Abstract: Based on three-coil detection technology under bidirectional excitation, a method to identify weld defect types building on weld state identification and feature subset optimization was proposed in this paper. Firstly, characteristics and complexity of weld signals were analyzed by utilizing non-linear characteristic analysis method and Lemper-Ziv complexity value so as to identify weld category. Then, extracted features were reduced in dimension based on improved maximum relevance minimum redundancy feature selection method. Finally, optimal feature subset under different weight factors were used as input of SVM to identify defects of tested signals in oil-gas station. The results indicate that Lemper-Ziv complexity-based distribution can help accurately distinguish weld types. And it is verified by tests that recognition method based on feature subset optimization has high recognition accuracy for both longitudinal defects and transverse defects of welds, and its overall accuracy can reach as high as 83.33%.

Key words: oil-gas station, weld defect, eddy current testing(ECT), Lemper-Ziv complexity value, feature selection, support vector machine(SVM)