Spot Welding Quality Classification Based On Support Vector Machine Model

The voltage, current and electrode displacement of resistance spot welding contains plenty of welding information, which associated with welding parameters, surface state of work pieces and welding stability, is the importance information resources for quality inspecting of weld spot. Aiming at joints quality classification, electrode voltage, current and displacement have been studied by means of modern signal analyze methods, time domain statistics factors of signals have been collected, constructing mult-information integration characters data set to describe the welding process, building the joint strength classification modern by use of SVM, CART, Hopfield network. Test result showed that SVM classification model owned high accurate rate, good expansion ability which produced practical value to apply the model to joint quality classification. The work have involved following factors.For gather electrode voltage, welding current and electrode displacement, a data collecting system have been constructed for the core of AC6115 AD card, Rogowski current transducer, DA-5differentical voltage transformer displacement sensor of direct current. Based Visual Basic6.0, a data gather software has been developed, which own the ability of data collecting, wave show, data show and connect with Matlab.By use of time-domain, frequency-domain methods, the characters of welding signals have been analyzed, the goal of which was to find a certain pattern of description of welding process from signal aspect. Result show that dynamic resistance and electrode displacement signals can reflect welding process. Signals characters in frequency domain and time frequency domain were not obvious, so factors would be extracted from time domain. From dynamic resistance and electrode displacement signals, 13 factors such as peak value, ascending and descending rate have been extracted, by relativity analyze methods, mult-information integration set has been built. By means of CART data mining, classification model has been built based on factor vectors from dynamic resistance and electrode displacement, which can show the complicated model with tree form, test result indicated that CART classification model can be applied in quality classification. Based on discrete factors set from dynamic resistance and electrode displacement, a character pattern has been built, by saving the pattern in Hopfield network, the pattern identify of unknown joint can be finished, by which the classification has been practiced.SVM is a new data mining method, based on dynamic resistance and electrodedisplacement signal factors set, SVM classification model have been constructed. By test between the two kinds of SVM, the result showed that they have same classification accuracy under the condition of same training set and match parameters. Under the match parameters condition, 4 kinds of kernel function have good classification result, the accuracy is above 95%, among of which the accuracy of liner kernel function has reached 100%. When the number of test sample is larger than that of training sample, RBF kernel function is better than others, under the same situation, compared with CART and Hopfield network, SVM own better classification ability.