Seminar Report

On

WELD QUALITY CHARACTERIZATION BY

VIBRATION ANALYSIS FOR ULTRASONIC
METAL WELDING PROCESSES
Submitted by

ABHIJITH.V (IEAUEME027)

In partial fulfilment for the award of degree of

BACHELOR OF TECHNOLOGY

IN

MECHANICAL ENGINEERING

DEPARTMENT OF MECHANICAL ENGINEERING

INSTITUTE OF ENGINEERING AND TECHNOLOGY

UNIVERSITY OF CALICUT

MALAPPURAM- 673636 KERALA

JUNE 2024

i
 CERTIFICATE

This is to certify that the seminar report entitled ‘Weld Quality

Characterization by Vibration Analysis for Ultrasonic Metal Welding
Processes’ submitted by Mr. ABHIJITH.V (IEAUEME027). Under our
supervision, in fulfilment of the requirements of the award of Degree of
Bachelor of Technology in Mechanical Engineering from Calicut University for
the year 2024.

SEMINAR GUIDE: SEMINAR COORDINATORS:

Mr. CHRISTO JOSE and
Mr. ARJUN O
Mrs. DHANYA P
Assistant Professor Assistant Professor
Department of ME Department of ME

HEAD OF DEPARTMENT:

Mr. ANIL JACOB

Assistant Professor

IETCU

ii
 ACKNOWLEDGEMENT

Dedicating this seminar to the almighty God whose abundant grace and
mercies enabled its successful completion, I would like to express my profound
gratitude to all the people who have inspired and motivated us to make this seminar a
success.

I am extremely grateful to our principal. Dr. Ranjith C, our HOD, Mr. Anil
Jacob Department of Mechanical Engineering, Institute of Engineering and
Technology, University of Calicut for their valuable suggestions and encouragement
that was a constant source of inspiration and motivation for during the course of this
seminar.

I am very much thankful to our seminar Coordinators Mr. Christo Jose and
Mrs. Dhanya P, Assistant Professor, Institute of Engineering and Technology,
University of Calicut, for their helpful hands during the course of my seminar.

I deem the privilege of thanking my Guide Mr. Arjun O, Assistant Professor,

Department of Mechanical engineering, Institute of Engineering and Technology,
University of Calicut, for his constant help, creative ideas, valuable advice and
guidance rendered us through the course of this seminar.

My profound gratitude to all staffs, for their valuable guidance, kind

cooperation and help provided for the successful completion of the seminar.

iii
 ABSTRACT

Ultrasonic metal welding is an efficient, solid-state welding process that

is well established in the production of electrical equipment. However, isolated
and untraceable failures occasionally occur in industrial production processes.
According to the state of the art, these failures are due to a multitude of
influencing variables, such as material and surface condition of the joining
partners, as well as tool and fixture wear. This study presents quality prediction
models based on measurements of mechanical vibrations of the welding
machine using laser triangulation sensors, sonotrode penetration depth
measurements, and machine internal welding power signals. The acquired data
are processed and transformed into a random forest model to estimate the shear
strength of the welds. The robustness of the prediction model has been
successfully validated by welding experiments with significant external
disturbances, such as surface roughness, contamination and material hardness.
Within the framework of this study, the development of a stable and industrially
applicable concept for process monitoring is demonstrated with a regression
model that achieves a mean relative estimation error of 4.30% and a R2 value of
0.964. Furthermore, a classification model that determines the external
disturbances for the individual welds was successfully validated, achieving a
micro F1 value of 94% and a macro F1 value of 95%.

Keywords: Ultrasonic metal welding, Vibration analysis, Process monitoring,

Vibration measurements, Random forest, Automated data processing, Machine
learning

iv
 CONTENTS
ACKNOWLEDGEMENT ..................................................................................... iii
ABSTRACT ............................................................................................................ iv
LIST OF FIGURES ............................................................................................... vi
LIST OF TABLES ................................................................................................ vii
LIST OF ABBREVIATIONS ................................................................................ ix
CHAPTER 1 ............................................................................................................ 1
INTRODUCTION ................................................................................................... 1
CHAPTER 2 ............................................................................................................ 3
ULTRASONIC WELDING PROCESS AND PROCESS INFLUENCES ........... 3
CHAPTER 3 ............................................................................................................ 8
MONITORING OF ULTRASONIC METAL WELDING PROCESS ................ 8
3.1 APPLICATION OF MACHINE LEARNING ................................ 9
CHAPTER 4 .......................................................................................................... 11
METHODOLOGY AND MATERIALS .............................................................. 11
CHAPTER 5 .......................................................................................................... 12
WELDING SETUP ............................................................................................... 12
CHAPTER 6 .......................................................................................................... 16
WELDING RESULTS OF THE TEST SERIES AND QUALITY FORECAST
BASED ON INTERNAL DATA ........................................................................... 16
CHAPTER 7 .......................................................................................................... 18
MEASUREMENT SETUP ................................................................................... 18
CHAPTER 8 .......................................................................................................... 20
DATA PROCESSING ........................................................................................... 20
8.1 METHODOLOGY FOR MAXIMUM TENSILE FORCE ........... 23
CHAPTER 9 .......................................................................................................... 24
THE RANDOM FOREST ALGORITHM ........................................................... 24
CHAPTER 10 ........................................................................................................ 25
FEATURE EXTRACTION .................................................................................. 25
CHAPTER 11 ........................................................................................................ 29
STRATIFIED K-FOLD CROSS-VALIDATION ................................................ 29
CHAPTER 12 ........................................................................................................ 30
RESULT AND DISCUSSION .............................................................................. 30

v
 12.1 PREDICTION OF MAXIMUM SHEAR FORCE ...................... 30
CHAPTER 13 ........................................................................................................ 33
CLASSIFICATION OF WELD DEFECTS ......................................................... 33
CHAPTER 14 ........................................................................................................ 35
CONCLUSION ..................................................................................................... 35
REFERENCES ...................................................................................................... 36

LIST OF FIGURES

Fig 1.1 Machine setup and quality prediction process 1

vi
Fig 2.1 Ultrasonic metal welding process 4

Fig 2.2 Process stages 4

Fig 4.1 Overview of variations of reference sample 11

properties

Fig 5.1 Work piece geometry and welding position 13

Fig 5.2 Formation and growth of joint area 14

Fig 6.1 Overview of the achieved shear strength and 16

weld time

Fig 7.1 Measurement setup 18

Fig 8.1 Exemplary time series of the anvil vibration 22

Fig 10.1 Feature extraction 25

Fig 11.1 Stratified K-fold cross validation 29

Fig 12.1 Prediction results based on anvil vibration or 32

penetration data

Fig 12.2 Prediction results for different combinations 32

Fig 13.1 Confusion matrix of weld defects 34

LIST OF TABLES

vii
Table 5.1 Configuration of welding machine 12

Table 5.2 Welding parameters 13

Table 5.3 Welding experiment test series configuration and 15

results

Table 7.1 Sensor specifications 19

Table 10.1 Welding experiment test series configuration and 28

results

Table 12.1 Estimation error for different combinations 30

Table 12.2 Random forest parameters 31

viii
 LIST OF ABBREVIATIONS

HH Half hard

ANV Anvil vibration

POW Power

PEN Penetration

STR Strain gauge

GPR Gaussian process regression

TP True positive

FN False negative

FP False positive

ix
 Weld Quality Characterization by Vibration Analysis for Ultrasonic Metal Welding
Processes

CHAPTER 1
INTRODUCTION

Modern vehicles, systems, and devices increasingly rely on electrified

components and digital controls. This increase is driven by technology trends, such
as electro mobility and autonomous driving. In these applications, the ability to
reliably transport and store energy and information is essential for success. As a
result, there is not only growing interest in, but also increasing demands on, suitable
manufacturing processes. One example of such a manufacturing process is ultrasonic
metal welding (USMW), which is used in the production of battery cells, high-power
electronics and wiring harnesses for modern automobiles. Ultrasonic metal welding
is a solid-state welding process that is used primarily in demanding electrical and
electronic applications, as it allows large joint cross sections and causes low heat
input. It is easy to automate, energy-efficient and allows the welding of dissimilar
metals with limited intermetallic formation, as well as the joining of multiple layers
of sheets and foils.

Fig 1.1 Machine setup and quality prediction process

Dept. of ME 1 IETCU
 Weld Quality Characterization by Vibration Analysis for Ultrasonic Metal Welding
Processes

The high joint quality precludes meaningful characterization of USMW joints

by electrical properties, such as contact resistance. Instead, joint quality is quantified
by destructive mechanical testing. Therefore, welding parameters for industrial
applications and research are typically optimized for the mechanical properties of the
joint. While USMW is standardized for certain applications, the joint formation and
the influences on the welding process are not yet fully understood. Furthermore, the
welding process is prone to a large number of disturbances, which lead to fluctuating
joint quality without apparent change in boundary conditions.

Due to the large number of influencing variables with simultaneously

increasing expectations on joint quality and process reliability, the development and
validation of suitable monitoring methods is necessary.

Dept. of ME 2 IETCU
 Weld Quality Characterization by Vibration Analysis for Ultrasonic Metal Welding
Processes

CHAPTER 2
ULTRASONIC WELDING PROCESS AND PROCESS
INFLUENCES

A simplified description of a USMW process can be as follows: Several work

pieces (metal sheets, foil stacks, stranded wires, wire terminals) are placed on top of
each other on the anvil. A sonotrode (also known as a horn) is then placed on top of
the work pieces and exerts a vertical force on them. This force is counteracted by the
anvil below. After reaching a predetermined process force, the sonotrode is excited
into its working vibration perpendicular to the process force. Typically, the
oscillation frequency is in the range of 20 kHz to 60 kHz and the amplitude is in the
range of 10 µm to 50 µm. The oscillation causes relative motion between the
sonotrode, the work pieces and the anvil, which causes friction, local heat generation
and plastic deformation. This leads to the formation and growth of micro-welds in
the joining area, as well as changes in the micro structure.

The sonotrode design is decided based on, among other considerations, the
accessibility of the weld area and the geometry and size of the required joint. The
anvil geometries and additional fixtures, such as clamping and alignment devices are
also adapted to the work pieces and joint type. Figure 2.1 shows a linear oscillation
system consisting of a horn, a booster, a transducer and the generator. The Figure
also includes the vertical process force F described previously and the lateral
standing wave (red line) characterizing the sonotrode oscillations. The oscillation is
defined by the generator, which produces a high-voltage alternating current. This
oscillation is converted into a mechanical oscillation in the transducer; modern
transducers accomplish this by using stacked piezoelectric ceramic elements. The
booster serves as a mechanical support for the oscillation system. Furthermore, it
stabilizes the oscillation and, if necessary, is used to amplify its amplitude according
to a factor determined by its shape. Finally, the sonotrode transforms the mechanical
oscillation of the booster to the working amplitude and transmits it to the work
pieces. All oscillating parts are tuned to match the working frequency of the
particular machine, e.g. a machined designed for a working frequency of 20 kHz
machine requires tools

Dept. of ME 3 IETCU
 Weld Quality Characterization by Vibration Analysis for Ultrasonic Metal Welding
Processes

with a natural frequency close to 20 kHz. The anvil is usually rigid and holds the
lower work piece in place.

Fig 2.1 Shows the setup for the welding experiments presented later. The variable F represents the
vertical process force F exerted by the sonotrode on the work pieces, perpendicular to its direction of
oscillation.

Fig 2.2 Process stages

Dept. of ME 4 IETCU
 Weld Quality Characterization by Vibration Analysis for Ultrasonic Metal Welding
Processes

Figure 2.2 shows the six-stage model of joint formation, based on and valid
for spot welds of two copper sheets. The relative movements of sonotrode, work
pieces and anvil change during the different phases of the welding process. The work
pieces remain in the solid state throughout the entire process. Typical process
temperatures are between 0.4 and 0.6 times the respective work piece melting
temperature , thus avoiding common geometrical and metallurgical problems of
fusion welding processes, such as porosity, brittle or cracking. In addition, the low
process temperatures result in the aforementioned minimal formation of brittle
intermetallic compounds in dissimilar metal welds. The six-phase model is based on
high-speed video measurements. The first phase of the process is characterized by
large interfacial motions between the sonotrode and the upper work piece, as well as
between the work pieces themselves. While there is a significant temperature rise,
cold deformation still predominates in the joining zone, which is limited to small
micro contacts. Analogous to the cleaning phases described in other literature, the
second process phase is characterized by shearing and cleaning . Here, elevated
temperatures lead to limited evaporation of impurities and oxide layers and shearing
of roughness peaks, resulting in bare metal contact. Subsequent phases three, four
and five are characterized by thermal softening and plastic deformation. Welding
energy is increasingly transferred in the areas between the sonotrode and the upper
work piece, as well as the lower work piece and the anvil. In phase five, the joint area
has reached its maximum. The differences between the phases are characterized by
changes in the microstructure of the joint zone. With increasing energy input, over-
welding occurs in the final, the sixth phase of the process . If too much energy is
introduced in the weld area it starts to soften and, from a certain point, will weaken
the joint.

USMW processes are controlled by a fixed set of welding parameters, which

include the pressure/force applied by the sonotrode, the amplitude of the sonotrode
oscillation and a weld termination criterion. The welding power (electrical power
required by the transducer) is time- dependent and affected by the damping of the
sonotrode oscillation, which in turn depends on the joint formation and friction
losses. The weld termination criterion is given either by the elapsed welding time, a
defined welding energy consumption, or by reaching a predefined penetration depth
of the sonotrode into the work pieces. The latter is due to softening and plastic

Dept. of ME 5 IETCU
 Weld Quality Characterization by Vibration Analysis for Ultrasonic Metal Welding
Processes

deformation. Since only one parameter is usually used to define the end of the weld,
the other two variables can be used to monitor the process. However, only
monitoring of, e.g., welding time and penetration depth in an energy controlled
process has proven to be insufficient to detect or even prevent relevant process
variations in industrial applications. The parameters for a specific welding process
are determined using statistical design of experiments, machine learning based
algorithms and especially concerning the definition of boundary conditions, such as
sonotrode and anvil knurling and clamping system skill and experience of the process
engineer.

A large number of variables that affect the USMW process and cause the
variations mentioned above have been described and studied previously. All
variables interact with each other in terms of joint formation and the weld that can be
achieved. For a better overview, the influences are divided into machine-dependent
and work piece-dependent influences.

Most machine-dependent influences do not change without operator

intervention. These include machine setup and welding parameters. Frequency and
maximum welding power are usually machine-specific as they depend on the
generator, the transducer and the natural frequencies of the booster and sonotrode.
Increasing the normal force applied by the sonotrode on the work pieces and welding
energy, will increase the strength of the joint. After reaching certain optimum values,
thermal and mechanical loads lead to over-welding phenomena and joint quality
decreases. When welding to a predetermined welding energy, the welding time is
strongly correlated with the welding force acting perpendicular to the sonotrode
oscillations. The frictional conditions in the joining zone and therefore the required
welding power are determined by this very force. It is also important to emphasize
the role of the sonotrode oscillation amplitude, which has a strong influence on the
relative velocities between the work pieces. The booster, sonotrode and anvil are part
of the machine setup and are changed periodically, either to set up the machine for
different work pieces or to replace worn tools. Tool wear can be considered a long-
term disturbance of the process, whereas the adhesion of material to the tool surface,
which occurs particularly with aluminum, is more of a short-term or sporadic
disturbance. Sonotrode and anvil are defined not only by their material and the size
and basic geometry of the weld surface, but also by the exact design of the knurl

Dept. of ME 6 IETCU
 Weld Quality Characterization by Vibration Analysis for Ultrasonic Metal Welding
Processes

pattern of the weld surface and any protective coatings. Knurling influences the
frictional conditions to such an extent that deviations in the welding process can be
observed, such as a variation in the welding time or, in the case of multi-layer welds,
the starting point of the joint formation. In the case of work piece influences,
deviations from the exact reference state represent disturbances in the welding
process, with a significant effect on joint formation and strength. Depending on the
selected welding configuration, even the smallest changes can have a considerable
influence on the achievable process variation as, e.g., critical natural frequencies of
the work pieces may be excited. Since the work piece characteristics usually cannot
be adjusted (exactly) for each individual weld, they represent a disturbance for each
individual welding process. The geometry of the upper work piece affects the
damping of the vibration by the thickness and mass. In addition, the resonance
behavior of the work piece, characterized by length, width and clamping,
significantly influences the welding process. Reducing the thickness of the upper
work piece usually results in a significantly reduced energy requirement, as
otherwise over-welding occurs. Similarly, the hardness of the work piece material
has a significant influence on the welding process and the strength of the joint. At
this point we would also like to refer to the models of joint formation, in which
elastic and plastic deformation play an important role. Since the first phases of the
USMW process are characterized by the relative motion between the work pieces and
the welding process as a whole is characterized by minimal mixing of the interfaces,
the smallest changes in the surface quality have a major impact on the joint quality.
To a certain extent, surface impurities and oxides are displaced from the joint zone
during the cleaning phase of the welding processes. Nevertheless, the surface
structure, cleanliness and overall layer structure, such as oxide layers, adsorbed water
and gases, (ingrained) rolling oil residues or coatings, play an important role. The
rolling direction can also have a significant influence on the welding process, which
does not necessarily affect the joint quality.

Dept. of ME 7 IETCU
 Weld Quality Characterization by Vibration Analysis for Ultrasonic Metal Welding
Processes

CHAPTER 3
MONITORING OF ULTRASONIC METAL
WELDING PROCESS

There are two main motivations for observing and measuring USMW
processes. On the one hand, there is a strong interest in basic research to investigate
and improve the understanding of joint formation, welding process and influence
interactions. On the other hand, there is a strong interest in industry- and application-
oriented research to predict the quality of a weld and to monitor ongoing production,
which is subject to a large number of influencing factors.

The different objectives are also reflected in the equipment used for process
monitoring. In serial production, the above-mentioned statistical evaluation and
monitoring of machine-internal parameters (e.g. resulting welding time and depth) is
well established. However, some researchers claim that internal monitoring solutions
do not provide accurate and reliable results. An earlier study shows that the
influences of material hardness, roughness and weld position overlap all influence
welding time and welding depth in an energy controlled process. Since these effects
can cancel each other out but can lead to very different weld qualities, monitoring the
USMW process only using these scalar values is not feasible.

For scientific studies of joint formation or the influence of individual

variables, such as the surface roughness, measurements using high-speed camera and
laser vibrometers are state of the art. Often two or more laser vibrometers are used to
determine the relative and absolute oscillations of sonotrode, anvil or work pieces
simultaneously. Investigations with laser vibrometers show, that during the welding
process oscillations occur not only at the excitation frequency, but also in other
frequency ranges. In addition to the expected harmonics, i.e. multiples of the
excitation frequency, isolated secondary frequency bands occur during the welding
process. Depending on the weld configuration, their formation indicates a change in
the relative motions. Sensors that measure directly in the process but influence the
bond formation or machine stiffness are thermocouples and force sensors. However,
since laser vibrometers and high-speed cameras require a complex setup, an
unobstructed view of the process (accessibility/installation space) and high
investments in material and personnel qualification, they are not suitable for

Dept. of ME 8 IETCU
 Weld Quality Characterization by Vibration Analysis for Ultrasonic Metal Welding
Processes

industrial process monitoring. In addition, several studies report poor signal-to-noise

ratios or interferences due to particle and vapor formation. Nevertheless, the studies
conducted using these instruments show that external sensors often provide more or
clearer information about the progress of the individual weld than internal sensors. In
particular, the anvil is presented as a candidate measurement position. The above
mentioned studies have shown first results on the relative motions during the
USMW, but some of them do not consider the entire mechanical system or contradict
each other, especially with regard to the description of the end of the process and the
phenomena of over-welding.

3.1 APPLICATION OF MACHINE LEARNING

Like the use of measurement equipment for process observation, the goal of
using machine learning tools in the context of USMW is twofold. Some researchers
successfully apply machine learning to model the interaction between process
parameters and achievable joint quality, replacing other statistical methods usually
applied in the form of design of experiments. For USMW of steel on aluminum,
Zhao used a neural network to model and investigate the relationship between
process parameters and achievable tensile strengths. The work of Mongan confirms
the ability of machine learning to predict the strength of welds between steel and
aluminum components. Meng demonstrate the ability of machine learning response
surface methods to simultaneously optimize welding parameters for multiple load
cases, such as shear and peel test forces . Considered the use of a deep neural
network architecture to predict the failure load of welded joints in ultrasonic welding
of composite materials (plastics).

Instead of time series data, the described neural network model uses five
process parameters as input data, including annealing temperature, surface condition,
welding energy, plunge speed, and trigger force. The model is used to estimate the
failure load of joints in a lap shear test as a measure of weld quality. The authors
report a maximum relative error of 4.5% on a test data set consisting of 54
specimens. Ahmed et al. used the CART variant of the decision tree algorithm to
predict the weld nugget width in resistance spot welding applications, using eleven
design and process parameters as inputs. They further leveraged the interpretability

Dept. of ME 9 IETCU
 Weld Quality Characterization by Vibration Analysis for Ultrasonic Metal Welding
Processes

of decision trees to extract informative weldability decision rules from the model.
The use of machine learning to monitor and predict the achieved weld quality for
each individual weld is not very established yet. For resistance spot welding Xing et
al. employed a random forest model to predict the quality of cold-rolled mild steel
welds. The authors considered time-series data consisting of the root mean square
welding resistance per half cycle, defined as the dynamic resistance, as inputs with
three weld quality types acting as classification targets. The resistance input was pre-
processed by extracting welding time stamps and resistance values from four
characteristic points in the resistance curve, as well as resistance gradients in the
intermediate segments. With the addition of the average dynamic resistance, its
standard deviation and the maximum gradient per curve, the model input vectors
comprised ten time series features per weld.

Furthermore, inclusion of the process parameters electrode force, welding

time and welding current as inputs was investigated. Using the established random
forest model, the authors report a classification accuracy of 93.6% without and
98.8% with the inclusion of electrode force, welding time and welding current. The
successful application of machine learning methods for the precise prediction of
welding conditions confirms the validity of this approach and suggests its use for the
assessment of weld quality in USMW based on measurements from external sensors.

Dept. of ME 10 IETCU
 Weld Quality Characterization by Vibration Analysis for Ultrasonic Metal Welding
Processes

CHAPTER 4
METHODOLOGY AND MATERIALS

In the analysis of their influences on process and joint quality, a distinction

can be made between defined process parameters (e.g., welding time or normal force)
and process disturbances (e.g., surface contamination, changes in microstructure).
The influences of the latter are particularly relevant for the monitoring of production
processes, as they are typically unknown and can change for every weld. Against this
background, this study aims to investigate whether an external, but industrially
applicable, sensor technology can detect and predict the influence of process
disturbances on the weld quality. In order to reproduce their effects on weld quality,
typical process disturbances are simulated by selecting appropriate materials, e. g., by
varying the rolling process to simulate changes in microstructure due to forming
processes, or by adjusting the condition of the work pieces. In a previous study, the
authors have already described the experimental setup used for the experiments.
However, it is supplemented by new investigations of the temporal development of
the joint formation. The influence of welding position, material hardness, rolling
direction, surface cleanliness and surface roughness on a spot weld of two copper
sheets is investigated in this study. Figure 4.1 lists the changes made starting from
the reference welding process.

Fig 4.1 Overview of variations of reference

Dept. of ME 11 IETCU
 Weld Quality Characterization by Vibration Analysis for Ultrasonic Metal Welding
Processes

CHAPTER 5
WELDING SETUP

The machine setup is shown in Figure 1.1, detailed information on the

machine configuration can be found in Table 5.1.

Table 5.1 Configuration of welding machine

COMPONENT CONFIGURATION

Transducer/Generator 4 kW maximum power / 20 kHz working

frequency

Booster 1 x 1:1 booster (newly procured for the

experiments, no wearout)

Sonotrode λ-half longitudinal, 8 mm x 8 mm

working surface (newly procured for the
experiments, no wear-out)

Sonotrode knurling 0.9 mm x 90∘ pyramids (diagonal to

oscillation)

Anvil knurling 0.5 mm x 90∘ pyramids (diagonal to

oscillation), truncated by

0.1 mm

The reference specimen consists of two Cu-ETP (CW004A) copper sheets of

equal size welded in overlap. Cu-ETP is a typical material for electrical and
electronic components. The reference material is acquired in the half-hard (HH)
rolling condition R240, which corresponds to a nominal tensile strength (RM) of 240
MPa (measured with RM =256.91 MPa). The dimensions of the 1 mm thick plates,
their rolling direction in the reference state (longitudinal) as well as the overlap and
welding position are shown in Figure 5.1. The figure also shows the Acrylonitrile
Butadiene Styrene (ABS) positioning mask used to ensure a repeatable position of
the lower work piece. No other clamping devices were used in the experiments; the

Dept. of ME 12 IETCU
 Weld Quality Characterization by Vibration Analysis for Ultrasonic Metal Welding
Processes

upper work piece was manually positioned based on markings on the anvil, allowing
the work piece to oscillate during the welding process.

Fig 5.1 Work piece geometry and position

The welding parameters are given in Table 5.2. They are optimized for the
reference configuration by a statistical design of experiments (central composite
design) for maximum shear force in a lap shear test. The parameter optimization was
conducted in two iterations. In a first iteration, the relevant parameter space was
identified by conducting preliminary welds with weld parameters selected based on
experience. In a second iteration, a full factorial parameter study was performed
based on the previously identified parameter space to determine the optimal welding
parameters. The welding parameters welding force and amplitude are given as
approximate values based on measurements, as on the welding machine used for the
experiments, they are only parametrized as cylinder pressure in bar and relative
transducer amplitude in percent.

Table 5.2 welding parameters

PARAMETER VALUE

Energy 2250 J

Force 3000 N

Amplitude 25 µm (At the tip of the sonotrode)

In the energy-controlled mode, the typical welding time of the established

reference process is about 1170 ms. The joint growth for the reference process is

Dept. of ME 13 IETCU
 Weld Quality Characterization by Vibration Analysis for Ultrasonic Metal Welding
Processes

shown in Figure 5.2. Each picture represents a weld time increment of 150 ms,
starting with (a) at 150 ms and ending with (h) at 1200 ms. A time-controlled process
was used to obtain these pictures, but the force and amplitude parameters were not
changed. The specimens were welded in a rotated overlap so that they could be tested
destructively by peel testing. This is necessary to obtain optically evaluable surfaces
without smearing of the weld surface in the joint zone, as is typical for shear testing.
The pictures show the joint surface in the form of peel fracture surfaces. For better
visualization, the contrast between the ruptured material and the unjoined surface
was increased by controlled oxidation of the welded specimen prior to the peel test.
This oxidation was achieved by briefly tempering the specimens in an oven under
ambient air. All images show the fracture area of the upper specimen beneath the
sonotrode. The depicted fracture surfaces support the stage models used to describe
the welding process (see Figure 2.2), with little joined area at the beginning of the
welding process.

Fig 5.2 Formation and growth of joint area

The first picture (a) at 150 ms welding time shows that, except for isolated
joint areas, only the cleaning process has started. The following pictures show the
further formation and large-area growth of micro- welds as described in the literature.
The images (a)–(d), taken at welding times of 150 ms to 600 ms, show fracture
surfaces with clear contrast. In addition, they show edge areas below the sonotrode
working surface that were created during the welding process. These areas are not

Dept. of ME 14 IETCU
 Weld Quality Characterization by Vibration Analysis for Ultrasonic Metal Welding
Processes

bonded, but the process brings the plate surfaces sufficiently close together to reduce
surface oxidation. It can be seen that the joint growth reaches a maximum at 1050 ms
(g) before the typical weld end occurs at 1170 ms, as shown in the image (h) taken at
1200 ms in Figure 2.2. The time-dependent growth of the joint surface suggests that
monitoring vibrations resulting from forces transmitted across the weld surface may
represent a promising approach for the weld quality characterization. Table 3 lists the
test series performed and analyzed for this paper. The test series abbreviations refer
to their configuration, i.e., the base material strength (e.g., HH - CW004 HH), the
rolling direction (e.g., L - longitudinal orientation with the long side of the work
piece as reference), the weld overlap (e.g., P1 - position 1 with 25 mm overlap), and
an optional suffix to describe the surface condition. The Table 5.3 lists all weld tests
series.

Table 5.3 Welding experiment test series configuration and result

Dept. of ME 15 IETCU
 Weld Quality Characterization by Vibration Analysis for Ultrasonic Metal Welding
Processes

CHAPTER 6
WELDING RESULTS OF THE TEST SERIES AND
QUALITY FORECAST BASED ON INTERNAL
DATA

Fig 6.1 Overview of achieved shear strength and weld time

The maximum shear strength of the welded specimens against the respective
welding time as a scatter plot. The welds are grouped according to their configuration
and the shear strength of the specimen material. In addition, an individual frequency
distribution chart for each group shows the occurrence and dispersion of strength and
welding time. While the welds with soft specimen material (S) show a small scatter
in both weld time and strength (combined with comparatively low mean values), the
welds with hard material (H) show widely scattered values. Some of the
configurations, H-L-P1 and H-P-P1, achieve particularly high strengths of 5.4 kN on
average, with low welding times and low scatter. In contrast, the strength of the H-L-
P2 welds varies between 5.2 kN and 2.2 kN with welding times ranging from 1.1 s to
2.6 s. The half-hard (HH) specimen welds lie between the clusters formed by the
hard and soft specimen welds. For all materials, the welds in position P1 are superior
to those in P2. Overall, the results show a very large scatter; therefore, from a process
engineering point of view, no process capability is given. Since the focus of this
paper is on weld quality estimation, a detailed discussion of the weld results and the

Dept. of ME 16 IETCU
 Weld Quality Characterization by Vibration Analysis for Ultrasonic Metal Welding
Processes

influencing factors is beyond the scope of this article. Table 5.3 shows the mean
values and the number of samples within each group, as well as the standard
deviations as a percentage of the mean strength, graphically presented in Figure 6.1.
The number of samples refers to the number of welds with a quality label, i.e. a valid
maximum shear strength value.

In a previous study, the authors proposed a process monitoring method aimed

at predicting the maximum tensile strength of welds using only machine-internal
data. This method is based on Gaussian process regression (GPR) with a quadratic
kernel using scalar values of weld time and depth as input data. Furthermore, it was
shown that a simple extraction of accessible machine-internal time series data allows
feature engineering and a GPR based on these features is able to increase the
prediction accuracy without the need for additional external sensor systems.
Specifically, the welding power and sonotrode height measurements sampled at 10
Hz from the welding machine itself were used as the basis for quality prediction.
Here the horn height is used as a measure of the penetration depth of the horn into the
upper work piece. Scalar features describing the time series data were extracted from
the measurements and used to train the GPR with a quadratic kernel. These features
include the weld time, the penetration depth, the time and height of the maximum
weld power, the curvature of the power curve at the maximum, the increase in weld
power during the first 60 ms, and the final welding power and its slope during the
last third of the weld. Training the GPR with randomized 2/3 of the data set allows
prediction of weld tensile strength with a lowest coefficient of determination R2 of
0.92 and a highest root mean square error (RMSE) of 266 N. These values represent
the worst cases of five random repetitions of data selection and training. The
corresponding average absolute mean error is 198 N, and the absolute median error is
150 N.

Dept. of ME 17 IETCU
 Weld Quality Characterization by Vibration Analysis for Ultrasonic Metal Welding
Processes

CHAPTER 7
MEASUREMENT SETUP
The measurement setup used for the experiments is shown in Figure 1.1. It
consists of two laser displacement sensors (i.e. laser triangulation sensors, LTS),
which measure the sonotrode and anvil vibrations (see figure 7.1). In addition, two
piezoelectric force sensors connected to a summing box are used to measure the
normal forces acting on the anvil. Due to poor accessibility and space limitations, a
displacement sensor with a larger measuring range has to be used at the anvil,
resulting in a lower resolution compared to the sensor used for the sonotrode. An
overview of the specifications is given in the Table 7.1. In addition to these external
measurements, a number of machine internal measurements are taken. These include
the welding power, the pressure on the pneumatic cylinder that generates the welding
force, the welding time and the penetration depth. Due to the poor temporal and
spatial resolution of the internal penetration depth sensor, an additional inductive
displacement sensor (i.e., a linear variable differential transformer) was mounted
directly on the sonotrode support to provide higher resolution measurements.

Fig 7.1 Measurement setup

The selection of this measurement setup is motivated by an cost- effective and

easy integration into existing welding machines as used in industry. Against this
background, additional measurement technologies have been tested as well, e.g., four
strain gauges installed as a full bridge on the backside of the anvil. Due to
spacelimitations, strain gauges cannot be attached on the sides of the anvil (see.
Fig.8). Installation of strain gauges on the knurled work surface at the top of the anvil
is also not possible. Thus, in this case normal and bending strains cannot be
distinguished (superimposed bending). However, for weld quality prediction,

Dept. of ME 18 IETCU
 Weld Quality Characterization by Vibration Analysis for Ultrasonic Metal Welding
Processes

vibration and penetration depth measurement have proven to be more suitable, as it

will be explained in the following.

Table 7.1 Sensor specifications

MEASURED DISTANCE NORMAL PENETRATION

QTY. FORCE DEPTH

Position Sonotrode, Anvil Anvil Sonotrode

Measuring 20 mm, 50 mm - -
distance

Measuring range 0.6 mm, 1.2 mm 0-26 KN 2 mm

Sample frequency 200 KHz 125 KHz 125 KHz

Tolerance ± 1.2 µm, ± 4 µm 1% ± 1.2 µm

Dept. of ME 19 IETCU
 Weld Quality Characterization by Vibration Analysis for Ultrasonic Metal Welding
Processes

CHAPTER 8
DATA PROCESSING

A high degree of automation in data processing is essential given the large

number of measurements and the goal of transferring the presented methods to
industrial applications. In the case of the welding machine used in our experiments, it
is not possible to have a precise signal to trigger data acquisition at the start of the
welding process. Only the signal indicating the beginning of the sonotrode’s
downward motion towards the work piece is available. However, the time between
the start of the descent and the start of the welding process is not constant. Therefore,
all measurements must be synchronized prior to the analysis. Due to the small
welding amplitude and the comparatively low resolution of the displacement sensors,
the measurement noise does not allow a gradient based method to determine the
beginning of the sonotrode oscillation amplitude increase and thus the beginning of
the welding period. Therefore, a method based on the amplitude of the raw signal is
used. To this end, an offset correction is performed by subtracting its mean value
from the measurement signal. Then, the start of all measurements is synchronized by
identifying the first measurement point with an absolute value greater than 80 % of
the final weld amplitude and truncating the signal 10 ms before this point. Since the
gradient of the sonotrode oscillation amplitude increase is controlled by the welding
machine and therefore mostly independent of the surface condition of the work
pieces, this method leads to sufficiently accurate results. Both the laser sensor for
measuring the oscillation of the sonotrode and the anvil are driven by the same
controller and start recording data at the same time. A similar technique is used for
all other sensors except the strain gauge and the penetration depth sensor. For the
latter, a gradient based method is used due to the low noise level and high variance in
amplitude.

Subsequently, the vibration measurements are transformed into the frequency

domain and the amplitude time series at the operating frequency (approximately 20
kHz) and its harmonics (40 kHz, 60 kHz and 80 kHz) are extracted. A steep, almost
impulsive increase in amplitude at the beginning of the welding cycle and rapid

fluctuations in the measured amplitudes make the continuous wavelet transform the
most appropriate method. When using a short-time Fourier transform, the choice of

Dept. of ME 20 IETCU
 Weld Quality Characterization by Vibration Analysis for Ultrasonic Metal Welding
Processes

the width of the windowing function has proven to be difficult, taking into account
the desired automation of the data processing and the variety of sensors, parameters,
frequencies and levels of variance of the measurements.

The continuous wavelet transform compares the signal to a translated and

dilated (compressed or stretched) version of a so-called mother wavelet . The dilation
factor is typically referred to as the scale. While there is a general correspondence
between scale and frequency, there is no precise mapping. Concepts such as the
center frequency, calculated by approximating the wavelet with a sine wave, are used
to transform from the scale to the frequency domain.

The exact sonotrode operating frequency is tuned by the welding machine

manufacturer to match the sonotrode’s natural frequency and depends on the overall
process conditions. Therefore, it deviates slightly from the specified working
frequency of 20 kHz and its exact value is unknown. In addition, measurements on
the same machine under similar conditions have shown frequency variations during
the welding process of 100 Hz around the set value. To account for the operating
frequency deviation and variation, a frequency range is defined around the spectral
line representing the operating frequency of the sonotrode (e.g. 17 kHz to 23 kHz).
For each scale corresponding to a frequency within this frequency range, the sum of
all elements in the associated wavelet transform is calculated. Here,
synchrosqueenzing is applied to improve frequency resolution. Synchrosquencing is
a reallocation method used to sharpen the time-frequency resolution of the wavelet
transform.

Then, the dominant scale with the highest sum is determined and its
maximum absolute value of the synchrosqueezed wavelet transform is evaluated. In
the next step, all neighboring scales with a maximum absolute value greater than
10% of the dominant scale’s maximum are determined and combined into a scale
band. Only consecutive scales are considered, i.e., those that are immediate
neighbors. A gap in the scale band indicates an additional vibration signal with a
frequency similar to the working frequency. Thus, all scales below/above this gap
would be neglected. Finally, the signal is reconstructed from its synchrosqueezed
wavelet transform within this scale band (as a complex pointer) and its amplitude is
extracted. The subsequent signal analysis has shown that in the case of the
application presented here, the amplitude time series at the working frequency is the

Dept. of ME 21 IETCU
 Weld Quality Characterization by Vibration Analysis for Ultrasonic Metal Welding
Processes

most meaningful for the characterization of the weld quality. Therefore, the
following studies will focus on the working frequency and neglect its harmonics.
Figure 8.1 shows a selection of the envelope curves of the anvil oscillations obtained
by the synchrosqueezed wavelet transformation. The curves are grouped by
configuration and the shear strength obtained for each weld is indicated by the line
color.

It shall be pointed out that the low amplitudes of the anvil vibrations in the
reference configuration (Fig. 8.1) at the beginning of the weld formation agree well
with the low level of joint formation from the first process steps and comply with the
observations in Fig. 2.2. The steep inclines of the vibration amplitude as to be
observed for the H-L-P1 configuration shown in Fig. 8.1 provide new impulses for
future work on the topic of joint formation and process stages under different
boundary conditions.

Fig 8.1 Exemplary time series

Dept. of ME 22 IETCU
 Weld Quality Characterization by Vibration Analysis for Ultrasonic Metal Welding
Processes

8.1 METHODOLOGY FOR MAXIMUM TENSILE FORCE

To predict the quality of a weld with unknown characteristics without

physical testing, a supervised machine learning algorithm, the random forest
algorithm, is used. Given annotated data from experimental evaluation, machine
learning algorithms can learn the complex relationship between sensor data and
expected weld quality, as well as the most likely type of process disturbance.

The approach presented below does not rely on the time series characteristics
specific to each type of sensor data. Furthermore, only sensor data that does not
require direct inspection of the welded parts was considered as input. Instead,
automatic feature extraction and filtering are used, as well as the ability of machine
learning algorithms to learn complex relationships.

Dept. of ME 23 IETCU
 Weld Quality Characterization by Vibration Analysis for Ultrasonic Metal Welding
Processes

CHAPTER 9
THE RANDOM FOREST ALGORITHM

The random forest algorithm belongs to the family of ensemble learning

methods in machine learning and can be used for both classification and regression
tasks. By combining a multitude of decision tree learners, the random forest
algorithm can produce predictions with reduced variance compared to a single
learner, without increasing prediction bias. Each decision tree makes an estimate of
the value of the output variable by learning a tree-like structure of simple decision
rules over the set of input data features, where each decision is represented by a node
in the tree. Given a collection of input vectors X and corresponding output vectors Y.

---------------------------- Eqn-9.1

----------------------------Eqn-9.2

The algorithm performs a binary split θN = (f,tN) of the collections of input-

output pairs C = (xi,yi)|i∈ {1,...,K} based on the value of some input feature f and a
threshold tN defined at Node N, resulting in a left and a right data partition.

------------------------------- Eqn 9.3

------------------------------- Eqn 9.4

This procedure is repeated recursively on the partitions CN,l,CN,r until a

termination criterion is met, such as the maximum depth of the tree, or each partition
contains only one sample. The random forest algorithm improves the performance of
decision trees by applying bootstrap aggregation to a set of tree models. For each
decision tree, a unique training data set is constructed by randomly sampling from
the original training data. This ensures that no data element is selected more than
once in a training data set. The model is trained using these newly constructed data
sets. The prediction of the random forest on a set of test data is then the average of all
individual predictions for regression problems, or the majority vote in the case of
classification.

Dept. of ME 24 IETCU
 Weld Quality Characterization by Vibration Analysis for Ultrasonic Metal Welding
Processes

CHAPTER 10
FEATURE EXTRACTION

The complexity of the relationship between input and output data is reduced
by extracting descriptive features from the time series measurement signals. Not only
does this reduce the computational costs of model training, but it also reduces the
size of the input data. These include geometrical and statistical measures such as the
mean and median values of the signal, the relative positions of significant change
points normalized to the welding time, or measures to access the symmetry of the
signal (Fig. 10.1a). A complete list of features is given in the Table 10.1. Each
feature is calculated for the original signal and its first two numerically calculated
derivatives. In addition, all features are accessed for the complete signals and for a
number of equally spaced sections of the signal (Fig. 10.1c), which greatly increases
the number of individual feature values and adds the ability to correlate feature
values with welding process phases. In the first step of the training phase, the random
forest is trained using all features of the provided inputs, followed by the
identification of the (model-dependent) most significant features on the training data
set using SHAP, a game- theoretic approach to explainable machine learning. The
model is then retrained using the set of significant features and validated on a
previously unseen test dataset where only the relevant features are extracted.

Fig 10.1 Change point positions (a) used as features extracted from each signal, numerical
differentiation of signals (b) and splitting of a signal into predefined segments (c)

Dept. of ME 25 IETCU
 Weld Quality Characterization by Vibration Analysis for Ultrasonic Metal Welding
Processes

The location of change points is computed by binary segmentation using the

Python library ruptures. The quadratic error loss is used as the cost function for the
segmentation, which computes the least squares deviation to detect shifts in the mean
of the signal. For curvature estimation, only the first section of the measurement data
series (by default 1/6 of the length of the series) is considered. The central point is
defined to be at the maximum in this section and the other two at a fixed distance
(±1/12 of the length of the series) from it in both directions. The hyper-parameter
optimization is conducted using a grid search. By far the best SHAP values are
attributed to the mean values of the equally spaced sections of the measurement data
series (here five sections were chosen) and the total sum over the entire data series.

Table10.1 Welding experiment test series configuration and results.

FEATURE INFLUENTIAL DESCRIPTION

PARAMETER

Length - Length of the

measurement vector. The

sampling frequencies are

held constant over all
experiments, thus the
number of measurement
points directly
corresponds to the weld
time.

Max. value - Maximum value of the

measurement vector

Min. value - Minimum value of the

measurement vector

Sum - Sum over the whole

measurement vector

Euclidean norm - Euclidean norm of the

measurement vector

Dept. of ME 26 IETCU
 Weld Quality Characterization by Vibration Analysis for Ultrasonic Metal Welding
Processes

Median - Median of the

measurement vector

Mean norm - Mean of the

measurement vector

Symmetry Threshold value r Symmetry of the

measurement vector x,
defined as Boolean
variable. True if:

(‖mean(x) − median(x)‖/

max(x) − min(x)) < r

No. of significant peaks Minimal peak height, Number of significant

minimal distance b/w 2 (positive and negative)
consecutive peaks peaks in the vector, after
lowpass filtering.

Position of change No. of points, detection Relative position of the

points method, parameters of points in the vector that
the detection method have been identified as

change points. Change

points can be used for
the segmentation of the
signal.

Dept. of ME 27 IETCU
 Weld Quality Characterization by Vibration Analysis for Ultrasonic Metal Welding
Processes

Curvature Choice of points, Using three points, a

smoothing method (here circle can be
running mean) parametrized, which is
used as an

approximation of the
curvature of the

measurement vector. This

proved to be more
meaningful than
characteristic

values based on, e.g., the

second derivative. The
curvature is evaluated

on the smoothed
measurement vector for
either, every change point
section or for a predefined
signal section. By default,
only the first section of
the vector is taken into
account, with the central
point at the maximum and
the two others at a fixed
distance from it in both
directions.

Dept. of ME 28 IETCU
 Weld Quality Characterization by Vibration Analysis for Ultrasonic Metal Welding
Processes

CHAPTER 11
STRATIFIED K-FOLD CROSS-VALIDATION

The most common model evaluation process involves two steps. First, the
model undergoes a training step in which its parameters are adjusted to achieve
optimal prediction performance on a training data set. In the subsequent validation
step, the performance of the optimized model is then evaluated on a previously
unseen test data set using predefined metrics. This is necessary to estimate how well
the model generalizes to new data. However, smaller data sets are often prone to
selection bias due to statistical variation, as the validation result may depend heavily
on the choice of training and test data.

To mitigate this problem, a stratified k-fold cross-validation procedure is

used for the validation presented in the latter. Here, the full data set containing all
independent samples is divided into k subsets of approximately equal size. In the
case of a classification task these subsets contain similar proportions of samples for
each target class. Iterating over all subsets, each is then used once for validation,
while the remaining subsets are used to train the model, as shown in Fig. 11.1. The
collection of k validation results can then be used to obtain an estimate of the mean
prediction performance according to the chosen metrics, allowing a more informative
assessment of the quality of the model and its ability to generalize.

Fig 11.1 K-fold cross-validation. Each subset is used exactly once for validation.

Dept. of ME 29 IETCU
 Weld Quality Characterization by Vibration Analysis for Ultrasonic Metal Welding
Processes

CHAPTER 12
RESULT AND DISCUSSION

This section discusses the quality prediction results obtained with the
previously presented data acquisition, processing, and machine learning strategies.
Two machine learning problems are addressed, quality prediction in the form of
tensile shear strength estimation and, for a smaller data set, a welding process
disturbance classification.

12.1 PREDICTION OF MAXIMUM SHEAR FORCE

Using the established training and prediction pipeline, the prediction of the
achieved weld strength, measured by the maximum tensile shear force, was
evaluated. Both the inclusion of a single measurement signal type and combinations
of multiple signal types were considered. The results of these experiments, obtained
using a 5-fold cross-validation method (4.3) to account for the relatively small size of
the data set, are summarized in Table. The parameters of the random forest
regression are listed.

Table 12.1 Estimation error for different combinations

DATA R² MEAN MEDIAN MEAN MEDIAN

TYPE ABS. ABS. REL. REL.
ERROR ERROR ERROR ERROR
(N) (%) (%)

ANV 0.906 209.21 136.12 5.75 3.73

POW 0.825 259.75 137.69 6.81 3.98

PEN 0.959 154.02 102.59 4.86 3.39

STR 0.595 258.24 215.12 8.62 7.56

ANV & 0.939 187.03 125.88 4.90 3.33

POW

Dept. of ME 30 IETCU
 Weld Quality Characterization by Vibration Analysis for Ultrasonic Metal Welding
Processes

As is evident from the numerical results, the anvil oscillation amplitude time
series signal (ANV) and the penetration data (PEN), with R2 values of 0.906 and
0.959, respectively, are particularly well suited for use as inputs. The graph of
measured shear forces against predicted shear forces for a representative test data set
in Fig. 12.1 confirms these results. However, it can be seen from Fig. 12.1b that, due
to the late inclusion of the high-resolution penetration sensor, there are few samples
in the mid-shear force range of 4000 N to 5000 N. As a result, the estimates based on
the anvil oscillation data are of greater significance. The results obtained using strain
gauge (STR) data are significantly worse and were not considered further.

Table 12.2 Random Forest parameters

PARAMETER CLASSIFICATION EGRESSION

No. of trees 1000 1000

No. of features No. of features Square root of the no. of

features

Maximal depth 1000 1000

Minimal no. of samples 2 2

for a split

Using different combinations of two of the three most relevant signal types
anvil vibration data, power data and penetration data, the previous results could be
improved, with the combination of welding power (POW) and anvil vibration data
achieving a R2 of 0.939 with results spanning the entire observed shear force
spectrum. A comparison of Fig. 12.2c with Fig.12.1a shows that this combination
significantly reduces the presence of outliers, resulting in lower mean absolute and
relative errors as shown in Table 12.1. Combinations of three or more signal types
were considered, but did not result in further improvements.

Dept. of ME 31 IETCU
 Weld Quality Characterization by Vibration Analysis for Ultrasonic Metal Welding
Processes

Fig 12.1 Prediction results based solely on anvil vibration data (a) or penetration data (b)

Fig 12.2 Prediction results for the combination of anvil vibration and penetration data (a), power and
penetration data (b) and power and anvil vibration data (c)

Dept. of ME 32 IETCU
 Weld Quality Characterization by Vibration Analysis for Ultrasonic Metal Welding
Processes

CHAPTER 13
CLASSIFICATION OF WELD DEFECTS

While predicting the maximum achievable shear force as an indicator of weld

quality is particularly relevant for automated sorting of welded work pieces, it does
not provide any information about the condition of the work pieces during the
welding process (e.g. surface condition, hardness). To evaluate whether it is possible
to predict the welding conditions from the collected sensor data, a random forest
classifier was trained on the features extracted from the anvil vibration data. Only a
subset of the original welding configuration variations (Table 5.3) was considered
(i.e. HH-L-P1, HH-L-P1-o, HH-L-P1-rf, HH-L- P1-sf, H-L-P1, S-L-P1), as including
data collected at the shifted welding position P2 resulted in minority classes with too
few samples. Further experiments would have to be conducted to allow classification
on the dataset welded at position P2. The parameters of the random forest classifier
are listed in Table 7. The average results obtained with this classification method,
again using a 5-fold cross-validation, are shown in Fig. 13.1a, in the form of a
confusion matrix normalized over the true classes.

The classifier was able to clearly separate the different welding conditions,
correctly identifying the presence of oil on the surface of the weld material and the
preparation of the surface with rough abrasive fabric (HH-L-P1-rf) in all cases. Only
the classification of the specimens prepared on the surface with light abrasive cloth
(HH-L-P1-sf) shows a classification accuracy below 90%. This can be explained by
the apparent similarities between the sensor data collected for this case and the data
for the reference case. To evaluate the overall classification accuracy, the macro and
micro averages of the F1 score are assessed. The F1 score is defined as,

F1 = TP/(TP +(1/2)(FP+FN)) -------------------------------Eqn-13.1

Here TP stands for the true positives, FN for the false negatives and FP for
the false positives. The macro-averaged F1 score is the arithmetic mean of all per-
class F1 scores, while the micro-averaged F1 score is the sum of all true positives,
false negatives and false positives. Overall, satisfactory macro and micro average F1
scores of 94% and 95%, respectively, were achieved. Furthermore, using a one-

Dept. of ME 33 IETCU
 Weld Quality Characterization by Vibration Analysis for Ultrasonic Metal Welding
Processes

against-all approach, the receiver operating characteristic (ROC) curve for each class,
which illustrates the diagnostic quality of a model as its decision threshold is varied,
was examined and a macro- and micro-average ROC curve over all classes was
computed. The area under the macro averaged curve was then used as an evaluation
metric, as it is independent of the chosen decision threshold and indicates how well
the classes are separated by the classifier. With a value of 0.99, the area under the
ROC curve further confirms the conclusion that the chosen approach is well suited
for the identification of unknown welding conditions from sensor data.

Fig 13.1 Confusion matrix of weld defect class prediction normalized over the true classes (a) and
averaged ROC curves (b) using the random forest algorithm.

Dept. of ME 34 IETCU
 Weld Quality Characterization by Vibration Analysis for Ultrasonic Metal Welding
Processes

CHAPTER 14
CONCLUSION

In summary, it can be concluded that it is possible and reasonable to monitor

the USMW using external measurements, especially when vibration and penetration
depth measurements are considered. Different monitoring methods have been
successfully tested. Possible sensor setups for different applications were validated
and characterized for their suitability for monitoring. Significant signal features were
identified and a method for automated signal processing was implemented. It should
be emphasized that the analysis of the oscillations at the operating frequency in the
range of 20 kHz already provides the relevant information for the formation of the
joint. The quality of the data obtained from the various measurement signals is
particularly noteworthy. The results indicate that welding power, penetration depth
and anvil vibration measurements are particularly valuable sources of information for
the welding configurations studied here.

Considering the anvil vibration and the penetration depth, a regression model
with a R2 value of 0.964 and a mean relative estimation error on the shear strength of
4.30% could be established. Taking into account the internal power curve and the
penetration depth, a slightly less accurate regression model with a R2 value of 0.963
and a mean relative estimation error of 4.45% could be obtained. These results far
exceed the usual established process monitoring methods on M-USS machines. In
this context, the laser triangulation sensor used to detect the anvil oscillation can also
be replaced by alternative measurement technologies, such as eddy current sensors.
The robustness of the quality prediction using random forest algorithms suggests that
the developed methods can be transferred to other welding configurations.

Validation of the selected measurement technology and adaptation of the data

preparation to the welding process is necessary for the particular application before
quality prediction or classification can be learned.

Dept. of ME 35 IETCU
 Weld Quality Characterization by Vibration Analysis for Ultrasonic Metal Welding
Processes

REFERENCES

• Agency, I. E.,. Global EV outlook 2021: accererating ambitions despite the

pandemic.https://iea.blob.core.windows.net/assets/ed5f4484-f556-4110-8c5c-
4ede8bcba637/GlobalEVOutlook2021.pdf.
• Ahmed, F., Kim, K.-Y., 2017. Data-driven weld nugget width prediction with
decision tree algorithm. Procedia Manuf. 10, 1009–1019.
https://doi.org/10.1016/j.promfg.2017.07.092.
• Balz, I., Raad, E.A., Rosenthal, E., Lohoff, R., Schiebahn, A., Reisgen, U.,
Vorlander, ¨ M.,2020. Process monitoring of ultrasonic metal welding of
battery tabs using external sensor data. J. Adv. Joining Process. 1, 100005.
https://doi.org/10.1016/j. jajp.2020.100005.
• Breiman, L., 1996. Bagging predictors. Mach. Learn. 24 (2), 123–140.
https://doi.org/10.1007/BF00058655.
• Deutsches Kupferinstitut, e. V.,. Werkstoff-datenblatter ¨ - cu-ETP.
• DiFinizio, T., 2005. Ultrasonic intelligence. Wire Cable Technol. Int. 33 (5),
88–90.
• Gordon, A.D., Breiman, L., Friedman, J.H., Olshen, R.A., Stone, C.J., 1984.
Classification and regression trees. Biometrics 40 (3), 874.
https://doi.org/10.2307/2530946. Greitmann, M.J., Adam, T., Holzweißig,
H.G., Stroh, D., Wagner, G., Wiesner, P., Züst, R., 2003. Gegenw¨artiger
stand und zukunftsaussichten der sonderschweißverfahren -
ultraschallschweißen. Schweißen und Schneiden 55 (6).306–308, 310, 312–
314 Harthoorn, J.L., 1978. Ultrasonic metal welding. Technische Hogeschool
Eindhoven,Eindhoven. Dissertation.

Dept. of ME 36 IETCU