基于电弧声信号的CO2焊接状态模式识别

CO2气体保护焊接电弧声信号与焊接参数和电弧状态密切相关,但由于存在高度的复杂性和非线性性,难以直接用于焊接过程监控.在对不同保护气流量和焊炬高度下电弧声信号频谱分析的基础上,采用线性预测编码(LPC)方法建立其参数化模型,利用LPC预测系数和反射系数构造特征向量,通过样本训练分别建立了RBF神经网络和支持向量机(SVM)模型,进行CO2气体保护焊接下气流量和焊炬高度识别和分类.测试结果表明,电弧声LPC预测系数和反射系数作为输入向量训练的RBF网络或SVM模型均能一定程度上实现保护气流量和焊炬高度的正确识别。其中采用LPC反射系数时结果优于预测系数;SVM模型的分类能力明显优于RBF网络,且不随训练样本的减少急剧下降.

Pattern recognition of CO2-gas shield welding based on arc sound signal

The frequency domain characteristics of arc sound signal in CO_2 shield welding is closely correlative with the welding parameters and arc behavior,which is difficult to be used directly for quality control and monitoring because the correlation among them presents high complexity and nonlinearity.On the basis of frequency spectrum analysis of the arc sound signals under different shield gas fluxes and torch heights,the LPC(linear prediction coding) was adopted for parametrically modeling the sound signals.The prediction coefficients and reflection coefficients were respectively utilized to construct the characteristics vectors,by which the RBF neural networks and SVM classifier models were built with the different training sample sets for classification of shield gas fluxes and torch heights in CO_2 shield welding.The test results showed that the RBF networks and SVM classifiers with the prediction or reflection coefficients input vectors were all capable of correct recognition for gas fluxes and torch heights to some extent,in which the classification exactitude rates of the models with the reflection coefficients input vectors were better than those with the prediction coefficients,the exactitude rates of the SVM classifiers obviously excelled those of the RBF networks models.The classification capability of the SVM classifiers was not debased quickly along with decrease of the training sample numbers.