 6 of 32

3. The System of Rehabilitation Training

3.1. Development Procedure
The subsequent subsections provide comprehensive insights into the design and
development particulars of the upper extremity rehabilitation robot outlined in this paper.
Furthermore, Table 2 compiles the design specifications and the chosen components for the
robot’s development.

Table 2. Design specifications and selected components of the rehabilitation robot.

Degrees of Freedom
Active Passive
3 3
Ranges of motion/joints’ limits (degrees)
Joint-J1 Joint-J2 Joint-J4 Joint-J5/3/20 Joint-J6 Joint-J60
−40 to 90 −45 to 90 0 to 135 −90 to 80 −60 to 40 −40 to 30
Fabrication
Material Aluminum 6061, stainless steel 304, plastic (photosensitive resins and polylactic acid)
Fabrication process Modeling, material selection, CNC machining, lathe turning, 3D printing, module assembly, circuit assembly
Actuators
Location Joint-J1 Joint-J2 Joint-J4
Speed reducer Speed reducer Steering engine DS5160,
Motors
DH-03X DH-03X 12.6 W
Operating voltage 12–24 V 12–24 V 8.4 V
Nominal speed 0.12–0.24 s/60◦ 0.12–0.24 s/60◦ 0.08 s/60◦
Nominal torque 11–38 N·m 11–38 N·m 7 N·m
Weight 595 g 595 g 158 g
Motor drivers KSMA-03X DHMCU -
Motor driver current rating 5–12 A -
Motor driver input 12–24 V DC power supply -
Motor driver feedback Current sense, Hall sensor pulses -
Damping hinges
Location Joint-J6/J60
Nominal torque 0.7 N·m
Material Stainless steel 304
Weight 25 g
Thicknesses 2 mm
Angle of activity 0–250◦
Upper arm rotating cuff assembly
Location J3/J20 J5
Inside diameter 65 mm 120 mm
External diameter 80 mm 140 mm
Pressure sensors RDF-6
Sensors 2023, 23, 8801 7 of 32

Table 2. Cont.

Embedded hardware circuitry

Circuit board STM32F407ZGT6
Hall sensors P3302(0.3%F.S.)
Sensors 2023, 23, x FOR PEER REVIEW 7 of 36
Position sensor MPU6050
Myoelectric sensors DFRobot in cooperation with OYMotion
Charger interface DC5.5 × 2.5
Charger interface DC5.5 × 2.5
Network interface Network interface RJ45 RJ45
Electrode patch Electrode patch X-1,X-1,
Xunda RadioRadio
Xunda Co.(Hangzhou, China) China)
Co.(Hangzhou,

Duringthe
During theprototype
prototypedesign
designphase,
phase,lightweight
lightweight materials
materials that
that were
were safe
safe and and reliable
reliable in
in terms of strength were required for the exoskeleton robot. In order to select
terms of strength were required for the exoskeleton robot. In order to select the appropriate the appro-
priate materials,
materials, a stress acheck
stress check
was was performed
performed on the non-standard
on the non-standard parts, and parts, and the stress
the maximum maxi-
mum stress was calculated as the basis for the material selection. The
was calculated as the basis for the material selection. The torque required for shouldertorque required for
shoulder
joint joint movement
movement was large to was large shaking
prevent to prevent shaking
of the of the base
base during during
training, andtraining, and
the material
selected for both the base and robot arm was chosen with patient safety and comfortand
the material selected for both the base and robot arm was chosen with patient safety in
comfortAlloy
mind. in mind.
steelAlloy steel was
was chosen forchosen for material,
the base the base material,
while 3Dwhile 3D printing
printing technologytechnol-
was
ogy was
used used
for the for the
robot arm; robot arm; specifically
specifically photosensitive
photosensitive resin andresin
PLA andwerePLA werefor
selected selected
their
for their light weight and high strength. The robot arm cross-section
light weight and high strength. The robot arm cross-section was rounded to better fitwas rounded to better
the
fit the patient’s
patient’s upper limbupper limb
and and increase
increase wearing wearing
comfort.comfort.
The rehabilitation
The rehabilitationrobotrobotsystem
systemisisdepicted
depictedin inFigures
Figures22and
and3.3.

Figure2.
Figure 2. The
The 3D
3D model
model of
of the
the upper
upper limb
limb rehabilitation
rehabilitationrobot.
robot.

Figure3.
Figure 3. Upper
Upper rehabilitation
rehabilitationrobot
robot(right)
(right)and
andwearing
wearingfigure
figure(left).
(left).

For the shoulder

For shouldermodule,
module,totoachieve
achieve both
bothinward/outward
inward/outward shoulder rotation
shoulder and flex-
rotation and
ion/extension, the robot’s
flexion/extension, shoulder
the robot’s structures
shoulder werewere
structures interconnected through
interconnected a tandem
through con-
a tandem
configuration,
figuration, as as depicted
depicted in in Figure
Figure 4. 4.InInthe
thecross-sectional
cross-sectionalview
viewof
ofthe
the frontal
frontal mechanism,
the primary structures, Link-1 and Link-2, were fabricated using 3D printing, with stand-
ard components such as geared motors, a D-shaped helm, and screws. The ergonomic
shoulder module was designed with two complementary structures responsible for ab-
duction internalization (Link-2) and internal rotation externalization (Link-1). The Link-2
module features a total of 12 threaded holes—6 on the inside and 6 on the outside. These
 Sensors 2023, 23, 8801 8 of 32

the primary structures, Link-1 and Link-2, were fabricated using 3D printing, with standard
components such as geared motors, a D-shaped helm, and screws. The ergonomic shoul-
der module was designed with two complementary structures responsible for abduction
Sensors
Sensors 2023,
2023, 23, 23, x FOR
x FOR PEER
PEER REVIEW
REVIEW internalization (Link-2) and internal rotation externalization (Link-1). The Link-28module of8 36
of 36
features a total of 12 threaded holes—6 on the inside and 6 on the outside. These correspond
to the 6 parallel holes recessed in the Link-1 module and can be securely fastened using M6
screws. Two rowspatients
of M6-threaded holes,shoulders.
spaced 10 The
cm apart, weretheadded to accommodate
to to accommodate
accommodate patients with with broader
broader shoulders. The design
design of of grooves
the grooves andand protru-
protru-
patients with broader shoulders. The and
design of the grooves and protrusions facilitated
sions facilitated the connection, fitting, and transmission of motion and force between thethe
sions facilitated the connection, fitting, transmission of motion and force between
the
two connection,
parts. The fitting, and
utilization transmission of motion and force between the twothe parts. The
two parts. The utilization of of a D-shaped
a D-shaped rudder
rudder plate
plate as as a connector
a connector increased
increased contact
the contact
utilization
area between of a D-shaped rudder plate as a connector increased the contact area between
area between thethe reduction
reduction gear
gear andandthethe rotating
rotating shaft,
shaft, enhancing
enhancing thethe stability.
stability.
the reduction gear and the rotating shaft, enhancing the stability.

Figure
Figure 4. Shoulder
4. Shoulder rotation
rotation module.
module.
Figure 4. Shoulder rotation module.
Figure
Figure
Figure 55 displays
displays
5 displays a cross-sectional
cross-sectional
a cross-sectional
a view
view
viewof of
thethe
of rotating
rotating
the rotating cuff
cuff assembly
assembly
cuff assembly forfor
thethe
for upper
upper
the upper
arm.
arm.
arm. To
To To enable
enable
enable the
thethe
largelarge
arm
large arm
armandand forearm
forearm
and forearm to rotate
to rotate internally
internally
to rotate internally and
andand externally,
externally, the
thethe
externally, outer
outer cuff
cuff
outer cuff
remained
remained
remained stationary
stationary
stationary while
while
while the
thethe inner
inner cuff
inner cuff rotated.
rotated.
cuff The
TheThe
rotated. outer
outer
outer cuff
cuff was
cuff was affixed
affixed
was to
to to
affixed the
thethe large
large
large
and
and
and small
small
small armsarms
arms of
thethe
of of the robot’s
robot’s
robot’s upper
upper
upper limblimb
limb using
using
using M3
M3M3 screws.
screws.
screws. Given
Given
Given that
thethe
that
that the axis
of of
axis
axis of rotation
rotation
rotation
was
was situated
situated at
at the
the humerus
humerus during
during forearm
forearm rotation,
rotation, the
the ergonomic
ergonomic
was situated at the humerus during forearm rotation, the ergonomic mounting position mounting
mounting position
position
ensured
ensured
ensured aa better
better
a better fit
fit fit around
around
around the
thethe upper
upper
upper limblimb
limb
andand
and reduced
reduced
reduced jitter
jitter
jitter during
during
during wear
wear
wear andand
and training.
training.
training.

Figure
Figure
Figure 5. Rotating
5. Rotating
5. Rotating sleeve
sleeve
sleeve assembly.
assembly.
assembly.

In In
thisthis study,
study, thethe rehabilitation
rehabilitation
rehabilitation robot
robot
robot waswas equipped
equipped with
with 140140
mm mmandand
80 80
mm mm rotating
rotating
sleeve
sleeve assemblies
assemblies
assemblies ononthethe large
large and and small
small arms,
arms, respectively.
respectively. BothBoth assemblies
assemblies shared
shared thethe
same parameters
parameters except
except for
for thethe varying
varying inner
inner and and outer
outer diameters.
diameters. The
same parameters except for the varying inner and outer diameters. The inner sleeve func- The
inner inner
sleeve sleeve
func-
tioned as a rotary axis, with the bearing design serving as a reference in
tioned as a rotary axis, with the bearing design serving as a reference in the design. The the design. The
user sleeve was divided into upper and lower sections, with a middle groove
user sleeve was divided into upper and lower sections, with a middle groove that fit into that fit into
thethe inner
inner convex
convex groove
groove of of
thethe outer
outer sleeve
sleeve band,
band, incorporating
incorporating a ball
a ball in in
thethe middle
middle to to
reduce friction. M3 screws were used to join the three parts into a single unit,
reduce friction. M3 screws were used to join the three parts into a single unit, with a total with a total
Sensors 2023, 23, 8801 9 of 32

functioned as a rotary axis, with the bearing design serving as a reference in the design.
The user sleeve was divided into upper and lower sections, with a middle groove that fit
Sensors 2023, 23, x FOR PEER REVIEWinto the inner convex groove of the outer sleeve band, incorporating a ball in the middle 9 of 36
to reduce friction. M3 screws were used to join the three parts into a single unit, with a
total height of 42 mm. The slide groove had a width of 5.5 mm, the convex groove was
5 mm wide, and both sides featured 1 mm rounded corners, with the ball measuring 3 mm
diameter. TheThe
in diameter. entire assembly
entire was was
assembly constructed fromfrom
constructed PLA material, with awith
PLA material, weight of 220 of
a weight g
for
220the large
g for the arm
largecuff
armassembly and 180
cuff assembly and g 180
for the
g forsmall arm cuff.
the small arm A pressure
cuff. sensor
A pressure was
sensor
inserted between the outer cuff and the user cuff, positioned at the uppermost
was inserted between the outer cuff and the user cuff, positioned at the uppermost part part of the
of
rotating cuff, primarily intended for measuring the force of the upper
the rotating cuff, primarily intended for measuring the force of the upper limb movement limb movement
along
alongthe
theanterior
anteriorsagittal
sagittalplane.
plane.
The
The wrist jointmodule
wrist joint moduleof ofthe
therobot
robotconsisted
consistedof oftwo
tworotary
rotaryjoints
jointsfor
forradial
radialoffset
offsetand
and
flexion/extension
flexion/extensionofofthe thewrist
wristjoint,
joint,asasdepicted
depictedininFigure
Figure6.6.Wrist
Wristlink
link11took
tookthetheform
formofof
an L-shaped structure, serving as an intermediary for the connection of
an L-shaped structure, serving as an intermediary for the connection of the wrist modules.the wrist modules.
Damping
Dampinghinges
hinges11linked
linkedWrist
Wristlink
link11to toWrist
Wrist link
link 2,
2, while
while Damping
Dampinghinges hinges22connected
connected
Wrist
Wrist link 1 to the primary structure of the robot. Both damping hingesshared
link 1 to the primary structure of the robot. Both damping hinges sharedaacommon
common
solid
solidpivot
pivotaxis
axiswith
withaa phosphor
phosphorbronze bronze pivot
pivot and
and offered
offered 0.7
0.7 N∙m
N·m of of damping.
damping. The Thekey
key
distinction
distinctionbetween
betweenthe thetwo
twowaswasthatthatDamping
Dampinghinges hinges11possessed
possessedaavertical
verticalstructure
structureon on
both
bothsides
sidesofofthe
thepivot
pivotaxis,
axis, whereas
whereasDamping
Dampinghinges hinges22featured
featured aa zigzag
zigzag structure
structure with
with
symmetrical
symmetricalrectangular
rectangular sections
sections onon either
either side.
side.

Figure
Figure6.
6.Wrist
Wristrotation
rotation assembly
assembly for
for the
the robot.
robot.

In
In the
the hand module,
module, we wedistinguished
distinguishedthreethreeprimary
primary components:
components: thethe wiring
wiring box,
box, the
the grip,
grip, andandthe the
base.base.
TheThe wiring
wiring box box housed
housed the MPU6050
the MPU6050 posture
posture sensor.
sensor. SinceSince
therethere
were
no holes
were in theinlid,
no holes thethe
lid,circuitry extended
the circuitry fromfrom
extended the bottom of the
the bottom boxbox
of the to the basebase
to the of the
of
grip. Base 1 was affixed to Base 2, and Base 2 was securely connected to
the grip. Base 1 was affixed to Base 2, and Base 2 was securely connected to Wrist link 2 Wrist link 2 using
M3 screws.
using M3 screws.
In this
In this paper,
paper, wewe utilized
utilized motor
motor drives
drives as as our
our method
method ofof choice.
choice. To To minimize
minimizeany any
burden on
burden on the
the patient,
patient, the
the entire
entire weight
weight ofof the
the robot
robot arm
arm was
was attached
attached to to the
the shoulder
shoulder
module, and
module, and the base waswas connected
connected to tothe
therobot
robotarm armtotocounteract
counteractthetheeffect
effectof of
gravity
gravityon
on the patient. The base’s underside featured a power on/off button, a charger port, and aa
the patient. The base’s underside featured a power on/off button, a charger port, and
networkcable
network cable port.
port.Additionally,
Additionally,to toaccommodate
accommodatepatientspatientsofofvarying
varyingheights,
heights,the thebase
base
was equipped with a height-adjustable mechanism that could be modified
was equipped with a height-adjustable mechanism that could be modified using the knob. using the knob.
In order
In order toto ensure
ensurethethesafety
safetyof of
upper
upperlimb training,
limb the motor
training, and the
the motor andcorresponding
the corresponding upper
limb mechanical joint were affixed to the movable joint through an intermediary
upper limb mechanical joint were affixed to the movable joint through an intermediary sleeve.
sleeve.
In the software of the upper limb rehabilitation robot, users could select between ac-
tive and passive modes and adjust the angular speed of movement, movement time, and
movement angle using the interface buttons.

3.2. Mass and Inertia Properties of the Proposed Exoskeleton Robot

Sensors 2023, 23, 8801 10 of 32

In the software of the upper limb rehabilitation robot, users could select between
active and passive modes and adjust the angular speed of movement, movement time, and
movement angle using the interface buttons.

3.2. Mass and Inertia Properties of the Proposed Exoskeleton Robot

The mass, center of gravity, and inertia tensor for each component of the proposed
exoskeleton robot were calculated using the Solidworks software (SolidWorks 2021). The
results are presented in Table 3, categorized by the modules.

Table 3. Mass inertia properties of the proposed exoskeleton system.

Length Weight Center of Gravity (mm) Moment of Inertia I at CG (kg·mm2 ) (103 )

Module
(mm) (kg) CGX CGY CGZ IXX IYY IZZ
Shoulder joint 247 3.2 −4.3 −187.4 −46.1 102.1 25.3 79.8
Elbow joint 330 1.8 3.7 −9.5 8.9 70.6 24.7 21.3
Wrist joint 15 ± 5 0.7 22.8 0 −70.9 0.2 0.8 0.8

3.3. Safety
Safety is of paramount importance in the design of upper limb exoskeleton robots, as
they operate in close proximity to the wearer. Human–robot interactions should be meticu-
lously engineered to ensure safe operation. Achieving this safety entails the incorporation
of safeguards at multiple levels, including mechanical, electronic, and control designs, to
guarantee the secure usage of the robot.
On a mechanical level, safety can be established by integrating physical blocking
devices within the robot’s structure. These devices serve to restrict the robot from surpass-
ing predetermined ranges of motion (ROM). Additionally, the design of links and robot
components should be such that adjacent links naturally act as physical blocking devices
within the defined ROM. Furthermore, it’s crucial to regulate the robot joints by setting
appropriate current and voltage limits to ensure they remain within the permissible ROM.
In terms of the electronic design, controls can be established to limit torque, force,
velocity, and position through saturation mechanisms. This ensures the wearer’s safety
in the event of a robot malfunction. The control algorithm employed in the rehabilitation
robot discussed in this paper includes predefined thresholds for the ROM, velocity, force,
and torque. These thresholds can be easily adjusted and set by users via the graphical user
interface embedded in the onboard software. Additionally, in the event of an emergency or
system failure, the onboard software and power button allow for a quick shutdown with a
single click. By implementing these safety measures across the mechanical, electronic, and
control domains, the safety of human–robot interactions is enhanced.

4. Kinematics, Jacobian, and Dynamics

4.1. Kinematic Analysis of Robots
For serial actuators, the commonly used approach by researchers is the Denavit–
Hartenberg (D–H) parametric method. This method is favored due to its simplicity and
ease of application in various scenarios, such as the development of forward kinematics,
inverse kinematics, Jacobian, and dynamic models.
As shown in Figure 7, the coordinate system of the robot base (base coordinate system)
is represented by X0 Y0 Z0 , which includes the coordinate system of joint one, followed by
the coordinate systems of joints two to six. The coordinate system of the end-effector of the
robot arm is represented by X6 Y6 Z6 .
 Hartenberg (D–H) parametric method. This method is favored due to its simplicity and
ease of application in various scenarios, such as the development of forward kinematics,
inverse kinematics, Jacobian, and dynamic models.
As shown in Figure 7, the coordinate system of the robot base (base coordinate sys-
tem) is represented by X0Y0Z0, which includes the coordinate system of joint one, followed
Sensors 2023, 23, 8801 11 of 32
by the coordinate systems of joints two to six. The coordinate system of the end-effector
of the robot arm is represented by X6Y6Z6.

Figure 7. Coordinate
Figure systems
7. Coordinate of theof6-DOF
systems robot (The
the 6-DOF robotred arrows
(The denote the
red arrows Z-axis,
denote thewhich
Z-axis,is which is
maintained perpendicular to the black arrows representing the X- and Y-axes.
maintained perpendicular to the black arrows representing the X- and Y-axes. The The blue double
blue double
arrows indicate the linkage lengths or offsets of the adjacent joints).
arrows indicate the linkage lengths or offsets of the adjacent joints).

After establishing the joint coordinate system, the D–H parameter (Table 4) of each
joint was determined using the adjacent joint coordinate system. The relationship between
two adjacent links was represented by four parameters: the joint angle θ, link length d, link
offset a, and twist angle α.

Table 4. DH parameter table.

i θi di ai αi
1 θ1 (−90◦ ) d1 0 90◦
2 θ 2 (−90◦ ) 0 0 90◦
3 θ 3 (180◦ ) d3 0 90◦
4 θ 4 (−90◦ ) 0 a4 −90◦
5 θ 5 (90◦ ) 0 0 90◦
6 θ 6 (0◦ ) d6 0 90◦

The general formula for the transformation between adjacent coordinates in the D–H
method, which describes the transformation of coordinate system i−1 with respect to
coordinate system i, is as follows.
i −1
i T = Ai = R(z, θi ) × T (0, 0, di ) × T ( ai , 0, 0) × R( x, αi ) (1)

where R(z,θ i ) and R(x,αi ) are the basic rotation transformation matrices around the z-axis
and x-axis, respectively, and T(0,0,di ) and T(αi ,0,0) are the basic translation transformation
matrices along the z-axis and x-axis, respectively. The four basic transformation matrices
are all invertible matrices, so Ai is also invertible. Ai can be further expressed as follows.

−sinθ i cosαi
 
cosθ i sinθ i sinαi ai cosθ i
 sinθ i cosθ i cosαi −cosθ i sinαi ai sinθ i 
Ai =   (2)
 0 sinαi cosαi di 
0 0 0 1
Sensors 2023, 23, 8801 12 of 32

The D–H parameters can be substituted into Equation (2), as follows.

   
c1 0 s1 0 c2 0 s2 0
0 T =  s1
 0 − c1 0 , 1 T =  s2 0 − c2 0,
 
1 0 1 0 d1  2  0 1 0 0
0 0 0 1 0 0 0 1
c4 0 − s4 a4 c4
   
c3 0 s3 0
2 T =  s3 0
 − c3 0  , 3 T =  s4 0
 c4 a4 s4 
,
3 0 1 0 d3  4  0 −1 0 0 
0 0 0 1 0 0 0 1
   
c5 0 s5 0 c6 0 s6 0
4 T =  s5
 0 − c5 0 , 5 T =  s6 0 − c6 0 
 
5 0 1 0 0  6  0 1 0 d6 
0 0 0 1 0 0 0 1

The matrix representation of the robot’s end position in the base coordinate system
(kinematic equations of the robot) is given by the following.
 
nx ox ax px
0 0 1 2 3 4 5  ny oy ay py 

6 T =1 T2 T3 T4 T5 T6 T =  n (3)

z oz az pz 
0 0 0 1

The end position matrix of the robot is as follows.

T
P T = Px , Py , Pz

(4)

In Equation (3),
 T
n x ny nz is the direction vector of the X6 -axis of the robot’s end (handle) coordi-
nate system in the base coordinate system.
 T
o x oy oz is the direction vector of the Y6 -axis of the robot’s end (handle) coordi-
nate system in the base coordinate system.
 T
a x ay az is the direction vector of the Z6 -axis of the robot’s end (handle) coordi-
nate system in the base coordinate system.

nx = (c4 c5 c6 − s4 s6 )(s1 s3 + c1 c2 c3 ) + s5 c6 (s1 c3 − c1 c2 s3 ) + c1 s2 (s4 c5 c6 + c4 s6 )

ny = (s4 s6 − c4 c5 c6 )(c1 s3 − s1 c2 c3 ) + s1 s2 (c4 s6 − s4 c5 c6 ) − s5 c6 (c1 c3 + s1 c2 s3 )

nz = −c6 c5 (c2 s4 − s2 c3 c4 ) − c6 s2 s3 s5 − s6 (c2 c4 + s2 c3 s4 )

ox = c4 s5 (s1 s3 + c1 c2 c3 ) + s5 c1 s2 s4 − c5 (s1 c3 − c1 c2 s3 )

oy = c5 (c1 c3 + s1 c2 s3 ) − s5 c4 (c1 s3 − s1 c2 c3 ) + s1 s2 s4 s5

oz = s2 s3 c5 − s5 (c2 s4 − s2 c3 c4 )

ax = (s4 c6 + c4 c5 s6 )(s1 s3 + c1 c2 c3 ) + s5 s6 (s1 c3 − c1 c2 s3 ) + c1 s2 (s4 c5 s6 − c4 c6 )

ay = (−s4 c6 − c4 c5 s6 )(c1 s3 − s1 c2 c3 ) − s5 s6 (c1 c3 + s1 c2 s3 ) + s1 s2 (s4 c5 s6 − c4 c6 )

az = c6 (c2 c4 + s2 c3 s4 ) − c5 s6 (c2 s4 − s2 c3 c4 ) − s2 s3 s5 s6

px = (a4 c4 + c4 s5 d6 )(s1 s3 + c1 c2 c3 ) + c1 s2 d3 − c5 d6 (s1 c3 − c1 c2 s3 ) + c1 s2 s4 s5 d6 + c1 s2 s4 a4

Sensors 2023, 23, 8801 13 of 32

py = (−a4 c4 − c4 d6 s5 )(c1 s3 − s1 c2 c3 ) + c5 d6 (c1 c3 + s1 c2 s3 ) + s1 s2 (s4 s5 d6 + d3 + a4 s4 )

pz = d1 − c2 d3 − s5 d6 (c2 s4 − s2 c3 c4 ) + s2 s3 c5 d6 − c2 a4 s4 + s2 c3 a4 c4
where si stands for sin θ i , ci stands for cos θ i , d1 = 80 mm, d3 = 330 mm, d6 = 4~20 mm, and
a4 = 400 mm.

4.2. Jacobian
In this paper, the linear velocity vector of the upper limb rehabilitation robot consisted of
velocities along the three Cartesian axes, and the rotational velocity vector included the angular
velocities around the Cartesian axes. In Equation (5), q1 . . .qn for the joint variables, x p represented an
output component of the multivariate function. ε was either 0 or 1; 0 when the joint was a rotating
−
joint and 1 when it was a moving joint, and ε was the inverse of ε. 0 Z1 . . .0 Zn represented the output
vector of the multivariate function. The Jacobian matrix of the robot, which was a 6 × 6 matrix, was
calculated using MATLAB (version R2018a, MathWorks, Natick, MA, USA) and based on Equation (5).
 
∂x p ∂x p ∂x p
· · ·
J = − ∂q1 −
∂q2
−
∂qn 
(5)
ε 1 0 Z1 ε 2 0 Z2 · · · ε n 0 Zn

4.3. Dynamics
The dynamic equations of the robot were derived from the Newton–Euler iterative formulation
as follows.
→ .
. .

τ = M(q) q + V q, q q + G (q) + F q, q (6)
.

where M(q) is the 6 × 6 mass matrix of the manipulator, V q, q is a 6 × 1 dimension vector
‘composed of the centrifugal and Coriolis terms, and G (q) is a 6 × 1 vector of the gravity terms. In
.

addition, F q, q is a 6 × 1 vector of nonlinear Coulomb
friction and can be expressed using the
. .
following relation with a coefficient of friction c. sgn(q is a signed function of q’s (generalized
velocity or velocity vector), indicating the direction of the velocity. The M, V, G, and F matrices were
computationally intensive and, therefore, are not explained in the article.
.
.

F q, q = c.sgn(q (7)

5. Elbow Joint Angle Prediction System

sEMG not only reflects the anatomical structure and physiological properties of the muscles
being monitored, but also contains information about their movement, such as the muscle moment,
movement speed, and movement angle. Since sEMG is non-invasive, rapid, and convenient, it
is commonly used in rehabilitation medicine to evaluate muscle function recovery. The field of
prosthetic control is also based on sEMG, which is used for motion recognition and classification to
control the movement of prostheses according to the patient’s intention [33–35].
In order to establish the relationship between sEMG and the elbow joint angle, it was necessary
to collect joint angle data during the flexion and extension movements of the elbow joint. This
information served two purposes: first, as a target value for model training and ongoing evaluation;
and second, as a metric to validate the model performance and prediction accuracy. The synchronized
signal acquisition system (Figures 8 and 9) consisted of two parts: hardware and software. In the
hardware part, the sEMG sensors were distributed around the elbow muscles and the angle sensors
were located on the rotation axis.
 Sensors2023,
Sensors 2023,23,
23,xxFOR
FORPEER
PEERREVIEW
REVIEW 15 of
15 of 36
36
Sensors 2023, 23, 8801 14 of 32

Figure8.
Figure
Figure 8.Synchronized
8. Synchronizedsignal
Synchronized signalacquisition
signal acquisitionequipment.
acquisition equipment.
equipment.

Figure9.
Figure
Figure 9.Synchronized
9. Synchronizedsignal
Synchronized signalacquisition
signal acquisitionsystem
acquisition systemupper
system uppercomputer
upper computersoftware.
computer software.
software.

The
Thesmall
The small
small arm arm
arm and
andlarge
and armarm
large
large of the
arm ofrehabilitation
of the rehabilitation
the devicedevice
rehabilitation were encompassed
device were encompassed
were by two modules,
encompassed byby two
two
each equipped
modules,
modules, each
each with six sEMGwith
equipped
equipped sensors.
with Data acquisition
six sEMG
six sEMG sensors. Data
sensors. was
Datafacilitated
acquisition
acquisition through
wasthe
was use of anthrough
facilitated
facilitated existing
through
data
the acquisition
the use
use of of an card,
an existing which
existing data
data performed the essential
acquisition
acquisition tasks ofperformed
card, which
card, which EMG signalthe
performed capture
the and digital-to-analog
essential
essential tasks of
tasks of EMG
EMG
conversion at a sampling frequency of 650 Hz.
signal capture
signal capture and and digital-to-analog
digital-to-analog conversionconversion at at aa sampling
sampling frequency
frequency of of 650
650 Hz.
Hz.
To operate the device, the white switch located at the junction of the large arm and the small
To operate
To operate the the device,
device, the the white
white switch
switch located
located at at the
the junction
junction of of the
the large
large armarm andand
arm was pressed. Upon doing so, a green light in a slow-flashing state became visible. Next, the
the
the small
small arm
arm was
was pressed.
pressed. Upon
Upon doing
doing so,
so, aa green
green light
light in
in
device was brought into proximity with the matching dongle to initiate the Bluetooth connection.aa slow-flashing
slow-flashing state
state became
became
visible.
visible.
Once Next, the
the Next,
connectionthe device
device was brought
was
was successfully brought into proximity
into
established, proximity
the indicator with
with the
the
light matchingfrom
matching
transitioned dongle
dongle to initiate
to
a slow initiate
flash to
the
athe Bluetooth
Bluetooth
steady, connection.
normalconnection.
light, signifying Once
Oncethat the
the connection
theconnection
Bluetooth pairingwas successfully
was successfully established,
process wasestablished,
complete. the indicator
the indicator
lightThe
light transitioned
transitioned
upper computer frominterface,
from aaslow
slowflash
flash toaasteady,
to
as depicted steady, normal
normal
in Figure light,
light,
9, served signifying
as signifying
the thatfor
that
central hub the
the Bluetooth
Bluetooth
the real-time
display,
pairingacquisition,
pairing process was
process and
was storage of sEMG and the joint angle data. This interface comprised eight
complete.
complete.
display The
The channels
upperand
upper was equipped
computer
computer interface,with
interface, asan
as operational
depicted
depicted button 9,
in Figure
in Figure situated
9, served
served atas
itsthe
as center.
the Of the
central
central hub
hub eight
for
for
channels, the
the real-time
the lower
real-time display, right corner
display, acquisition, featured
acquisition, and channel
and storage 8,
storage of which
of sEMG was
sEMG and dedicated
and the to
the joint joint
joint angleangle
angle data.information,
data. This
This in-in-
presenting data regarding the bending angle of the arm. The remaining seven channels served as
terface comprised eight display channels and was equipped
terface comprised eight display channels and was equipped with an operational button with an operational button
myoelectric signal channels, presenting real-time waveforms of the sEMG.
situated at
situated at its
its center.
center. Of Of thethe eight
eight channels,
channels, the the lower
lower rightright corner
corner featured
featured channel
channel 8, 8,
The central operational button held the key functions for initiating data acquisition and saving
which
which was
was dedicated
dedicated to
to joint
joint angle
angle information,
information, presenting
presenting
for both sEMG and the joint angle data. Prior to commencing the data acquisition, this button data
data regarding
regarding the
the bending
bending
angle
angle
was of the
of
displayedthe arm.
arm. The
as theThe remaining
remaining
“save” seven
button. seven
By clickingchannels
channels servedbutton,
served
the “save” as myoelectric
as myoelectric signal channels,
signal
the data acquisition channels,
began
presenting
presenting real-time
real-time waveforms
waveforms of
of the
the sEMG.
sEMG.
instantaneously, and the button transformed into a “stop” button. When the required data amount
The central
The
was collected, central operational
operational
the “stop” button was button
button held
held
clicked the
tothe key
haltkey functions
functions
the data for initiating
for
acquisition, initiating data acquisition
data
and the generated acquisition
data file
was
and saved
and saving in
saving for bin format
for both on
both sEMG the
sEMG and computer.
and thethe joint
joint angle
angle data.
data. Prior
Prior to to commencing
commencing the the data
data acqui-
acqui-
sition, this
sition, this button
button was
was displayed
displayed as
as the
the “save”
“save” button.
button. By
By clicking
clicking the
the “save”
“save” button,
button, the
the
Sensors 2023, 23, 8801 15 of 32

5.1. Analytical Study of Signal Preprocessing

In order to establish a quantitative relationship between sEMG and the elbow joint angle, it was
essential to collect joint angle data during elbow flexion and extension movements. This data served
two purposes: one was to be used as the target value for model training and continuous optimization,
and the other was to be used as an index for evaluating the model’s performance and verifying its
prediction effect.
Common joint angle measurement methods fall into two categories: direct and indirect. The
advantage of the direct measurement method is its low cost and simplicity of operation. However,
the device needs to be bound to the limb in direct contact, which may limit limb movement. For
this study, the direct measurement method was adopted using an angle sensor (Shanghai Aoyi
Information Technology Co., Ltd., Shanghai, China). This sensor contained a circular rotating varistor
that changed resistance value when the angle rotated, allowing for the rotation angle to be obtained
by measuring the dividing voltage of one of the segments of the varistor.
EMG signals are divided into two types based on the type of electrodes used for collection:
nEMG and sEMG. For this study, the surface electrode acquisition method (sEMG) was used, where
surface electrodes were attached to the skin surface to pick up the superposition of the action
potentials from different motor units in time and space. This method has the advantage of being non-
invasive and simple to operate, and the measured EMG changes can reflect the functional changes of
the whole muscle, making it suitable for sports training applications [36].
For this experiment, the selected electrode patch was a single-use ECG electrode. The EMG
sensor used in this paper integrated the filtering and amplification circuits. It amplified weak human
sEMG signals within the range of ±1.5 mV by 1000 times and effectively suppressed noise, especially
industrial frequency interference, through differential input and analog filtering circuits. The output
signal was in analog form, with 1.5 V as the reference voltage, and the output range was 0~3.0 V. The
signal amplification was consistent and unchanged for all the subjects in the experiments herein.

5.2. Preprocessing of the EMG Signals

The acquired sEMG and joint angle signals were susceptible to noise and required preprocess-
ing. sEMG, like other physiological electrical signals, has a low amplitude and demands precise
experimental equipment and controlled environmental conditions. In this chapter, we employed the
Fourier transform for sEMG noise reduction and wavelet transform for the joint angle signal noise
reduction. We also analyzed the process and outcomes of signal preprocessing.
The Fourier transform noise reduction process began with transforming the time domain signal
using the Fourier analysis. The transformed signal was then filtered to eliminate noise frequencies.
Finally, the filtered signal underwent Fourier inversion to yield a noise-reduced signal. In the case
of sEMG acquisition, the noise primarily arose from industrial frequency interference related to the
equipment and inherent cardiac interference from the human body. A breakdown of the primary
sources and their associated frequency ranges of noise can be found in Table 5.

Table 5. Noise source and frequency band.

Number Noise Sources Frequency (Hz) Affect

1 Industrial frequency interference 50 One of the main noises
2 Cardiac interference 0–30 One of the main noises
3 Non-muscle collection 0–15 Very low frequency interference

As shown in the table, the noises mixed into the sEMG were primarily low-frequency noises
with frequencies below 30 Hz, with the exception of the 50 Hz industrial frequency interference.
The sampling frequency of the synchronous signal acquisition system was 650 Hz, which meant the
acquired signal frequency range was 0–325 Hz.
To filter out noise while retaining the useful information of the original signal, we designed
both a 30–300 Hz bandpass filter and a 50 Hz IF trap using the MATLAB(MATLAB R2018b) toolbox
Filter Design and Analysis Tool with the ‘filterDesigner’ command. The default setting for the filter
was an infinite impulse response (IIR) filter.
This paper selected the Butterworth filter for both the 30–300 Hz bandpass filter and the 50 Hz
IIR trap. For the 30–300 Hz bandpass filter, we set the stopband cutoff frequencies to 10 Hz and
320 Hz, the passband cutoff frequencies to 30 Hz and 300 Hz, the stopband attenuation to 40 dB, and
the passband attenuation to 1 dB. This resulted in a 10th order Butterworth bandpass filter.
 This paper selected the Butterworth filter for both the 30–300 Hz bandpass filter and
the 50 Hz IIR trap. For the 30–300 Hz bandpass filter, we set the stopband cutoff frequen-
cies to 10 Hz and 320 Hz, the passband cutoff frequencies to 30 Hz and 300 Hz, the stop-
band attenuation to 40 dB, and the passband attenuation to 1 dB. This resulted in a 10th
Sensors 2023, 23, 8801 order Butterworth bandpass filter. 16 of 32
For the 50 Hz trap, we set the passband cutoff frequencies to 47 Hz and 53 Hz, the
stopband cutoff frequencies to 49.5 Hz and 50.5 Hz, the two passband attenuations to 5
dB and 1 dB, respectively,
For the 50 Hz trap, and thethestopband
we set passbandattenuation to 30 to
cutoff frequencies dB.
47 This
Hz and yielded
53 Hz, athe
sixth
stopband
order cutoff
industrial frequency trap.
frequencies to 49.5 Hz and 50.5 Hz, the two passband attenuations to 5 dB and 1 dB, respectively,
Using
and the the parameters
stopband described
attenuation to 30above, the
dB. This signala was
yielded sixthfirst
orderbandpass
industrialfiltered andtrap.
frequency then
Using the parameters described above, the signal was first bandpass
50 Hz trap filtered. Figure 10a displays the waveforms of the sEMG signal from the biceps filtered and then 50 Hz trap
brachii muscle before and after filtering, while Figure 10b exhibits the waveform of the before
filtered. Figure 10a displays the waveforms of the sEMG signal from the biceps brachii muscle
bicepsandsEMGaftersignal
filtering, while
after Figure 10b exhibits the waveform of the biceps sEMG signal after filtering.
filtering.

FigureFigure 10. Comparison

10. Comparison of the signal
of the EMG EMG signal waveforms
waveforms beforebefore andfiltering.
and after after filtering.

As seen in Figure 10, the filter removed the bias voltage and the noise-induced spikes in the
As seen in
original Figure
sEMG, 10, thein
resulting filter removed
a smoother the
and bias voltage
cleaner and the noise-induced
signal waveform with the baseline spikes
amplitude
in thefluctuating
original sEMG,
above and below 0. As shown in Figure 11, after filtering, the power spectrumbase-
resulting in a smoother and cleaner signal waveform with the outside the
Sensors 2023, 23, x FOR PEER REVIEW
line amplitude
30–300 Hz fluctuating above
band registered at 0,and
andbelow 0. As
the power at shown in Figurewas
50 Hz frequency 11,significantly
after filtering, 18 ofThis
the
reduced. 36
powerdemonstrated
spectrum outside
that thethe 30–300
designed Hzeffectively
filter band registered
suppressedat the
0, and the power
influence at 50
of various Hz fre- noises
frequency
quencyon was
the sEMG.
significantly reduced. This demonstrated that the designed filter effectively
suppressed the influence of various frequency noises on the sEMG.

Figure 11.
Figure Comparison of
11. Comparison of the
the signal
signal power
power spectrum
spectrum before
before and
and after
after filtering.
filtering.

For preprocessing the joint angle signals, we employed wavelet transform denoising in MAT-
LAB.For preprocessing
While MATLAB offers the joint angle
a range signals,denoising
of wavelet we employed wavelet
functions, transform
our paper denoising
utilized the one-
in MATLAB. While MATLAB offers a range of wavelet denoising functions,
dimensional automatic noise reduction function, ‘wden’. To determine the threshold value our paper
and
utilized
choose the the one-dimensional
appropriate automatic
threshold function, wenoise reduction
made the followingfunction, ‘wden’.
settings. We opted To determine
for the heuristic
the threshold
threshold value method
calculation and choose the appropriate
to compute threshold
the threshold value andfunction, we soft
selected the made the following
threshold function
for filtering
settings. Wethe wavelet
opted system.
for the heuristic threshold calculation method to compute the threshold
valueInand order to achieve
selected optimal
the soft denoising
threshold results,
function for we assessed
filtering the the denoising
wavelet effect using the
system.
‘smoothness
In orderr’toindex.
achieveThis index, set
optimal at a decomposition
denoising results, we level of four
assessed thelayers, was effect
denoising based using
on the
Coiflet-4 (coif4) wavelet basis. The results of this denoising process are presented in Figure 12.
the ‘smoothness r’ index. This index, set at a decomposition level of four layers, was based
on the Coiflet-4 (coif4) wavelet basis. The results of this denoising process are presented
in Figure 12.
 the threshold value and choose the appropriate threshold function, we made the following
settings. We opted for the heuristic threshold calculation method to compute the threshold
value and selected the soft threshold function for filtering the wavelet system.
In order to achieve optimal denoising results, we assessed the denoising effect using
the ‘smoothness r’ index. This index, set at a decomposition level of four layers, was based
Sensors 2023, 23, 8801 17 of 32
on the Coiflet-4 (coif4) wavelet basis. The results of this denoising process are presented
in Figure 12.

Figure 12.
Figure Comparison of
12. Comparison of the
the joint
joint angles
angles before
before and
and after
after denoising.
denoising.

5.3. Experimental Data Acquisition

5.3. Experimental Data Acquisition
First, the acquired sEMG and joint angle signals underwent preprocessing to cancel out any
First, the acquired sEMG and joint angle signals underwent preprocessing to cancel
noise. Since most of the noise frequencies in the sEMG were below 30 Hz, the Fourier transform was
out anyfor
utilized noise. Since using
denoising, mostaof the noise
bandpass filterfrequencies
with a rangeinof the sEMG
30–300 Hz and were below 30frequency
an industrial Hz, the
Fourier
trap of 50transform was utilized
Hz to effectively filter outfor
thedenoising,
noise in theusing
sEMG.a bandpass filter with a range of 30–
300 Hz Forand an industrial
the joint frequency
angle signal, trap of 50
wavelet threshold Hz to effectively
denoising filterThe
was employed. outnumber
the noise in the
of wavelet
sEMG.
decomposition layers was set to four, and the coif4 function was used as the wavelet basis. This
approach successfully
For the eliminated
joint angle signal,noise in thethreshold
wavelet joint angledenoising
signal usingwas
the one-dimensional
employed. The automatic
number
noise
of cancellation
wavelet function “wden”.
decomposition layers was set to four, and the coif4 function was used as the
waveletFurthermore,
basis. Thistaking advantage
approach of the synchronous
successfully eliminatedacquisition
noise in theof joint
the sEMG
angleand jointusing
signal angle
signals, the starting and ending points of the muscle activity segments were detected based on
the one-dimensional automatic noise cancellation function “wden”.
the angle signal waveforms. This step laid the foundation for the subsequent extraction of the
Furthermore,
characteristics of thetaking
muscleadvantage of the synchronous acquisition of the sEMG and joint
activity segments.
angleThesignals, the starting and ending points ofwere
enrolled participants in the experiment the muscle
requiredactivity
to be insegments were
good health, detected
without any
neuromuscular diseases, and possess flexible and obstacle-free movement in all their joints. To
ensure the validity of the experiment, the participants were not allowed to engage in strenuous
exercise within 24 h before the data collection began, ensuring that their upper limbs were in a normal
condition during the experiment. There were six enrolled participants in this study, comprising four
males (coded as M1~M4) and two females (coded as W1~W2). The basic information about the
enrolled participants is presented in Table 6.

Table 6. Heights and weights of the experimenters.

Members Height (cm) Weight (Kg) Age

M1 174 75 25
M2 176 82 35
M3 180 71 26
M4 183 75 25
W1 166 57 22
W2 170 54 24

In this paper, the upper limb muscles associated with elbow flexion/extension that were
measured included the triceps brachii, biceps brachii, elbow, and brachioradialis muscles. To collect
the EMG signals and angle signals during elbow flexion and extension, three surface electrode patches
were affixed to each muscle: positive, negative, and reference electrodes.
Before placing the electrodes, it was important to enhance the skin’s adherence to the electrodes
by preparing the measurement area. This involved cleaning the area with mild soap and water
to eliminate dirt, sweat, sebum, and other impurities. If the measurement area was covered by
thick hair, it was advisable to trim or shave the hair. Long hairs could obstruct the electrodes from
making close contact with the skin, potentially affecting the signal quality. Since the surface electrode
patch primarily comprised a non-woven backing, pressure-sensitive adhesive, electrode core, and
 Before placing the electrodes, it was important to enhance the skin’s adherence to the
electrodes by preparing the measurement area. This involved cleaning the area with mild
soap and water to eliminate dirt, sweat, sebum, and other impurities. If the measurement
area was covered by thick hair, it was advisable to trim or shave the hair. Long hairs could
Sensors 2023, 23, 8801 obstruct the electrodes from making close contact with the skin, potentially affecting 18 ofthe
32
signal quality. Since the surface electrode patch primarily comprised a non-woven back-
ing, pressure-sensitive adhesive, electrode core, and conductive adhesive, no additional
skin preparations were required. Once the measurement area was completely dry, the
conductive adhesive, no additional skin preparations were required. Once the measurement area
electrodes
was were
completely attached
dry, to thewere
the electrodes skin,attached
as illustrated in Figure
to the skin, 13.
as illustrated in Figure 13.

Sensors 2023, 23, x FOR PEER REVIEW 20 of 36

Figure13.
Figure 13. Surface
Surfaceelectrode
electrodesticking
stickingposition.
position.

The
Thepreparatory
preparatory posture
posture at the beginning
at the beginningof each experiment
of each was set
experiment as follows.
was The subject
set as follows. The
stood
the in an
hand
subject upright
clenched
stood position,
in an in
uprightwith
a fist arms
with thenaturally
position, palm hanging
with facing
arms down, the
upward.
naturally Theforearm and upper
synchronized
hanging down, arm forming
the signal
forearmacqui-
and
asition
straight
upper lineforming
system
arm perpendicular
was worn ontothe
a straight the horizontal
subject’s
line plane,
right
perpendicular and
arm, theshown
as
to the palm of inthe
horizontal hand 14.
Figure
plane, clenched
andPrior in
to aeach
the palm fist
of
with the palm facing upward. The synchronized signal acquisition system was worn on the subject’s
experiment, angular calibration was performed. The calibration method involved having
right arm, as shown in Figure 14. Prior to each experiment, angular calibration was performed. The
the subject relax and naturally hold the preparatory posture for 10 s before the start of the
calibration method involved having the subject relax and naturally hold the preparatory posture for
experiment.
10 s before theThestartangle
of thesignal during
experiment. this
The preparatory
angle posture
signal during was saved and
this preparatory subsequently
posture was saved
denoised and averaged to obtain the baseline value of the angle signal.
and subsequently denoised and averaged to obtain the baseline value of the angle signal.

Figure14.
Figure 14. Experimenter’s
Experimenter’s wear
wear chart.
chart. The
The subject
subject stood
stood in
in an
an upright
upright position,
position, with
with arms
arms naturally
naturally
hanging down, the forearm and upper arm forming a straight line perpendicular to the horizontal
hanging down, the forearm and upper arm forming a straight line perpendicular to the horizontal
plane, and the palm of the hand clenched in a fist with the palm facing upward.
plane, and the palm of the hand clenched in a fist with the palm facing upward.

Theangle
The angledata
datacollected
collected during
during thethe elbow
elbow flexion/extension
flexion/extension movement
movement was referred
was referred to as
to as
the the “original
“original angle”.
angle”. To obtainTotheobtain the actual
actual angle duringangle during
the elbow the elbow flexion/extension
flexion/extension movement, the
movement,
baseline anglethe
wasbaseline
subtractedangle
fromwas subtracted
the original anglefrom
data.the
Thisoriginal
process angle data.
enabled This process
the determination
of the real angle during the elbow flexion/extension movement for further analysis.
enabled the determination of the real angle during the elbow flexion/extension movement
for further analysis.
5.4. Feature Extraction and Quantitative Modeling
The purpose
5.4. Feature of this chapter
Extraction was to extract
and Quantitative suitable EMG signal eigenvalues to lay the foundation
Modeling
for the subsequent modeling. There were two evaluation indices for the prediction effect of the model.
One wasThethepurpose of thiserror
mean square chapter
(MSE)was to extract
between suitable value
the predicted EMG and
signal
the eigenvalues
real value. The to smaller
lay the
foundation for the subsequent modeling. There were two
the MSE value, the better the prediction effect, as shown in Equation (8). evaluation indices for the pre-
diction effect of the model. One was the mean square error (MSE) between the
Another indicator is the regression coefficient (R), as shown in Equation (9). R represents the predicted
value andbetween
correlation the realthe value. The and
predicted smaller
actualthe MSEAn
values. value, theofbetter
R value the prediction
1 indicates effect, as
a strong relationship,
while
shown an in
R value of 0 indicates
Equation (8). a random relationship [36].
In summary,
Another a smaller
indicator MSEregression
is the value andcoefficient
a larger R value
(R), asindicated
shown in a better prediction
Equation effect,
(9). R repre-
.
as described
sents in Equation
the correlation (9). Where
between θ i is the predicted
the predicted angle
and actual value of
values. Anthe model,ofθi1isindicates
R value the actuala
strong relationship, while an R value of 0 indicates a random relationship [36] .
In summary, a smaller MSE value and a larger R value indicated a better prediction
effect, as described in Equation (9). Where i is the predicted angle value of the model,
i is the actual measured angle value, and n is the number of sampling points of the
Sensors 2023, 23, 8801 19 of 32

−
. −
measured angle value, and n is the number of sampling points of the sample. θ and θ are the average
of the predicted angle and the average of the actual measured angle, respectively.
n . 2
∑ θ i − θi
MSE = i=1 (8)
n
Sensors
Sensors 2023,
2023, 23,
23, xx FOR
FOR PEER
PEER REVIEW
REVIEW 21
21 of
of 36
36
−
!
n − . .
 
1
n ∑ θi − θ θi − θ i
i =1
R= s s (9)
By
By comparing
comparing the the root
root mean
mean 1square
n  . (RMS),
square . 2 mean
(RMS), meann
 absolute
− 2 value
absolute

value (MAV),
(MAV), integral
integral ab-
ab-
1
solute ∑ θ − θ ∑ θ − θ
solute value
value (IAV),
(IAV), wavewave length
length (WL),
n(WL), Wilson
i =1
i i
Wilson amplitude
n
amplitude
i =1
i (WAMP),
(WAMP), zero
zero crossing
crossing (ZC),
(ZC),
slope
slopeBysign
sign change
change (SSC),
(SSC), mean
mean power
power frequency
frequency (MPF),
(MPF),value peak frequency
peak(MAV),
frequency (PKF),
(PKF), median
median
comparing the root mean square (RMS), mean absolute integral absolute value
frequency
frequency (MF),
(MF), total
total power
power (TTP),
(TTP), spectral
spectral moments
moments (SM),
(SM), auto
auto regressive
regressive (AR),
(AR), and
and
(IAV), wave length (WL), Wilson amplitude (WAMP), zero crossing (ZC), slope sign change (SSC),
SampEn
SampEn eigenvalues
eigenvalues of
of the
the model
model prediction
prediction effect,
effect, we
we finally
finally selected
selected
mean power frequency (MPF), peak frequency (PKF), median frequency (MF), total power (TTP), the
the fifth
fifth order
order AR
AR
model
model coefficients
spectral coefficients
moments (SM), as the
the input
as auto input values
values
regressive (AR),of the
ofand BP
BP neural
theSampEn neural network.
network.
eigenvalues ofThe
The prediction
prediction
the model merits
merits
prediction of
of
effect,
the
we elbow
finally joint
selected angle
the were
fifth orderthen
AR analyzed
model using
coefficients MATLAB
as the input
the elbow joint angle were then analyzed using MATLAB with error back propagation with
values error
of theback
BP propagation
neural network.
neural
The
neural network
prediction
network (BPNN),
merits of thegeneralized
(BPNN), generalized regression
elbow joint angle were then
regression neural
analyzed
neural network
using(GRNN),
network MATLABsupport
(GRNN), with error
support vector
back
vector
machine
propagation regression
neural (SVR),
network regression
(BPNN), tree
generalized (RT), and
regression ensemble
neural regression
network
machine regression (SVR), regression tree (RT), and ensemble regression learning (ERL) (GRNN), learning
support (ERL)
vector
machine
regression regression (SVR), regression tree (RT), and ensemble regression learning (ERL) regression
algorithms.
regression algorithms.
algorithms.
The
The chosen
chosen window
window width width for
for the
the superimposed
superimposed window window in in this
this paper
paper was was 130
130 sig-
sig-
The chosen window width for the superimposed window in this paper was 130 signal points.
nal points. The results are depicted in Figures
nal points. The results are depicted in Figures 15–20. 15–20.
The results are depicted in Figures 15–20.

Figure 15.
15. Linear
Figure 15.
Figure Linear regression
Linear R
regression R
regression values
R values for
values for the
for the different
the different eigenvalues.
differenteigenvalues.
eigenvalues.

Figure 16. Mean

Figure 16. Mean square
square error
error for
for the
the different
differenteigenvalues.
eigenvalues.
Sensors
Sensors 2023,
2023, 23,
23, xx FOR
FOR PEER
PEER REVIEW
REVIEW
Sensors 2023, 23, x FOR PEER REVIEW
Sensors 2023, 23, 8801 20 of 32

Figure 17. Linear regression R values for the different AR model orders.
Figure 17. Linear
Linear regression
regression R for
R values values for the different
the different AR model orders.
Figure
Figure17.17. Linear regression R values for the AR model orders.
different AR model orders.

Figure
Figure
Figure 18.18.
18. Mean
Mean
Mean square
square
square error
the for
error
error for for the
the different
different AR modelAR
different AR model
model orders.
orders. orders.
Figure 18. Mean square error for the different AR model orders.

Figure 19. Linear regression R values for the different regression algorithms.
Figure
Figure 19.
19. Linear
Linear regression
regression R
R values
values for
for the
the different
different regression
regression algorithms.
algorithms.
Figure 19. Linear regression R values for the different regression algorithms.
SensorsSensors
2023, 2023,
23, x 23,
FOR x FOR PEER
PEER REVIEW
REVIEW 23 of 36
Sensors 2023, 23, 8801 21 of 32

Figure 20. Mean square error of the different regression algorithms.

5.5. BPNN
FigureAn Mean square
20. artificial errornetwork
neural of the different
(ANN)regression algorithms.
is an algorithm designed and developed to sim-
Figure 20. Mean square error of the different regression algorithms.
ulate
5.5. BPNNthe structure and function of the human brain’s neural network. It consists of neu-
rons connected by adjustable weights and exhibits a good adaptive learning ability.
5.5. An
BPNN artificial neural network (ANN) is an algorithm designed and developed to simulate the
Among various neural network algorithms, the back propagation neural network (BPNN)
structure and function of the human brain’s neural network. It consists of neurons connected by
is theAn
adjustablemost widely
artificial
weights andapplied.
exhibits Its
neural structure
network
a good is illustrated
(ANN)
adaptive is an
learning in Figure
algorithm
ability. Among 21,various
whichneural
designed consists
and of an
developed
network
input layer,
algorithms, hidden layer, andneural
output layer. (BPNN)
The input signal iswidely
propagated inItsthe forward
ulate the structure and function of the human brain’s neural network. It isconsists
the back propagation network is the most applied. structure
direction in
illustrated through
Figure 21,thewhich
layersconsists
with adjustable
of an input weights, while
layer, hidden theand
layer, error between
output layer.the
Theoutput
input
rons
and the
signal
connected
desired pattern
is propagated
by adjustable
obtained
in the forward
weights
by thethrough
direction outputthe
and
layer
exhibits
is with
layers
a good
transmitted back
adjustable
adaptive
in thewhile
weights,
learning
reverse
the
Among
error
direction various
between through neural
the output
each network
andlayer
the desired algorithms,
using pattern obtainedthe
its quantized by
lossback
the propagation
output
function.layer neuralback
is transmitted
Simultaneously, network
the
in the reverse
isweights
the most direction through
widely applied.
and thresholds each layer
Its are
of each layer using
structureits quantized
modifiedisand loss
illustrated function.
updated toinobtain Simultaneously,
Figure 21, which
the output the
clos- consis
weights and thresholds of each layer are modified and updated to obtain the output closest to the
est to the
input labelhidden
layer, [37]. layer, and output layer. The input signal is propagated in the f
label [37].
direction through the layers with adjustable weights, while the error between the
and the desired pattern obtained by the output layer is transmitted back in the
direction through each layer using its quantized loss function. Simultaneou
weights and thresholds of each layer are modified and updated to obtain the outp
est to the label [37].

Figure 21.BPNN
Figure21. BPNNschematic.
schematic.

6. Control
6. Control
According to the grading of muscle strength, limbs with a muscle strength below grade two are
According to the grading of muscle strength, limbs with a muscle strength below
unable to resist gravity, especially in the initial stage after a stroke where the patient’s upper limb
grade twomuscles
nerves and are unable
may be todamaged,
resist gravity, especially
resulting in a lack in the initial
of active stage
power. after
In this a stroke
early where
rehabilitation
the patient’s upper limb nerves and muscles may be damaged, resulting in
stage, passive training is prioritized to improve muscle tone, reduce muscle spasms, and guidea lack of active
the
power.limb
affected In this early active
towards rehabilitation stage,The
rehabilitation. passive training
process is prioritized
of passive to improve
control involves muscle
operating the
tone, reduce
system muscle
as planned, spasms,
processing the and guide
output the affected
information, limb towards
and solving activethat
any deviations rehabilitation.
may occur.
The As the stroke
process progresses,
of passive the upper
control limboperating
involves may regain thesome motor
system asability,
planned,and processing
the upper limbthe
rehabilitation
Figure
output 21. robot transitions
BPNN
information, schematic. from leading
and solving the movements
any deviations that mayto following
occur. the patient’s movements.
The focus then shifts to active rehabilitation, with an emphasis on human–machine interaction. Estab-
lishing a closed-loop pathway between the human and the machine can improve the effectiveness of
6. Control in the later stages and restore normal physiological activities of the upper limb.
rehabilitation
According to the grading of muscle strength, limbs with a muscle strength
grade two are unable to resist gravity, especially in the initial stage after a strok
the patient’s upper limb nerves and muscles may be damaged, resulting in a lack o
power. In this early rehabilitation stage, passive training is prioritized to improve
 As the stroke progresses, the upper limb may regain some motor ability, and th
per limb rehabilitation robot transitions from leading the movements to following th
tient’s movements. The focus then shifts to active rehabilitation, with an emphasis on
Sensors 2023, 23, 8801 man–machine interaction. Establishing a closed-loop pathway between 22 of the
32 human
the machine can improve the effectiveness of rehabilitation in the later stages and re
normal physiological activities of the upper limb.
Based
Based on the above on the above
rehabilitation rehabilitation
perspective, perspective,
this paper this paper
proposed three controlproposed
modes, as three co
modes,
depicted in the Figure 22.as depicted in the Figure 22.

Figure
Figure 22. Schematic of 22.
the Schematic of thestrategy.
overall control overall control strategy.

1. Passive mode1. controlled by thecontrolled

Passive mode PID (proportional-integral-derivative) to guide the initial to guid
by the PID (proportional-integral-derivative)
rehabilitation process.
initial rehabilitation process.
2. Active rehabilitation,
2. Activewhere the affected
rehabilitation, side was
where driven byside
the affected the was
healthy side,byallowing
driven for side, a
the healthy
coordinated movement between both sides.
ing for coordinated movement between both sides.
3. Interactive control, enabling the patient to move autonomously, fostering more independent
3. Interactive control, enabling the patient to move autonomously, fostering more i
and self-driven rehabilitation activities.
pendent and self-driven rehabilitation activities.
6.1. PID
6.1. PID
The PID control is known for its simplicity in structure, stable operation, and widespread
application. It has beenThe PID control
extensively used is
in known
the fields for
of its simplicity robotics
rehabilitation in structure, stable operation,
and industrial control and w
spread application. It has been extensively
and remains one of the earliest and most popular control methods [38–40]. used in the fields of rehabilitation robotics
industrial control
One of the advantages and remains
of PID control is that it one
doesofnotthe earliest
heavily relyandon most popular
the model of thecontrol
control methods
object. Satisfactory
40].results can often be achieved by tuning the system parameters appropriately.
In this study, a simple OnePIDofcontroller based on
the advantages of state feedback,
PID control as shown
is that it doesinnot
Equation
heavily(10),
relywas
on the mod
implemented to test
the control object. Satisfactory results can often be achieved by tuning the of
the desired functionality of the developed exoskeleton robot. The trajectory system pa
motion was controlled by manipulatingIn
eters appropriately. thethis
jointstudy,
torques a calculated
simple PID by controller
the PID controller.
based on Thestate
PID feedbac
controller effectively regulated the robot’s movements and ensured its desired performance during
shown in Equation (10), was implemented to test the desired functionality of the d
the rehabilitation exercises.
oped exoskeleton robot. The trajectory of motion was controlled by manipulating the
Z
.
τ = K p e + KV e + K I edt (10)
torques calculated by the PID controller. The PID controller effectively regulated th
where bot’s movements and ensured its desired performance during the rehabilitation exer
. . .
e = qd − q, e = qd − q
 =K
qd − q is the error between the desired joint position
. .
and + Kactual
p ethe Ve + K I edt
joint 
position.
qd − q is the error between the desired joint vector and the actual joint vector.
KP , KI , andwhere
KV are the diagonal matrices for the proportional, integral gains, and derivative,
respectively.
e = qd − selection
The reliability of such a controller depends on the proper
q, e = qdof− the
q proportional (K ),
P
derivative gains (KV ), and integral (KI ). In this paper, the PID gains were set for the controller. The
gains chosen for the experiments were KP = diag (2200, 2000, 2200), KI = diag (50, 40, 50), and KV =
diag (20, 18, 20).
As depicted in Figure 23, the exoskeleton upper limb rehabilitation robot employed joint-based
control, wherein a predetermined position and velocity were input to the controller. Feedback
from the position sensors was obtained to determine the actual position and velocity of the robot,
allowing for the calculation of the error. The controller then estimated the necessary torque, angle,
and angular velocity based on this information. Subsequently, using specific formulas, the obtained
 KI = diag (50, 40, 50), and KV = diag (20, 18, 20).
As depicted in Figure 23, the exoskeleton upper limb rehabilitation robot employed
joint-based control, wherein a predetermined position and velocity were input to the con-
troller. Feedback from the position sensors was obtained to determine the actual position
and velocity of the robot, allowing for the calculation of the error. The controller then
Sensors 2023, 23, 8801 estimated the necessary torque, angle, and angular velocity based on this information. 23 of 32
Subsequently, using specific formulas, the obtained data were transformed into a motor
current, which was referred to as the desired current. This control mechanism ensured
that were
data the robot’s movements
transformed were
into a motor accurately
current, controlled
which was referredand
to asaligned with
the desired the desired
current. tra-
This control
jectory during
mechanism the rehabilitation
ensured that the robot’sprocess.
movements were accurately controlled and aligned with the
desired trajectory during the rehabilitation process.

Figure 23. Principles of the PID control for upper limb rehabilitation robots.
Figure 23. Principles of the PID control for upper limb rehabilitation robots.
6.2. Bilateral Control
6.2. Bilateral
Bilateral Control
coordinated movement of the upper limb is an effective method for neural remodeling.
StrokeBilateral
patients coordinated
often experience unilateralofhemiparesis,
movement the upper limb and using the healthy
is an effective limb tofor
method guide the
neural
rehabilitation training of the affected limb has shown positive results [41–43]. In a unilateral robotic
remodeling. Stroke patients often experience unilateral hemiparesis, and using the
system, the movement parameters of the healthy side are captured by sensors and used for mirror
healthy limb to guide the rehabilitation training of the affected limb has shown positive
control of the affected limb. The electromyographic signals obtained from the healthy side provide
results [41–43].
valuable informationIn about
a unilateral
the motionrobotic system,
intention of thethe movement
upper parameters
limb. By utilizing a BP of the network,
neural healthy
side are captured by sensors and used for mirror control of the affected limb.
the features of the collected electromyographic signals can be extracted to establish a joint prediction The electro-
myographic
model. signals
This model obtained
effectively from
predicts thethe healthy
motion, side control
enabling provide valuable
of the mirror information about
image of the affected
the motion intention of the upper limb. By utilizing a BP neural network, the features of
limb.
In this study,
the collected a Hall sensor wassignals
electromyographic employed cantobeextract the angle
extracted information
to establish from the
a joint healthy
prediction
side’s
model. elbow
Thisjoint
model (theeffectively
circuit principle
predicts is depicted
the motion,in Figure 24). control
enabling The BP neural
of the network (BPNN)
mirror image of
was
the then utilized
affected limb.to predict the angle of the elbow joint. The pre-scaler register (PSC) value was
Sensors 2023, 23, x FOR PEER REVIEW
set to (7200−1 ), with an internal clock (INT) of 72 MHz, resulting in a counting time of 0.1 ms. 26 of 36
In this study, a Hall sensor was employed to − 1 extract the angle information fromThe the
auto-load register (ARR) value was configured as (200 ), generating a counting interrupt period of
healthy side’s elbow joint (the circuit principle is depicted in Figure 24). The BP neural
20 ms, which aligned with the required period for the servo drive. By using the output comparison
networkin(BPNN)
function the PWM was
mode,thentheutilized
duty to predict
cycle the angle byof the elbow the joint. The
of thepre-scaler
CCR value
register was
(PSC) derived
value was from
set the
to Hall−1could
(7200 sensor
), with
bethrough
adjusted
an internal
modifying
digital-to-analog
clock (INT) of
value
conversion
72 MHz,
condition
(ADC)
resulting
code register (CCR). The CCR value was derived from the Hall sensor through digital-to-analog
using a specific
in a counting formula
time of 0.1 transformation.
ms. The auto-load register (ARR) value was configured as (200−1),
conversion (ADC) using a specific formula transformation.
generating a counting interrupt period of 20 ms, which aligned with the required period
for the servo drive. By using the output comparison function in the PWM mode, the duty
cycle could be adjusted by modifying the value of the condition code register (CCR). The

Figure 24. The healthy side drove the affected side control
control principle.
principle.

6.3. Interaction Control

6.3. Interaction Control
Human–robot interaction plays a crucial role in the field of rehabilitation robotics, as it facilitates
Human–robot interaction plays a crucial role in the field of rehabilitation robotics, as
information exchange and interaction between humans and machines. This interaction becomes even
it facilitates information exchange and interaction between humans and machines. This
more important as robots are required to work around people, enabling collaboration and seamless
interactionIn
interaction. becomes
essence, even more important
human–robot interactionasaims
robots are required
to enhance to work
machines’ abilityaround people,
to communicate
enabling collaboration
effectively with humans. Inand
theseamless
context of interaction.
rehabilitationIn essence,
robot human–robot
training, human–robotinteraction
interaction
aims to
leads to increased
enhance patient
machines’ ability to
engagement communicate
and effectively
visualized training data. with humans. In the context
of rehabilitation robot training, human–robot interaction leads to increased patient en-
gagement and visualized training data.
Virtual reality technology is utilized to achieve human–robot interaction in rehabili-
tation robotics. This approach allows for more flexible and natural interactions between
 interaction becomes even more important as robots are required to work around people,
enabling collaboration and seamless interaction. In essence, human–robot interaction
aims to enhance machines’ ability to communicate effectively with humans. In the context
of rehabilitation robot training, human–robot interaction leads to increased patient en-
gagement and visualized training data.
Sensors 2023, 23, 8801 Virtual reality technology is utilized to achieve human–robot interaction in rehabili- 24 of 32
tation robotics. This approach allows for more flexible and natural interactions between
robots and humans, thereby enhancing the robot’s intelligence. As robot technology con-
tinues to advance,
Virtual human–robot
reality technology interaction
is utilized is sethuman–robot
to achieve to become a interaction
significant in
focus in future
rehabilitation
robot development
robotics. This approach[44].allows for more flexible and natural interactions between robots and humans,
thereby Inenhancing
this paper, thewe
robot’s intelligence.
developed As robot
a game scenetechnology continues
using Unity3D andto incorporated
advance, human–robot
six-axis
interaction is set to become a significant focus in future robot development [44].
sensors to facilitate communication between the serial port and Unity3D. Through these
In this
sensors, gamepaper, we developed
objects a game scene
in the Unity3D sceneusing
couldUnity3D and incorporated
be controlled, enablingsix-axis sensors
specific actionsto
facilitate communication between the serial port and Unity3D. Through these sensors, game objects
to be performed. Importantly, the patient retained full control during this process, and the
in the Unity3D scene could be controlled, enabling specific actions to be performed. Importantly,
robot did not provide assistance. The six-axis sensor was fixed at the end of the patient’s
the patient retained full control during this process, and the robot did not provide assistance. The
arm, and
six-axis the was
sensor armfixed
position
at thewas
end ofmapped to thearm,
the patient’s screen, allowing
and the the training
arm position to be tocom-
was mapped the
pleted by extracting the end pose, as depicted in Figure 25.
screen, allowing the training to be completed by extracting the end pose, as depicted in Figure 25.

Sensors 2023, 23, x FOR PEER REVIEW 27 of 36

Figure25.
Figure 25. Unity3D-based
Unity3D-basedgame
gameinteractive
interactiveinterface.
interface.
7.1. Results of the Elbow Joint Angle Measurement
7.
7. Results
Results
Combined with the conclusions of the feature extraction in Chapter 5, the fifth order
7.1. Results coefficients
AR model of the Elbow were
Joint Angle Measurement
extracted from the collected sEMG. The BP neural network
Combined
algorithm was with
thenthe conclusions
utilized of quantitative
for the the feature extraction
modelingin Chapter 5, the fifth
of the sEMG andorder AR angle,
elbow model
coefficients were extracted from the collected sEMG. The BP neural network
with the number of neurons in the hidden layer set to 40. The waveforms of the actualalgorithm was then
utilized for the quantitative modeling of the sEMG and elbow angle, with the number of neurons in
(predicted) elbow angle and the reference (ideal) angle are depicted in Figure 26, where
the hidden layer set to 40. The waveforms of the actual (predicted) elbow angle and the reference
the red solid line represents the actual angle of the elbow joint, and the green solid line
(ideal) angle are depicted in Figure 26, where the red solid line represents the actual angle of the
represents
elbow joint, the
and reference angle.
the green solid line represents the reference angle.

Figure 26. Actual vs. reference

reference values
values for
for the elbow angle (where the red solid line represents the
actual angle of the elbow joint predicted with
actual angle of the elbow joint predicted with the
the BPNN
BPNN and
and the
the green solid line
green solid line represents
represents the
the
reference angle).
reference angle).

The linear
The linearregression
regression relationship
relationship between
between the refence
the refence value
value and the and the
actual actual
value valuea
showed
showed
high a high
R value R valueAdditionally,
of 94.41%. of 94.41%. Additionally, thebetween
the average error average error
the two between
angles wasthe two angles
approximately
was approximately 4° with a MSE of 7.2°^2.When compared with the other literature that
reported angle prediction errors in the range of 5~8° [45], the model established in this
paper demonstrated a better prediction effect.

7.2. Control Results

 The linear regression relationship between the refence value and the actual
showed a high R value of 94.41%. Additionally, the average error between the two a
was approximately 4° with a MSE of 7.2°^2.When compared with the other literatur
reported angle prediction errors in the range of 5~8° [45], the model established in
Sensors 2023, 23, 8801 25 of 32
paper demonstrated a better prediction effect.

7.2.4◦Control Results
with a MSE of 7.2◦ 2 .When compared with the other literature that reported angle prediction errors
This
in the range of 5~8◦ [45], the
experiment aimed
modelto analyzeinthe
established thismotion of a single
paper demonstrated joint,prediction
a better and in effect.
order to val
the7.2.
experimental
Control Results
results, it was analyzed with respect to PID control and healthy
drivenThisaffected sideaimed
experiment control. The the
to analyze robot kinematic
motion of a single model
joint, andwas established
in order to validate using
the the
and Serial Link
experimental functions
results, in thewith
it was analyzed MATLAB Robotics
respect to PID Toolbox.
control and The link affected
healthy side-driven parameters o
side control.
upper The robot kinematic
limb rehabilitation model
robot was established
were input to using
obtainthe its
Linkkinematic
and Serial Link functions
space, as illustrat
in the MATLAB Robotics Toolbox. The link parameters of the upper limb rehabilitation robot were
Figure 27,
input to anditsthe
obtain results
kinematic indicated
space, that
as illustrated in the joints
Figure could
27, and achieve
the results sufficient
indicated movemen
that the joints
could achieve sufficient movement.

Sensors 2023, 23, x FOR PEER REVIEW 28 of 36

Sensors 2023, 23, x FOR PEER REVIEW 28 of 36

Figure 27. Robot spatial motion trajectory.

FigureFor
27. the PID
Robot control
spatial (Figures
motion 28–37), the plots of the joint position versus time and the
trajectory.
error plots of the reference position versus the actual position are provided. The red
dashedFor27.lines
Figure the represent
PID
Robot control
spatial the reference
(Figures
motion values
28–37),
trajectory. for the
the plots position
of the and velocity
joint position versus tracking,
time andwhile
the
the blue
error plots solid
of lines indicate the
the reference actualversus
position values.the actual position are provided. The red
For the PID control (Figures 28–37), the plots of the joint position versus time and the error plots
dashed lines represent
of the reference the the
position versus reference valuesarefor
actual position the position
provided. The red and velocity
dashed tracking,
lines represent the while
the blue solid lines indicate the actual values.
reference values for the position and velocity tracking, while the blue solid lines indicate the actual
values.

Figure 28. Positional error plot of the single-joint flexion and extension motion of the shoulder
joint.
Figure 28. Positional error plot of the single-joint flexion and extension motion of the shoulder joint.
Figure 28. Positional error plot of the single-joint flexion and extension motion of the shoulder
joint.

29. Velocity–time
Figure 29.
Figure Velocity–timediagram
diagramof of
thethe
shoulder unicompartmental
shoulder flexionflexion
unicompartmental and extension.
and extension.

Figure 29. Velocity–time diagram of the shoulder unicompartmental flexion and extension.
 Sensors 2023, 23, 8801 26 of 32

Figure 29. Velocity–time diagram of the shoulder unicompartmental flexion and extension.

Sensors 2023, 23, x FOR PEER REVIEW 29 of 36

Sensors 2023, 23, x FOR PEER REVIEW 29 of 36
Sensors 2023, 23, x FOR PEER REVIEWFigure30.
Figure 30.Motor
Motoroutput
outputtorque
torqueand
andlimb
limb force
force during
during thethe shoulder
shoulder joint
joint single-joint
single-joint flexion–exten-
29 of 36
flexion–extension
sion exercise
exercise (from (from
top to top to bottom:
bottom: the output
the motor motor output
torque, torque, force measured
force measured by the shoulder
by the shoulder cuff
cuff module,
module, and wrist force).
and wrist force).

Figure 31. Shoulder internal and external rotation single-joint training position error chart.
Figure 31. Shoulder internal and external rotation single-joint training position error chart.
Figure31.
Figure Shoulderinternal
31. Shoulder internaland
andexternal
externalrotation
rotation single-joint
single-joint training
training position
position error
error chart.
chart.

Figure 32. Single-joint velocity–time diagrams for shoulder internal and external rotation.
Figure
Figure32.
32.Single-joint
Single-jointvelocity–time
velocity–timediagrams
diagramsfor
forshoulder
shoulderinternal
internaland
andexternal
externalrotation.
rotation.
Figure 32. Single-joint velocity–time diagrams for shoulder internal and external rotation.

Figure33.
Figure Motionoutput
33.Motion outputtorque
torqueandandlimb
limbstrength
strengthduring
duringthe
theunicompartmental
unicompartmentalrotational
rotationalinternal
internal
Figure 33. Motion output torque and limb strength during the unicompartmental rotational internal
and external
Figure movements
33. Motion outputofof the
torque shoulder
and limb joint
strength(from top
during to
the bottom: the motion output torque,
and external
and external movements ofthe shoulder
the shoulderjoint (from
joint top to
(from top tounicompartmental
bottom: the motion
bottom: rotational
output
the motion torque, internal
output strength
torque,
strength
and measured
external
measured by the cuff
by movements
the rotator rotator
the cuff
of module, module,
shoulder and
joint
and wrist wristtop
(from
strength).strength).
to bottom: the motion output torque,
strength measured by the rotator cuff module, and wrist strength).
strength measured by the rotator cuff module, and wrist strength).
 Figure 33. Motion output torque and limb strength during the unicompartmental rotational internal
Sensors 2023, 23, 8801 and external movements of the shoulder joint (from top to bottom: the motion output torque, 27 of 32
strength measured by the rotator cuff module, and wrist strength).

ors 2023, 23,Sensors

x FOR2023,
PEER23, x FOR PEER REVIEW
REVIEW 30 of 36 30 of 36

Figure 34.34.
Figure Elbow flexion
Elbow extension
flexion single-joint
extension exercise.
single-joint exercise.

Figureprofiles
Figure
Figure 35. Velocity 35. Velocity
35. Velocity profiles
profiles
and limb and limb
and
strength limb strength
strength
during during
during
the elbow the elbow
the elbow
uni-joint uni-joint flexion–extension
uni-joint flexion–extension
flexion–extension move- move-
ments
ments (from ments (from top
(from top
top to bottom: to bottom: the
tovelocity–time
the velocity–time
bottom: the velocity–time relationships,
relationships,relationships, strength measured
strength measured by measured by
the rotatorby the rotator cuff
the rotator cuff
cuff
module,
module, and module, and wrist strength).
and wrist strength).
wrist strength).

Figure
Figure
Figure 36. Sagittal 36. Sagittal
36.
plane Sagittal plane reach
plane reach
reach trajectory trajectory diagram.
trajectory
diagram. diagram.
Sensors2023,
Sensors 2023,23,
23,8801
x FOR PEER REVIEW 31
28 of 36
of 32

Figure 37. Receiving

Figure 37. Receiving object
object trajectory
trajectory movement
movement diagram.
diagram.

For
Forthe theshoulder
shoulder flexion–extension
flexion–extension movement, the jointthe
movement, wasjoint
initiallywaspositioned
initially at zero degrees
positioned at
and then extended to 85◦ before returning to 0◦ . The same movement was repeated once more. From
zero degrees and then extended to 85° before returning to 0°. The same movement was
the top curve shown in Figures 28 and 29, it was evident that the actual position closely aligned with
repeated once more. From the top curve shown in Figures 28 and 29, it was evident that
the reference position, indicating that the proposed exoskeleton robot accurately followed the given
the actual
reference position
position. Theclosely
maximum aligned
error inwith the reference
position tracking was position,
only 1.02indicating that the
◦ , demonstrating thatpro-
the
posed exoskeleton
controller exhibited anrobot accurately
excellent trackingfollowed
performance.the given reference
Additionally, the position.
maximumThe maximum
velocity during
error
the in position
repetitive motion tracking was only 1.02°,
was approximately demonstrating
25.3 degrees per second.that the controller exhibited an
excellent
Figure tracking
29 depicts performance. Additionally,
the velocity–time the maximum
curves during velocity during
the unicompartmental the repetitive
movements of the
shoulder
motion was joint.approximately
In both flexion–extension
25.3 degrees motions, the velocity change in the first round-trip motion
per second.
was greater
Figurethan in the second
29 depicts one. Throughout
the velocity–time the during
curves movement, there were fluctuations,
the unicompartmental but these
movements
were characterized by minimal errors.
of the shoulder joint. In both flexion–extension motions, the velocity change in the first
As shown in Figure 30, the top image illustrates the output torque of the reducer during this
round-trip motion was greater than in the second one. Throughout the movement, there
movement. The middle image displays the measured force of the rotating cuff at the shoulder joint,
were fluctuations, but these were characterized by minimal errors.
while the bottom image shows the measured force of the rotating cuff at the wrist joint. Notably, the
gearboxAsoperated
shown within
in Figure 30, the
the rated toprange
torque image illustrates
throughout the output
the process. torque
The force of the
sensor reducer
recorded the
during this movement. The middle image displays the measured force
highest interaction between the subject and the system during the vertical flexion of the shoulder joint. of the rotating cuff
at theForshoulder
the inward joint,
and while
outward theshoulder
bottomrotation,
image shows the measured
the experiment began with forceallof
thethe rotating
joints in the
◦ and returned to 0◦ . This
cuff position.
zero at the wrist joint. Notably,
The shoulder the gearboxrotated
was then horizontally operated
outwardwithin to 45the rated torque range
motion was repeated
throughout at the same
the process. Thespeed.
force The experimental
sensor recordedresults are presented
the highest in Figures
interaction 31–33. The
between the
maximum error ◦ , indicating a good tracking performance
subject and theinsystem
positionduring
trackingthe
wasvertical
approximately
flexion1.21 of the shoulder joint.
of the controller. The maximum speed during the repetitive motion was 13 degrees per second.
For the inward and outward shoulder rotation, the experiment began with all the
However, it was noteworthy that from the error plot, it could be observed that the position sensor
joints in the zero position. The shoulder was then horizontally rotated outward to 45° and
data fluctuated significantly from the ideal data error when the subject performed rotational motion
returned
on to 0°. This
the horizontal plane.motion wasone
Since only repeated at the same
force transducer was speed.
situatedThe experimental
in the results are
rotating cuff assembly to
presented in Figures 31–33. The maximum error in position tracking
measure the forces in the anterior sagittal plane, the interaction between the force transducer and the was approximately
1.21°, indicating
subject was minimized a goodwhentracking performance
the shoulder joint wasof the controller.
rotated inward and The maximum speed dur-
outward.
ing the repetitive motion was 13 degrees per second.
For the elbow flexion–extension motion, the elbow joint was repeatedly However, it was to 95◦ that
noteworthy
extended and
then returned to 0 ◦ . The experimental results are displayed in Figures 34 and 35. From the top curve
from the error plot, it could be observed that the position sensor data fluctuated signifi-
in the figure,
cantly fromitthe was evident
ideal datathat the when
error actual position
the subjectandperformed
the referencerotational
position almost
motioncoincided
on the with
hor-
each other, and the maximum error in position tracking was 0.93 ◦ , indicating an excellent tracking
izontal plane. Since only one force transducer was situated in the rotating cuff assembly
performance of the controller. The maximum speed during the repetitive motion was 26 degrees
to measure the forces in the anterior sagittal plane, the interaction between the force trans-
per second. Since the rudder was chosen as the drive for the elbow joint, the rated torque was
ducer and the subject was minimized when the shoulder joint was rotated inward and
continuously output throughout the operation. The force transducer of the rotating cuff interacted
outward.
most with the subject when the elbow joint was vertically flexed.
For the elbow flexion–extension motion, the elbow joint was repeatedly extended to
95° and then returned to 0°. The experimental results are displayed in Figures 34 and 35.
 the operation. The force transducer of the rotating cuff interacted most with the subject
when the elbow joint was vertically flexed.
For simultaneous shoulder and elbow joint movements, the first movement repli-
cated a forward-reaching motion, starting from the initial position (all the joints at 0° and
Sensors 2023, 23, 8801the elbow at 90°), moving to the reaching position (the shoulder extended to 90° and the 29 of 32
elbow extended to 90°), and then returning to the initial position. As shown in Figure 36,
the results indicated that the developed exoskeleton robot followed the reference trajec-
tory. From the figure, it can be seen that the positional errors of all the joints remained
For simultaneous shoulder and elbow joint movements, the first movement replicated a forward-
below 2°, with the elbow joint having the largest error (1.65°).
reaching motion, starting from the initial position (all the joints at 0◦ and the elbow at 90◦ ), moving
The diagonal
to the catch
reachingmovement
position (the consisted
shoulder ofextended
flexion–extension
to 90◦ and the andelbowthe inward
extendedand to 90◦ ), and then
outward rotation of thetoshoulder
returning the initialjoint, as well
position. As as flexion–extension
shown in Figure 36, the of results
the elbow joint. that
indicated The the developed
movement began with the
exoskeleton elbow
robot and the
followed shoulder
reference joints internally
trajectory. From and externally
the figure, it can berotated at the positional
seen that
0°. First, the shoulder joint
errors of all the was
jointsinternally
remained belowrotated 2◦ ,to 45°the
with and thenjoint
elbow washaving
maintained error (1.65◦ ).
at this
the largest
angle until the endThe diagonal
of the catch movement
movement (located atconsisted
Point 1).of flexion–extension
Subsequently, and the inward
the shoulder joint and outward
rotation of the shoulder joint, as well as flexion–extension
was internally retracted from 90° to 0°, and the elbow joint was rotated from◦ 0° to 90° of the elbow joint. The movement began
with the elbow and shoulder joints internally and externally rotated at 0 . First, the shoulder joint
before the movement concluded. The experimental results are shown in Figure 37. It was
was internally rotated to 45◦ and then was maintained at this angle until the end of the movement
observed that the maximum error between the reference and actual positions of joint two◦
(located at Point 1). Subsequently, the shoulder joint was internally retracted from 90 to 0◦ , and the
was approximately 1.83°.
elbow joint was rotated from 0◦ to 90◦ before the movement concluded. The experimental results
Due to the unilateral disability
are shown in Figure 37. Itofwashemiplegic
observed that patients, special error
the maximum consideration
between the should
reference and actual
be given to the mutualofcoordination
positions of bilateral limbs
joint two was approximately 1.83 in
◦ . bilateral rehabilitation training.
In this study, we chose
Due totothe use the healthy
unilateral upper
disability limb to drive
of hemiplegic the affected
patients, side as a reha-
special consideration should be given to
bilitation taskthe mutual coordination
to promote of bilateral limbs
bilateral coordination. Theingoal
bilateral rehabilitation
of this task was totraining.
achieveInbetter
this study, we chose
coordination to byuse the healthy
equalizing theupper
angle limbof to drive the
motion affected
of the elbow sidejoints
as a rehabilitation
of both upper tasklimbs.
to promote bilateral
However, for hemiplegic patients, accomplishing the equivalent output force of the af- the angle of
coordination. The goal of this task was to achieve better coordination by equalizing
fected elbow motion of the elbow joints of both upper limbs. However, for hemiplegic patients, accomplishing
joint without any assistance or feedback was challenging.
the equivalent output force of the affected elbow joint without any assistance or feedback was
It was previously mentioned that there were still errors in angle prediction using
challenging.
BPNN, and there was It was some delay mentioned
previously in the system. Therefore,
that there sEMG
were still errorswas not introduced
in angle prediction using as BPNN, and
a parameter for the healthy side to drive the affected side’s movement. It was
there was some delay in the system. Therefore, sEMG was not introduced as a parameter for the more reliable
to have the patient
healthywear
sidethe elbowthe
to drive rehabilitator
affected side’s on movement.
the healthyItside wasandmore use the Hall
reliable sensor
to have the patient wear
to obtain the the
jointelbow
angle. To demonstrate
rehabilitator the stability
on the healthy side andof this
use approach,
the Hall sensor the subjects
to obtainwere the joint angle. To
demonstrate
asked to complete a singletheelbow
stability of this
joint approach,
exercise the subjects
repeatedly for were
1000askeds. The to results
complete ofa the
single elbow joint
experiment are exercise
shown repeatedly
in Figure for38,
1000 s. The
with results
a mean of the
error ofexperiment
1.56 degrees. are shown in Figure 38,
This indicated a with a mean
relatively smallerror of 1.56 degrees.
deviation between Thisthe
indicated
desired a relatively
and actual smallelbow
deviation
jointbetween
anglesthe desired
during theand actual elbow
joint angles during the repetitive motion, suggesting that the task was stable and achievable for the
repetitive motion, suggesting that the task was stable and achievable for the participants.
participants.

ors 2023, 23, x FOR PEER REVIEW 33 of 36

Figure 38. TheFigure

healthy
38.side
Theof the elbow
healthy sidedrove
of the the affected
elbow droveside.
the affected side.

8. Conclusions
8. Conclusions
Through research on rehabilitation robots in recent years, significant breakthroughs have been
Through research on rehabilitation robots in recent years, significant breakthroughs
made by research institutes and universities in the field of upper limb rehabilitation robots. The
have been made by research institutes and universities in the field of upper limb rehabil-
development trend has shifted from low degrees of freedom to high degrees of freedom, and from
itation robots. The development trend has shifted from low degrees of freedom to high
degrees of freedom, and from traditional control to artificial intelligence control. In de-
signing upper limb rehabilitation robots, it is essential to consider the precise torque re-
quired for limb operation, taking into account the differences in the structure, control, and
specific treatments.
Sensors 2023, 23, 8801 30 of 32

traditional control to artificial intelligence control. In designing upper limb rehabilitation robots,
it is essential to consider the precise torque required for limb operation, taking into account the
differences in the structure, control, and specific treatments.
This study presented the development of a six-degree-of-freedom (6-DOF) upper limb rehabili-
tation robot with three control modes, an end-pose extraction module, and a mechanical joint capable
of rotation around the center. A hybrid approach was utilized to derive the kinematic equations of
the developed robot. The implementation of PID control algorithms allowed for passive control of
the patient during the prehabilitation period.
The experimental results demonstrated that the robot exhibited low tracking error in joint
motion. For patient active rehabilitation in the middle stage, bilateral control was employed, enabling
patients to actively participate and improve their training efficiency. Before this, an eight-channel
elbow angle prediction system was developed, and the experiments confirmed the effectiveness
of using BPNN for prediction. The patients wore an elbow rehabilitator on the healthy side and
the upper extremity rehabilitation robot on the affected side, utilizing the healthy side to drive the
affected side. During repetitive tracking training, the affected side’s upper limb was able to achieve
basic following, although a slight delay was observed in the data transmission from the Hall sensor
to the register.
For the interaction control, a Unity3D-based game system was developed, enabling the patients
to participate in the game based on the position of their end arm. This approach increased the patients’
interest in rehabilitation and allowed for a better visualization of the rehabilitation training data.
Future work may involve the experimental validation of active rehabilitation and the incorporation
of impedance in the interaction control.
To address the issue of delayed data transmission, several strategies can be employed. Firstly,
elevating the sampling rate can enhance the data density, yielding finer and more detailed information,
consequently bolstering the temporal resolution of events. Secondly, data interpolation and filtering
can be employed. While increasing the sampling rate may augment the data volume, filtering out
high-frequency noise and erratic data can alleviate delays. Thirdly, advanced control algorithms,
such as ahead-lag control or model predictive control, can be implemented. These algorithms, by
predicting future events, enable proactive system responses, reducing reliance on real-time data and
mitigating data transmission delays.
In the future, our research will emphasize the comprehensive evaluation of the rehabilitation
robot efficacy on hemiplegic patients, both in the short-term and long-term rehabilitation phases. We
aim to investigate the effectiveness of upper limb rehabilitation robots combined with multimodal
control for patients across various life stages. Our focus will also extend to enhancing the interaction
design for Unity3D, with the intention of creating games grounded in the patients’ daily activities,
allowing us to assess their impact on the rehabilitation process and overall effectiveness.

Author Contributions: Conceptualization, H.R.; methodology, H.R.; writing—original draft prepa-

ration, H.R.; writing—review and editing, H.R. and T.L.; project administration, J.W.; funding
acquisition, J.W. All authors have read and agreed to the published version of the manuscript.
Funding: This research was supported by the Key Research and Development Program of the
Ministry of Science and Technology of the People’s Republic of China (No. 2020YFC2008700), the
Translational Medicine National Major Science and Technology Infrastructure (Shanghai) Open
Subject Fund (No. TKS-2021-140), the Shanghai Jiao Tong University School of Medicine, Geogao
University Double Hundred Program (No. 20152224), the Shanghai Jiao Tong University School of
Medicine, Translational Medicine Innovation Fund Grant (No. TM201915), the Clinical Research MDT
Program, Ninth People’s Hospital, Shanghai Jiao Tong University School of Medicine (No. 201914),
the National Natural Science Foundation of China (82301158), and the Project of the Shanghai Science
and Technology Commission (22015820100).
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Informed consent was obtained from all the subjects involved in the study.
Data Availability Statement: All the test data mentioned in this paper will be made available upon
request from the corresponding author with appropriate justification.
Acknowledgments: Thank you for the support from Yuling Shen.
Conflicts of Interest: The authors declare no conflict of interest.
Sensors 2023, 23, 8801 31 of 32

References
1. Benjamin, E.J.; Muntner, P.; Alonso, A.; Bittencourt, M.S.; Callaway, C.W.; Carson, A.; Chamberlain, A.M.; Chang, A.R.; Cheng, S.;
Das, S.R.; et al. Heart Disease and Stroke Statistics—2019 Update: A Report from the American Heart Association. Circulation
2019, 139, e56–e528. [CrossRef] [PubMed]
2. Mathew, M.; Thomas, M.J.; Navaneeth, M.; Sulaiman, S.; Amudhan, A.; Sudheer, A. A systematic review of technological
advancements in signal sensing, actuation, control and training methods in robotic exoskeletons for rehabilitation. Ind. Robot. Int.
J. Robot. Res. Appl. 2023, 50, 432–455. [CrossRef]
3. Qassim, H.M.; Wan, H.W. A review on upper limb rehabilitation robots. Appl. Sci. 2020, 10, 6976. [CrossRef]
4. Zhang, L.; Guo, S.; Sun, Q. Development and Assist-As-Needed Control of an End-Effector Upper Limb Rehabilitation Robot.
Appl. Sci. 2020, 10, 6684. [CrossRef]
5. Lee, S.H.; Park, G.; Cho, D.Y.; Kim, H.Y.; Lee, J.-Y.; Kim, S.; Park, S.-B.; Shin, J.-H. Comparisons between end-effector and
exoskeleton rehabilitation robots regarding upper extremity function among chronic stroke patients with moderate-to-severe
upper limb impairment. Sci. Rep. 2020, 10, 1806–1808. [CrossRef] [PubMed]
6. Amici, C.; Ragni, F.; Ghidoni, M.; Fausti, D.; Bissolotti, L.; Tiboni, M. Multi-Sensor Validation Approach of an End-Effector-Based
Robot for the Rehabilitation of the Upper and Lower Limb. Electronics 2020, 9, 1751. [CrossRef]
7. Meng, Q.; Jiao, Z.; Yu, H. Design and evaluation of a novel upper limb rehabilitation robot with space training based on an end
effector. Mech. Sci. 2021, 12, 639–648. [CrossRef]
8. Qian, C.; Li, W.; Jia, T.; Li, C.; Lin, P.-J.; Yang, Y.; Ji, L. Quantitative assessment of motor function by an end-effector upper
limbrehabilitation robot based on admittance control. Appl. Sci. 2021, 11, 6854. [CrossRef]
9. Yuan, R.; Qiao, X.; Tang, C.; Zhou, T.; Chen, W.; Song, R.; Jiang, Y.; Reinhardt, J.D.; Wang, H. Effects of Uni- vs. Bilateral Upper
Limb Robot-Assisted Rehabilitation on Motor Function, Activities of Daily Living, and Electromyography in Hemiplegic Stroke:
A Single-Blinded Three-Arm Randomized Controlled Trial. J. Clin. Med. 2023, 12, 2950. [CrossRef]
10. Brahmi, B.; Saad, M.; Brahmi, A.; Luna, C.O.; Rahman, M.H. Compliant control for wearable exoskeleton robot based on human
inverse kinematics. Int. J. Adv. Robot. Syst. 2018, 15, 6. [CrossRef]
11. Bouteraa, Y.; Ben Abdallah, I.; Elmogy, A.M. Training of Hand Rehabilitation Using Low Cost Exoskeleton and Vision-Based
Game Interface. J. Intell. Robot. Syst. 2019, 96, 31–47. [CrossRef]
12. Islam, M.R.; Brahmi, B.; Ahmed, T.; Assad-Uz-Zaman, M.; Rahman, M.H. Chapter 9—Exoskeletons in upper limb rehabilitation:
A review to find key challenges to improve functionality. In Control Theory in Biomedical Engineering; Boubaker, O., Ed.; Academic
Press: Cambridge, MA, USA, 2020; pp. 235–265.
13. Dalla, G.S.; Roveda, L.; Pedrocchi, A. Review on patient-cooperative control strategies for upper-limb rehabilitation exoskeletons.
Front. Robot. AI 2021, 8, 745018. [CrossRef] [PubMed]
14. De Caro, J.S.; Islam, R.; Montenegro, E.M.; Brahmi, B.; Rahman, M. Inverse Kinematic solution of u-Rob4 an hybrid exoskeleton
for stroke rehabilitation. In Proceedings of the 2021 18th International Multi-Conference on Systems, Signals & Devices (SSD),
Monastir, Tunisia, 22–25 March 2021; pp. 755–764.
15. Bao, G.; Pan, L.; Fang, H.; Wu, X.; Yu, H. Academic Review and Perspectives on Robotic Exoskeletons. IEEE Trans. Neural Syst.
Rehabil. Eng. 2019, 27, 2294–2304. [CrossRef] [PubMed]
16. Gasser, B.W.; Martinez, A.; Sasso-Lance, E.; Kandilakis, C.; Durrough, C.M.; Goldfarb, M. Preliminary Assessment of a Hand and
Arm Exoskeleton for Enabling Bimanual Tasks for Individuals with Hemiparesis. IEEE Trans. Neural Syst. Rehabil. Eng. 2020, 28,
2214–2223. [CrossRef] [PubMed]
17. Sanjuan, J.; Castillo, A.; Padilla, M.; Quintero, M.; Gutierrez, E.; Sampayo, I.; Hernandez, J.; Rahman, M. Cable driven exoskeleton
for upper-limb rehabilitation: A design review. Robot. Auton. Syst. 2020, 126, 103445. [CrossRef]
18. Sanchez-Villamañan, M.D.C.; Gonzalez-Vargas, J.; Torricelli, D.; Moreno, J.C.; Pons, J.L. Compliant lower limb exoskeletons: A
comprehensive review on mechanical design principles. J. Neuroeng. Rehabil. 2019, 16, 55. [CrossRef] [PubMed]
19. Islam, M.R.; Assad-Uz-Zaman, M.; Brahmi, B.; Bouteraa, Y.; Wang, I.; Rahman, M.H. Design and Development of an Upper Limb
Rehabilitative Robot with Dual Functionality. Micromachines 2021, 12, 870. [CrossRef] [PubMed]
20. Paolucci, T.; Agostini, F.; Mangone, M. Robotic rehabilitation for end-effector device and botulinum toxin in upper limb
rehabilitation in chronic post-stroke patients: An integrated rehabilitative approach. J. Neurol. Sci. 2021, 1, 11. [CrossRef]
21. Zhang, W.S.; Yu, H.L.; Huang, X.H. A novel four-degree-of-freedom upper limb rehabilitation robot. J. Chin. Rehabil. Theory Pract.
2019, 10, 1202–1208.
22. Tommaso, P.; Ciaran, O.N.; Lucas, G. Restoring arm function with a soft robotic wearable for individuals with amyotrophic lateral
sclerosis. Sci. Transl. Med. 2023, 15, 681.
23. Zachar, Y.; Nicholas, K.; Christina, C.M. Design of a High-Speed Prosthetic Finger Driven by Peano-HASEL Actuators. Front.
Robot. AI 2020, 7, 586216.
24. Panagiotis, P.; Zheng, W.; Kevin, C.G. Soft robotic glove for combined assistance and at-home rehabilitation. Robot. Auton. Syst.
2015, 73, 135–143.
25. Wang, J.; Wang, W.; Hou, Z.-G. Toward Improving Engagement in Neural Rehabilitation: Attention Enhancement Based on
Brain–Computer Interface and Audiovisual Feedback. IEEE Trans. Cogn. Dev. Syst. 2020, 12, 787–796. [CrossRef]
26. Toledo-Pérez, D.C.; Martínez-Prado, M.A.; Gómez-Loenzo, R.A.; Paredes-García, W.J.; Rodríguez-Reséndiz, J. A Study of
Movement Classification of the Lower Limb Based on up to 4-EMG Channels. Electronics 2019, 8, 259. [CrossRef]
Sensors 2023, 23, 8801 32 of 32

27. Jiang, N.; Dosen, S.; Mueller, K.-R.; Farina, D. Myoelectric control of artificial limbs—Is there a need to change focus? [In the
Spotlight]. IEEE Signal Process. Mag. 2012, 29, 150–152. [CrossRef]
28. McDonald, C.G.; Sullivan, J.L.; Dennis, T.A.; O’Malley, M.K. A Myoelectric Control Interface for Upper-Limb Robotic Rehabilita-
tion Following Spinal Cord Injury. IEEE Trans. Neural Syst. Rehabil. Eng. 2020, 28, 978–987. [CrossRef] [PubMed]
29. Robertson, J.W.; Englehart, K.B.; Member, S.; Erik, J. Effects of Confidence-Based Rejection on Usability and Error in Pattern
Recognition- Based Myoelectric Control. IEEE J. Biomed. Health Inform. 2019, 23, 2002–2008. [CrossRef]
30. Ma, X.; Liu, Y.; Song, Q.; Wang, C. Continuous Estimation of Knee Joint Angle Based on Surface Electromyography Using a Long
Short-Term Memory Neural Network and Time-Advanced Feature. Sensors 2020, 20, 4966. [CrossRef]
31. Namdari, S.; Yagnik, G.; Ebaugh, D.D.; Nagda, S.; Ramsey, M.L.; Williams, G.R.; Mehta, S. Defining functional shoulder range of
motion for activities of daily living. J. Shoulder Elb. Surg. 2012, 21, 1177–1183. [CrossRef]
32. Eckstrand, G.A.; Van Cott, H.P.; Kinkade, R.G. Human Engineering Guide to Equipment Design. Am. J. Psychol. 1973, 86, 891.
[CrossRef]
33. Yang, Z.; Guo, S.; Hirata, H.; Kawanishi, M. A Mirror Bilateral Neuro-Rehabilitation Robot System with the sEMG-Based
Real-Time Patient Active Participant Assessment. Life 2021, 11, 1290. [CrossRef]
34. Muguro, J.K.; Laksono, P.W.; Rahmaniar, W.; Njeri, W.; Sasatake, Y.; Suhaimi, M.S.A.b.; Matsushita, K.; Sasaki, M.; Sulowicz, M.;
Caesarendra, W. Development of Surface EMG Game Control Interface for Persons with Upper Limb Functional Impairments.
Signals 2021, 2, 834–851. [CrossRef]
35. Bu, D.; Guo, S.; Guo, J.; Li, H.; Wang, H. Low-Density sEMG-Based Pattern Recognition of Unrelated Movements Rejection for
Wrist Joint Rehabilitation. Micromachines 2023, 14, 555. [CrossRef]
36. Anastasiev, A.; Kadone, H.; Marushima, A.; Watanabe, H.; Zaboronok, A.; Watanabe, S.; Matsumura, A.; Suzuki, K.; Matsumaru,
Y.; Ishikawa, E. Supervised Myoelectrical Hand Gesture Recognition in Post-Acute Stroke Patients with Upper Limb Paresis on
Affected and Non-Affected Sides. Sensors 2022, 22, 8733. [CrossRef] [PubMed]
37. Zhang, X.; Hu, Y.; Luo, R.; Li, C.; Tang, Z. The Impact of Load Style Variation on Gait Recognition Based on sEMG Images Using a
Convolutional Neural Network. Sensors 2021, 21, 8365. [CrossRef] [PubMed]
38. Ai, Q.; Liu, Z.; Meng, W.; Liu, Q.; Xie, S.Q. Machine Learning in Robot Assisted Upper Limb Rehabilitation: A Focused Review.
IEEE Trans. Cogn. Dev. Syst. 2021. [CrossRef]
39. Luo, S.; Androwis, G.; Adamovich, S.; Nunez, E.; Su, H.; Zhou, X. Robust walking control of a lower limb rehabilitation
exoskeleton coupled with a musculoskeletal model via deep reinforcement learning. J. Neuroeng. Rehabil. 2023, 20, 34. [CrossRef]
40. Kao, P.-C.; Lewis, C.L.; Ferris, D.P. Invariant ankle moment patterns when walking with and without a robotic ankle exoskeleton.
J. Biomech. 2010, 43, 203–209. [CrossRef]
41. Yang, Z.; Guo, S.; Liu, Y.; Hirata, H.; Tamiya, T. An intention-based online bilateral training system for upper limb motor
rehabilitation. Microsyst. Technol. 2021, 27, 211–222. [CrossRef]
42. Deconinck, F.J.A.; Smorenburg, A.R.P.; Benham, A.; Ledebt, A.; Feltham, M.G.; Savelsbergh, G.J.P. Reflections on mirror therapy:
A systematic review of the effect of mirror visual feedback on the brain. Neurorehabilit. Neural Repair 2015, 29, 349–361. [CrossRef]
43. Shi, D.; Zhang, W.; Zhang, W.; Ding, X. Assist-as-needed attitude control in three-dimensional space for robotic rehabilitation.
Mech. Mach. Theory 2020, 154, 104044. [CrossRef]
44. Song, Z.; Guo, S.; Xiao, N.; Gao, B.; Shi, L. Implementation of human-machine synchronization control for active rehabilitation
using an inertia sensor. Sensors 2012, 12, 16046–16059. [CrossRef]
45. Qin, Q.L. Based on SEMG assisted MIMU Research on Personal Localization Technology. Master’s Thesis, Chongqing University,
Chongqing, China, 2017.

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.