2018 Volume 58 Issue 4 Pages 721-726
In the present study, a quality monitoring system in small scale resistance spot welding based on dynamic resistance and neural network was proposed. A precise dynamic resistance curve was obtained directly from the welding machine. Dynamic resistance and nugget development were found sensitive to the change of welding current. The second resistance peak gradually disappeared as welding current increased. Weld expulsion occurrence could be detected through the obvious fluctuation in dynamic electrical signals. The resistance valley and end resistance were found highly correlated with the nugget development. Features extracted from the dynamic resistance curve were selected as inputs in the neural network analysis. A reliable weld quality estimation through the neural network approach could be finally achieved.
Resistance spot welding has been an important jointing technique since the 1940s. The overlapped sheets are jointed together by welding current and electrode force. Weld nugget is formed by resistance heating at the faying surface between plates. It is a crucial task for weld quality monitoring in production line. Weld quality can be assessed by destructive and non-destructive methods. Destructive inspections such as tensile-shear and cross-tension tests require much time and money. The overall situation may not be reflected just by testing on a sample basis. More attentions are now being paid to the non-destructive method. Non-destructive quality detection in traditional large scale resistance spot welding (LSRSW) is generally concentrated on the usage of electrode force,1,2,3,4) electrode displacement,5,6,7) ultrasonic signal,8,9,10) dynamic resistance,11,12,13,14) etc.
Park et al.1) developed different neural network models to evaluate the weld quality through classified electrode force patterns. Podrzaj et al.2) showed that the electrode force was a good indicator of weld expulsion occurrence. El-Banna et al.3) classified cold, normal, and expulsion welds using the non-intrusive force sensor. Wang et al.5) developed the monitoring procedure based on electrode displacement curve using moving range chart. Wang et al.6) adopted the electrode vibration signal for quality monitoring under various welding conditions. Zhang et al.7) incorporated the measured electrode displacement with radar chart method for weld quality estimation. Thornton et al.8) found that the weld quality detected by ultrasonic C-scan showed good agreement with peeled samples and metallographic observation. Martin et al.9) used the pattern recognition technique to classify ultrasonic oscillograms of welded joints. Liu et al.10) found that the ultrasonic C-scan image of frequency domain characteristics signal could be used for an accurate estimation of nugget dimension.
As to the dynamic resistance based weld quality estimation in LSRSW, the dynamic resistance was generally obtained from the secondary circuit in early studies.11) The induced voltage can be eliminated in this way when the dynamic resistance is calculated at the peak of welding current. Cho et al.12) developed a dynamic resistance based monitoring system combining the primary circuit. Ling et al.14) adopted real part of the input electrical impedance as the monitoring signature, the resulting continuous dynamic resistance could be well related to the nugget development.
Small scale resistance spot welding (SSRSW) is a recently developed jointing technique, by which sheet metals generally thinner than 0.5 mm are jointed together. The SSRSW is widely applied in fields of medical products and electronic components.15) Joints are formed between workpieces through resistance heating, which is the same fundamental mechanism as that in LSRSW. However, there exist significant differences between the SSRSW and LSRSW, including the magnitude of welding parameters, cooling way of electrodes, and materials to be welded.
Less attentions have been paid to the weld quality monitoring in SSRSW. Farson et al.16) found that the detection of electrode voltage spike in SSRSW could be used for expulsion monitoring. Zhao et al.17) developed a quality monitoring model in SSRSW using the electrode voltage signal. Tseng et al.18) stated that the maximum electrode displacement in SSRSW was useful for the estimation on nugget development. Chen et al.19) found that the nugget size in SSRSW of stainless steel was inversely proportional to the minimum dynamic resistance obtained numerically. Tan et al.20) showed that the dynamic resistance in SSRSW of bare Ni sheets was useful for weld quality control, but further related work was not mentioned.
The precise value of dynamic resistance can be obtained conveniently in SSRSW, on condition that the power source is provided at high frequency inverter mode. However, by now, limited work has been conducted on weld quality monitoring research in SSRSW by using the dynamic resistance. The variation mechanism of dynamic resistance should possess specific characteristics compared with that in LSRSW. Therefore, applicability of the quality monitoring approach based on the dynamic resistance in SSRSW deserves to be investigated.
This study aims at developing a reliable weld quality monitoring system in SSRSW of titanium alloy. Continuous dynamic resistance was obtained using the measured electrode voltage and welding current. The variation characteristics of dynamic resistance curve were associated with nugget development and weld quality. A simple and effective quality monitoring approach adopting features extracted from the dynamic resistance curve and neural network was finally proposed.
Titanium alloy is a promising material due to the excellent properties of low density, high strength to weight ratio, and good corrosion resistance. Here, the base metal was selected as TC2 titanium alloy with dimensions of 100 mm × 30 mm × 0.4 mm. Chemical composition of the TC2 is listed in Table 1. The spot weld was generated at the centre of overlapped region (30 mm × 30 mm), as illustrated in Fig. 1.
| Alloying Elements | Impurities (not higher than) | |||||||
|---|---|---|---|---|---|---|---|---|
| Al | Mn | Ti | Fe | C | N | H | O | Others |
| 3.5–5.0 | 0.8–2.0 | Bal. | 0.30 | 0.10 | 0.05 | 0.012 | 0.15 | 0.40 |
Specimen dimensions for SSRSW.
Although the antioxidation capacity of titanium alloy at high temperature is relatively poor, resistance spot welding is still supposed as a useful jointing technique, considering that an effective atmospheric isolation in the welding zone can be provided by the electrode pressure.21) However, the SSRSW of titanium alloy should still be taken carefully to maintain a stable weld quality. Here, the mechanical and chemical cleaning on samples were both conducted before welding. Coarse oxides and other contaminants were first removed by a hard brush. The samples were then etched with a mixed solution of nitric acid (45%), hydrofluoric acid (20%), and water (35%) for several minutes. Afterwards, the samples were placed in a ventilated environment after cleaned by running water.
The SSRSW machine was provided by the Miyahci Unitek Corporation, as shown in Fig. 2. No cooling water was supplied during the welding process. Diameter of the electrode tip was selected as 3.0 mm. Welding was carried out using the high frequency inverter power supply at the constant welding current mode. Welding parameter combinations were arranged randomly and uniformly as much as possible, and as follows: the electrode force (F) was varied from 75 N to 200 N at an interval of 25 N, the welding current (I) was varied from 1.0 kA to 2.4 kA at an interval of 0.2 kA, and the welding time (T) was varied from 4 ms to 12 ms at an interval of 2 ms.
The SSRSW machine. (Online version in color.)
The electrode voltage was directly measured by signal leads attached to the electrode. The welding current was measured using the Rogowski coil. Both signals were sampled at a fixed frequency. The dynamic resistance could be obtained by dividing the electrode voltage by corresponding welding current.
The quasi-static tensile-shear test was performed on spot welded samples. An Instron universal testing machine was adopted at a constant cross-head speed of 1.0 mm/min. Failure load was extracted from the load-displacement curve. Nugget size was obtained from the fractured surface after tensile-shear test. The nugget size was measured by a vernier caliper for several times from different points of view. The average value was used as the nugget size adopted finally.
Effects of welding current and electrode force on the dynamic resistance curve are shown in Figs. 3(a) and 3(b), respectively. Here, the variation characteristic of a representative curve in Fig. 3(a) under the condition of F = 100 N, I = 1.0 kA, and T = 10 ms is first analyzed. A rapid increase in dynamic resistance is detected at the initial stage, the peak of which is termed here as the γ resistance peak (Rγ). The Rγ can be generally interpreted by the asperity heating.20) The asperity resistivity at incomplete contact interfaces increases with temperature, meanwhile the heat generation in bulk material is limited. The dynamic resistance drops dramatically after the Rγ, which may be due to the surface asperity breakdown, material softening, and contact area increasing. As the welding continues, temperature of the bulk material increases consistently. The resulting increase in bulk resistance gradually dominates the variation of dynamic resistance, which leads to a second increasing tendency of the dynamic resistance. The α resistance valley (Rα) is formed as a consequence.
Effects of welding parameters on dynamic resistance variation.
The contact resistance at the sheet/sheet interface gradually disappears after the nugget formation. The cross-sectional area increases as the nugget size enlarges. Meanwhile, the flow path for the welding current is further shortened due to the mechanical collapse. The passage of welding current is thus promoted. Moreover, variation of the bulk resistance is restricted due to the limitation of further increase in temperature. The β resistance peak (Rβ) is formed as a result of above mentioned competitive mechanism. Afterwards, the dynamic resistance decreases sustainedly until the end of welding.
The electrode force is maintained at 100 N in Fig. 3(a). Variation of the dynamic resistance before Rα is relatively complicated. It seems that a lower welding current promotes a higher initial resistance, but it is not all the cases. The difference in surface pretreatment quality should account for this phenomenon. The Rβ is found sensitive to the change of welding current. The Rβ gradually moves forward and disappears as the welding current increases, which is due to the fact that the extent of surface breakdown, material softening, nugget development, and mechanical collapse are all promoted under the condition of a higher welding current, while the increase in bulk resistance is limited. The increasing rate in nugget size and mechanical collapse gradually slows down as welding current increases. The overall decreasing trend of the dynamic resistance is therefore reduced. Additionally, the occurrence of weld expulsion can be detected under the condition of 2.4 kA welding current. An abnormal fluctuation in dynamic resistance curve can be observed then. Besides, the Rα is detected at nearly an identical time under all welding conditions.
The welding current is kept constant at 1.6 kA in Fig. 3(b). A lower electrode force generally indicates a larger initial Rγ, which is probably due to the change in initial contact asperity and resistance heating at contact interfaces. The Rβ can not be detected when the electrode force is 75 N, which is related to the premature nugget formation caused by a higher initial resistance. The difference in dynamic resistance after Rα under 100 N, 125 N, and 150 N welding conditions can be found very limited. The overall variation of dynamic resistance curve under 175 N and 200 N electrode force conditions is very close to each other. In addition, the Rα is detected at nearly an identical time under all welding conditions.
3.2. Effects of Welding Parameters on Weld QualityThe nugget size in SSRSW is much smaller than that in LSRSW. Reliability of the weld quality represented by nugget size should be analyzed first. In both welding processes, the failure load can be adopted as the quality indicator. Here, result of the correlation analysis between nugget size and failure load in SSRSW shows that the correlation coefficient is roughly at 0.96. There exists a good linear relationship between them. The nugget size is proved useful in characterizing the weld quality in SSRSW.
Effects of welding parameters on nugget size is depicted in Fig. 4. The electrode force is set as 100 N in Fig. 4(a). The welding current is kept constant at 1.6 kA in Fig. 4(b). As can be seen, the nugget size increases with welding time considering the sustained heat input. Variation of the nugget size is found sensitive to the welding current and insensitive to the electrode force. The nugget size increases obviously when the welding current is increased from 1.2 kA to 2.0 kA. The increasing rate of nugget size is reduced when the welding current is further increased from 2.0 kA to 2.4 kA, which may be attributed to the increasing heat dissipation. The small variation in nugget size under various conditions of electrode forces can be interpreted by the limited effect of electrode force on the heat generation. Effects of welding parameters on the nugget size is found similar to the variation of dynamic resistance curve after Rα.
Effects of welding parameters on nugget size.
A group of dynamic electrical signals under the condition of weld expulsion are illustrated in Fig. 5. The electrode force and welding current are kept constant at 75 N and 2.4 kA, respectively. The welding mode of constant welding current can be clearly identified from Fig. 5(a). The curve of electrode voltage signal first experiences a peak value, and then decreases sustainedly until the welding end. The difference in dynamic resistance at the initial stage in Fig. 5(c) may be attributed to inconsistent surface pretreatment qualities. The disappearance of Rβ in Fig. 5(c) represents a very high rate of heat generation and nugget growth.
Detection of the weld expulsion occurrence (F = 75 N, I = 2.4 kA).
As to the condition when welding time is larger than 4 ms, weld expulsion can be detected due to the combination of low electrode force and high welding current, under which condition the thermal expansion of liquid metal is too much to be contained in the fusion zone. Obvious fluctuation in both electrode voltage and dynamic resistance can be detected on expulsion occurrence, as seen from the insets of Figs. 5(b) and 5(c). Both dynamic electrical signals are found decreased steadily again after weld expulsion. The occurrence of weld expulsion can thus be deduced from the abnormal fluctuation in electrode voltage and dynamic resistance.
3.4. Back Propagation Neural Network (BPNN) ApproachA statistical analysis on effects of welding current and features extracted from the dynamic resistance curve on the nugget size is shown in Fig. 6. As indicated previously, the Rα and Rβ are insignificant under very high welding current condition. Considering that the Rα can be detected at nearly the identical time, therefore the Rα is extracted here at a specified average time. The Rβ is not used here due to the measurement difficulty under very high welding current condition. The correlation coefficient between Rα and nugget size is found approximately at −0.79. The correlation coefficient between end resistance (Re) and nugget size is found approximately at −0.90. The Rα and Re will be adopted as inputs in the neural network approach.
Effects of extracted features and welding current on nugget size.
Artificial neural network is a computer program simulating the working mechanism of human brain in information processing. The strong capabilities of self-adaptive, self-learning, and nonlinear mapping make it very promising. BPNN is one of the most widely used neural network models capable of approximating continuous function. The BPNN is a multilayer feed-forward network including input, hidden, and output layers. A typical BPNN approach generally consists of model training and model validation.
An error back propagation algorithm is adopted here in the model training. The training procedure is composed of forward propagation and error back propagation stages. Initial weights connecting neurons are first randomly generated. Input vectors of training samples are fed into the input layer. Corresponding outputs are compared with target values through an error function (Eq. (1)). The weights are then updated in the error back propagation stage using gradient descent algorithm (Eq. (2)). Training is stopped when the error function converges to a predefined small value or the maximum iteration number is attained.
| (1) |
| (2) |
The neural network toolbox in MATLAB software is used here for model training and model validation. The F, I, T, Rα, and Re are selected as network inputs. The network output is designed as the nugget size (D). The tansig and purelin transfer functions are utilized in hidden and output layers, respectively. The Levenberg-Marquardt algorithm is used to update network weights.
Determination of the network architecture is an important task in BPNN analysis. There is no strict criterion in determining the number of hidden layers and neurons in each hidden layer. One hidden layer is generally enough to achieve most continuous mapping. The trial and error attempt method is adopted here to establish the structure of hidden layer. One hidden layer with five neurons is found the most appropriate after many attempts. Results of the model training and model validation are shown in Fig. 7. Seventy individuals are randomly selected from all samples in the training procedure. The remaining sixteen samples are used for model validation.
Results of the BPNN model training and model validation. (Online version in color.)
In practice, knowing the information of weld quality level is enough for production engineers. Here, the weld quality can be classified into three levels. The threshold nugget size for different levels is selected as 1.34 mm and 2.15 mm, respectively. For spot weld at level I (D > 2.15 mm), weld expulsion can be detected easily despite the large nugget size. As to spot weld at level III (D < 1.34 mm), the unsatisfactory interfacial failure is the common case. The spot weld at level II (1.34 mm ≤ D ≤ 2.15 mm) is supposed satisfactory at both surface quality and mechanical performance.
As can be seen from the comparison between linear fit and reference line in Fig. 7(a), the predicted nugget size shows good agreement with measured size. The root mean square error is only 0.09 mm. Result of the weld quality classification is also depicted in Fig. 7(a). The level of weld quality can be found well predicted. Additionally, the root mean square error in model validation is about 0.12 mm, which is a little larger than that in the training procedure. Only the No. 4 sample is inappropriately classified in model validation, which may be due to that the corresponding nugget size is close to the threshold value of 1.34 mm. The inappropriate classification is related to the BPNN prediction accuracy.
Based on above analysis, the quality level classification combining BPNN can be used to achieve a relatively high accuracy rate in weld quality prediction. The data processing in BPNN can be accomplished completely before next welding operation is started. Measurement of dynamic electrical signals is also very convenient. A simple and effective quality monitoring approach can thus be established.
An effective weld quality monitoring approach in SSRSW of titanium alloy was proposed in this study. The dynamic resistance signal and BPNN model were used to build the quality monitoring system. The precise dynamic resistance curve was directly derived from the welding machine. Effects of welding parameters on the dynamic resistance and nugget size were investigated. Features extracted from the dynamic resistance curve were adopted as neural network inputs. Weld quality was estimated using the proposed BPNN model. Following conclusions can be obtained:
• The occurrence of weld expulsion can be deduced from the obvious fluctuation in dynamic electrical signals.
• The resistance valley and end resistance are highly correlated with the nugget size.
• A reliable weld quality estimation can be achieved by the BPNN model combined with quality level classification.
This research work was funded by the National Natural Science Foundation of China (No. 11072083 and 51705180) and the Doctoral Dissertation Innovation Fund of Huazhong University of Science and Technology (No. 0118240059). The authors would like to thank Mr. Lin and Mr. Sheng from the Miyachi Unitek Corporation for providing the welding machine. Experimental support from the Analysis and Testing Centre of Huazhong University of Science and Technology is also acknowledged.
