SVM classifier for wire extension monitoring using arc sound signal in GMAW

To find the approach of monitoring welding quality by arc sound,the frequency spectral characteristics of the arc sound signals in short circuit GMAW process were analyzed.The concept of tone channel of welding arc was introduced,which was considered a time dependent distributed parameters system influenced by welding parameters,arc behavior and the other factors.The LPC,(linear prediction coding) model of the arc sound signal was an estimation of transmission properties of the tone channel.The spectrum analyses indicated that the frequency characteristics of the arc sound signal were closely related to the wire extension,but the correlation presented high complexity and nonlinearity.The classifiers based on SVM,(support vector machine) for monitoring wire extensions were established,in which the input vectors of sample sets were built with the predictor coefficients and reflection coefficients of LPC model of the arc sound signals.The training and testing results showed that the SVM classifiers with different kernels are all capable of classifying the wire extension,whose performances were obviously better than that of the RBF,(radial basis function) neural networks under the same condition,and present good capability of generalization at small sample set.The classifiers with 3rd order polynomial kernel trained with reflection coefficients input vectors has the best accuracy,i.e.over 98%.The study indicated that forming characteristic vectors by the LPC coefficients of arc sound to build SVM pattern recognition model is a feasible way for welding parameters monitoring.