key: cord-0319335-x4s7pi7e
authors: Lee, Jiwoong; Noh, Inwoong; Jeong, Sang Ik; Lee, Yongho; Lee, Sang Won
title: Development of Real-time Diagnosis Framework for Angular Misalignment of Robot Spot-welding System Based on Machine Learning
date: 2020-12-31
journal: Procedia Manufacturing
DOI: 10.1016/j.promfg.2020.05.140
sha: 7f902ad1ad7d86c2d48c0caa127b7f49a4ffe565
doc_id: 319335
cord_uid: x4s7pi7e

Abstract This paper focuses on the real-time online monitoring and diagnosis framework for the angular misalignment of the robot spot-welding system, which can result in significant quality degradation of a weld nugget such as porosity. The data-driven approach is applied by installing the voltage and current sensors, collecting the associated mass data and processing them under normal and abnormal (angular misalignment) conditions. Two categories of features are extracted from the dynamic resistance (DR) and the voltage and current ones that are decomposed by wavelet transform (WT). The DR features are extracted from the DR profile and some critical features are selected by a t-test methodology. In the case of the WT-based features, the critical ones are selected by a max-relevance and min-redundancy (mRMR) and a sequential backward selection (SBS) wrapper. Consequently, three types of critical feature sets, such as DR features, WT features, and hybrid features combining those, are prepared to train machine learning-based models. Support vector machine (SVM) and probabilistic neural network (PNN) are applied to establish the diagnosis models, and the diagnostic accuracy and robustness are evaluated. Finally, the software for the on-line monitoring and diagnosis for angular misalignment of robot spot-welding system is developed and demonstrates its real-time applicability in an industrial site.

In modern manufacturing, as an automated production and a high product quality are simultaneously required, the need of advanced technologies for intelligent maintenance has emerged [1, 2] . In this respect, the robot-based spot-welding process, one of the most highly automated processes in automotive industry, has also required such improved quality management system to guarantee high quality of automobiles, since more than 3,000 weld-spots per one vehicle production are usually needed [3, 4] . Hence, there is much interest in a robust quality monitoring system for spot-welding process.

Methods for monitoring of spot-weld quality are classified into two types: off-line and on-line. The off-line method is to directly inspect the weld area, so that relatively clear information on the weld quality can be obtained. However, it is not suitable to be used for the in-situ monitoring system because of high cost and interruption of the process. On the other hand, the on-line method indirectly estimates the weld quality using various sensor signals, which is suitable for the in-situ quality monitoring. However, additional signal processing techniques are required to establish the correlation models between the signals and the weld quality [5, 6] .

The recent studies on the on-line monitoring for spotwelding process show that there have been two approaches. The first approach focused on improving the field applicability of non-destructive testing (NDT) methods by reducing the time delay and the cost. Ultrasonic signal and several signal processing techniques such as wavelet packet transform (WPT) are widely used in this approach [7] [8] [9] . These methods are promising and powerful for inspecting spot-weld quality, but the signals for testing should be

In modern manufacturing, as an automated production and a high product quality are simultaneously required, the need of advanced technologies for intelligent maintenance has emerged [1, 2] . In this respect, the robot-based spot-welding process, one of the most highly automated processes in automotive industry, has also required such improved quality management system to guarantee high quality of automobiles, since more than 3,000 weld-spots per one vehicle production are usually needed [3, 4] . Hence, there is much interest in a robust quality monitoring system for spot-welding process.

Methods for monitoring of spot-weld quality are classified into two types: off-line and on-line. The off-line method is to directly inspect the weld area, so that relatively clear information on the weld quality can be obtained. However, it is not suitable to be used for the in-situ monitoring system because of high cost and interruption of the process. On the other hand, the on-line method indirectly estimates the weld quality using various sensor signals, which is suitable for the in-situ quality monitoring. However, additional signal processing techniques are required to establish the correlation models between the signals and the weld quality [5, 6] .

The recent studies on the on-line monitoring for spotwelding process show that there have been two approaches. The first approach focused on improving the field applicability of non-destructive testing (NDT) methods by reducing the time delay and the cost. Ultrasonic signal and several signal processing techniques such as wavelet packet transform (WPT) are widely used in this approach [7] [8] [9] . These methods are promising and powerful for inspecting spot-weld quality, but the signals for testing should be

In modern manufacturing, as an automated production and a high product quality are simultaneously required, the need of advanced technologies for intelligent maintenance has emerged [1, 2] . In this respect, the robot-based spot-welding process, one of the most highly automated processes in automotive industry, has also required such improved quality management system to guarantee high quality of automobiles, since more than 3,000 weld-spots per one vehicle production are usually needed [3, 4] . Hence, there is much interest in a robust quality monitoring system for spot-welding process.

Methods for monitoring of spot-weld quality are classified into two types: off-line and on-line. The off-line method is to directly inspect the weld area, so that relatively clear information on the weld quality can be obtained. However, it is not suitable to be used for the in-situ monitoring system because of high cost and interruption of the process. On the other hand, the on-line method indirectly estimates the weld quality using various sensor signals, which is suitable for the in-situ quality monitoring. However, additional signal processing techniques are required to establish the correlation models between the signals and the weld quality [5, 6] .

The recent studies on the on-line monitoring for spotwelding process show that there have been two approaches. The first approach focused on improving the field applicability of non-destructive testing (NDT) methods by reducing the time delay and the cost. Ultrasonic signal and several signal processing techniques such as wavelet packet transform (WPT) are widely used in this approach [7] [8] [9] . These methods are promising and powerful for inspecting spot-weld quality, but the signals for testing should be

In modern manufacturing, as an automated production and a high product quality are simultaneously required, the need of advanced technologies for intelligent maintenance has emerged [1, 2] . In this respect, the robot-based spot-welding process, one of the most highly automated processes in automotive industry, has also required such improved quality management system to guarantee high quality of automobiles, since more than 3,000 weld-spots per one vehicle production are usually needed [3, 4] . Hence, there is much interest in a robust quality monitoring system for spot-welding process.

Methods for monitoring of spot-weld quality are classified into two types: off-line and on-line. The off-line method is to directly inspect the weld area, so that relatively clear information on the weld quality can be obtained. However, it is not suitable to be used for the in-situ monitoring system because of high cost and interruption of the process. On the other hand, the on-line method indirectly estimates the weld quality using various sensor signals, which is suitable for the in-situ quality monitoring. However, additional signal processing techniques are required to establish the correlation models between the signals and the weld quality [5, 6] .

The recent studies on the on-line monitoring for spotwelding process show that there have been two approaches. The first approach focused on improving the field applicability of non-destructive testing (NDT) methods by reducing the time delay and the cost. Ultrasonic signal and several signal processing techniques such as wavelet packet transform (WPT) are widely used in this approach [7] [8] [9] . These methods are promising and powerful for inspecting spot-weld quality, but the signals for testing should be 48th SME North American Manufacturing Research Conference, NAMRC 48 (Cancelled due to sampled after process done. Therefore, they still seem difficult to be applied to real-time monitoring systems for a robot spot-welding process.

In the second approach, the sensor data are collected during a real-time welding process for the on-line quality monitoring. Current, voltage, force, and displacement sensor data were mainly used in this approach, and machine learning techniques were frequently applied to modeling of the correlation between data and weld quality [10] [11] [12] [13] [14] [15] [16] . As a prime example, Cullen et al. presented the on-line monitoring model predicting the nugget size based on the neural network algorithm and multi-sensor fusion technique [17] . Also, much research on diagnosing the defects within a weld nugget or predicting the nugget dimensions has been conducted by sensor data based approach [18] [19] [20] [21] [22] [23] [24] .

Meanwhile, most of above studies tend to rely on typical features from sensor data such as the dynamic resistance (DR) for investigating the correlation between the sensor signal and the physical behavior of nugget formation. This is efficient for interpreting the physical meaning of sensor signal when the amount of data is small, but it is difficult to derive the robust diagnostic model which can be applied in real-time spotwelding process where the numerous variables affect the sensor data. Therefore, the additional processes of extracting various features from sensor data and selecting the critical ones through statistical analysis are required based on collecting the large amount of data.

In this paper, the angular misalignment of electrode, one of major failure modes of robot spot-welding process, is regarded as a target fault to cause the quality defects. In addition, the main goal is to develop an improved diagnostic system that performs robustly in real-time condition. Hence, at the first, the large amount of data was collected from the voltage and current sensors, and two types of features were extracted: physics-based features by dynamic resistance (DR) and data-driven features by wavelet transform (WT). Each type of features was analyzed by suitable statistical method and final features were selected. Then, two machine learning algorithmssupport vector machine (SVM) and probabilistic neural network (PNN)were applied to build the spot-weld quality diagnosis model. The trained models were evaluated against not only their diagnostic accuracy but also the robustness using additional test data. Finally, the real-time software is also developed to demonstrate applicability of the proposed spot-weld quality diagnostic framework in an industrial site.

The robot spot-welding experiment was firstly designed to investigate the effect of the angular misalignment on the weld quality. Then, the quality measurement was performed on the welded nuggets. In addition, the quality deterioration factor was identified by the dynamic resistance analysis.

The robot spot-welding experiments were conducted using the AC servo motor-based welding gun (OBARA) and the industrial 6-axis robot (Hyundai Robotics). The experimental welding jig was designed to simulate the shape of vehicle body and mount the welding specimens at 7 positions. The material of welding specimen was the low carbon steel which consists of 0.018% of carbon, and the width, length and thickness of the specimen were 300mm, 100mm and 0.8mm respectively. A spot-welding experiment was carried out by joining two specimens and each spot was welded at 20mm intervals to avoid the shunting effect [25] . Fig. 1 . shows the experimental testbed and specimen. The appropriate process parameters, such as welding current, electrode force, welding time and holding time, were generally determined according to the fastest process time while they satisfy the condition that the diameter of welded nugget without any porosity or expulsion is longer than 5√t (where t is thickness of specimen) [26] . Based on this, the preliminary experiments for setting the proper parameters were carried out, then the parameters are finally defined as Table 1 .

Meanwhile, the current and voltage signals are closely related to welding behavior because the spot-welding operates according to the Joule heating principle [27] , so they can provide the significant indices for process result. Hence, the Rogowski coil and Differential Voltage probe were installed on the servo gun to gather the current and voltage data from spot-welding experiments, where the sampling frequency was 12.8kHz. Fig. 1 . also shows the sensor setup position with experimental testbed. After the experimental preparation, the spot-welding experiments were performed for different electrode angles to investigate the results of angular misalignment. As shown in Fig. 2 . the vertical angle between the welding specimen and the electrode was set to 0°, and the robot was controlled to increase the angle by 2.5°, 5.0° and 7.5°, respectively. The experiments were conducted 300 spots for each electrode angle and the sensor data were also sampled for every spot. 

The quality measurement methods for spot-welding process typically include the peel test, compression shear test and cross section test [28] [29] [30] [31] . The peel test and the compression shear test can quickly provide the results of tensile strength and shear strength respectively, but it is difficult to figure out the defects inside the welded nugget. Hence, in this paper, the cross-section test was chosen to accurately identify the effect of angular misalignment. The cross sections of welded spots were prepared through the sequential processes of wire-EDM cutting, surface polishing and etching. Then the welded nuggets on the cross section were captured by optical microscope. Fig. 3 shows the optical images of cross section and measured nugget diameters for four different electrode angles.

The result of quality measurement showed that the "porosity" occurred inside the welded nugget if the electrode is misaligned more than 5°. This indicates that the abnormality happened in the process of shrinkage of nugget after its melting [32] . Also, it was confirmed that the diameters of nuggets corresponding to the misalignment of 5° and 7.5° were slightly shorter than the reference value of 5√t (4.47mm, where t=0.8mm). Therefore, as a result, the angular misalignment over 5° was determined to be the target failure mode to monitor and diagnose. 

The dynamic resistance is the ratio of change in voltage to the change in current, and it is often considered as the significant index to describe the process of welded nugget formation. In this paper, the dynamic resistance was extracted from the collected voltage and current sensor data. To extract the dynamic resistance, the half cycles of the current and voltage data should be identified by referring to the points of crest, trough and transition for each signal, as shown in Fig. 4 . Then, the dynamic resistance (DR) was finally derived from Eq. (1) using the root mean square (RMS) values of voltage and current signals within each half cycle.

(1) Fig. 5 shows a typical DR curve with two inflection points called α peak and β peak. The first section from initial point to α peak is considered as the stage of weld preparation. At this stage, the DR curve sharply decreases due to the increase of current application and the enlargement of contact surface. The short region between α peak and β peak indicates the stage of nugget creation. The melting nugget is produced between the specimens at this stage, and the DR curve rises because the bulk resistance increases with the temperature. The last section after the β peak is the stage of nugget formation, the DR curve gradually falls as the nugget formation progresses [33] . On the other hand, the DR curves of sensor data for normal and misalignment (5° ~ 7.5°) conditions are shown in Fig. 6 (a). As can be seen in this chart, the α peak and the end point of both normal and abnormal conditions are relatively similar, whereas the resistance values of initial point and the time of β peak occurrence are noticeably different according to the condition.

Those differences can be explained by physical behavior of angular misalignment. When the spot-welding is progressing, the contact surface between angular-misaligned electrode and specimen increases larger than in normal condition because the deformation occurs in the specimen which become ductile due to the Joule heating ( Fig. 6 (b) ). Enlarged contact surface reduces the current density and it causes the decline of bulk resistance at the initial point and delays the occurrence of β peak. The delayed β peak, meaning the shortening of nugget formation time, leads to the reduction in nugget diameter and the porosity generation. It supports the result of quality measurement in section 2.2. a b Fig. 6 . Changes DR curve and geometry due to angular misalignment, (a) DR curves for normal and misalignment condition; (b) Schematic of geometric change due to misalignment

Through the robot spot-welding experiments, the current and voltage sensor data were gathered for 300 spots of normal state and 600 spots of abnormal state (angular misalignment for 5° and 7.5°). Each data had 2,560 samples as it was collected for 0.2 seconds at sampling frequency of 12.8kHz. The data with excessively high dimensions compared to the number of data can cause the overfitting and delay the training time of machine learning models, so the signal processing step for extracting significant features from raw data was required. In this paper, two different types of features were extracted. The first type of features was extracted from the dynamic resistance (DR) analyzed in section 2.3. These features have the advantage of being physically descriptive in the spot-welding process. The second type was the data-driven features which were extracted by wavelet transform (WT) method. The data-driven features are hard to be assigned physical meaning to each, but they can provide robust indicators that are difficult for human analyzer to find. After extracting each type of features, some critical features had been selected based on the appropriate statistical methodologies.

The DR features were extracted based on the 4 major points, such as initial point, α peak, β peak and end point, on the DR curve as shown in Fig. 7 . The descriptions of each DR feature are given in the Table 2 . Slope between initial point and α peak K2 Slope between α peak and β peak K3 Slope between β peak and end point A1 Average of resistance from β peak to end point S1 Standard deviation from β peak to end point Every DR feature has physical meaning for nugget formation and spot-welding quality: (1) R1 and K1 indicate the effect of initial application of welding current, so the geometric error or erroneous setting for process parameters can be detected by these features. (2) T2, R2 and K2 provide the information of local melting and nugget creation. (3) T3, R3, K3, R4, A1 and S1 are significant indices for monitoring the nugget formation which is directly related to weld quality [34, 35] .

Feature selection is essential process to improve the classification performance and shorten the training time of machine learning algorithm. For selecting the valid features that better distinguish between normal and abnormal state, a statistical method is needed to compare the distribution between two data groups. There are generally two types of test methods for statistically comparing multiple groups: Ztest and T-test. However, the Z-test is only available if the population distribution is known, and it is difficult to identify the population using the sampled sensor data. Therefore, the T-test method was applied to select DR features that better distinguish between normal and abnormal (angular misalignment) state.

Before implementing the T-test, the assumption that the population variances of two groups are equal should be satisfied because the groups with unequal variances may have differences in the criteria to reject the null hypothesis [36] . Hence, the satisfaction of equal variance assumption for the spot-welding data was evaluated by Levene's test, which is the typical method, and it was satisfied. Through the T-test, the P-values, the values of significance level for t-statistics, for each DR feature were calculated as in Table 3 . The null hypothesis, meaning that the mean values of two groups are equal, is commonly rejected when the significance level is less than 5% or 1%, and the 1% was determined as the criteria in this feature selection process. As a result, 6 DR features with the P-value less than 0.01 were selected as the final features to be used in machine learning. Fig. 8 is the probabilistic density functions (PDF) with histograms for topranked (R1 and T2) and bottom-ranked (R4 and A1) DR features. The charts show the possibility of each feature that distinguishes between normal and abnormal data.

The wavelet transform (WT) is the widely used signal processing technique for decomposing and reconstructing the dynamic signals using the mother wavelet. It is specialized for non-stationary signals whose frequency information changes with time because the density values of wavelength in narrow sections are continuously obtained according to parallel translation of mother wavelet [37] .

Eq. (2) is the function of wavelet coefficient for discrete time signal, where φ is the mother wavelet, b is the time delay of wavelet, x(k) is the raw signal and n is the number of samples.

The mother wavelet consists of high pass filter and low pass filter , which decompose the original signal into high frequency and low frequency signal as shown in Fig. 9 . Every time the signal is decomposed, the frequency of raw signal is reduced in half. After decomposition, extracted high frequency signal is left as a coefficient of the corresponding level, and the low frequency signal is decomposed again to next level. From this way, high frequency coefficient can obtain the higher resolution signal than low frequency coefficient. The most widely used mother wavelet is the "Daubechies", and its modified form to increase the symmetry is the "Symlet" [38] which is applied in this research. The level of decomposition was set to 12 because it seems difficult to extract the valid information from lower frequency than 3.125Hz which is the level 12 frequency for spot-welding data. Fig. 10 shows the sample of wavelet decomposition result of current sensor data for both normal and abnormal state. D1 is the coefficient of level 1, which is the highest frequency signal, and A1 is the lowest frequency signal. It was very hard to find a notable difference between normal and misalignment data at any coefficient. Therefore, the sophisticated datadriven techniques for extracting and selecting features were required. Table 4 . Consequently, a total of 168 features were extracted since 7 features were calculated for each of the 12wavelet transform (WT) levels using the 2-sensor data. There are two popularly used methodologies for datadriven feature selection: filter and wrapper. The filter methodology is used for filtering out the non-critical features via evaluating each feature by certain characteristic value. The P-value discussed in section 3.1.2 was also considered as the value for filtering DR features. As DR features with P-value greater than 0.01 had been filtered, the specific criteria for selecting features should be determined in this method. The second method, wrapper observes the most effective subset of original features by assessing the performance of machine learning model trained by each subset. This method considers the model architecture so the powerful features can be selected, but it is ineffective for applying to large number of features because it is time consuming and computationally expensive [39] . Therefore, in this paper, both filter and wrapper are used sequentially for selecting robust features from extracted 168 WT features.

At the first step, the max-relevance and min-redundancy (mRMR) technique, which is one of filter methods, was applied to filtering out the relatively less important WT features. The mRMR reorders the original features in order of high relevance and low redundancy based on mutual information. The mutual information of two discrete variables, , is derived as Eq. (3) using probabilistic density functions , and [40] . To apply the collected spot-welding data into Eq. (3), each datum was discretized as following rules: (1) the datum was altered to 1 when it is greater than sum of the mean value (μ) and standard deviation (σ); μ + σ, (2) altered to -1 when it is smaller than μ − σ, (3) 0 for other cases. Using the mutual information calculated from Eq. (3), the relevance and the redundancy can be maximized and minimized respectively, through Eq. (4) and Eq. (5), where S is the objective subset of original features, ⌈S⌉ is the number of S, is the feature in S, and C is the targeted class.

(4)

As the result of this step, the 30 top-ranked features were selected according to the mutual information quotient (MIQ), which is calculated by maximizing ( ). Since the number of features was reduced through the mRMR, the wrapper could be easily performed as the second step of feature selection. The wrapper is generally divided into two methods: sequential forward selection (SFS) and sequential backward selection (SBS). Both methods iterate the classification test with a subset of features and a certain classifier until classification error no longer decreases. The SFS wrapper starts with an empty subset and adds a feature for each iteration and, conversely, the SBS wrapper starts with the entire features and removes the least important feature. In this paper, the SBS method was chosen and the support vector machine (SVM) algorithm was applied for the test classifier.

Through the 2 steps of feature selection process, mRMR and SBS wrapper, the 16 WT features were finally selected as shown in Table 5 . The result of selection showed that more significant factors had been extracted from the voltage signal than current signal, and the mean values of WT coefficient were mainly selected as the important indices for classifying the normal and angular misalignment condition. It means that there is a close relationship between the angular misalignment and the energy variation of the voltage signal. 

To develop the system for monitoring the angular misalignment of robot spot-welding process, the machine learning based diagnosis models were established using the selected features and validated. Also, the graphical user interface (GUI) based software for real-time implementation was developed.

The two representative machine learning algorithms for classification, support vector machine (SVM) and probabilistic neural network (PNN), were applied to construct angular misalignment diagnosis model. For training each machine learning model, three types of features were prepared as the input variables: (1) 6 physics-based features from dynamic resistance (DR), (2) 16 data-driven features based on wavelet transform (WT), (3) 22 hybrid (HB) features combining the physics-based and data-driven features. Then, as the output responses, the numerical values of 0 and 1 were allocated to normal state and abnormal (angular misalignment) state, respectively.

The first model was constructed based on the support vector machine (SVM). SVM is robust classifier designed to maximize the margin between two classes, unlike other classifiers, including neural network, designed for the purpose of minimizing the error rate. SVM considers each element of data as a vector, and the vectors at the boundary between two classes are determined as the support vectors, so the margin can be calculated by the distance between the support vectors.

Meanwhile, to measure the distance for non-linear data, such as the features of spot-welding data, the radial basis function (RBF) kernel is widely used [41, 42] . The RBF kernel is also referred to the Gaussian kernel, and is expressed as in Eq. (6) , where , are the data points and γ is the hyper-parameter controlling the kernel scale. (6) In this study, the kernel scale γ was determined to be 0.5, which yielded the best performance of model. Fig. 11 shows the modeling process for classifying spot-welding conditions based on SVM. The second model was built based on probabilistic neural network (PNN). The PNN consists of 4 layers. The first layer is the input layer that the input variables are supplied to PNN. The second layer, named pattern layer, receives the sum of the input variables multiplied by the weights , and outputs the result of activation function for the input variables. The sigmoid function was applied as the activation function in this process. The third layer is the summation layer, and the number of nodes in this layer is determined by the number of output classes, which is two in this study. Each node of summation layer is only connected to the nodes, belonging to the same class, among the pattern layers. After that, it receives the sum of the outputs from connected nodes of pattern layer as the input value and delivers the input value multiplied by parameter , defined by user, to the next layer. The last layer, as the output layer, finally yields the output result for input data [43, 44] .

Unlike other neural network-based algorithms, the architecture of PNN is determined by the modeling environment. Firstly, the number of nodes in input layer is same as the dimension of input data, so there were 6 nodes for training the DR features, 16 nodes for the WT features and 22 nodes for the hybrid features. The number of nodes in pattern layer was 900 because it is identical to the number of input data (300 data for normal condition and 600 for abnormal). Lastly, the number of nodes in summation layer and output layer was 2, as mentioned above paragraph.

After training the PNN, the model starts to diagnose the angular misalignment based on Bayesian classifier, which determines the output class by selecting the class having a maximum result value among the nodes in the output layer. Fig. 12 shows the schematic architecture for PNN based diagnosis model. 

The performance evaluation of the models was conducted using the 900 collected data set consisting of 300 data for normal state and 600 data for angular misalignment (5° or 7.5°). At this time, the models were sorted into 6 types according to the combination of 3 kinds of features (DR, WT and hybrid (HB)) and 2 classifiers (SVM and PNN). The kfold cross validation method was adopted to estimate the accuracy of each model. In this method, the original data is randomly partitioned into k sub-data sets of equal size. Then, a single subset of each fold is considered as the validation data, and the remaining k-1 subsets are used for training the model, so that the limited number of data can be used more efficiently than a conventional method. Therefore, the 5-fold cross validation was carried out in this section, and the accuracy of each fold was calculated by counting the number of times for correctly diagnosing the spot-welding state among 180 validation data. The validation results are shown in Table 6 , and the 6 models were compared through the average accuracy. After validation, additional experiments were conducted for further testing the models with practical data. In this test experiments, the 30 data were collected as the normal data where the electrode angle is 0°. In addition, the 60 abnormal data were sampled at the condition in which the electrode angles were randomly distributed between 5° and 7.5°. The purpose of the test with the data where the sampling conditions are not exactly same as the trained data is to assess the robustness of the model to variations in data and the sampling environment. Thus, each of 6 models had been evaluated for diagnostic accuracy, and the results of test had been compared to the validation result as shown in Fig. 13 . From the results, it was found that the type of feature data used for training has a greater influence on the model performance than the type of the machine learning algorithm. The models trained by DR features, physically descriptive and typically regarded as the main indicators in other spotwelding monitoring studies, showed high diagnostic accuracy when verifying with the data collected under the same conditions, but significantly lower performance in robustness tests. On the other hand, the models trained with WT features extracted from data-driven method showed relatively robust results for data variation, although the diagnostic accuracy was not very high. Consequently, it had been confirmed that the hybrid (HB) feature data, which has both high accuracy and robustness, maximizes the performance of the angular misalignment diagnosis model. Therefore, the models trained with HB features were chosen as the final models for monitoring and diagnosing the angular misalignment of robot spot-welding process.

The software system was developed based on MATLAB program to monitor and diagnose the robot spot-welding condition during the real-time process. The graphical user interface (GUI) was also designed as Fig. 14 . for visualizing the monitoring information and the diagnosis results. It shows the functions of the real-time data sampling and the signal monitoring as well as the failure diagnosis.

The system diagnoses the spot-welding condition based on both SVM and PNN model trained by HB features. Fig. 15 shows the flowchart of system from data sampling to failure diagnosis and how to determine the diagnosis result. As can be seen in Fig. 15 ., through inputting the HB features of data to the SVM and PNN model sequentially, the final diagnosis result is decided as following rules:

(1) if both models output the '0' representing the normal state: "Normal" state,

(2) if both models output '1' meaning the angular misalignment: "Abnormal" state, (3) if the results of 2 models are different: "Caution" state. From these rules, the robust diagnosis could be performed at the real-time implementation. In addition, for an intuitive information transfer, green, red and yellow were used as the color informing diagnosis results corresponding to "Normal", "Abnormal" and "Caution" respectively, as shown in Fig. 14 . 

In this paper, the on-line monitoring and diagnosis system for the angular misalignment of the welding gun during the robot-based spot-welding process, which can result in significant quality degradation in weld nuggets, were developed based on the data-driven approach. The collected mass data of voltage and current signals were processed to obtain the critical features of DR and WT coefficients. The selected features were used as inputs to the diagnosis model that was created by the machine learning algorithms such SVM and PNN. In addition, the software was developed to demonstrate applicability of the proposed framework in an industrial site in a real-time fashion. More detailed conclusions of each section are as follows.

In section 2, the effect of angular misalignment on weld quality is investigated. It was confirmed that the "porosity" occurred within welded nugget where the electrode was misaligned over 5°, and the physical behavior of this quality deterioration was identified by delayed β peak of dynamic resistance (DR) curve. Then, the current and voltage sensor data for each normal and abnormal state were collected through the robot spot-welding experiments.

In section 3, the physics-based and the data-driven features were extracted based on dynamic resistance (DR) and wavelet transform (WT) respectively. Also, the critical features of each type were selected using the t-test, max-relevance and min-redundancy (mRMR) and sequential backward selection (SBS) wrapper method.

In section 4, support vector machine (SVM) and probabilistic neural network (PNN) based diagnosis models were established by training the three types of feature data, such as 6 DR features, 16 WT features and 22 hybrid (HB) features. Then, the diagnostic accuracy of the 6 models was validated by 5-fold cross validation method and the robustness was also tested with the additional data collected under different sampling condition. As the result, the feature type was more influential in both diagnostic accuracy and robustness of model than the type of machine learning algorithm, and the HB features were selected as the final features to be applied to real-time monitoring system. Lastly, the on-line monitoring system based on graphical user interface (GUI) was developed and implemented in a realtime manner.

Smart manufacturing: Past research, present findings, and future directions

Manufacturing: Design, Production, Automation, and Integration

Resistance Spot Weldability of High Strength Steel Sheets for Automobiles and the Quality Assurance of Joints

Dependence of overload performance on weld attributes for resistance spot welded galvanized low carbon steel

Condition Monitoring and Control for Intelligent Manufacturing

Development and analysis of an online tool condition monitoring and diagnosis system for a milling process and its real-time implementation

Ultrasonic nondestructive evaluation of spot welds for zinc-coated high strength steel sheet based on wavelet packet analysis

Assessment of resistance spot welding quality based on ultrasonic testing and tree-based techniques

Magnetic Nondestructive Test for Resistance Spot Welds Using Magnetic Flux Penetration and Eddy Current Methods

Online monitoring and evaluation of the weld quality of resistance spot welded titanium alloy

Multisensor-based real-time quality monitoring by means of feature extraction, selection and modeling for Al alloy in arc welding

Real-time image processing for monitoring of free weld pool surface

Modeling of human welder response to 3D weld pool surface: Part I-Principles

Characterization of three-dimensional weld pool surface in GTAW

Sensor systems for real-time monitoring of laser weld quality

Expulsion detection system for resistance spot welding based on a neural network

Multisensor fusion for online monitoring of the quality of spot welding in automotive industry

On-line monitoring method for electrode invalidation during spot welding of zinc-coated steel

Study on effect of electrode force on resistance spot welding process

Monitoring of expulsion in small scale resistance spot welding

Expulsion characterization of stainlesssteel resistance spot welding based on dynamic resistance signal

Welding quality evaluation of resistance spot welding using the time-varying inductive reactance signal

Online nugget diameter control system for resistance spot welding

Quality monitoring of spot welding with advanced signal processing and datadriven techniques

Characteristics of shunting effect in resistance spot welding in mild steel based on electrode displacement

Recommended practices for test methods for evaluating the resistance spot welding behavior of automotive sheet steel materials

Resistance welding: fundamentals and applications

The effect of nucleus size on mechanical properties in electrical resistance spot welding of sheets used in automotive industry

Strength assessment of spot-welded sheets of interstitial free steels

Failure analysis of dissimilar thickness resistance spot welded joints in dual-phase steels during tensile shear test

Resistance spot welding of galvannealed high strength interstitial free steel

Modeling dynamic electrical resistance during resistance spot welding

Quality assessment of resistance spot welding process based on dynamic resistance signal and random forest based

Weld quality monitoring research in small scale resistance spot welding by dynamic resistance and neural network

The unequal variance t-test is an underused alternative to Student's t-test and the Mann-Whitney U test

Application of mother wavelet functions for automatic gear and bearing fault diagnosis

Outer race defect width measurement in taper roller bearing using discrete wavelet transform of vibration signal

Fault Diagnosis of Planetary Gear Carrier Packs: A Class Imbalance and Multiclass Classification Problem

Feature Selection Based on Mutual Information: Criteria of Max-Dependency, Max-Relevance, and Min-Redundancy

What is a support vector machine?

A quantitative estimation technique for welding quality using local mean decomposition and support vector machine

Probabilistic neural networks

Application of the discrete wavelet transform and probabilistic neural networks in IC engine fault diagnostics