A new method for nondestructive quality evaluation of ...

Int J Adv Manuf Technol DOI 10.1007/s00170-014-6654-1

ORIGINAL ARTICLE

A new method for nondestructive quality evaluation of the resistance spot welding based on the radar chart method and the decision tree classifier Hongjie Zhang & Yanyan Hou & Jianye Zhang & Xiangyang Qi & Fujun Wang

Received: 22 May 2014 / Accepted: 25 November 2014 # Springer-Verlag London 2014

Abstract To develop an effective nondestructive evaluation method for the welding quality of the resistance spot welding, the electrode displacement signal during the resistance spot welding process is monitored, and the acquisition data of the signal are innovatively presented as the radar chart format. Some geometric features of the radar charts are extracted to reflect the welding quality. The decision tree classification technique is adopted to build a classifier and to provide a visible and intuitive diagnostic procedure for welding quality assessment. Test results of the decision tree classifier show that it is feasible and reliable to evaluate weld quality based on the graphics features of the radar chart. The features and the weld quality are closely related, and their extraction avoids complex algorithm. Meanwhile, when there are small samples, the decision tree classifier can identify good or bad weld accurately and rapidly, even though the weld is from an abnormal welding process, such as expulsion, current shunting, greasy surface, and small edge distance condition.

Keywords Resistance spot welding . Nondestructive quality evaluation . The electrode displacement signal . Radar chart . Decision tree classifier H. Zhang (*) : J. Zhang : X. Qi Tianjin Key Laboratory of Modern Mechatronics Equipment Technology, School of Mechanical Engineering, Tianjin Polytechnic University, Tianjin 300387, China e-mail: [email protected] Y. Hou Department of Architecture Engineering, Hebei Academy of Governance, Shijiazhuang 050031, China F. Wang Tianjin Key Laboratory of Equipment Design and Manufacturing Technology, School of Mechanical Engineering, Tianjin University, Tianjin 300072, China

1 Introduction The resistance spot welding (RSW) is widely used owing to its high speed and excellent adaptability for automation in high volume production. However, RSW suffers from a significant problem of inconsistent quality. Traditionally, destructive test methods can only ensure the qualities of the specific weld specimens, and they are time-consuming and costly. Therefore, nondestructive quality evaluation methods based on the feature analysis of the dynamic signals during RSW process attract special attentions [1–4]. The dynamic resistance and electrode displacement are the most commonly used parameters or variables for weld quality estimation [5, 6]. The electrode displacement is mainly caused by the thermal expansion of the weld nugget, and it can make rapid responses to small changes of any variables affecting the welding quality [7–9]. It is believed that the amount of the thermal expansion, melting, expulsion, and electrode wear can be reflected by the slope, magnitude, and fluctuation of the displacement curve [10–12]. Based on the features extracted from the dynamic signals, the multiple regression analysis [13–16] and the neural network-based methods are used to build the correlation between the features and welding quality indicators, such as weld nugget size and strength. Cho et al. [17] used the back-propagation neural network to predict tensile shear strength of the weld. Zhao et al. [18] selected linear vector quantization neutral network to evaluate the welding quality. The same network is proposed by EL-Banna et al. [19]. In his research, the welding quality was grouped into three classes, including cold, expulsion, and normal. Zhang et al. [20] developed a weld quality classifier by using the associative memory network. Gong et al. [21] used the Bayesian belief network to determine the optimal welding parameters and to estimate the weld quality online. Zhang et al. [14] adopted the radial basis function neural network to predict the tensile shear strength of the weld.

Int J Adv Manuf Technol Fig. 1 The schematic representation of the experimental setup

Some other methods of the machine learning technologies have been proposed as well. Zhang et al. [22] combined the principal component analysis with support vector machine technique to build the mathematical model between monitored features of the displacement signal and weld strength. Li et al. [23] achieved some quality assessment rules by using attribute reduction algorithm of the rough set theory. Zhang et al. [24] discriminated good or bad welding based on the kernel fisher discriminant analysis technique, and Podržaj et al. [25] estimated the weld quality on the basis of the fuzzy logic algorithm. In general, welding quality assessment methods used currently mainly focus on developing the relation between the selected quality indicators and various features or parameters from the monitored signals. Feature extraction and selection excessively depend on experiences and theory analysis which often involve complicated algorithms and heavy calculations. In addition, weld quality diagnostic procedures based on machine learning techniques are generally invisible, just as black boxes, which results in the difficulties in understanding and interpreting the evaluation process. In order to improve the efficiency of acquiring monitored features and present a new quality evaluation method, the electrode displacement signal of RSW process is monitored, and the acquisition data

Fig. 2 a The original electrode displacement waveform. b The electrode displacement curve through noise reduction

of the signal are presented as the radar chart format, and then a nondestructive quality evaluation method for RSW based on the geometric features of the radar chart and the decision tree technology is discussed.

2 Measurement of the electrode displacement signal The measurement system for the electrode displacement is schematically shown in Fig. 1, which consists of the AC resistance spot welding machine, sensors, signal acquisition device, and computer. When the welding current flows through the workpieces, a volume of material at the faying surfaces between workpieces melts owing to the joule heat, and the weld nugget is formed whose size defines the degree of thermal expansion and the weld strength. A commercial linear variable differential transformer displacement sensor is used to measure the movement of the upper electrode, and its resolution is 0.01 μm. The welding current signal is monitored

Fig. 3 The diagrammatic sketch for the radar chart with n indices

Int J Adv Manuf Technol

Fig. 4 a Data reduction result for the electrode displacement curve. b The corresponding radar chart

by a Rogowsky coil which is used to calculate the dynamic effective value of the welding current during RSW process. These two signals are synchronously gathered by a signal acquisition device NI USB6366 with the sampling rate of 400 kHz. The lapping welding experiments are performed on a 0.7mm-thick sheet of uncoated low-carbon steel which is cut into 30×120-mm coupons. The face diameter of the cone electrode is 6 mm. The welding time and electrode force are set as 20 cycles and 2.0 kN, respectively. When the welding current is increased at 0.2-kA intervals starting from 3.2 up to 4.4 kA, some weld samples are obtained. Ten weld samples are selected from each welding current parameter; moreover, ten expulsion welds are considered as well, thus, 80 weld samples in all are viewed as a training set to explore the welding quality evaluation method. An original waveform of the electrode displacement signal from 4.4 kA welding current is shown in Fig. 2a, and the signal length is limited as 25 welding cycles. To remove the high-frequency noise in the original signal, wavelet de-noising method is adopted, and the result is displayed in Fig. 2b where the electrode displacement curve experiences four stages. In stage I, the displacement curve presents a transient drop owing to the electrode force brings an improved contact. With the accumulation of joule heat, the weld nugget forms and grows, and the volume change of solid-to-liquid cause the amplitude of the displacement to rise to its maximum value rapidly in stage II. When the nugget size exceeds the electrode tip diameter and reaches saturation, the nugget center closes to

Fig. 5 a The electrode displacement curve from expulsion. b The reduced displacement curve from expulsion, and c the corresponding radar chart

Fig. 6 The schematic diagram for feature extraction


Fig. 7 The scatter diagrams for the features: a Gravity and D_max, b Area and c Girth

3.1 Radar chart of the displacement curve

above, thus, 25 periodic average values of the electrode displacement can be calculated. By replacing the electrode displacement data of each welding cycle with the periodic average value of corresponding cycle, the data quantity of the signal can be reduced as 25. Figure 4a displays the reduction result of the displacement signal shown in Fig. 2. Then, these 25 periodic averages are regarded as index variables to draw a radar chart, and the result is shown in Fig. 4b where a circle is averagely divided into 24 circular sectors with the central angle of 15°. The numbered index axes are arranged as counterclockwise rotation, and the first and the 25th indices share the same index axis. The radius of the circle is set as the

The radar chart, also known as spider chart, is a data visualization approach to describe the multidimensional data in the two-dimensional plane [26, 27]. A typical radar chart is shown in Fig. 3 where a circle is averagely divided into n circular sectors according to the number of the index variables (Z1 to Zn). The radiuses of circular sectors are selected as index axes of the radar chart, and each index axis represents one index variable. Through connecting the data points in the different index axes with lines, a closed polygon can be obtained. If the electrode displacement curve can be presented as the radar chart format, some geometric features of the closed polygon will possibly provide help for welding quality evaluation. For this purpose, data reduction is conducted on the acquisition data of the electrode displacement signal. In this research, segmented average method is adopted. The time sequence of the displacement values can be split into 25 welding cycles according to the signal length as mentioned

Fig. 8 Radar charts of the electrode displacement curves from different welding current parameters

the electrode tip surface, which results in the heat dissipation more than generation. The displacement curve begins to drop again in stage III till the thermal equilibrium is reached, and the displacement amplitude approximates to constant. In stage IV, the welding current is cut off, and the curve drops rapidly.

3 Radar chart of the displacement curve and feature extraction


maximum value of the displacement amplitude of the 80 weld samples mentioned above. During RSW process, the weld nugget does not always reach its optimal size under proper welding parameters. In fact, the nugget will collapse if the solid material around the nugget is no longer able to withstand the pressure of the electrodes. This phenomenon is called expulsion, and it should be monitored owing to the fact that it decreases the energy absorption capability of the weld as well as the corrosion resistance of the coated materials [28, 29]. An electrode displacement curve from expulsion weld is shown in Fig. 5a, and a sudden drop can be observed when the expulsion occurs. As shown in Fig. 5b, under expulsion condition, the periodic average values of the displacement amplitude are possibly negative in stage III of the welding process. Owing to the fact that the index variable displayed in the radar chart is generally nonnegative, a treatment method for the negative periodic averages of the displacement is proposed. Once a negative periodic average value occurs in stage III, the value and the periodic average values in the remaining welding cycles are all set as zero. Figure 5c displays the radar chart of the expulsion weld. 3.2 Features extraction from the radar chart In the radar chart, the reduced electrode displacement curve is changed to a closed polygon as the shaded part shown in Fig. 6 where the periodic average values of the electrode displacement are defined as Z1 to Z25. In order to utilize the radar chart to evaluate the welding quality, some important geometric features of the closed polygon are extracted. The primary geometric features are marked in Fig. 6, including the area, girth, and gravity of the closed polygon. In addition, the maximum value of the periodic averages is considered as well, and it is defined as D_max.

For the closed polygon shown in Fig. 3, its area can be calculated from the area summation of some triangles ! n−1 X Xn 1 Area ¼ S j ¼ sinðαÞ Z j Z jþ1 þ Z n Z 1 ð1Þ 2 j¼1 j¼1 where n is the number of the index variables and α corresponds to the central angle which is equal to 2π/n. The girth of the closed polygon can be obtained from the summation of a series of distances between the adjacent index variables sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi n−1 X 2π Girth ¼ Z 2j þ Z 2jþ1 −2Z j Z jþ1 cos n j¼1 sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ffi 2π 2 2 þ Z n þ Z 1 −2Z 1 Z n cos n

As mentioned above, in the radar chart of the displacement curve, the first and the 25th indices share the same index axis, therefore, the area and the girth of the closed polygon can be deduced as ! 24 24 X 1 π X Area ¼ S j ¼ sin Z j Z jþ1 ð3Þ 2 12 j¼1 j¼1 rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi π ffi 2 2 þ jZ 25 −Z 1 j ð4Þ Z j þ Z jþ1 −2Z j Z jþ1 cos Girth ¼ 12 j¼1 24 X

The areal coordinates of the closed polygon in radar chart can be obtained from the computational formula for the barycenter of the rigid body.

8 n−1 n−1 n−1 n−1 X X X X > > > x2i yiþ1 þ x2n y1 − x2iþ1 yi −x21 yn þ xi xiþ1 yiþ1 þ xn x1 y1 − xi yiþ1 yi þ xn y1 yn > > > i¼1 i¼1 i¼1 i¼1 > > Gravity x ¼ ! > n−1 n−1 > X X > > > 3 x y þ x y − xiþ1 yi −x1 yn > i n iþ1 1 < i¼1

i¼1

n−1 n−1 n−1 n−1 > X X X X > > > xi y2iþ1 þ xn y21 − xiþ1 y2i −x1 y2n þ xi yi yiþ1 þ xn yn y1 − xi xiþ1 yi þ xn x1 yn > > > i¼1 i¼1 i¼1 i¼1 > > y ¼ Gravity ! > n−1 n−1 > X X > > > 3 x y þ x y − x y −x y : i iþ1 n 1 iþ1 i 1 n i¼1

ð2Þ

ð5Þ

i¼1

where (Gravity_x,Gravity_y) corresponds to the areal coordinates of the polygon. n is the number of the side of the

polygon, and here, it is 25. (xi,yi) are the coordinates of the index variable Zi in the radar chart. So far, for each weld


[30, 31] but has not yet been applied in welding quality evaluation for RSW. A standard decision tree consists of one root node, a number of internal nodes, and leaves. As schematically shown in Fig. 10, the chains from the root node to leaves form the branches of the tree. Each node in the branch, except for leaf, involves a classification rule which defines a feature-based comparison between a numeric feature and a threshold value. The leaves are the terminal nodes, and each of them represents one category of the samples. The ID3 algorithm is one of the most representative methods to generate decision tree [32]. The most important task of this algorithm is to decide which feature the node should test. To realize feature selection, a parameter named information gain is defined, and it measures how well a given feature separates training set according to its target classification. The entropy of the training set is used to achieve the information gain. If a training set S has T samples and these samples can be classified as M classes, the entropy of the training set is given by Fig. 9 a The tensile shear strengths of the weld samples. b The nugget diameters of the weld samples

EntropyðS Þ ¼ −

XM i¼1

pi log2 ðpi Þ

ð6Þ

sample in the aforementioned training set, graphics features from corresponding radar charts can be obtained. In Fig. 7, the scatter diagrams for these features are provided (the features Gravity and D_max are shown in Fig. 7a, and the Area and Girth in Fig. 7b and c, respectively).

where pi refers to the proportion of the sample T belonging to the class i, and it can be calculated through Ni/T, where Ni is the amount of the samples in the class i. If the feature A of the training set has k values {a1,a2,⋯,aj, ⋯,ak}, then it can partition the training set into k subsets {S1, S2,⋯,Sj,⋯,Sk}, where Sj ={x|x∈S,x.A=aj}. Thus, the information gain of feature A is defined as

4 Decision tree classifier and welding quality evaluation

GainðX ; AÞ ¼ EntropyðX Þ−

4.1 Decision tree classifier The decision tree is a tree based on the knowledge methodology used to represent classification rules, and it has been widely utilized to develop the classification models

Table 1 The variation ranges of the tensile shear strength and diameter of the welds

Welding current (kA)

3.2 3.4 3.6 3.8 4.0 4.2 4.4

Xk T j ⋅Entropy S j N j¼1

ð7Þ

where Tj represents the number of the samples in Sj for which feature A has a value of aj. Based on the Eq. 7, the information gain for each feature of the training set can be achieved, where

Tensile shear strength (kN)

Weld diameter (mm)

Maximum

Minimum

Average

Maximum

Minimum

Average

1.6190 2.1482 2.7910 3.0282 3.3712 3.4500 3.5623

1.4998 1.9453 2.6705 2.9012 3.2691 3.4108 3.5084

1.5758 2.0375 2.7276 2.9728 3.3202 3.4343 3.5256

2.18 2.88 3.63 4.23 4.93 5.42 5.90

1.91 2.61 3.39 3.97 4.57 5.23 5.69

2.06 2.68 3.49 4.11 4.76 5.34 5.79


Fig. 10 The optimal decision tree for weld quality classification

the feature with maximum value of information gain is selected as test feature at the node of the decision tree.

4.2 Welding quality assessment The construction of decision tree requires supervised training, so, before building the decision tree, the categories of the training samples must be confirmed first. Figure 8 provides three radar charts randomly selected from the ten weld samples under each welding current parameter and obvious differences can be observed among the samples from different welding current parameters. The tensile shear strengths of the weld samples in training set, which are examined as the welding quality indicator, are shown in Fig. 9a. In order to know the level of melting activity at the interface between two separate sheets in the different strength ranges, another 70 weld coupons are sectioned and followed the standard metallurgical examination procedures, and then the weld nugget diameters are measured on the sectioned specimens using an optical microscope. The results are shown in Fig. 9b. Table 1 lists the maximum, minimum, and average values of the tensile shear strength and diameter of the weld nuggets.

Fig. 11 Validation test results of the decision tree classifier

Generally, the weld strengths and diameters increase with the increasing of the welding current, if other welding parameters remain constant. Accordingly, the welding quality can be summarized as seven classes according to the categories of the welding current parameters, except for expulsion welds which are defined as the eighth class. In Fig. 9a, the strength criterion of 2.89 kN is expressed as a dot line which decides good or bad welding quality. When the welding current is smaller than 3.8 kA, all welds present poor strength. With the welding current increasing to 3.8 kA, the welds strengths are 2.90– 3.03 kN, and they are considered as good welds. Other classes show adequate strength for any circumstance except class 8. Thus, a guideline for welding quality evaluation is proposed that it is considered as bad welding if the weld is assorted into classes 1, 2, 3, and 8. It is good welding if the weld is classified as other categories. Figure 10 plots the decision tree classifier for weld quality evaluation based on the features of the radar chart. The pruning method for decision tree classifier adopts pos-pruning algorithm. The classification rules expressed by inequalities are provided beside the root and internal nodes. If the numeric feature from the radar chart of a certain weld meets the classification rule, the left branch is selected; otherwise, the right branch is selected. The assorting process starts at the root of the tree and continues till a leaf node is encountered. Based on the decision tree, human can observe diagnostic procedure for weld quality intuitively. To test the classification performance of the decision tree, 120 test weld samples are selected, and the test results are shown in Fig. 11, where the sign ‘o’ represents the specified class according to the monitored welding current parameter and sign ‘*’ represents the classification result of the decision tree. Table 1 provides the numeric features of some test samples, and they can help the person to reproduce the assorting processes of these weld samples. It can be seen that all test welds can be classified correctly except for one case that the test sample 46 is classified as class 3, despite that its specified class is 4 according to welding current. However, this incorrect classification can be ignored because numeric features of the test weld 46 listed in Table 2 present big difference from the features of other test welds

Int J Adv Manuf Technol Table 2

Numeric features from radar charts of the test weld samples marked in Fig. 10

Test sample

7 16 28 37 41 46 57 68 87 100 109 116

Current (kA)

3.2 3.4 3.6 3.6 3.8 3.8 3.8 4.0 4.2 4.4 4.4 4.4

Characteristics Area (μm2)

Gravity_x (μm)

Gravity_y (μm)

D_max (μm)

428.72 624.49 732.01 721.33 815.80 736.77 797.34 1,044.69 1,172.54 1,413.74 71.04 96.84

−2.8806 −3.5092 −3.7355 −4.0436 −4.4059 −3.8632 −3.8233 −4.3172 −4.1816 −3.5774 8.3411 8.3822

−1.5789 −1.2652 −0.1991 0.0836 1.1378 −0.1386 0.9517 1.8886 3.4983 4.3454 6.7083 7.9963

14.36 17.04 19.00 18.61 20.50 18.93 20.06 23.25 25.77 27.72 21.38 21.91

under the same welding current. It is justified in thinking that some reasons influence the quality of the test weld during its welding process. It also can be confirmed from its strength 2.7910 kN which is smaller than the strengths of other test welds from 3.8 kA. It is notable that all expulsion weld samples are identified correctly which means the classifier can recognize expulsion well.

5 Quality evaluation under abnormal welding conditions In industrial applications of RSW, the welding quality is often influenced by various abnormal welding conditions, and many of them can influence the current density of the welding

Fig. 12 a Schematic of shunting in welding, b the electrode displacement curves and their radar charts from normal welding and shunting

Strength (kN)

Recognition class

1.5624 2.0452 2.6998 2.7301 2.9941 2.7909 2.9394 3.2936 3.4107 3.5135 Expulsion Expulsion

1 2 3 3 4 3 4 5 6 7 8 8

area such as current shunting, greasy surface, and small edge distance. Small current density in welding area will result in small heat generation, small nugget size, and poor welding quality; therefore, it is crucial for a reliable evaluation method to correctly assess the weld quality from the abnormal welding process. In this research, shunting, greasy surface, and the small edge distance welding are considered. Shunting in RSW is the diversion of the welding current from the weld to be made (the shunted weld) to a nearby existing weld (the shunt weld). The distribution of the welding current in shunting is illustrated in Fig. 12a. The proportion of the diverted current is determined by the relative electrical resistance values in the shunting and welding paths. As the applied electric current is shared by the shunt welds and shunted weld. The heat generation in the shunted weld may not be sufficient for it to grow to the designed size. The electrode displacement curves from normal welding and shunting welding, together with their respective radar charts, are compared in Fig. 12b, where the curves (1), (2), and (3) correspond to the normal, two-, and four-weld shunting welding, respectively. The schematic diagrams for designed shunting welding specimens are shown in Fig. 13a and b. Before the shunted weld ‘E’ is welded, other shunt welds around it with fixed spacing have already formed, thus, shunting phenomenon will occur during the welding process of the shunted weld ‘E.’ Three curves in Fig. 12 are all from 4.4 kA welding current.

Fig. 13 The schematic diagrams for the weld specimens: a two-weld shunting, b four-weld shunting, and c small edge distance


Fig. 14 The electrode displacement curves and their radar charts under normal and greasy surface welding

Compared with curve (1), the displacement amplitude of curves (2) and (3) reduces obviously. There are obvious differences in their radar charts. Greasy surface decreases the real contact area of the faying surface at the early stage of the RSW process, which causes the local current density increase; as a result, a quick increase of the heat generation induces the faying surfaces of the workpieces to soften rapidly. Thus, the contact resistance and the heat generation decrease quickly. With the welding process going on, the greasy surface will disappear, owing to the carbonization or gasification, and then heat generation caused by bulk resistance promotes the nugget growth just as it does for the normal weld. The electrode displacement curves (1) and (2) in Fig. 14 correspond to normal welding and greasy welding, respectively, and these two curves are acquired at the same welding current. Compared with the displacement curve from normal welding, the displacement amplitude of the greasy welding increases relatively slowly after a quick rising process and its peak value time also relatively lags. Moreover, the maximum amplitude of the greasy welding is small than normal welding. The obvious differences also can be observed from their radar charts. The schematic diagram for the specimen of small edge distance welding is shown in Fig. 13c. When the workpieces are welded with a small edge distance, the area for the heat dissipation is smaller than the normal welding; therefore, the

Fig. 15 The electrode displacement curves and their radar charts under normal welding and small edge distance condition

temperature of the weld increases quickly at the early stage of the welding process. However, without the constraint of the surrounding cold metal, the expansion of the nugget is much easier than the normal weld as the nugget grows to the stages II and III. In this case, the displacement amplitude curve from small edge distance welding is less than the normal welding just as the Fig. 15 shows (These two displacement curves are both from 4.4 kA). In addition, their radar charts also present notable differences. In order to test the validation of the classifier to estimate the weld quality from abnormal welding process, 40 test welds from shunting, greasy surface, small edge distance, and expulsion conditions are selected. The test results are shown in Fig. 16a where signs ‘○’ and ‘*’ represent specified class and recognition class, respectively. For the test welds from the small edge distance and greasy surface welding, the tensile shear strengths are measured and listed in Fig. 16b. For the

Fig. 16 a Recognition results of test welds from abnormal welding process, b the tensile shear strength of the welds from small edge distance and greasy surface welding, c the weld nugget diameters of the test welds from shunting welding

Int J Adv Manuf Technol Table 3

Numeric features from radar charts of the selected test samples

Test samples

4 9 12 16 18 23 26 34 37

Current (kA)

3.8 4.4 4.4 4.4 4.2 4.4 3.8 3.8 4.2

Characteristics Area (μm2)

Gravity_x (μm)

Gravity_y (μm)

D_max (μm)

770.56 66.60 54.02 66.42 1,139.56 1,135.93 696.11 540.24 790.72

−4.0877 8.2306 8.9676 8.7370 −4.6103 −4.2143 −3.6241 −2.9507 −4.05

−0.0361 6.5377 6.8157 6.9368 2.2165 2.9213 −0.0135 −0.9969 1.0822

19.40 20.98 21.39 22.11 24.11 25.30 18.09 15.39 20.19

shunting welding, measuring the shunted weld strength has proven difficult, and therefore, the weld nugget diameters of the shunted welds are measured, and the results are provided in Fig. 16c. By referring to the Table 1, it can be concluded that all test welds are classified properly by the decision tree classifier. Nine test welds marked in Fig. 16 are selected to reproduce the classification process, and Table 3 provides the numeric features of these test welds. As mentioned above that welds of 3.8 kA from normal welding process present enough strength, the weld quality, however, is deteriorated if it is form abnormal welding process. For example, the test weld 4 (with the tensile shear strength of 2.79 kN) from small edge distance welding and test weld 26 (with the weld diameter of 3.55 mm) from shunting welding are classified as class 3, although they are made at 3.8 kA. Referring to Table 1, the test results are proven reasonable. For test weld 9 from small edge distance and the test weld 16 from greasy welding, they are assorted to class 8 because expulsion occurs during their welding processes.

6 Conclusions Different from previous intelligent methods for welding quality evaluation focuses on extracting features directly from the acquisition data of monitored signal during RSW process; this paper proposed a novel approach to present the electrode displacement curve as radar chart format, and some geometric characteristics of the radar chart are selected to reflect the welding quality. By using the decision tree classification technology, a classifier for welding quality classification is developed, and the following results have been obtained. 1. The geometric characteristics of the radar chart from the electrode displacement signal closely relate to the weld quality, and it is feasible and reliable to use them to assess the welding quality non-destructively.

Specified class

Recognition class

4 7 7 7 6 7 4 4 6

3 8 8 8 5 6 3 2 4

2. Compared with the complex algorithm for feature extraction and selection directly from the monitored signal, achieving the geometric features from the radar chart is simple, fast, and efficient. 3. Performance test results of the decision tree classifier show that the classifier can identify good or bad weld accurately and rapidly, even though the weld is from abnormal welding process, in particular, the diagnostic procedure for welding quality is visible and intuitive, and it is easily understood and interpreted. 4. In this research, only mild steel material is selected to investigate the performance of the proposed method, and the performance of the method for some modern steels will be further studied.

Acknowledgments The support of this work by the Science and Techn o l o g y C o m m i s s i o n o f Ti a n j i n M u n i c i p a l i t y ( g r a n t n o . 13JCQNJC04100) is gratefully acknowledged.

References 1. Podržaj P, Polajnar I, Diaci J, Kariž Z (2008) Overview of resistance spot welding control. Sci Technol Weld Join 13:215–224 2. Gedeon SA, Sorenson CD, Ulrich KT, Eagar TW (1987) Measurement of dynamic electrical and mechanical properties of resistance spot welds. Weld J 66:378s–385s 3. Lee H-T, Wang M, Maev R, Maeva E (2003) A study on using scanning acoustic microscopy and neural network techniques to evaluate the quality of resistance spot welding. Int J Adv Manuf Technol 22:727–732 4. Ruisz J, Biber J, Loipetsberger M (2007) Quality evaluation in resistance spot welding by analyzing the weld fingerprint on metal bands by computer vision. Int J Adv Manuf Technol 33:952–960 5. Zhou K, Cai LL (2013) Online nugget diameter control system for resistance spot welding. Int J Adv Manuf Technol 68:2571–2588 6. Farson DF, Chen JK, Ely K, Frech T (2004) Monitoring resistance spot nugget size by electrode displacement. J Manuf Sci Eng 126: 391–394

Int J Adv Manuf Technol 7. Chen JZ, Farson DF, Ely K, Frech T (2006) Modeling small-scale resistance spot welding machine dynamics fro process control. Int J Adv Manuf Technol 27:672–676 8. Simončič S, Podržaj P (2012) Image-based electrode tip displacement in resistance spot welding. Meas Sci Technol 23:1–7 9. Kučšer L, Polajnar I, Diaci J (2011) A method for measuring displacement and deformation of electrode during resistance spot welding. Meas Sci Technol 15:592–598 10. Tang H, Hou W, Hu SJ (2003) Influence of welding machine mechanical characteristics on resistance spot welding process and weld quality. Weld J 85:116–124 11. Jou M (2003) Real time monitoring weld quality of resistance spot welding for the fabrication of sheet metal assemblies. J Mater Process Technol 132:102–113 12. Zhang YS, Wang H, Chen GL (2007) Monitoring and intelligent control of electrode wear based on a measured electrode displacement curve in resistance spot welding. Meas Sci Technol 18:867–876 13. Cho Y, Rhee S (2002) Primary circuit dynamic resistance monitoring and its application to quality estimation during resistance spot welding. Weld J 81:104s–111s 14. Zhang PX, Zhang HJ, Chen JH (2007) Quality monitoring of resistance spot welding based on electrode displacement characteristics analysis. Front Mech Eng China 2:330–335 15. Luo Y, Liu JH, Xu HB, Xiong CZ, Liu L (2009) Regression modeling and process analysis of resistance spot welding on galvanized steel sheet. Mater Des 30:2547–2555 16. Li W (2003) Manufacturing process diagnosis using functional regression. J Mater Process Technol 186:323–330 17. Cho Y, Rhee S (2000) New technology for measuring dynamic resistance and estimating strength in resistance spot welding. Meas Sci Technol 11:1173–1178 18. Zhao DW, Wang YX, Lin ZG, Sheng SN (2013) An effective quality assessment method for small scale resistance spot welding based on process parameters. NDT & E Int 55:36–41 19. EL-Banna M, Filev D, Chinnam RB (2008) Online qualitative nugget classification by using a linear vector quantization neural network for resistance spot welding. Int J Adv Manuf Technol 36:237–248

20. Zhang HJ, Hou YY, Sui XW (2013) Quality estimation of the resistance spot welding based on pattern feature of the electrode displacement signal. China Weld Eng Ed 22:53–58 21. Gong L, Liu CL, Li YM (2012) Control criteria determination and quality inference for resistance spot welding through monitoring the electrode displacement using Bayesian belief networks. J Adv Mech Des Syst Manuf 6:432–444 22. Zhang HJ, Hou YY (2009) Quality estimation of the resistance spot welding based on PCA-SVM. Hanjie Xuebao 30:97–100 23. Li DT, Pan CH, Du SM, Guo SL (2009) The multi-information fusion quality judgment of spot welding based on rough sets. Hanjie Xuebao 30:21–24 24. Zhang HJ, Hou YY (2011) Quality estimation of resistance spot welding based on kernel fisher discriminant analysis. Hanjie Xuebao 32:105–108 25. Podržaj P, Simončič S (2011) Resistance spot welding control based on fuzzy logic. Int Adv Manuf Technol 52:959–567 26. Wan Y, Cheng K, Liu ZQ, Ye HT (2013) An investigation on machinability assessment of difficult-to-cut materials based on radar charts. Proc IME B J Eng Manufact 227: 1916–1920 27. Kalonia C, Kumru OS, Kim JH, Middauqh CR, Volkin DB (2013) Radar chart array analysis to visualize effects of formulation variables on IgG1 particle formation as measured by multiple analytical techniques. J Pharm Sci 102: 4256–4267 28. Kimchi M (1984) Spot weld properties when welding with expulsion—a comparative study. Weld J 63:58s–63s 29. Pouranvari M, Abedi A, Marashi P (2008) Effect of expulsion on peak load and energy absorption of low carbon steel resistance spot welds. Sci Technol Weld Join 18:39–43 30. Kotsiantis SB (2013) Decision trees: a recent overview. Artif Intell Rev 39:261–283 31. Amarnath M, Sugumaran V, Kumar H (2013) Exploiting sound signals for fault diagnosis of bearings using decision tree. Measurement 46:1250–1256 32. Quinlan JR (1986) Induction of decision tree. Mach Learn 1:81–106