Jorge F.

Arinez
Manufacturing Systems Research Laboratory,
GM Research & Development,
Warren, MI 48092
Artificial Intelligence in Advanced
e-mail: [email protected]

Qing Chang
Manufacturing: Current Status
Department of Mechanical and
Aerospace Engineering,
and Future Outlook
University of Virginia,
Charlottesville, VA 22904 Today’s manufacturing systems are becoming increasingly complex, dynamic, and con-
e-mail: [email protected] nected. The factory operations face challenges of highly nonlinear and stochastic activity
due to the countless uncertainties and interdependencies that exist. Recent developments

Downloaded from http://asmedigitalcollection.asme.org/manufacturingscience/article-pdf/142/11/110804/6594922/manu_142_11_110804.pdf by guest on 01 March 2024

Robert X. Gao in artificial intelligence (AI), especially Machine Learning (ML) have shown great potential
Department of Mechanical and to transform the manufacturing domain through advanced analytics tools for processing the
Aerospace Engineering, vast amounts of manufacturing data generated, known as Big Data. The focus of this paper
Case Western Reserve University, is threefold: (1) review the state-of-the-art applications of AI to representative manufactur-
Cleveland, OH 44106 ing problems, (2) provide a systematic view for analyzing data and process dependencies at
e-mail: [email protected] multiple levels that AI must comprehend, and (3) identify challenges and opportunities to
not only further leverage AI for manufacturing, but also influence the future development
Chengying Xu of AI to better meet the needs of manufacturing. To satisfy these objectives, the paper
Department of Mechanical and Aerospace adopts the hierarchical organization widely practiced in manufacturing plants in examining
Engineering, the interdependencies from the overall system level to the more detailed granular level of
North Carolina State University, incoming material process streams. In doing so, the paper considers a wide range of
Raleigh, NC 27695 topics from throughput and quality, supervisory control in human–robotic collaboration,
e-mail: [email protected] process monitoring, diagnosis, and prognosis, finally to advances in materials engineering
to achieve desired material property in process modeling and control.
Jianjing Zhang [DOI: 10.1115/1.4047855]
Department of Mechanical and
Keywords: computer-integrated manufacturing, control and automation, modeling and
Aerospace Engineering,
simulation, plant engineering and maintenance, process planning, production systems
Case Western Reserve University,
optimization, sensing, monitoring and diagnostics, sensors
Cleveland, OH 44106
e-mail: [email protected]

1 Introduction 1.1 Evolution of Artificial Intelligence. AI can be defined as

the “ability of computers to perform cognitive functions associated
Manufacturing is entering a period of substantial innovation and
with human minds, such as perceiving, reasoning, learning, and
change driven by the increased integration of sensors and the
problem solving” [3]. The term “AI” dates back to the 1950s
Internet-of-things (IoT), increased data availability, and advances
with the invention in 1956 of the perceptron, a neural network
in robotics and automaton. This leads to pervasive digitalization
(NN) structure designed to simulate a human neural system by uti-
of the factory and challenges manufacturing enterprises to recon-
lizing a weighted sum of inputs [2,4]. Despite its roots in human
sider, reexamine, and reevaluate their present operations and
learning mechanisms and widespread anticipation that human-like
future strategic directions in the new era known as Smart Manufac-
cognitive AI was well within reach, the development of perceptron
turing and Industry 4.0 [1]. These innovations representing a future
was hampered by its inability to process even the simplest logic
vision of manufacturing motivate the central theme of this paper,
and intractable computational complexity [4]. This quickly led to
namely to examine how artificial intelligence (AI) can play a vital
reduced enthusiasm for supporting AI-related research and resulted
role in realizing these opportunities to advance the manufacturing
in the beginning of the first AI winter. The recovery from the first
industry in a profound new manner. To date, the implementation
AI winter was enabled by expert systems that correspondingly
of AI in modern manufacturing has been built on the progressive
needed handcrafting by experts culminating in extensive series of
development of a series of techniques over many decades, such as
“if-then” rules. This prevailing wisdom caused many to believe
machine learning (ML) [2]. Recent advances in computational hard-
that expert knowledge was the best means to create AI.
ware as well as in sensing technology for the collection of critical
However, the era of knowledge-based AI turned out to be rather
process/machine data have made the application of these AI tech-
short-lived, as the expert systems became prohibitive in the
niques feasible in a practical sense and led to a great interest in
design and maintenance of inherently complex logic, consequently
the capabilities and benefits that they offer. Further, a review of
leading to the second AI winter. ML contributed to the recovery of
state-of-the-art AI applications helps to identify some unique man-
AI from the second winter, especially the invention of deep learn-
ufacturing problems where AI techniques might provide solutions
ing (DL), along with the advances in sensing and computational
and thus significantly improve productivity, quality, flexibility,
infrastructure that have allowed ML/DL models to be established
safety, and cost. Such knowledge and understanding are of great
[4,5]. These techniques are based upon generalization, meaning
benefit to the practical implementation of AI in today’s highly
that they infer the general description of a class by observing the
complex industrial environments that each has its own individual
behavior of individuals in that class [4,5]. The expectation of AI
requirements and context.
has also been realigned to the more focused perspective of
enabling analytics, whereby AI tools are a complement to the
Manuscript received October 23, 2019; final manuscript received July 6, 2020; domain knowledge of human experts in the factory rather than
published online August 13, 2020. Assoc. Editor: Laine Mears. their replacement [6–13].

Journal of Manufacturing Science and Engineering NOVEMBER 2020, Vol. 142 / 110804-1
Copyright © 2020 by ASME
 1.2 Opportunities for Artificial Intelligence in manufacturing, providing analysis at a more granular level with
Manufacturing. The main driving factors and requirements in support from more detailed examples. For example, Refs. [11,12]
most daily manufacturing operations across industries are those of discussed AI for machine condition monitoring and fault diagnosis,
meeting throughput, quality, and cost objectives while ensure a while Ref. [13] provides a comprehensive review of AI in the
safe working environment for all. However, meeting these goals emerging field of human–robot collaboration (HRC).
has become increasingly difficult with the multitude of demands Furthermore, since the topic of AI in manufacturing has wit-
stemming from growing product and process complexity, higher nessed this rapid growth in the amount of published research, to
variability in customer demand and preferences, along with relent- write a review paper such as this requires an organizing construct
less competitive pressures from others in the marketplace to remain as a framework to help assess, review, and synthesize the vast liter-
profitable. Seen from a positive perspective, this tough operating ature in this field of ongoing research, development, and industrial
climate facing the majority of manufacturers provides an opportu- implementation. Therefore, the review presented herein aims to
nity for the unique capabilities of AI over conventional tools and strike a balance between a high-level overview of the advanced
approaches. In particular, the commonplace activity of problem- manufacturing systems and the specific AI techniques as represen-
solving which involves looking for root causes lends itself well to tative examples. The uniqueness of this paper is that it follows a

Downloaded from http://asmedigitalcollection.asme.org/manufacturingscience/article-pdf/142/11/110804/6594922/manu_142_11_110804.pdf by guest on 01 March 2024

AI tools capable of identifying and classifying multivariate, nonlin- hierarchical view to classifying plant systems and processes since
ear patterns in operational and performance data that are hidden to this is how plants are commonly organized, both physically and
the plant engineer. Nowadays, huge amounts of continuously gen- functionally, as illustrated in Fig. 2. This hierarchical view as an
erated data are produced by machines, ambient sensors, controllers, organizational construct is thus a valuable concept from the per-
and labor records, etc. The data could be categorized as follows: spective of system requirements when reviewing the suitability of
(1) environmental data collected from ambient sensors, e.g., room any AI technology to ensure global goals are met and sub-
temperature and humidity, (2) process data collected from sensors optimization is avoided. The hierarchical, three-level decomposi-
on process machines or stations, e.g., machining and grinding tion of the manufacturing system (system-process-material) also
coolant temperatures, power, and heat treat temperature/energy, draws parallels from the ISA-95 framework that defines from the
(3) production operation data recorded in controller systems, e.g., low-level physical processes (for which modeling of the material
timestamps or elapsed time of each part in each operation station, property plays an important role), to the mid-level sensing/monitor-
machine downtime, starvation/blockage, idle time, and shift ing of the processes (the foundation of maintenance and prognostics
scheduling, (4) measurement or check data from product quality and health management, or PHM) and finally, the top-level manag-
inspections, e.g., product diameter, form, and balance. Embedded ing manufacturing operation (the system-level modeling and perfor-
in all this big data are unprecedented opportunities for pattern dis- mance analysis).1
covery that can contain important clues to solve tough problems Given that the field of AI as described above is very broad with a
while offering a complementary understanding of the physical long history and theoretical underpinnings in statistics, optimiza-
meaning of parameters to other physical characteristics of a tion, and computer science, there is an overwhelming amount of lit-
system or process. Coupled with the ability to comprehend high erature to survey, and therefore, this paper has been scoped with
dimensional data, AI provides the ability to transform large three major objectives: (1) review the state-of-the-art applications
amounts of complex manufacturing data, which has become com- of AI to important manufacturing problems, (2) provide and use a
monplace in today’s factory, into actionable and insightful informa- system-level view with which to understand data and process
tion [9–13]. dependencies at the multiple levels that AI must comprehend, and
(3) identify challenges and opportunities to not only further lever-
age AI, but to influence the future development of AI to better
1.3 Hierarchical Approach to Manufacturing Systems. The meet the needs of manufacturing.
topic of AI in manufacturing has attracted much attention in the With these goals in mind, this paper is organized following the
scientific community with the number of publications steadily hierarchical logic of Fig. 2. First, beginning from the overall
growing over the past 40 years, as shown in Fig. 1. Recently, system level, we examine AI in areas such as throughput and
several review publications have also been published on this quality optimization in Sec. 2. Next, we look at a specific
topic. In Refs. [7,8], a high-level, general framework and key ele- example of system control, namely that of human supervisory
ments in smart manufacturing systems and governmental initia- control in the context of HRC in Sec. 3. Following this, we
tives around the globe are presented. The constituent technologies proceed to the process level in Sec. 4, in which the analysis of
such as IoT, cyber-physical systems (CPS), cloud computing, big signals from machines and processes provides new opportunities
data analytics, and information and communications technology to advance the field of diagnosis and prognosis. Section 5 considers
(ICT) and their interrelationships are discussed. A review of the the opportunities provided by AI to improve material processing
ML and DL techniques and their applications in manufacturing is and characterization which are the fundamental building blocks
found in Refs. [9,10], respectively. In these publications, the for reducing the uncertainty of incoming material streams. Future
focus was on the survey of the techniques themselves instead challenges and opportunities are then summarized in Sec. 6, and
of the requirements derived from the manufacturing system. overall conclusions are drawn at the end.
Several other review papers have focused on specific aspects in

2 Artificial Intelligence for Manufacturing System

Optimization
A manufacturing system can mean many things, depending on
the viewpoint taken. In this paper, manufacturing systems comprise
machines, robots, conveyors, and supporting activities such as
maintenance and material handling arranged to produce the
desired product, as shown in Fig. 3. Factory operations are highly
nonlinear and stochastic due to countless uncertainties and interde-
pendencies [14,15]. The performance (hence the global competi-
tiveness) of such modern manufacturing systems is critically

Fig. 1 Growing scientific publications for AI in manufacturing 1

isa-95.com/

110804-2 / Vol. 142, NOVEMBER 2020 Transactions of the ASME

 Downloaded from http://asmedigitalcollection.asme.org/manufacturingscience/article-pdf/142/11/110804/6594922/manu_142_11_110804.pdf by guest on 01 March 2024
Fig. 2 Hierarchical plant organization

dependent on the “optimal control” of material flow through the However, the traditional analytical modeling methods based on
work cells. This section aims to review the recent advancement in queuing theory and Markov Chains [17,18,20,21] suffer from
AI approaches as it applies to manufacturing systems including two considerable disadvantages. The adoption of data analytics
system modeling and performance analysis, and optimal system- and ML techniques offer great potential to compensate for these
level control and decision-making. shortcomings.
First, the analytical methods for estimating throughput, either
exact or approximate, are limited to simple system structures
2.1 Modeling and Performance Analysis. Production system under strict assumptions. For more complex systems, simulation
performance evaluation, diagnosis, and prognosis in terms of turns out to be the only feasible approach to evaluate long-term
productivity, quality, and efficiency are of great importance. system performance. ML methods can be applied to generalize
However, unreliable machines and finite buffers make the material the results from simulations, to avoid repetitive simulation runs
flow in manufacturing systems difficult to model and analyze since when the production system parameters are changed. In Ref. [22],
the former makes it stochastic and the latter nonlinear. the gray model and neural network are combined to predict
system throughput for a multi-product production line considering
2.1.1 Throughput Evaluation. Throughput analysis is aimed rework loops. In Ref. [23], a single hidden layer neural network
at evaluating long-term or short-term productivity of manufactur- is trained to predict makespan and throughput for multi-product
ing systems, which could facilitate system design, performance manufacturing systems considering stochastic cycle times.
improvement, and daily operation of production systems. Sub- Second, conventional throughput improvement approaches focus
stantial amounts of research have been devoted to the analysis mainly on long-term steady-state performance analysis, which are
of manufacturing system dynamics and performance [15–19]. not applicable to real-time throughput prediction and production

Fig. 3 Schematic of a modern manufacturing system

Journal of Manufacturing Science and Engineering NOVEMBER 2020, Vol. 142 / 110804-3
control. Further, they are unable to take full advantage of today’s product quality. However, due to the fact that system-level
vastly superior sensor readings. Short-term system performance, quality diagnosis requires intensive expert knowledge, the
inferred from buffer states and machine status, has proven to be applications of pure ML methods in this context are
useful in real-time control for complex production systems limited. Bayesian networks are one of the methods that can
[24,25]. Improving production system throughput is an important be used to conduct root cause diagnosis of quality issues as
task of production line management and control. In practice, it is it is convenient to infer fault causes from observed quality
often accomplished by identifying the bottleneck machine and deviations. In Ref. [36], a new method of Bayesian
improving its operation. Data-driven methods in identifying network construction is proposed, using a framework
throughput bottlenecks have been extensively studied. Earlier designed for root cause diagnosis in an automotive body
works [15,18] develop an indirect index to identify production bot- assembly process. The proposed method can work with
tlenecks. In Ref. [26], an event-based method is discussed to evalu- small data sets and medium-level noise in measurements.
ate permanent production loss in battery manufacturing. In Ref. [27], In Ref. [37], a distributed diagnostic framework based on a
a data-driven model is developed using available sensor data, and a multi-agent Bayesian approach is proposed and implemented
system diagnostic method is proposed to identify real-time produc- in a lab-based modular assembly system.

Downloaded from http://asmedigitalcollection.asme.org/manufacturingscience/article-pdf/142/11/110804/6594922/manu_142_11_110804.pdf by guest on 01 March 2024

tion constraints and bottlenecks. In Ref. [28], a recency-weighted
stochastic learning method is proposed to predict the system produc- 2.2 Manufacturing System Decision-Making and Control.
tion losses of serial production lines in a small look-ahead window. Manufacturing system control is concerned with decision-making
and controlling of the physical activities in a factory in order to
2.1.2 Quality Analysis. The complex interactions among dif- improve productivity, quality, and overall efficiency. Algorithms
ferent processes in multistage manufacturing systems have pre- at the system level are used to decide what and how much to
sented significant challenges for quality analysis in the system produce, when production is to be finished, how and when to use
level. Classical statistics-based approaches, e.g., control charts and the resources or make them available, when to release jobs
run charts, which mainly focus on process-level quality assurance, and which jobs to release, job routing, and job/operation sequencing
might not be applicable to system-level analysis, since the quality [38]. The approaches can be categorized into centralized or hierar-
deviation in a particular stage is not only determined by that stage chical control approaches and distributed multi-agent-based
but also by upstream stages or sub-assemblies. Model-based control approaches. Traditional centralized manufacturing control
methods for quality analysis, e.g., the stream-of-variation model approaches and software packages are developed and adapted case
[29], provide an analytical approach to examine quality issues by case and lack flexibility, expandability, agility, and reconfigur-
through modeling the generation and propagation of quality varia- ability [38]. On the other hand, multi-agent-based control
tions. However, model-based methods often require substantial approaches derived from distributed AI techniques provide several
domain knowledge that is not always readily available. ML-based important benefits such as robustness, reconfigurability, and respon-
quality analysis is well suited to situations where there is limited if siveness [39].
any, domain knowledge regarding the process physics and system Smart manufacturing seeks to increase factory productivity and
dynamics. These analysis approaches have been found to achieve the efficient utilization of resources in real-time [40]. To achieve
one or more of the following objectives by applying the AI/ these objectives, manufacturing systems need to transform large
ML-based methods. amounts of data into manufacturing knowledge and useful actions
in order to become more responsive to market changes and
(1) Early detection of quality defects. For multistage manufac- random disruption events. It is crucial to consider adaptive plan-
turing systems, undesirable quality deviations might propa- ning, scheduling, and control for dynamic manufacturing environ-
gate along stages and result in defects of final products. It ments as key research issues in smart manufacturing management.
is of great benefit to detect quality deviations in earlier
stages, which would save production resources and reduce 2.2.1 Job Dispatching and Scheduling. In flexible manufactur-
the overall cost by promptly reworking defective parts. Con- ing systems, the job dispatching problem arises when there are mul-
ventionally, statistical process control (SPC) is commonly tiple product orders awaiting processing during the same time
used to monitor the quality of work-in-process parts by window. It is critical to determine which order should be prioritized
inspecting several important variables according to given and which machine should undertake the job. In this context, the job
specification ranges. Despite its widespread usage, SPC has dispatching problem seeks to find the optimal job sequences and job
its drawbacks in that, for example, the data are assumed to routes considering various internal factors, e.g., machine capability
be unimodal, which is not always true. More importantly, and availability, and external factors, e.g., order due dates, such that
SPC is dependent on clear, known relationships between one or more of the following minimizing objectives can be
inspected variables and final product quality. With the vast achieved: mean tardiness, mean stay time, and energy consumption.
amounts of time-series data, some ML-based approaches Over the past few decades, the job dispatching problem has been
[30–35] are proposed to effectively overcome some of the widely studied in the literature with many different reported
deficiencies in conventional process control methods. In approaches. Exact approaches, e.g., branch-and-bound (B&B) and
Ref. [34], a case study is presented on the quality control integer programming, are feasible for small-scale job dispatching
of work-in-progress products in the powder metallurgy problems, while heuristic-based methods, e.g., simulated annealing
process based on autoencoders and recurrent neural networks and genetic algorithm, might deliver dispatching rules that lack
(RNNs). In Ref. [35], a combination of cluster analysis and adaptiveness to system dynamics [41]. To cope with the increasing
support vector machines (SVM) is introduced to achieve complexity of manufacturing systems, ML-based methods, includ-
the goal of improved quality monitoring. Compared with ing both supervised learning and reinforcement learning (RL), have
quality prediction, the early detection of quality defects is been extensively applied in this area.
more challenging since the latter needs to discover the rela- Supervised learning involves learning from labeled examples
tionship between incomplete observations up to specific provided by an expert with domain or application knowledge.
stages and quality characteristics of the final product with Moreover, supervised ML is frequently used due to the data-rich
time-series data. but knowledge-sparse nature of problems in the manufacturing
(2) Root cause diagnosis of quality issues. In multistage manu- environment. Conventionally, simple heuristic dispatching rules
facturing, one defective product might result from one or are used in production scheduling, such as SPT (shortest process
multiple imperfect quality stages. Identifying the root cause time), COVERT (cost over time), EDD (earliest due date), and
for quality issues is of great practical value in industries FIFO (first in, first out) [42]. However, adhering to a single dis-
since corrective actions can be taken promptly to improve patching rule does not necessarily deliver better performance than

110804-4 / Vol. 142, NOVEMBER 2020 Transactions of the ASME

dynamically switching dispatching rules according to system states. human operators, and machines. Resource allocation is an optimiza-
Inspired by observation, classification problems are formulated by tion process whereby limited manufacturing resources are allocated
taking the relevant system dynamic variables as inputs, and over time among parallel and sequential activities in order to achieve
simple dispatching rules as outputs. Commonly used ML algo- a desirable system performance. Resource allocation problems may
rithms in this context include Decision Tree [43–45], Neural arise in both the design stage and daily operation. A typical resource
Network [46–48], SVM [41,49,50], and ensemble learning allocation problem in the design stage is the buffer allocation
methods [41]. Despite the ML algorithms, the authenticity of train- problem [63], i.e., allocating buffer spaces to a production line
ing data is the prerequisite to reliable production scheduling. given the total available spaces. In Ref. [64], a decision support
Although simulation (e.g., Refs. [41,46–49]) is a typical source system is proposed for buffer allocation, where a neural network is
for training data, it suffers from the disadvantage that data might trained to quickly predict system throughput given system design
be biased if the simulation is incapable of representing real opera- parameters. The resource allocation problems in daily operations
tions. Therefore, researchers in Refs. [43,44,50] attempt to avoid are further complicated by the highly uncertain and fast-evolving
such bias by aggregating multiple data sources, including simula- factory environments, compared with that in the design stage. In
tion, historical data, and expert knowledge. In Ref. [45], a real-time, daily operations, resource allocation is often coupled with task

Downloaded from http://asmedigitalcollection.asme.org/manufacturingscience/article-pdf/142/11/110804/6594922/manu_142_11_110804.pdf by guest on 01 March 2024

big data framework is established to collect, process, and store assignments. RL has been widely applied to tackle problems in
actual data from the shop floor upon which a real-time production this area. In Ref. [65], an RL-based method is proposed for dispatch-
scheduling and rescheduling method are implemented. ing material handling dolly trains in a general assembly line, wherein
RL is suitable for a model-free problem with delayed conse- the dolly train delivers materials to workstations and carries multiple
quences, in which the model dynamics are unknown and must be esti- types of parts at a time. In Ref. [66], a gantry assignment problem in
mated through interactions of agents with the environment. The production lines is also formulated as an RL problem and solved by
applications of RL methods in job dispatching problems are not the Q-learning algorithm. In both studies, random factors, such as
new since job dispatching is a sequential decision-making problem machine failures in Ref. [66] and product queue lengths in
in dynamic environments. Extensive RL-based research, including Ref. [65], drive the transition of the system states, which are difficult
Refs. [51–54], has been reported for distinct applicable scenarios. to obtain the complete state transition models. RL fits such sequen-
In Ref. [51], by adopting the decentralized Markov decision tial decision-making problems well and can solve them in a model-
process (DEC-MDP) framework, processing machines are modeled free way with various algorithms. Nonetheless, RL problem formu-
as distributed RL agents and a policy gradient algorithm is used to dis- lation needs careful analysis and a thorough understanding of the
cover near-optimal dispatching rules. In Ref. [52], a relational RL system dynamics. In Ref. [67], a comprehensive simulation study
approach is proposed to obtain policies for efficiently rescheduling is conducted on the same gantry assignment problem presented in
production plans, which is able to handle abnormal and unplanned Ref. [66] with a focus on comparing different reward settings,
events such as inserting an arriving order. In Ref. [53], a scheduling which demonstrates that the knowledge-guided reward setting out-
method based on variable neighborhood search (VNS) is proposed to performed all other four heuristic-based reward settings.
obtain job dispatching decisions considering random job arrivals and With the advancement of industrial IoT, AI and specifically ML
machine random failures, in which RL is used to obtain proper param- will see increasing utilization penetrating the entire manufacturing
eter for VNS at a rescheduling decision point. There are dozens of system. ML algorithms potentially provide powerful tools to
RL-based research approaches tackling the job dispatching reduce cycle time and scrap, improve quality, and improve resource
problem where slightly over 80% of this work adopts the tabular utilization in certain NP-hard manufacturing problems [68].
Q-learning algorithm, a powerful, but hard to scale, off-policy RL
algorithm. With the rapid advancement of RL in recent years, a
great deal of novel algorithms have emerged, including the deep 3 Artificial Intelligence for Manufacturing
Q-network (DQN) [55], which is able to solve large-scale RL prob-
lems by integrating deep neural networks (DNN). In Ref. [54], the Applications of Human–Robot Collaboration
DQN is applied to solving a large-scale scheduling problem in a semi- This section specifically examines the utilization of AI in the
conductor manufacturing system. The novel RL algorithms like context of HRC. Consistent with this paper’s use of a hierarchical
DQN, which have largely enhanced learning efficiency and scalabil- framework to structure and organize the review of AI in manufactur-
ity, are expected to help solve more sophisticated and practical pro- ing systems, the topic of HRC is one that uniquely straddles both the
duction scheduling problems in the future. system and machine/process layers. HRC may be defined [13] as “a
Maintenance activities aim to keep machines in desirable reliabil- state in which a purposely designed robot system and an operator
ity levels or to quickly recover them from random failures. As main- work on simultaneous tasks within a collaborative workspace, i.e.,
tenance work orders can be modeled as a sequential decision- where the robot system and a human can perform tasks concurrently
making problem given machine states and system states, RL has or even jointly.” The impact of AI on HRC is particularly important to
also been used in Refs. [56–60] to obtain near-optimal maintenance review given the continually increasing diversity and quantity of
policies for manufacturing systems. Many studies [57,58,60] inte- robotic applications in manufacturing. Traditionally, the objective
grate multi-agent-based learning and control in their methods, of applying robotics in manufacturing has been to leverage the advan-
helping dismantle complex structural and operational dependencies tages robots have over humans such as repeatability, endurance,
among components. Compared to traditional maintenance policies strength, ability to operate in hazardous environments, etc. Industry
[61], e.g., periodic policy and age-dependent policy, the RL-based has long recognized that utilizing robots to perform these types of
maintenance policies are more adaptive to the manufacturing tasks frees human operators to more efficiently focus on their inherent
system dynamics and therefore yield better system performance. natural advantages, namely, those related to cognition, adaptability,
The adaptiveness of the RL-based maintenance policies provides a ambiguity, and flexibility. However, with the advent of increasingly
great opportunity to build real-time maintenance decision-making more powerful and capable AI/ML tools, this relative balance of
systems by making full use of real-time data analytics on the plant skills, functionality, and types of capabilities is continuously shifting
floor. In Ref. [62], a conceptual framework for maintenance sched- and becoming less clear as the sophistication and computational
uling is proposed to schedule condition-based maintenance in power and application domain of these technologies increase.
smart manufacturing systems. Correspondingly, the body of work in HRC is a very broad set of
industrial research areas that drives an equally diverse set of poten-
2.2.2 Resource Allocation. In a manufacturing system, tially applicable robot-related AI technologies. Reviewing such a
resource allocation is the process of assigning manufacturing resulting breadth and depth of literature is challenging without a
resources for specific time periods to the set of manufacturing pro- roadmap that can offer guidance as to what has been developed,
cesses, where resources include but are not limited to robots, what is state of the art, and what important problem areas remain

Journal of Manufacturing Science and Engineering NOVEMBER 2020, Vol. 142 / 110804-5
to be solved. While there can be no definitive or absolute categori- of human–robot relationships that maybe used to describe particular
zation of such a vast field of work, as we have done for the overall human–robot operational environments.
paper, we also propose a framework adapted from human supervi-
sory control theory and combine it with a classification scheme for
3.3 Human Supervisory Control at the Intersection of
describing the characteristics of HRC. In this manner, the resulting
Human–Robot Collaboration. Sheridan [77,78] defines a para-
descriptive framework is used to cross-reference AI-based technol-
digm composed of human supervisory control functions wherein
ogies with key aspects of HRC and therefore highlight foundational
a human supervises a generalized machine whether it be an entire
work and the problem domain for which AI/ML is either already
automation system, a robot, a computer, or any other physical
being deployed, or in the future may be well suited to solving
device capable of receiving commands and carrying out actions cor-
other more difficult problems.
respondingly. The specific functions consist of planning, teaching,
monitoring, intervening, and learning. However, in our paper, we
augment these five functions with a sixth function, that of operating
3.1 Overview of Human–Robot Collaboration. Over the
a system to explicitly call out task execution activities by both the
decades since robots were first introduced into industrial environ-
human and robot in HRC systems. Though this section discusses

Downloaded from http://asmedigitalcollection.asme.org/manufacturingscience/article-pdf/142/11/110804/6594922/manu_142_11_110804.pdf by guest on 01 March 2024

ments, manufacturers have sought to gain ever more efficiencies
a collaboration between a human and a robot, there is no loss in gen-
and capabilities from these significant capital investments. Simulta-
erality by considering the human as a supervisor in this collabora-
neously, the role of the human operators and the nature of the inter-
tive setting since Sheridan’s paradigm allows for varying degrees
action with robots and robotic systems have evolved in conjunction
of supervision between humans and automation. Each of these six
with the increasing functionality and application space where they
functions is potential area for the application of a wide variety of
have been installed. However, with recent developments in many
AI/ML tools since they are all fundamentally performed by a
robot capabilities such as sensing, perception, control, teaching,
human supervisor. In this regard, Sheridan’s classical paradigm of
and learning the very nature of the human–robot relationship is
human supervisory control provides a suitable perspective
being examined with a focus toward attaining greater productivity,
through which to analyze human-collaborative robotic systems in
flexibility, and adaptability [69,70]. These benefits are driving man-
manufacturing that seek to leverage recent advances in AI/ML.
ufacturing firms to move their human–robotic systems from one of
Given this well-established foundation, we reexamine the paradigm
coexistence toward collaboration.
in light of today’s advances in robotics to consider these six func-
However, despite the attraction of the potential benefits that HRC
tions extended into the robotic domain as described in greater
promises, there are still key technical challenges and requirements
detail in Sec. 3.4.
that need to be addressed and overcome before improvements in
productivity can be realized. Chief among these are the industrial
safety requirements that prohibit human operators and robots 3.4 Integrated Perspective for Surveying Human–Robot
from sharing a common workspace, and as such is an ongoing Collaboration and Artificial Intelligence Manufacturing
area of research [71–73]. Other challenges concern how robots Literature. Using the above foundation, we make use of Fig. 4
understand and interpret the uncertainty in the behavior, actions, as an organizing framework to map AI/ML technologies to existing
and intentions of the human(s) that comprise an HRC system and potential industrial HRC applications and find common themes
[74]. Nevertheless, advances are being made by some robot manu- across problem types and corresponding AI/ML solutions.
facturers to develop commercial robotic solutions that can begin to
meet some of these challenges such as in industries having 3.4.1 Plan. Planning [77,78] comprises three elements and
light-duty, low-powered, or low-payload applications. Though activities: system modeling, setting and meeting objectives, and for-
firms are developing innovative solutions to solve practical indus- mulating strategies. In an example of planning from an assembly
trial problems, HRC still represents a large area of ongoing research application [79], Bayesian sequential decision-making is investi-
before industry-wide adoption can occur. However, advances in AI gated to determine the optimal allocation of sensing modes
technologies may enable and accelerate the development and imple- between purely autonomous and manual states. A shared vision
mentation of HRC into the industry. system is considered wherein a combination of observational
modes exists ranging from solely human-based, mixed human–
robot based, and solely robot based with the goal being to determine
3.2 Characterizing the Human–Robot Relationship. When the optimal allocation decision among these three modes. In the
reviewing the literature, the type of AI techniques that have been development of “intelligent agents” [80], vision technology and
developed and implemented for HRC depending on the specific planning activities are examined for how AI can integrate these
details of the HRC manufacturing application such as the complex- two areas. Another facet of HRC planning [81], considers the
ity and type of an assembly operation and on the nature of the assignment of roles and allocating tasks among the humans and
human–robot relationship. The latter is particularly important as robots in a collaborative system. Under the Sheridan supervisory
it governs how a manufacturing operation is executed jointly by control paradigm, it is given that the balance between roles and
the human and robot and warrants further examination. In this tasks is specified in the original system design. In Ref. [81], a frame-
regard, four types of HRC have been defined [75] wherein the work is proposed for lab-based experiments of a human and
degree of collaboration is classified according to how closely robot collaborative system to assemble a basic household item of
humans and robots work together on a specific manufacturing oper- furniture that enables reasoning to make such types of decisions.
ation comprised of processes and workpieces. These range from (i) This reasoning framework utilizes partially observable Markov
complete independence of different processes and workpieces, (ii) decision processes (POMDP) capable of performing in uncertain
synchronization of different processes with a common workpiece, environments thus providing the ability to plan adaptively. In
(iii) simultaneous different processes and common workpiece, other work [82], an “anticipatory control” method is developed to
and (iv) simultaneous collaboration on common process and proactively plan robot actions based on anticipation of the human
workpiece. operator’s intention during task performance. Here, the AI tech-
Further work [76] classifies various human–robot relationships nique employed is that of an SVM to predict gaze patterns of the
from the perspective of sharing within five types of situations or operator and in turn predict task intentions.
scenarios, namely that of workspace, direct contact, work task,
simultaneous process, and sequential process. These scenarios are 3.4.2 Teach. Secondly, teaching or training, involves commu-
then mapped to four different types of human–robot relationships: nicating commands to a robot(s) execute specific actions. In an
coexistence, interaction, cooperation, and collaboration. This HRC assembly application [83], a machine learning framework is
tabular construct then permits a rich and comprehensive description proposed to identify successful snap-fit assembly operations and

110804-6 / Vol. 142, NOVEMBER 2020 Transactions of the ASME

 Downloaded from http://asmedigitalcollection.asme.org/manufacturingscience/article-pdf/142/11/110804/6594922/manu_142_11_110804.pdf by guest on 01 March 2024
Fig. 4 Illustrative sets of actions, functions, requirements, characteristics, and behaviors for
human operators and robots which are candidates for application of AI/ML approaches

transfer them from human to robotic operations. This framework of the robots’ actions and progress toward completion of all tasks
utilizes force profile features to identify types of snap-fits based and respond to any anomalous events that may arise during opera-
on data sets generated via HRC having high variability enabling tion. Observing or monitoring includes analysis of the robot system
machine learning, in this case, an SVM classifier was trained. data from many data sources and types while considering the overall
This is an example of where HRC is used for training to develop system state as part of maintaining situational awareness. Consider-
autonomous robotic operations. In Ref. [84], a system based on ing robotic monitoring capabilities, the term “Internet of Robot
natural language is proposed to improve the teaching and robot pro- Things” (IoRT) [86] describes intelligent devices that can monitor
gramming task. The system is built on a semantic network as the events, fuse sensor data from a variety of sources, use local and dis-
basis for a natural language processing system comprised of auto- tributed intelligence to determine the best course of action. A
matic speech recognition, visual simulation environment, and rea- system architecture that enables monitoring [87] can integrate
soning. Human voice commands are inputs to the reasoning observational factors such as work patterns during operation (dis-
system which also takes inputs from the natural language processor cussed previously). RL is used in the estimation of factors such
to search for a reasonable set of robotic actions. In this work, the as human kinematics via a recursive least-square (RLS) algorithm.
semantic hierarchical network improves system robustness by pre- In a more comprehensive use of AI in monitoring [88], a deep
venting incorrect or irrelevant robotic tasks from occurring such as a learning-based multimodal fusion architecture is developed for
robot picking up a worktable instead of a workpiece thus yielding the robust operation of an HRC manufacturing system. Three
faster object referencing. modes of communication between humans and robots are consid-
ered in this work: voice, hand motion, and body motion. These par-
3.4.3 Operate. Operation concerns the required active steps, ticular modalities for monitoring provide an intuitive approach for
processes, tasks that are completed independently, sequentially, human supervisors or operators to interact with robots. Further,
or simultaneously by the human(s) and the robot(s) in HRC an architecture integrating these modes provides a more flexible
systems to achieve planned objectives. Depending on the type of alternative to the more prescriptive approaches of Refs. [81,89]
application and system, these operational steps or processes may and more closely represents the uncertain, stochastic nature of mon-
or may not be pre-planned which will accordingly drive the itoring human performance interacting with robotic systems. More-
requirements of a given AI technique chosen to improve opera- over, each of these individual modalities is addressed by a unique
tional efficiencies. Moreover, the lack or inability to pre-plan AI approach before multimodal fusion is carried out. Voice
HRC operations notwithstanding, an even greater challenge command recognition is formulated as a CNN, hand motions as
facing HRC is the inherent uncertainty and variability in human an LSTM (RNN), and body motion makes use of MLP transfer
task performance times. Though it may be argued that such vari- learning.
ability is the cost of flexibility provided by human operators, this
characteristic of HRC nonetheless requires the development of 3.4.5 Intervene. The fifth element of HRC relates to the inter-
approaches to ensure productivity is maintained. An online AI vention of operations and considers what alternative steps or correc-
technique [85] developed for automatic and unsupervised cluster- tive actions must be pursued if, in the course of monitoring, the
ing of basic HRC operational steps uses real-time force/torque human or robot observes the occurrence of a fault or anomaly. In
data to address the challenge in human cycle time variability. human–robotic systems with limited interaction, a human supervi-
The specific AI method developed and experimentally tested sor will need to alter the system operation by overriding with
uses dynamically trained one-class support vector machines manual controls or by stopping and re-teaching the system to
(OCSVMs) to discover states of manufacturing process steps. prevent reoccurrence of the fault. An important aspect of interven-
This type of online algorithm demonstrates the ability to realize tion in the control of the HRC system is the capability for real-time
real-time performance without the penalty of requiring labeled decision-making. In complex and ambiguous situations, human
data from training phases. operators have a relative advantage compared to robots owing to
human cognitive abilities by virtue of experience. The requirement
3.4.4 Monitor. Next, monitoring requires the human operator for complex event processing on the robotic side of the collabora-
or supervisor to continuously observe and evaluate the execution tion is an opportunity that AI can enhance. For example,

Journal of Manufacturing Science and Engineering NOVEMBER 2020, Vol. 142 / 110804-7
Ref. [90] presents a layered architecture for HRC connected to industrial revolution witnessed the expansion of data collection
higher-level global plant functions and supporting the use of AI and the adoption of AI methods [95]. These methods, especially
tools. This framework allows for data collected at the level of an ML, can inductively learn useful data patterns and relate to process-
HRC work cell to train for unknown/unplanned scenarios that can ing condition and performance, with the underlying concept that
be detected during operation via supervised learning algorithms data are the reflection of the physical machine characteristics. As
that detect known categories of failure scenarios. the complexity of manufacturing processes and data continues to
grow, it became evident that the traditional ML techniques, which
3.4.6 Learn. Finally, the sixth is that of learning which Sheri- rely on high-quality feature engineering that is subjective by
dan [77] considers broadly to be the human cognitive activity of nature, are inherently limited. Data in modern manufacturing,
gaining knowledge from observing the results of the system’s per- such as images for surface defect detection and sequences that
formance. Other investigators [74] have also similarly observed that underlie machine degradation, are increasingly requiring progres-
as humans and robots work more collaboratively in industrial sive, hierarchical representation which can decompose the com-
systems, a convergence and need emerges for improved teaming plexity into small pieces for more effective handling [5]. The
and integration of human actions with robot training and learning. current wave of the fourth industrial revolution sees the ML land-

Downloaded from http://asmedigitalcollection.asme.org/manufacturingscience/article-pdf/142/11/110804/6594922/manu_142_11_110804.pdf by guest on 01 March 2024

In the telerobotic paradigm [77], teaching and learning are directed scape transformed into deep learning (DL), which leverages
from the human supervisor to the robot. In contrast, a robot teaming advanced computational infrastructure, such as cloud computing
model [74] is integrated with a human operator mental model by [96,97], and cascade of network layers that corresponds to different
means of a Markov decision process (MDP). This work highlights levels of abstraction of machine condition-related patterns [5],
how learning can extend across the collaboration space and can be thereby improving the effectiveness of defect diagnosis and RUL
bi-directional in nature. An example [89], of integrating planning, estimation [11]. The classification of diagnosis and prognosis
teaching, and learning into robot functions alleviates the human methods is illustrated in Fig. 5. The objective of this section is to
from these supervisory functions. In this work, a learning approach present a summary of major AI techniques that have advanced
is developed based on symbolic AI (inductive logic programming, the state of monitoring, diagnosis, and prognosis, which are of crit-
or ILP) for task execution in cognitive robots. ILP enables planning, ical value to modern manufacturing.
execution, and learning framework where a set of hypotheses are
constructed, updated, or discarded as the robot gains new knowl- 4.1 Artificial Intelligence for Condition-Based Maintenance.
edge in the form of further observations. Specifically, experiential Manufacturing systems are subject to faults caused by a broad
physical learning is combined with an adaptive planning strategy range of causes, from heat generation to corrosion and fatigue.
so that feedback is provided to a mobile robot (in lab environment) Fault diagnosis is intended to pinpoint the source of fault generation
to improve future task performance thereby robustness. Another and provide an estimate of the severity of the fault, which forms the
key need in advancing HRC is being able to understand and learn basis for condition-based maintenance. The idea of AI-based diag-
the wide range of activities performed by the human operator. nosis is to formulate the task as classification and associate the
This ability involves being able to infer human intentions along information embedded in data to the fault types and severity
with the myriad of complexities that this objective entails. In a levels. Related research works are presented in this section.
very focused study [91], an algorithm is developed to model nonlin-
ear human motions using an artificial neural network (ANN) based
4.1.1 Associating Features With Faults. ML-based fault diag-
on position and velocity data with online learning. In Ref. [92], an
RNN-based human motion trajectory predictive model parses the nosis represents the earliest adoption of AI in manufacturing indus-
try [98], and the success of the ML techniques is largely attributed
interaction among human body parts for more accurate trajectory
prediction. Further, Monte-Carlo dropout was investigated to to the good understanding of certain machine and process based on
which the physics-informed features from data, such as those in
measure the prediction uncertainty and improve the model robust-
time, frequency, and time-frequency domain, can be effectively
ness. A broader approach [93] is developed whereby a three-
pronged, integrated teaching, learning, operating strategy is extracted [11]. With good quality features, ML has demonstrated
its capability in effective fault diagnosis, as reviewed in this
adopted. This approach consists of the human first teaching the
section. The most common research area for ML-based diagnosis
robot via natural language instructions, and thereafter, the robot
is rotating machinery [11].
learns from human assembly demonstrations via an RL algorithm.
In Ref. [99], the performance of different ML algorithms, includ-
Once the teaching-learning phase is completed, this learned knowl-
edge is used during the operation to actively assist during collabo- ing SVM, k-nearest neighbor (KNN), and random forest (RF), has
been comparatively studied for induction motor fault diagnosis.
rative assembly tasks.
Four different sensing signals were evaluated as input: vibration,

4 Artificial Intelligence For Process Monitoring,

Diagnostics, Prognostics
The history of manufacturing is closely associated with the
history of improving the reliability of machine equipment to
reduce unexpected downtime. Research on machine condition mon-
itoring, defect root cause diagnosis, and remaining useful life (RUL)
prognosis establishes the core competence for maintenance and
PHM. The goal is to enable timely detection and isolation of the pre-
cursors or incipient faults of components, predict their progression,
and support rational decision-making. The significance of the main-
tenance and PHM system in manufacturing lies in revealing the
health state of individual machines and/or the production system
in real-time, providing a diagnosis of anomalies’ root cause, and
preventing the occurrence of a failure, all to ultimately achieve near-
zero downtime [94]. Research on PHM started to take shape in the
20th century [95], with the initial focus on the physics-based under-
standing of the working mechanisms of machines and processes. Fig. 5 Classification of diagnosis and prognosis methods: D,
The shift from manual to automated production in the third diagnosis and P, prognosis.

110804-8 / Vol. 142, NOVEMBER 2020 Transactions of the ASME

current, voltage, and flux. Features were extracted from time and fre- textures, it has shown that the process faults can be accurately
quency domain as well as the parameters from a statistical regression detected in a timely manner in order to avoid producing low-quality
model. The results have shown that RF has the most robust perfor- products.
mance across different types of sensing signals while SVM has the Beyond images captured by vision systems, DCNNs have been
best overall fault detection accuracy. Authors in Ref. [100] reviewed increasingly extended to analyze signals that are in other forms
Naive Bayes, KNN, and SVM for induction motor fault diagnosis by nature. In Ref. [109], a diagnostic system has been built to deter-
based on the current sensor signal. Differing from prior work, mine the gearbox fault severity level. Specifically, the 1D vibration
current envelope analysis has been carried out first to reveal the signals were transformed into time-frequency images in order to
amplitude-modulated nature of the motor current signal to generate better leverage the DCNN architecture for fault-related pattern rec-
candidate features. Then, the minimum redundancy maximum rele- ognition. The flowchart of this research is shown in Fig. 6. In
vancy (mRMR) method selected the most “dominant” feature subset Ref. [110], the authors reported a novel method for both bearing
as input. It was shown that each technique has a different level of fault type and severity level recognition. One-dimensional vibration
sensitivity to different features. The authors also confirmed that signals from the sensors were first converted to images using
the best performance is achieved by SVM. wavelet packet transform. Then, they were analyzed by DCNN

Downloaded from http://asmedigitalcollection.asme.org/manufacturingscience/article-pdf/142/11/110804/6594922/manu_142_11_110804.pdf by guest on 01 March 2024

In Ref. [101], a novel method for bearing fault diagnosis based on for fault-related pattern recognition. Three fault types and four
visual words and sparse classifiers has been developed. The wavelet severity levels were evaluated using the developed method, and a
time-frequency spectra of vibration signals are first analyzed for near-perfect accuracy was achieved. A similar work has been
energy-related feature extraction. Then, each signal was encoded reported in Ref. [111]. In Ref. [112], a 1D DCNN-based method
by visual words or feature clusters, which were used as input to a for bearing fault diagnosis has been developed that takes advantage
sparse classifier to determine bearing fault type. The sparse classi- of the shift-invariance of the convolution operation in order to elim-
fier has also been investigated for gearbox fault severity level rec- inate the need for time-domain signal alignment. This allows the
ognition [102]. The contribution of this work also includes a DCNN to directly analyze the collected sensing signals, without
novel multi-sensor fusion method based on the covariance matrix, having to rely on feature extraction or transformation to the fre-
which allows pair-wise correlation among sensing signals to be esti- quency domain. More recently, researchers have begun to focus
mated and incorporated into the analysis. In both works, the authors on understanding DL mechanisms with the aim of facilitating the
reported that the diagnosis accuracy from a sparse classifier is com- broad acceptance of the technique. An early work has been reported
parable to an SVM with reduced computational time. Sparse classi- for DCNN-based motor diagnosis in which the layer-wise relevance
fiers have also been investigated for wind turbine condition propagation (LRP) has been investigated to visualize the frequency
monitoring and fault diagnosis [103]. band that the DCNN is focused on when distinguishing different
Beyond rotating machinery, ML techniques also contributed to motor structural faults [113].
improved fault recognition capability in other manufacturing equip- Besides DCNNs, machine fault diagnosis based on deep belief
ment. In Ref. [104], faulty tonnage in a stamping machine is networks (DBN) has also been reported. The main idea of DBNs
detected by SVM using vibration signal features extracted by a is to first use the stack of restricted Boltzmann machines to progres-
recurrent plot (RP) method. In Ref. [105], a novel method for sively improve the feature discriminability. Then, the obtained fea-
detecting filament breakage and nozzle clogging in fused deposition tures pass through a multi-layer perceptron for fault classification.
modeling (FDM) has been developed, with an SVM as the condi- In Ref. [114], a DBN has been investigated to classify bearing
tion classifier. The contribution of the work is the Bayesian Dirich- health conditions over the bearing lifecycle. The ant colony optimi-
let method to effectively characterize the sensing signals. zation (ACO) algorithm was investigated in order to find the
optimal DBN parameters that maximize the classification accuracy.
4.1.2 Learning Features From Data. With increased process
and data complexity, manual feature extraction becomes difficult.
DNN-based ML, which allows automated learning of features spe- 4.2 Artificial Intelligence-Based Prognosis for Predictive
cific to each task, has attracted increasing attention. Accordingly, Maintenance. Prognosis aims at predicting the temporal progres-
AI-based diagnosis has evolved in the sense that (1) tasks can be sion of machine performance degradation, from its current state to
increasingly customized to fit specific needs, such as simultaneous final functional failure. Reliable RUL estimation contributes to
identification of fault type and fault severity level, and (2) sensing timely maintenance. In general, AI-based prognosis is part of the
data are increasingly transformed into proper forms suitable for dif- data-driven method that relies on establishing a machine perfor-
ferent NN architectures, leading to novel diagnosis concepts. mance evolution model to predict future machine performance
As sensing technology has continually evolved and adopted on based on its current and past status. To estimate RUL,
the factory floors, image data has been increasingly gathered and one-step-ahead prediction is iteratively carried out until the pre-
analyzed for process monitoring and fault diagnosis. Compared to dicted value passes a failure threshold [97]. In case that the
time-series data that has been traditionally measured, images have
a higher information density and their use has been shown to be
advantageous in gaining insight into the objects being monitored.
How to effectively use image data has been the main research
focus of DL. One of the frontier applications of image analysis is
an additive manufacturing (AM), given the spatial information
that is inherently embedded during the layer-by-layer printing
process. Image analysis has shown to reveal critical information
associated with the quality of the printed part [106]. In
Ref. [107], a deep ANN model is reported to detect laser power
deviation based on melt-pool images in selective laser melting
(SLM). Due to the localization of the melt pool, the classification
model has shown to be capable of inferring the location of potential
microstructural defects of the printed part. In Ref. [108], a vision
system based on deep convolutional neural network (DCNN) has
been built for process fault detection in SLM. As surface textures
caused by various process faults may share similar local features,
higher-level abstractions are required to effectively distinguish dif-
ferent fault-related patterns. By using DCNN to analyze the surface Fig. 6 DCNN-based gearbox fault severity level diagnosis

Journal of Manufacturing Science and Engineering NOVEMBER 2020, Vol. 142 / 110804-9
machine performance is difficult to measure directly, an artificial ineffective. In such scenarios, the NN architectures, especially RNN
health index (HI) is often created from sensor data to represent and its variant long short-term memory (LSTM), offer several
the machine performance [115]. The HI-based prognosis and advantages in comparison: (1) they can explicitly model the inner
RUL estimation are illustrated in Fig. 7. Related works for relationship among machine performance at different time steps;
AI-based machine prognosis are presented in this section. (2) they can decompose the complex temporal patterns into a
series of simple components which, individually, can be approxi-
4.2.1 Recognizing Degradation Patterns. Common ML tech- mated by a single regression at each network layer, before being
niques such as SVM and RF, although initially developed for the assembled to fully describe the degradation trend. These advantages
task of classification, can be effectively formulated as regression. have made RNN/LSTM an attractive option for machine prognosis.
By taking performance from current and past time steps as ML The research highlights are presented as follows.
input, they can be associated with the future performance at In Ref. [121], an RNN-based bearing prognosis method has been
output and therefore establishing the relationship underlying the developed. Vibration signals from a defective rolling bearing were
machine degradation. Out of all ML techniques, SVM has been transformed using continuous wavelet transform. Statistical param-
the most commonly used for machine prognosis. eters computed from both the raw data and the pre-processed data

Downloaded from http://asmedigitalcollection.asme.org/manufacturingscience/article-pdf/142/11/110804/6594922/manu_142_11_110804.pdf by guest on 01 March 2024

In Ref. [116], SVM was investigated for bearing prognosis. The were then utilized as candidate inputs to an RNN. Analysis has
vibration, temperature, and sound sensing signals were first trans- shown that the developed method is accurate in predicting
formed by wavelet packet decomposition into features in the time- bearing defect progression. In Ref. [122], a bi-directional LSTM
frequency domain. Then, an isometric feature mapping reduction for aircraft engine RUL estimation has been developed. The HI is
technique was applied to fuse these features into HI, representing constructed by a single-layer NN which fuses the on-board
the status of the bearing. Finally, the temporal pattern among HI sensing signals to represent the engine performance. The
at different time steps was analyzed by SVM for future status pre- bi-directional LSTM allows the information to flow forward for pre-
diction and RUL estimation. In Ref. [117], an SVM-based diction and backward for disturbance smoothing. The developed
method for compressor degradation prognosis has been developed. method has shown to improve RUL prediction accuracy as com-
The HI of the compressor was quantified by vibration, temperature, pared to uni-directional LSTM and traditional ML techniques
pressure, and moisture sensing signals, as well as the parameters of such as SVM. A similar work has been reported in Ref. [123] in
an autoregressive moving-average (ARMA) model. The HI which improved RUL estimation for the lithium-ion battery has
sequence was then analyzed by SVM to predict future progression been achieved using the LSTM-based approach. Specifically, the
of compressor performance. A similar work has been reported for Monte Carlo method has been investigated to construct an ensemble
lithium-ion battery prognosis in which the loss of rated capacity of LSTMs, in contrast to using one single LSTM. This approach
was used as the HI of battery, with SVM as the ML technique [118]. allows the RUL estimation to be displayed as a probability distribu-
In contrast to the aforementioned works in which an evolution tion, rather than a deterministic value. Furthermore, it has shown to
model is first constructed to reveal the temporal relationship be more robust under noisy sensing signals. In addition to revealing
among machine performance at different time steps and subse- degradation patterns, LSTM has also been investigated to model the
quently, leveraged to predict future performance and estimate relationship between sensing signals and machine performance for
RUL, other researchers have investigated the method of directly HI construction. Related works have been reported in Ref. [124].
associating current/past performance to RUL. In Ref. [119], an Recently, LSTMs have also been investigated in conjunction
SVM has been investigated for direct RUL estimation of aircraft with model-based techniques, such as particle filters (PF), in
engine based on sensing signal features from current and past order to alleviate the limitation of insufficient observations for
time steps and therefore, bypassing the steps of generating HI. In degradation model parameter estimation. PFs have been widely
Ref. [115], an ensemble RUL estimation method has been devel- studied for the prognosis of complex engineering systems such as
oped for induction motor prognosis. During the operation, 13 chan- HVAC [125] and aircraft engines [126] in which the degradation
nels of signals from seven sensors were collected, representing model can be continuously fine-tuned by the incoming sensing
voltage, current, vibration, load, speed, temperature, and sound. observations based on Bayesian inference. In Ref. [127], an inte-
Five single-layer NNs with different initialization constitutes the grated method has been developed for fuel cell RUL estimation.
ensemble algorithms. A similar work based has been reported for For an on-going degradation sequence, an LSTM trained on histor-
aircraft engine performance prognosis in Ref. [120], in which the ical data was leveraged to provide degradation path forecast,
ensemble algorithms involve an SVM, a relevance vector beyond what has been observed from the actual fuel cell. These
machine (RVM), an exponential model, and a quadratic model. additional forecast data help PFs more accurately estimate the
To adaptively weigh the contribution from each algorithm, an degradation model parameters, leading to improved estimation
optimization-based fusion method has been developed and shown accuracy.
to be able to improve the estimation accuracy and robustness. As computational capability and infrastructure continue to
advance, it can be expected that advanced AI algorithms will con-
4.2.2 Analyzing Degradation Patterns. When the relationship tinue to emerge in the era of the fourth Industrial Revolution.
among degradation time steps becomes too complex to be character- Prior research in this area has established a solid foundation for
ized by a single regression model, traditional ML techniques become further advancement toward the realization of smart manufacturing.

Fig. 7 HI-based machine prognosis and RUL estimation

110804-10 / Vol. 142, NOVEMBER 2020 Transactions of the ASME

5 Artificial Intelligence for Manufacturing Process different combinations of tooling parameters. This enables tool
Control life modeling using the ANN to accurately predict future wear
[144].
This last section deals with the lower levels of the manufacturing
system hierarchy, namely, the machine/equipment level as it extends
into the material level domain. While the previous section treated the 5.1.2 Artificial Intelligence for Online Control of
condition and state of the machine/process as it relates to reducing Manufacturing Processes. Manufacturing processes are generally
equipment failures, this section focuses on modeling and control multivariable and time-varying by nature. As an example, grinding
of the process to achieve desired process quality under the assump- is highly nonlinear and its multi-modal relationships are difficult to
tion that equipment failures are not causing quality issues. Further- be defined accurately via mathematical equations since output var-
more, this section specifically explores the role AI techniques can iables are not only dependent on input variables but also dependent
provide to significantly increase understanding of the material-to- on other output or intermediate variables. Furthermore, most exist-
process relationship as an important new opportunity for enhanced ing analytical models for complex grinding processes are rather
process control, efficiency, quality, and productivity [128,129]. limited and can only describe partial relationships between design
(control) variables and process (output) variables, while all the

Downloaded from http://asmedigitalcollection.asme.org/manufacturingscience/article-pdf/142/11/110804/6594922/manu_142_11_110804.pdf by guest on 01 March 2024

The impact of an improved understanding of this relationship
spans the classical manufacturing process control and materials pro- parameters must be considered simultaneously for optimal opera-
cessing communities. The former is provided with newfound vari- tion of such processes. In such cases, online process control faces
ables and means to better optimize process control and the latter a major challenge. Conventional feedback control or optimization
will be potentially dramatically changed in the context of designing techniques encounter severe limitations in dealing with such prob-
new materials and materials processing to more directly optimize lems. Because of these difficulties, when a new process must be
downstream manufacturing processes. designed or an existing process must be controlled, an engineer
attempts to utilize all other available resources besides analytical
models, such as expert knowledge, experimental data, handbooks,
5.1 Artificial Intelligence for Modeling and Control of and vendor information. Even if pertinent information is available,
Manufacturing Processes. To date, the initial AI techniques that integrating all the heterogeneous information and designing an
have been applied in manufacturing process control have been optimal condition is not an easy task, often requiring a long lead
focused on applications where there are limited formal or analytical time and much trial-and-error experimentation. In view of the
models due to their complexity or the resource-intensive computa- complex nature of the grinding process and stringent finish, accu-
tions required to develop tunable process controllers. However, as racy, and part surface integrity requirements, the current practice
with other areas in the factory, the increasing amount of instrumen- of designing or controlling grinding processes leave an opportunity
tation and the ability to collect large amounts of data along with the for improvement using more robust approaches. This has led to
ability to extract necessary process signals and features will con- the development of an “intelligent” approach to online control of
tinue to facilitate these types of applications which do not require the grinding processes [146,147] and other strategies that use
an accurate mathematical model nor formal physics-based simula- AI-based controllers based on grinding forces or other in process
tion of the system [130–134]. A survey of traditional manufacturing data such as tooling deflection.
processes such as machining, joining, stamping, and molding indi-
cate this type of benefit may be obtained by using AI techniques
such as SVM, RF, DNN, and genetic algorithms [135–139]. 5.2 Modeling the Effect of Manufacturing Conditions on
Material Property. One of the challenges faced in manufacturing
processes is determining what parameters to adjust and what are the
5.1.1 Artificial Intelligence for Modeling and Optimization of levels to adjust. The selection of these parameters can be a major
Manufacturing Processes. An overarching objective of implement- contributing factor to the quality of the final manufactured parts.
ing an AI tool in manufacturing process control is to produce high- The ability to collect a vast amount of data on these changes can
quality parts cost-effectively [140]. The implication is that critical help in increasing efficiency and quality [148,149]. AI has been suc-
process parameters need to be captured in the ML model. For cessfully implemented when there is a large pool of data to be
example, a laser manufacturing study [141] used an ANN with trained on. When enough of these data points are collected, they
parameters pertaining to laser power, cutting speed, and pulse fre- can be used as a data-driven way to predict properties and results
quency which was critical model factors in determining process of experiments in a fraction of the time. AI’s applications can be
success. In order to optimize these parameters, the ANN was devel- applied to both macroscopic and microscopic properties for predic-
oped to predict laser cutting quality expressed as explicit non-linear tion, covering the whole spectrum of possibilities. For example, the
functions. The results demonstrated that the ANN successfully properties of materials, such as hardness, melting point, and molec-
determined a set of parameters that best optimized the quality of ular atomization energy, can be classified and described at either the
the cutting process and predicted outputs based on given inputs to macroscopic or microscopic level [150]. In most cases when the
the laser. macroscopic performance of materials is studied, the focal point
In other applications of AI to metal cutting, specifically milling is geared toward the structure-performance relationship [151]. AI
process control [142,143], the use of predictive modeling is benefi- applications in microscopic property prediction can concentrate
cial in two ways. The first being the improvement of process effi- on several aspects, including and are not limited to the microstruc-
ciencies [144]. This encompasses maintaining a real-time ture, the lattice constant, electron affinity, and molecular atomiza-
knowledge of the current mill conditions and creating a stable envi- tion energy [150,152–155]. Material’s microstructure can be
ronment for the tool, thus increasing tool life. In conventional characterized through image data such as scanning electron micro-
milling processes, a major task is to ensure desired surface rough- scope (SEM) as well as transmission electron microscope (TEM).
ness which is a parameter that normally degrades with increased
tool wear. Typically, routine tool schedules change worn cutting
inserts before a given surface quality threshold is exceeded [145].
A variety of ML tools have been used to model predictions in 6 Challenges and Opportunities for Future Research
surface roughness changes. In this manner, tool wear and supplier Factories are getting smarter as companies are increasingly able
power can be considered based on surface roughness deviations. to leverage AI to transform information from various aspects of
The other major improvement from ML practices is the prediction the manufacturing system into actionable insights. However,
of wear time. Process control via ML can then proactively adjust many gaps still exist that should be addressed to ensure that AI
the process parameters to increase its life span. Notably from the can be seamlessly integrated into factory operations. Five related
data perspective, ML requires a large database to account for topics are summarized here as recommendations for future research.

Journal of Manufacturing Science and Engineering NOVEMBER 2020, Vol. 142 / 110804-11
 6.1 System-Level Analysis. ML has seen increasing utiliza- Transfer learning includes a series of techniques to adapt a model
tion across all levels of the manufacturing system hierarchy. that has been well established in a source domain in which training
However, compared with the successes of ML in specific applica- data are sufficient, to a related, target domain in which data is
tions of process monitoring, optimization, and PHM, utilization is scarce. For condition monitoring and fault diagnosis/prognosis, the
limited at the system level of decision-making [68]. This is primar- most noted advantage of transfer learning is to allow experimental
ily attributable to the stochastic and non-linear dynamical nature of data obtained from machines without faulty conditions for model
manufacturing systems and the complex multi-stage processes and construction on real production lines when faculty conditions may
dependencies among vast amounts of heterogeneous data generated occur [157]. Reported research on this topic this field can be charac-
therein. Furthermore, although the general advantages of ML lie in terized into three categories: (1) model transfer among different
its ability to handle NP-complete problems, typical of intelligent working conditions [158,159], (2) model transfer among different
optimization problems, the appropriate selection of techniques machines [160], and (3) model transfer among different fault loca-
and algorithms remains challenging. An in-depth understanding tions or types [161]. Further examples of transfer learning include
of a problem, its causes, consequences, and desired solution state human action recognition [162] for human–robot collaboration in
must be known or well investigated to improve the likelihood of assembly. These examples illustrate that, through the proper transfor-

Downloaded from http://asmedigitalcollection.asme.org/manufacturingscience/article-pdf/142/11/110804/6594922/manu_142_11_110804.pdf by guest on 01 March 2024

effective AI tool selection and subsequent model building, data mation that corresponds to the specific manufacturing scenarios,
analysis, and interpretation. These matters all deal with the need transfer learning can address model scalability, opening up the pos-
to have adequate domain expertise during the problem definition sibility of establishing a general rather than a limited solution to a
phase which is vital to ensure that all aspects of the problem are specific manufacturing aspect. On the other hand, it is noted that
well understood and no key data or assumption is overlooked. transfer learning at the current stage is still limited to trial-and-error,
Additionally, at the highest level in the manufacturing hierarchy, ad-hoc approaches. More rigorous theoretical research on the topics
there is a variety of interacting plant control systems that govern such as transferability of data remains essential to facilitating the
overall plant performance. Though single ML approaches have broad applicability and acceptance of transfer learning-based
been developed, no single AI tool or even suite of tools have yet techniques.
to integrate and bridge all of the performance objectives of these Data synthesis refers to generating synthetic data that highly
control systems. RL has been suggested as an approach. resembles the real data by sufficiently learning data characteristics,
However, practical difficulties to train an RL algorithm in a real which allows us to bypass the limitation of data availability. Tradi-
operating system where productivity cannot be jeopardized tionally, data synthesis often relies on interpolation (e.g., Synthetic
remain a challenge. Minority Over-sampling Technique or SMOTE), which cannot
capture complex data characteristics [163]. A major breakthrough
came with Generative Adversarial Networks (GAN) [164], a DL
6.2 Data Quality. The increasing availability of heteroge-
method that is able to learn salient features and synthesize data
neous data in manufacturing systems presents new challenges.
with high fidelity. The basic concept of GAN is a competition
First, the data can contain a high degree of irrelevant and redundant
between a generator, which analyzes real data to produce synthetic
information while the relevant part may be missing. These data cura-
ones, and a discriminator, which distinguishes the synthetic data
tion issues present a challenge for the application of ML algorithms
from real ones. These two are trained iteratively to improve the
as the availability and quality of the manufacturing data have a
capability of each. The final result is an equilibrium state, resulting
strong influence on the performance and suitability of AI algorithms
in a synthesis of the manufacturing data with high-fidelity. Early
relative to expected results. Second, the quality of any knowledge
works of GAN implementation have been reported. For example,
generated from data analysis depends on the context developed
Ref. [165] investigated using synthetic vibration signals to establish
when collecting and managing the data itself. Given the variety of
the gearbox fault diagnostic model. Spectral analysis has shown that
viewpoints in production, it is critical that data collection and man-
the synthetic data effectively captured the fault-related signal fea-
agement approaches support multiple viewpoints for different appli-
tures, such as the characteristic frequencies. As with other DL tech-
cations by dynamically linking different data, information, and
niques, a deeper analysis of the theories related to GAN remains
models [156]. One future opportunity associated with improved
essential to fully establishing the technique as a viable solution to
data quality is to converge local correlation-based predictions
the issue of data limitation.
(e.g., machine tool health status and individual process status) to
facilitate a causal understanding of the dynamic and complex man-
ufacturing operations in a factory. As described in Fig. 2, any 6.4 Modeling Material-Processing-Property Relationships.
system-level decision-making and control can be traced to individ- To ensure the desired performance of the final manufactured
ual process-level control action, and any process-level optimization parts, a comprehensive understanding of the material-processing-
or parameter changes can be examined for its real impact on the property relationship is required. Conventional modeling and
entire connected system. Therefore, ensuring the quality of local control schemes have been developed and applied to achieve man-
data is the key to enable a causal analysis of the manufacturing ufacturing performance in the presence of variations in process
system. dynamics and unpredicted uncertainties. However, these controls
are usually difficult to design and computationally intensive when
the processes are highly nonlinear. In addition, automatically updat-
6.3 Transfer Learning and Data Synthesis. While various
ing the necessary parameters from the modeled process remains a
AI techniques have demonstrated the capability to accurately
challenge. Furthermore, a priori information on the structure of
model and optimize system performance, interpret human motion
the process dynamics and model uncertainty bounds is usually
to realize human–robot collaboration, detect and classify defect
unavailable. In such cases, AI techniques have the potential to
and predict future machine condition, research published so far
avoid the complexity of modeling the complete material-
has been largely built upon the assumption that sufficient data are
processing-property relationship for improving prediction accuracy
available for model training and validation. However, in real appli-
and thus productivity in a variety of manufacturing processes.
cations, obtaining data from different operating conditions or man-
ufacturing configurations in a systematic fashion is often times
infeasible, since data collection that encompasses all possible sce- 6.5 Promoting Trust in Artificial Intelligence. As AI contin-
narios is financially and operationally prohibitive. Furthermore, ues to evolve and advance, understanding and interpreting the
data exchange/accessibility through public/private networks can output of AI tools and related technical details become increasingly
be restricted due to security concerns. Two approaches have been exclusive to data scientists and similar professionals with special-
investigated in the scientific community to overcome these chal- ized skills in this domain. Often times, this creates a knowledge
lenges: transfer learning and data synthesis. gap leaving plant managers and production engineers lacking this

110804-12 / Vol. 142, NOVEMBER 2020 Transactions of the ASME

background with the ability to comprehend and appropriately inter- are being used to successfully aid in the communication of intent
pret the meaning of the results from AI models in a manner consis- between human and robot, based on voice, gesture, gaze, and
tent with the context of the problem area. This highlights a explicit commands. On the other hand, a higher cognition level of
significant challenge of AI in manufacturing, namely the impor- interactions focused on shared workpiece activities still requires
tance of proper interpretation of AI analysis to decision-makers further research to develop robots with reliable mental models of
who may not be experts in AI. Without a proper grasp of the anal- their human counterpart to dynamically generate the corresponding
ysis that relates to the fundamental physics, users would have no motion-adaptive to human behavior.
basis to trust and accept the analysis results. Since manufacturing In regard to process monitoring, diagnostics, and prognostics, the
operations are based on what is known physically, not on what deployment of AI tools has been more extensive owing to the rich
might probabilistically occur as indicated by AI models, more trans- stream of data emanating from processes, sensors, and equipment.
parent, physics-guided process models are required. This has given Conventional ML methods rely on high-quality data to extract rel-
rise to research into explainable AI models, commonly referred to as evant features. Thus, AI-based diagnosis has found popular usage to
XAI, which facilitates mental-support models to assist users of AI formulate classifications and associate data to corresponding fault
technologies more readily adopt them as powerful tools to enable types and severity levels.

Downloaded from http://asmedigitalcollection.asme.org/manufacturingscience/article-pdf/142/11/110804/6594922/manu_142_11_110804.pdf by guest on 01 March 2024

“smart” manufacturing. As AI research continues to evolve, it can This work also examined the research into the application of AI
be expected that the topic of trust in AI will assume an increasingly to facilitate improved understanding of material properties as used
important role and attract more intense research activities. in manufacturing process monitoring and modeling. For example,
AI has been used to predict material properties and experimental
results in a fraction of the time that would otherwise be spent via
6.6 Practical Implementation of Artificial Intelligence in
conventional methods. Such prediction of material properties
Manufacturing. While the majority of the AI research in manufac-
applies to both macro and micro levels wherein properties such as
turing has been conducted in a laboratory environment (such as
hardness, the melting point can be represented through simulation,
using test rigs), companies have begun to adapt and implement
providing valuable input to complex modeling of processes that use
the state-of-the-art into daily operations to improve production effi-
highly time-dependent material properties.
ciency, flexibility, and reduce cost. As an example, a technical
The four domains summarized above also point to the new chal-
report published in 2018 [166] described the implementation of
lenges and opportunities for future research, as identified in the area
DCNN as a machine vision solution to facilitate early detection
of system-level analysis, data quality, model and knowledge trans-
and classification of defects, such as epoxy staining and excursion
fer, modeling material-processing-property relationships, and inter-
related to ball application, in a wafer production line. The system
pretation of results based on AI techniques. Despite the varying
consists of a classification inference cluster to support large-scale
degrees of applicability and research gaps that exist and need to
production and a model-building unit that builds and refines the
be overcome in each of the four domains, the trend is undeniably
DCNN model. During wafer production, when the DCNN result
one of the increased implementation of AI-based analytical tools.
does not meet the required confidence level, which is specified
That there remain significant data, and problem formulation chal-
per individual production stages, the wafer unit is transferred to
lenges to be solved does not limit the already demonstrated oppor-
manual classification, which in turn allows to augment the existing
tunity for AI to transform manufacturing as we know it today.
training data and continuously improves the DCNN accuracy and
robustness. It is reported that the system can now automatically
and accurately detect and classify any defect that would otherwise
require a manual procedure, and is able to reduce about 80% of Acknowledgment
total workers’ time. This work is partially supported by the National Science Foun-
In Ref. [167], a novel vision-based system is developed for fabric dation under awards CMMI-1830295, CMMI-1351160, and
texture identification and implemented in a company that produces CMMI-1853454.
automotive interior parts. In an industrial environment, the unstable
ambient illumination poses considerable challenges in applying
existing image classification techniques. The images captured in Conflict of Interest
unstable lighting conditions are pre-processed with Laws and
Sobel filters to extract features, which are then fed to an SVM clas- There are no conflicts of interest.
sifier enhanced by pyramid analysis. The proposed technique
reaches a texture classification accuracy of 98% while satisfying
the computation time requirements in a massive production setting. References
It is important to note that more effort is needed to promote AI [1] Monostori, L., Kádár, B., Bauernhansl, T., Kondoh, S., Kumara, S., Reinhart,
from the perspective of the industry and facilitate the broad accep- G., Sauer, O., Schuh, G., Sihn, W., and Ueda, K., 2016, “Cyber-Physical
tance of AI techniques. Systems in Manufacturing,” CIRP Ann., 65(2), pp. 621–641.
[2] Wang, L., 2019, “From Intelligence Science to Intelligent Manufacturing,”
Engineering, 5(4), pp. 615–618.
[3] Chui, L., Kamalnath, V., and McCarthy, B., 2018, An Executive’s Guide to AI,
7 Conclusions McKinsey, https://www.mckinsey.com/business-functions/mckinsey-analytics/
our-insights/an-executives-guide-to-ai.
In this work, we have broadly reviewed the current use of AI and [4] Cardon, D., Cointet, J. P., and Mazières, A., 2018, “Neurons Spike Back: The
its potential for further opportunities in manufacturing systems and Invention of Inductive Machines and the Artificial Intelligence Controversy,”
processes across multiple hierarchical levels. The survey of the lit- Reseaux, 5(211), pp. 173–220.
[5] Lecun, Y., Bengio, Y., and Hinton, G., 2015, “Deep Learning,” Nature,
erature showed that a wide array of AI tools has already been imple- 521(7553), pp. 436–444.
mented to address a diverse set of problems throughout a plant [6] Lee, J., Davari, H., Singh, J., and Pandhare, V., 2018, “Industrial Artificial
hierarchy. However, despite this widespread use, there have been Intelligence for Industry 4.0-Based Manufacturing Systems,” Manuf. Lett., 18,
varying degrees of success and corresponding challenges that pp. 20–23.
[7] Li, B., Hou, B., Yu, W., Lu, X., and Yang, C., 2017, “Applications of Artificial
have been identified in implementing these tools. For example, Intelligence in Intelligent Manufacturing: A Review,” Front. Inf. Technol.
supervised learning in manufacturing system control benefits from Electron. Eng., 18(1), pp. 86–96.
the high richness of labeled data, yet the problem is knowledge- [8] Zhong, R. Y., Xu, X., Klotz, E., and Newman, S. T., 2017, “Intelligent
sparse. Simulation of manufacturing systems is a path to knowledge Manufacturing in the Context of Industry 4.0: A Review,” Engineering, 3(5),
pp. 616–630.
through the ease by which training data may be generated. [9] Sharp, M., Ak, R., and Hedberg, T., 2018, “A Survey of the Advancing Use and
However, this is limited by the model’s ability to reflect reality Development of Machine Learning in Smart Manufacturing,” J. Manuf. Syst.,
with a high fidelity. In the area of HRC, many AI technologies 48(Part C), pp. 170–179.

Journal of Manufacturing Science and Engineering NOVEMBER 2020, Vol. 142 / 110804-13
[10] Wang, J., Ma, Y., Zhang, L., Gao, R., and Wu, D., 2018, “Deep Learning for [40] Wang, S., Wan, J., Li, D., and Zhang, C., 2016, “Implementing Smart Factory of
Smart Manufacturing: Methods and Applications,” J. Manuf. Syst., 48(Part Industrie 4.0: An Outlook,” Int. J. Distrib. Sens. Networks, 12(1), p. 3159805.
C), pp. 144–156. [41] Priore, P., Ponte, B., Puente, J., and Gómez, A., 2018, “Learning-Based
[11] Liu, R., Yang, B., Zio, E., and Chen, X., 2018, “Artificial Intelligence for Fault Scheduling of Flexible Manufacturing Systems Using Ensemble Methods,”
Diagnosis of Rotating Machinery: A Review,” Mech. Syst. Signal Process., 108, Comput. Ind. Eng., 126, pp. 282–291.
pp. 33–47. [42] Rajendran, C., and Holthaus, O., 1999, “Comparative Study of Dispatching
[12] Zhao, R., Yan, R., Chen, Z., Mao, K., Wang, P., and Gao, R., 2019, “Deep Rules in Dynamic Flowshops and Jobshops,” Eur. J. Oper. Res., 116(1),
Learning and Its Applications to Machine Health Monitoring,” Mech. Syst. pp. 156–170.
Signal Process., 115, pp. 213–237. [43] Shahzad, A., and Mebarki, N., 2012, “Data Mining Based Job Dispatching
[13] Wang, L., Gao, R., Váncza, J., Krüger, J., Wang, X. V., Makris, S., and Using Hybrid Simulation-Optimization Approach for Shop Scheduling
Chryssolouris, G., 2019, “Symbiotic Human-Robot Collaborative Assembly,” Problem,” Eng. Appl. Artif. Intell., 25(6), pp. 1173–1181.
CIRP Ann., 68(2), pp. 701–726. [44] Doh, H. H., Yu, J. M., Kwon, Y. J., Shin, J. H., Kim, H. W., Nam, S. H., and
[14] Chang, Q., Ni, J., Bandyopadhyay, P., Biller, S., and Xiao, G., 2007, Lee, D. H., 2014, “Decision Tree Based Scheduling for Flexible job Shops
“Supervisory Factory Control Based on Real-Time Production Feedback,” with Multiple Process Plans,” Int. J. Mech. Aero. Ind. Mechatronic Manuf.
ASME J. Manuf. Sci. Eng., 129(3), pp. 653–660. Eng., 8(3), pp. 621–627.
[15] Li, J., and Meerkov, S. M., 2009, Production Systems Engineering, Springer, US. [45] Ji, W., and Wang, L., 2017, “Big Data Analytics Based Fault Prediction for Shop
[16] Wu, K., and McGinnis, L., 2012, “Performance Evaluation for General Floor Scheduling,” J. Manuf. Syst., 43(Part 1), pp. 187–194.
[46] El-Bouri, A., and Shah, P., 2006, “A Neural Network for Dispatching

Downloaded from http://asmedigitalcollection.asme.org/manufacturingscience/article-pdf/142/11/110804/6594922/manu_142_11_110804.pdf by guest on 01 March 2024

Queueing Networks in Manufacturing Systems: Characterizing the Trade-off
Between Queue Time and Utilization,” Eur. J. Oper. Res., 221(2), pp. 328–339. Rule Selection in a Job Shop,” Int. J. Adv. Manuf. Technol., 31(3–4),
[17] Tan, B., and Gershwin, S. B., 2009, “Analysis of a General Markovian pp. 342–349.
Two-Stage Continuous-Flow Production System with a Finite Buffer,” [47] Azadeh, A., Jeihoonian, M., Shoja, B. M., and Seyedmahmoudi, S. H.,
Int. J. Prod. Econ., 120(2), pp. 327–339. 2012, “An Integrated Neural Network-Simulation Algorithm for Performance
[18] Gershwin, S. B., and Werner, L. M., 2007, “An Approximate Analytical Method Optimisation of the Bi-Criteria Two-Stage Assembly Flow-Shop
for Evaluating the Performance of Closed-Loop Flow Systems With Unreliable Scheduling Problem with Stochastic Activities,” Int. J. Prod. Res., 50(24),
Machines and Finite Buffers,” Int. J. Prod. Res., 45(14), pp. 3085–3111. pp. 7271–7284.
[19] Negahban, A., and Smith, J. S., 2014, “Simulation for Manufacturing System [48] Hammami, Z., Mouelhi, W., and Said, L. B., 2015, “A Self Adaptive
Design and Operation: Literature Review and Analysis,” J. Manuf. Syst., Neural Agent Based Decision Support System for Solving Dynamic
33(2), pp. 241–261. Real Time Scheduling Problems,” Proceedings of the 10th International
[20] Colledani, M., and Gershwin, S. B., 2013, “A Decomposition Method for Conference on Intelligence System Knowledge Engineering, ISKE 2015,
Approximate Evaluation of Continuous Flow Multi-Stage Lines With General Taipei, pp. 494–501.
Markovian Machines,” Ann. Oper. Res., 209(1), pp. 5–40. [49] Liu, Y. H., Huang, H. P., and Lin, Y. S., 2005, “Dynamic Scheduling of Flexible
[21] Liu, Y., Li, J., and Chiang, S. Y., 2012, “Re-Entrant Lines with Unreliable Manufacturing System Using Support Vector Machines,” Proceedings of the
Asynchronous Machines and Finite Buffers: Performance Approximation and IEEE Conference on Automation Science and Engineering, Edmonton, Alta,
Bottleneck Identification,” Int. J. Prod. Res., 50(4), pp. 977–990. Canada, pp. 387–392.
[22] Li, C., Wang, H., and Li, B., 2016, “Performance Prediction of a Production Line [50] Shiue, Y. R., 2009, “Data-Mining-Based Dynamic Dispatching Rule Selection
with Variability Based on Grey Model Artificial Neural Network,” Chinese Mechanism for Shop Floor Control Systems Using a Support Vector Machine
Control Conference, CCC, pp. 9582–9587. Approach,” Int. J. Prod. Res., 47(13), pp. 3669–3690.
[23] Yang, F., 2010, “Neural Network Metamodeling for Cycle Time-Throughput [51] Gabel, T., and Riedmiller, M., 2012, “Distributed Policy Search Reinforcement
Profiles in Manufacturing,” Eur. J. Oper. Res., 205(1), pp. 172–185. Learning for Job-Shop Scheduling Tasks,” Int. J. Prod. Res., 50(1), pp. 41–61.
[24] Huang, J., Chang, Q., Zou, J., and Arinez, J., 2018, “A Real-Time Maintenance [52] Palombarini, J., and Martínez, E., 2012, “SmartGantt—An Intelligent System
Policy for Multi-Stage Manufacturing Systems Considering Imperfect for Real Time Rescheduling Based on Relational Reinforcement Learning,”
Maintenance Effects,” IEEE Access, 6, pp. 62174–62183. Expert Syst. Appl., 39(11), pp. 10251–10268.
[25] Zou, J., Chang, Q., Arinez, J., and Xiao, G., 2017, “Data-Driven Modeling and [53] Shahrabi, J., Adibi, M. A., and Mahootchi, M., 2017, “A Reinforcement
Real-Time Distributed Control for Energy Efficient Manufacturing Systems,” Learning Approach to Parameter Estimation in Dynamic Job Shop
Energy, 127, pp. 247–257. Scheduling,” Comput. Ind. Eng., 110, pp. 75–82.
[26] Li, Y., Chang, Q., Biller, S., and Xiao, G., 2014, “Event-Based Modelling of [54] Waschneck, B., Reichstaller, A., Belzner, L., Altenmüller, T., Bauernhansl, T.,
Distributed Sensor Networks in Battery Manufacturing,” Int. J. Prod. Res., Knapp, A., and Kyek, A., 2018, “Optimization of Global Production Scheduling
52(14), pp. 4239–4252. with Deep Reinforcement Learning,” Proc. CIRP, 72, pp. 1264–1269.
[27] Zou, J., Chang, Q., Lei, Y., and Arinez, J., 2018, “Production System [55] Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A. A., Veness, J., Bellemare,
Performance Identification Using Sensor Data,” IEEE Trans. Syst. Man M. G., Graves, A., Riedmiller, M., Fidjeland, A. K., Ostrovski, G., Petersen,
Cybern. Syst., 48(2), pp. 255–264. S., Beattie, C., Sadik, A., Antonoglou, I., King, H., Kumaran, D., Wierstra,
[28] Zou, J., Chang, Q., Arinez, J., and Xiao, G., 2018, “Production Performance D., Legg, S., and Hassabis, D., 2015, “Human-Level Control Through Deep
Prognostics Through Model-Based Analytical Method and Recency-Weighted Reinforcement Learning,” Nature, 518(7540), pp. 529–533.
Stochastic Approximation Method,” J. Manuf. Syst., 47, pp. 107–114. [56] Huang, J., Chang, Q., and Chakraborty, N., 2019, “Machine Preventive
[29] Shi, J., and Zhou, S., 2009, “Quality Control and Improvement for Multistage Replacement Policy for Serial Production Lines Based on Reinforcement
Systems: A Survey,” IIE Trans. (Inst. Ind. Eng)., 41(9), pp. 744–753. Learning,” Proceedings of the IEEE Conference on Automation Science and
[30] Carbery, C. M., Woods, R., and Marshall, A. H., 2018, “A Bayesian Network Engineering, Vancouver, BC, Canada, pp. 523–528.
Based Learning System for Modelling Faults in Large-Scale Manufacturing,” [57] Aissani, N., Beldjilali, B., and Trentesaux, D., 2009, “Dynamic Scheduling of
Proceedings of the IEEE International Conference on Industrial Technology, Maintenance Tasks in the Petroleum Industry: A Reinforcement Approach,”
Lyon, France, pp. 1357–1362. Eng. Appl. Artif. Intell., 22(7), pp. 1089–1103.
[31] Carbery, C. M., Woods, R., and Marshall, A. H., 2019, “A New Data Analytics [58] Wang, X., Wang, H., and Qi, C., 2016, “Multi-Agent Reinforcement Learning
Framework Emphasising Preprocessing of Data to Generate Insights Into Based Maintenance Policy for a Resource Constrained Flow Line System,”
Complex Manufacturing Systems,” Proc. Inst. Mech. Eng. Part C J. Mech. J. Intell. Manuf., 27(2), pp. 325–333.
Eng. Sci., 233(19–20), pp. 6713–6726. [59] Xanthopoulos, A. S., Kiatipis, A., Koulouriotis, D. E., and Stieger, S., 2017,
[32] Peres, R. S., Barata, J., Leitao, P., and Garcia, G., 2019, “Multistage Quality “Reinforcement Learning-Based and Parametric Production-Maintenance
Control Using Machine Learning in the Automotive Industry,” IEEE Access, Control Policies for a Deteriorating Manufacturing System,” IEEE Access, 6,
7, pp. 79908–79916. pp. 576–588.
[33] Lieber, D., Stolpe, M., Konrad, B., Deuse, J., and Morik, K., 2013, “Quality [60] Kuhnle, A., Jakubik, J., and Lanza, G., 2019, “Reinforcement Learning for
Prediction in Interlinked Manufacturing Processes Based on Supervised & Opportunistic Maintenance Optimization,” Prod. Eng., 13(1), pp. 33–41.
Unsupervised Machine Learning,” Proc. CIRP, 7, pp. 193–198. [61] Wang, H., 2002, “A Survey of Maintenance Policies of Deteriorating Systems,”
[34] Wang, G., Ledwoch, A., Hasani, R. M., Grosu, R., and Brintrup, A., 2019, “A Eur. J. Oper. Res., 139(3), pp. 469–489.
Generative Neural Network Model for the Quality Prediction of Work in [62] Kumar, A., Shankar, R., and Thakur, L. S., 2018, “A Big Data Driven
Progress Products,” Appl. Soft Comput. J., 85, p. 105683. Sustainable Manufacturing Framework for Condition-Based Maintenance
[35] Wuest, T., Irgens, C., and Thoben, K. D., 2014, “An Approach to Monitoring Prediction,” J. Comput. Sci., 27, pp. 428–439.
Quality in Manufacturing Using Supervised Machine Learning on Product [63] Demir, L., Tunali, S., and Eliiyi, D. T., 2014, “The State of the Art on Buffer
State Data,” J. Intell. Manuf., 25(5), pp. 1167–1180. Allocation Problem: A Comprehensive Survey,” J. Intell. Manuf., 25(3),
[36] Liu, Y., and Jin, S., 2013, “Application of Bayesian Networks for Diagnostics in pp. 371–392.
the Assembly Process by Considering Small Measurement Data Sets,” [64] Tsadiras, A. K., Papadopoulos, C. T., and O’Kelly, M. E. J., 2013, “An Artificial
Int. J. Adv. Manuf. Technol., 65(9–12), pp. 1229–1237. Neural Network Based Decision Support System for Solving the Buffer
[37] Sayed, M. S., and Lohse, N., 2013, “Distributed Bayesian Diagnosis for Allocation Problem in Reliable Production Lines,” Comput. Ind. Eng., 66(4),
Modular Assembly Systems—A Case Study,” J. Manuf. Syst., 32(3), pp. 1150–1162.
pp. 480–488. [65] Chen, C., Xia, B., Zhou, B. H., and Xi, L., 2015, “A Reinforcement Learning
[38] Leitao, P., 2009, “Agent-Based Distributed Manufacturing Control: A Based Approach for a Multiple-Load Carrier Scheduling Problem,” J. Intell.
State-of-the-Art Survey,” Eng. Appl. Artif. Intell., 22(7), pp. 979–991. Manuf., 26(6), pp. 1233–1245.
[39] Shen, W., Wang, L., and Hao, Q., 2006, “Agent-Based Distributed [66] Ou, X., Chang, Q., Arinez, J., and Zou, J., 2018, “Gantry Work Cell Scheduling
Manufacturing Process Planning and Scheduling: A State-of-the-Art Survey,” Through Reinforcement Learning with Knowledge-Guided Reward Setting,”
IEEE Trans. Syst. Man Cybern. Part C Appl. Rev., 36(4), pp. 563–577. IEEE Access, 6, pp. 14699–14709.

110804-14 / Vol. 142, NOVEMBER 2020 Transactions of the ASME

[67] Ou, X., Chang, Q., and Chakraborty, N., 2019, “Simulation Study on Reward From Human Demonstrations,” IEEE Trans. Autom. Sci. Eng., 16(2),
Function of Reinforcement Learning in Gantry Work Cell Scheduling,” pp. 640–653.
J. Manuf. Syst., 50, pp. 1–8. [94] Lee, J., Wu, F., Zhao, W., Ghaffari, M., Liao, L., and Siegel, D., 2014,
[68] Wuest, T., Weimer, D., Irgens, C., and Thoben, K. D., 2016, “Machine Learning “Prognostics and Health Management Design for Rotary Machinery Systems
in Manufacturing: Advantages, Challenges, and Applications,” Prod. Manuf. —Reviews, Methodology and Applications,” Mech. Syst. Signal Process.,
Res., 4(1), pp. 23–45. 42(1–2), pp. 314–334.
[69] Green, S. A., Billinghurst, M., Chen, X., and Chase, J. G., 2008, “Human-Robot [95] Tao, F., Qi, Q., Liu, A., and Kusiak, A., 2018, “Data-Driven Smart
Collaboration: A Literature Review and Augmented Reality Approach in Manufacturing,” J. Manuf. Syst., 48(Part C), pp. 157–169.
Design,” Int. J. Adv. Robot. Syst., 5(1), pp. 1–18. [96] Wang, P., Gao, R., and Fan, Z., 2015, “Cloud Computing for Cloud
[70] Verma, A., and Kumar, S., 2018, “Cognitive Robotics in Artificial Intelligence,” Manufacturing: Benefits and Limitations,” ASME J. Manuf. Sci. Eng. Trans.,
Proceedings of the International Conference on Cloud Computing, Data Science 137(4), p. 044002.
and Engineering, Noida, India, pp. 65–70. [97] Gao, R., Wang, L., Teti, R., Dornfeld, D., Kumara, S., Mori, M., and Helu, M.,
[71] Duan, F., Tan, J. T. C., and Arai, T., 2011, “A New Human-Robot Collaboration 2015, “Cloud-Enabled Prognosis for Manufacturing,” CIRP Ann., 64(2),
Assembly System for Cellular Manufacturing,” Proceedings of the Chinese pp. 749–772.
Control Conference, Yantai, China, pp. 5468–5473. [98] Malkoff, D. B., 1987, “A Framework for Real-Time Fault Detection and
[72] Marvel, J. A., Falco, J., and Marstio, I., 2015, “Characterizing Task-Based Diagnosis Using Temporal Data,” Artif. Intell. Eng., 2(2), pp. 97–111.
Human-Robot Collaboration Safety in Manufacturing,” IEEE Trans. Syst. [99] Gang, N., Son, J. D., Widodo, A., Yang, B. S., Hwang, D. H., and Kang,
D. S., 2007, “A Comparison of Classifier Performance for Fault Diagnosis of

Downloaded from http://asmedigitalcollection.asme.org/manufacturingscience/article-pdf/142/11/110804/6594922/manu_142_11_110804.pdf by guest on 01 March 2024

Man, Cybern. Syst., 45(2), pp. 260–275.
[73] Robla-Gomez, S., Becerra, V. M., Llata, J. R., Gonzalez-Sarabia, E., Induction Motor Using Multi-Type Signals,” Struct. Health Monit., 6(3),
Torre-Ferrero, C., and Perez-Oria, J., 2017, “Working Together: A Review on pp. 215–229.
Safe Human-Robot Collaboration in Industrial Environments,” IEEE Access, [100] Wang, J., Liu, S., Gao, R., and Yan, R., 2012, “Current Envelope Analysis for
5, pp. 26754–26773. Defect Identification and Diagnosis in Induction Motors,” J. Manuf. Syst., 31(4),
[74] Nikolaidis, S., Lasota, P., Rossano, G., Martinez, C., Fuhlbrigge, T., and Shah, pp. 380–387.
J., 2013, “Human-Robot Collaboration in Manufacturing: Quantitative [101] Zhang, J., Wang, P., Gao, R., and Yan, R., 2018, “An Image Processing
Evaluation of Predictable, Convergent Joint Action,” International Symposium Approach to Machine Fault Diagnosis Based on Visual Words
on Robotics, ISR 2013, Seoul, South Korea. Representation,” Proc. Manuf., 19, pp. 42–49.
[75] Helms, E., Sehraft, R. D., and Hägele, M., 2002, “Rob@Work: Robot Assistant in [102] Sun, C., Wang, P., Yan, R., and Gao, R., 2016, “A Sparse Approach to Fault
Industrial Environments,” Proceedings of the IEEE International Workshop on Severity Classification for Gearbox Monitoring,” Proceedings of the
Robot and Human Interactive Communication, Berlin, Germany, pp. 399–404. International Conference on Information Fusion, Heidelberg, Germany, pp.
[76] Wang, X. V., Seira, A., and Wang, L., 2018, “Classification, Personalised Safety 2303–2308.
Framework and Strategy for Human-Robot Collaboration,” Proceedings of the [103] Du, Z., Chen, X., Zhang, H., and Yan, R., 2015, “Sparse Feature Identification
International Conference on Computers and Industrial Engineering., Based on Union of Redundant Dictionary for Wind Turbine Gearbox Fault
Auckland, New Zealand. Diagnosis,” IEEE Trans. Ind. Electron., 62(10), pp. 6594–6605.
[77] Sheridan, T. B., 1992, Telerobotics, Automation, and Human Supervisory [104] Zhou, C., Liu, K., Zhang, X., Zhang, W., and Shi, J., 2016, “An Automatic
Control, MIT Press, Cambridge, MA. Process Monitoring Method Using Recurrence Plot in Progressive Stamping
[78] Sheridan, T. B., 2016, “Human-Robot Interaction,” Hum. Factors, 58(4), Processes,” IEEE Trans. Autom. Sci. Eng., 13(2), pp. 1102–1111.
pp. 525–532. [105] Rao, P. K., Liu, J., Roberson, D., Kong, Z., and Williams, C., 2015, “Online
[79] Rahman, S. M. M., Liao, Z., Jiang, L., and Wang, Y., 2016, “A Regret-Based Real-Time Quality Monitoring in Additive Manufacturing Processes Using
Autonomy Allocation Scheme for Human-Robot Shared Vision Systems in Heterogeneous Sensors,” ASME J. Manuf. Sci. Eng., 137(6), p. 061007.
Collaborative Assembly in Manufacturing,” IEEE International Conference on [106] Everton, S. K., Hirsch, M., Stavroulakis, P. I., Leach, R. K., and Clare, A. T.,
Automation and Science Engineering, Fort Worth, TX, pp. 897–902. 2016, “Review of In-Situ Process Monitoring and in-Situ Metrology for Metal
[80] Brady, M., 1985, “Artificial Intelligence and Robotics,” Artificial Intelligence, Additive Manufacturing,” Mater. Des., 95, pp. 431–445.
26(1), pp. 79–121. [107] Kwon, O., Kim, H. G., Ham, M. J., Kim, W., Kim, G. H., Cho, J. H., Kim,
[81] Roncone, A., Mangin, O., and Scassellati, B., 2017, “Transparent Role N. I. L., and Kim, K., 2020, “A Deep Neural Network for Classification of
Assignment and Task Allocation in Human Robot Collaboration,” Melt-Pool Images in Metal Additive Manufacturing,” J. Intell. Manuf., 31(2),
Proceedings of the IEEE International Conference on Robotics Automation, pp. 375–386.
Singapore, pp. 1014–1021. [108] Caggiano, A., Zhang, J., Alfieri, V., Caiazzo, F., Gao, R., and Teti, R., 2019,
[82] Huang, C. M., and Mutlu, B., 2016, “Anticipatory Robot Control for Efficient “Machine Learning-Based Image Processing for on-Line Defect Recognition
Human-Robot Collaboration,” ACM/IEEE International Conference on in Additive Manufacturing,” CIRP Ann., 68(1), pp. 451–454.
Human-Robot Interaction, Christchurch, New Zealand, pp. 83–90. [109] Wang, P., Ananya, Yan, R., and Gao, R. X., 2017, “Virtualization and
[83] Doltsinis, S., Krestenitis, M., and Doulgeri, Z., 2020, “A Machine Learning Deep Recognition for System Fault Classification,” J. Manuf. Syst., 44(Part 2),
Framework for Real-Time Identification of Successful Snap-Fit Assemblies,” pp. 310–316.
IEEE Trans. Autom. Sci. Eng., 17(1), pp. 513–523. [110] Ding, X., and He, Q., 2017, “Energy-Fluctuated Multiscale Feature Learning
[84] Akan, B., Curuklu, B., Spampinato, G., and Asplund, L., 2010, with Deep ConvNet for Intelligent Spindle Bearing Fault Diagnosis,” IEEE
“Towards Robust Human Robot Collaboration in Industrial Environments,” Trans. Instrum. Meas., 66(8), pp. 1926–1935.
ACM/IEEE International Conference on Human-Robot Interaction, Osaka, [111] Sun, W., Yao, B., Zeng, N., Chen, B., He, Y., Cao, X., and He, W., 2017, “An
Japan, pp. 71–72. Intelligent Gear Fault Diagnosis Methodology Using a Complex Wavelet
[85] Károly, A. I., Kuti, J., and Galambos, P., 2018, “Unsupervised Real-Time Enhanced Convolutional Neural Network,” Materials (Basel), 10(7), pp. 790.
Classification of Cycle Stages in Collaborative Robot Applications,” [112] Eren, L., Ince, T., and Kiranyaz, S., 2019, “A Generic Intelligent Bearing Fault
Proceedings of the IEEE Symposium on Applied Machine Intelligence and Diagnosis System Using Compact Adaptive 1D CNN Classifier,” J. Signal
Informatics, Kosice, Slovakia, pp. 97–102. Process. Syst., 91(2), pp. 179–189.
[86] Batth, R. S., Nayyar, A., and Nagpal, A., 2019, “Internet of Robotic Things: [113] Grezmak, J., Zhang, J., Wang, P., Loparo, K. A., and Gao, R., 2020,
Driving Intelligent Robotics of Future—Concept, Architecture, Applications “Interpretable Convolutional Neural Network Through Layer-Wise
and Technologies,” Proceedings of the International Conference on Computer Relevance Propagation for Machine Fault Diagnosis,” IEEE Sens. J., 20(6),
Science, Jalandhar, India, pp. 151–160. pp. 3172–3181.
[87] Sadrfaridpour, B., and Wang, Y., 2018, “Collaborative Assembly in Hybrid [114] Ma, M., Sun, C., and Chen, X., 2017, “Discriminative Deep Belief Networks
Manufacturing Cells: An Integrated Framework for Human-Robot with Ant Colony Optimization for Health Status Assessment of Machine,”
Interaction,” IEEE Trans. Autom. Sci. Eng., 15(3), pp. 1178–1192. IEEE Trans. Instrum. Meas., 66(12), pp. 3115–3125.
[88] Liu, H., Fang, T., Zhou, T., and Wang, L., 2018, “Towards Robust [115] Yang, F., Habibullah, M. S., Zhang, T., Xu, Z., Lim, P., and Nadarajan, S., 2016,
Human-Robot Collaborative Manufacturing: Multimodal Fusion,” IEEE “Health Index-Based Prognostics for Remaining Useful Life Predictions in
Access, 6, pp. 74762–74771. Electrical Machines,” IEEE Trans. Ind. Electron., 63(4), pp. 2633–2644.
[89] Sariel, S., Yildiz, P., Karapinar, S., Altan, D., and Kapotoglu, M., 2015, [116] Benkedjouh, T., Medjaher, K., Zerhouni, N., and Rechak, S., 2013, “Remaining
“Robust Task Execution Through Experience-Based Guidance for Cognitive Useful Life Estimation Based on Nonlinear Feature Reduction and Support
Robots,” Proceedings of the International Conference on Advanced Robotics, Vector Regression,” Eng. Appl. Artif. Intell., 26(7), pp. 1751–1760.
Istanbul, Turkey, pp. 663–668. [117] Tran, V. T., Thom Pham, H., Yang, B. S., and Tien Nguyen, T., 2012, “Machine
[90] Reimann, J., and Sziebig, G., 2019, “The Intelligent Factory Space—A Concept Performance Degradation Assessment and Remaining Useful Life Prediction
for Observing, Learning and Communicating in the Digitalized Factory,” IEEE Using Proportional Hazard Model and Support Vector Machine,” Mech. Syst.
Access, 7, pp. 70891–70900. Signal Process., 32, pp. 320–330.
[91] Ravichandar, H. C., and Dani, A., 2015, “Human Intention Inference and [118] Nuhic, A., Terzimehic, T., Soczka-Guth, T., Buchholz, M., and Dietmayer, K.,
Motion Modeling Using Approximate E-M with Online Learning,” IEEE 2013, “Health Diagnosis and Remaining Useful Life Prognostics of Lithium-Ion
International Conference Intelligence on Robots Systems, Hamburg, Germany, Batteries Using Data-Driven Methods,” J. Power Sources, 239, pp. 680–688.
pp. 1819–1824. [119] Khelif, R., Chebel-Morello, B., Malinowski, S., Laajili, E., Fnaiech, F., and
[92] Zhang, J., Liu, H., Chang, Q., Wang, L., and Gao, R., 2020, “Recurrent Neural Zerhouni, N., 2017, “Direct Remaining Useful Life Estimation Based on
Network for Motion Trajectory Prediction in Human-Robot Collaborative Support Vector Regression,” IEEE Trans. Ind. Electron., 64(3), pp. 2276–2285.
Assembly,” CIRP Ann., 69(1), pp. 9–12. [120] Hu, C., Youn, B. D., Wang, P., and Taek Yoon, J., 2012, “Ensemble of
[93] Wang, W., Li, R., Chen, Y., Diekel, Z. M., and Jia, Y., 2019, “Facilitating Data-Driven Prognostic Algorithms for Robust Prediction of Remaining
Human-Robot Collaborative Tasks by Teaching-Learning-Collaboration Useful Life,” Reliab. Eng. Syst. Saf., 103, pp. 120–135.

Journal of Manufacturing Science and Engineering NOVEMBER 2020, Vol. 142 / 110804-15
[121] Malhi, A., Yan, R., and Gao, R., 2011, “Prognosis of Defect Propagation [145] Pimenov, D. Y., Bustillo, A., and Mikolajczyk, T., 2018, “Artificial Intelligence
Based on Recurrent Neural Networks,” IEEE Trans. Instrum. Meas., 60(3), for Automatic Prediction of Required Surface Roughness by Monitoring Wear
pp. 703–711. on Face Mill Teeth,” J. Intell. Manuf., 29(5), pp. 1045–1061.
[122] Zhang, J., Wang, P., Yan, R., and Gao, R., 2018, “Long Short-Term Memory for [146] Xu, C., and Shin, Y. C., 2007, “Control of Cutting Force for Creep-Feed
Machine Remaining Life Prediction,” J. Manuf. Syst., 48, pp. 78–86. Grinding Processes Using a Multi-Level Fuzzy Controller,” ASME J. Dyn.
[123] Zhang, Y., Xiong, R., He, H., and Pecht, M. G., 2018, “Long Short-Term Syst. Meas. Control. Trans, 129(4), pp. 480–492.
Memory Recurrent Neural Network for Remaining Useful Life Prediction of [147] Xu, C., Tang, Y., and Jackson, M. J., 2009, “Survey on Various Control
Lithium-Ion Batteries,” IEEE Trans. Veh. Technol., 67(7), pp. 5695–5705. Techniques in Micro Grinding Processes,” Int. J. Nanomanuf., 3(4), pp. 398–
[124] Guo, L., Li, N., Jia, F., Lei, Y., and Lin, J., 2017, “A Recurrent Neural Network 408.
Based Health Indicator for Remaining Useful Life Prediction of Bearings,” [148] Sadeghi, B. H. M., 2000, “BP-Neural Network Predictor Model for
Neurocomputing, 240, pp. 98–109. Plastic Injection Molding Process,” J. Mater. Process. Technol., 103(3),
[125] Wang, P., and Gao, R., 2017, “Automated Performance Tracking for pp. 411–416.
Heat Exchangers in HVAC,” IEEE Trans. Autom. Sci. Eng., 14(2), pp. 634–645. [149] Saleh Meiabadi, M., Vafaeesefat, A., and Sharifi, F., 2013, “Optimization of
[126] Wang, P., and Gao, R., 2016, “Markov Nonlinear System Estimation for Engine Plastic Injection Molding Process by Combination of Artificial Neural
Performance Tracking,” ASME J. Eng. Gas Turbines Power., 138(9), p. 091201. Network and Genetic Algorithm,” J. Optim. Ind. Eng., 13, pp. 49–54.
[127] Yang, C., Li, Z., Liang, B., Lu, W., Wang, X., and Liu, H., 2017, “A Particle [150] Liu, Y., Zhao, T., Ju, W., Shi, S., Shi, S., and Shi, S., 2017, “Materials Discovery
Filter and Long Short Term Memory Fusion Algorithm for Failure Prognostic and Design Using Machine Learning,” J. Mater., 3(3), pp. 159–177.

Downloaded from http://asmedigitalcollection.asme.org/manufacturingscience/article-pdf/142/11/110804/6594922/manu_142_11_110804.pdf by guest on 01 March 2024

of Proton Exchange Membrane Fuel Cells,” Proceedings of the Chinese [151] Bertinetto, C., Duce, C., Micheli, A., Solaro, R., Starita, A., and Tiné, M. R.,
Control Decision Conference, Chongqing, China, pp. 5646–5651. 2009, “Evaluation of Hierarchical Structured Representations for QSPR
[128] Shaw, M. J., and Whinston, A. B., 1989, “An Artificial Intelligence Approach to Studies of Small Molecules and Polymers by Recursive Neural Networks,”
the Scheduling of Flexible Manufacturing Systems,” IIE Trans. (Institute Ind. J. Mol. Graph. Model., 27(7), pp. 797–802.
Eng), 21(2), pp. 170–183. [152] Carrera, G. V. S. M., Branco, L. C., Aires-de-Sousa, J., and Afonso, C. A. M.,
[129] Hutchinson, G.K., 1984, “Flexibility is Key to Economic Feasibility of 2008, “Exploration of Quantitative Structure-Property Relationships (QSPR) for
Automated Small Batch Manufacturing,” Ind. Eng., 16(6), pp. 77–86. the Design of New Guanidinium Ionic Liquids,” Tetrahedron, 64(9), pp. 2216–
[130] Shin, Y. C., and Xu, C., 2008, Intelligent Systems: Modeling, Optimization and 2224.
Control, CRC Press, Boca Raton, FL. [153] Jia, Y., Chowdhury, A., and Xu, C., 2020, “Electromagentic Property of
[131] Xu, C., and Shin, Y. C., 2008, “A Fuzzy Inverse Model Construction Method for Polymer Derived SiC-C Solid Solution Formed at Ultra-High Temperature,”
General Monotonic Multi-Input-Single-Output (MISO) Systems,” IEEE Trans. Carbon, 162, pp. 74–85.
Fuzzy Syst., 16(5), pp. 1216–1231. [154] Chowdhury, M. A. R., Wang, K., Jia, Y., and Xu, C., 2020,
[132] Xu, C., and Shin, Y. C., 2011, “A Self-Tuning Fuzzy Controller for a Class of “Semiconductor-Conductor Transition of Pristine Polymer-Derived Ceramics
Multi-Input Multi-Output Nonlinear Systems,” Eng. Appl. Artif. Intell., 24(2), SiC Pyrolyzed at Temperature Range From 1200 °C to 1800 °C,” J. Am.
pp. 238–250. Ceram. Soc., 103(4), pp. 2630–2642.
[133] Liu, J., Li, J., Ji, Y., and Xu, C., 2011, “Investigation on the Effect of SiC [155] Chowdhury, M. A., Wang, K., Jia, Y., and Xu, C., 2020, “Electrical
Nanoparticles on Cutting Forces for Micro-Milling Magnesium Matrix Conductivity and Structural Evolution of Polymer Derived SiC
Composites,” ASME 2011 International Conference on Manufacturing Ceramics Pyrolyzed From 1200 °C to 1800 °C,” ASME J. Micro
Science and Engineering, MSEC, Corvallis, OR, pp. 525–536. Nano-Manuf., 8(2), p. 024502.
[134] Liu, J., Li, J., and Xu, C., 2013, “Cutting Force Prediction on Micromilling [156] Regli, W., Rossignac, J., Shapiro, V., and Srinivasan, V., 2016, “The New
Magnesium Metal Matrix Composites with Nanoreinforcements,” ASME Frontiers in Computational Modeling of Material Structures,” CAD Comput.
J. Micro Nano-Manuf., 1(1), p. 011010. Aided Des., 77, pp. 73–85.
[135] Zhang, X. W., Ding, Y. Q., Lv, Y. Y., Shi, A. Y., and Liang, R. Y., 2011, “A [157] Yan, R., Shen, F., Sun, C., and Chen, X., 2019, “Knowledge Transfer for Rotary
Vision Inspection System for the Surface Defects of Strongly Reflected Metal Machine Fault Diagnosis,” IEEE Sens. J., 20(15), pp. 8374–8393.
Based on Multi-Class SVM,” Expert Syst. Appl., 38(5), pp. 5930–5939. [158] Wang, J., Zhao, R., and Gao, R., 2020, “Probabilistic Transfer Factor Analysis
[136] Zhang, H. C., and Huang, S. H., 1995, “Applications of Neural Networks for Machinery Autonomous Diagnosis Cross Various Operating Conditions,”
in Manufacturing: A State-of-the-Art Survey,” Int. J. Prod. Res., 33(3), IEEE Trans. Instrum. Meas., 69(8), pp. 5335–5344.
pp. 705–728. [159] Hasan, M. J., and Kim, J.-M., 2018, “Bearing Fault Diagnosis Under Variable
[137] Ma, G., and Zhang, F., 2012, “Genetic Algorithms for Manufacturing Process Rotational Speeds Using Stockwell Transform-Based Vibration Imaging and
Planning,” Variants of Evolutionary Algorithms for Real-World Applications, Transfer Learning,” Appl. Sci., 8(12), p. 2357.
R. Chiong, T. Weise, and Z. Michalewicz, eds., Springer, Berlin, Heidelberg, [160] Wang, P., and Gao, R., 2020, “Transfer Learning for Enhanced Machine Fault
pp. 205–244. Diagnosis in Manufacturing,” CIRP Ann., 69(1), pp. 413–416.
[138] Wu, D., Jennings, C., Terpenny, J., Gao, R., and Kumara, S., 2017, “A [161] Kim, H., and Youn, B. D., 2019, “A New Parameter Repurposing
Comparative Study on Machine Learning Algorithms for Smart Method for Parameter Transfer with Small Dataset and Its Application in
Manufacturing: Tool Wear Prediction Using Random Forests,” ASME Fault Diagnosis of Rolling Element Bearings,” IEEE Access, 7, pp. 46917–
J. Manuf. Sci. Eng., 139(7), p. 071018. 46930.
[139] Udo, G. J., 1992, “Neural Networks Applications in Manufacturing Processes,” [162] Xiong, Q., Zhang, J., Wang, P., Liu, D., and Gao, R., 2020, “Transferable
Comput. Ind. Eng., 23(1–4), pp. 97–100. Two-Stream Convolutional Neural Network for Human Action Recognition,”
[140] Wacker, J. G., 1996, “A Theoretical Model of Manufacturing Lead Times and J. Manuf. Syst., In Press
Their Relationship to a Manufacturing Goal Hierarchy,” Decis. Sci., 27(3), [163] Kozjek, D., Vrabič , R., Kralj, D., and Butala, P., 2017, “Interpretative
pp. 483–517. Identification of the Faulty Conditions in a Cyclic Manufacturing Process,”
[141] Chaudhari, P. S., Patel, P. D. M., and Juneja, J. L., 2012, “Artificial Intelligence J. Manuf. Syst., 43(Part 2), pp. 214–224.
Apply for Prediction of Laser Cutting Process—A Review,” Int. J. Eng. Res. [164] Goodfellow, I. J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D.,
Appl., 2(4), pp. 1025–1028. Ozair, S., Courville, A., and Bengio, Y., 2014, “Generative Adversarial Nets,”
[142] Xu, C., and Shin, Y. C., 2008, “An Adaptive Fuzzy Controller for Constant Adv. Neural Inf. Process. Syst., 3, pp. 2672–2680.
Cutting Force in End-Milling Processes,” ASME J. Manuf. Sci. Eng., 130(3), [165] Wang, Z., Wang, J., and Wang, Y., 2018, “An Intelligent Diagnosis
p. 031001. Scheme Based on Generative Adversarial Learning Deep Neural Networks
[143] Tang, Y., and Xu, C., 2010, “Online Tool Deflection Compensation in End and Its Application to Planetary Gearbox Fault Pattern Recognition,”
Milling of Curved Workpiece,” Int. J Electron. Comput. Eng. Edu., 1(1), Neurocomputing, 310, pp. 213–222.
pp. 23–28. [166] Tuv, E., Guven, M., Ennis, P., and Lee, D. H. L., 2018, White Paper.
[144] Khorasani, A. M., Reza Soleymani Yazdi, M., and Safizadeh, M. S., 2011, “Tool [167] Malaca, P., Rocha, L. F., Gomes, D., Silva, J., and Veiga, G., 2019, “Online
Life Prediction in Face Milling Machining of 7075 Al by Using Artificial Neural Inspection System Based on Machine Learning Techniques: Real Case Study
Networks (ANN) and Taguchi Design of Experiment (DOE),” Int. J. Eng. of Fabric Textures Classification for the Automotive Industry,” J. Intell.
Technol., 3(1), pp. 30–35. Manuf., 30(1), pp. 351–361.

110804-16 / Vol. 142, NOVEMBER 2020 Transactions of the ASME