Rail defect identification method based on pulsed eddy current detection and Small-DCNN

Accurate and rapid identification of defect was an important guarantee for the safe operation of equipment. In order to avoid the shortcomings of time-consuming and subjective factors in the feature extraction process, a defect recognition pipeline based on pulsed eddy current testing (PECT) and Small Deep Convolutional Neural Network (S-DCNN) was proposed. Firstly, the one dimensional (1-D) defect signal was obtained from the PECT experiment. Secondly, the above-mentioned 1-D signal was transformed into a two-dimensional time-frequency representation (2-D TFR) by using the Smooth Pseudo Wille-Velle Distribution with Contour (SPWVD-C) method. Lastly, the above-mentioned 2-D TFR was used as the input of the S-DCNN, which consisted of four convolutional layers (Cons) and a fully connected layer (FC). Experimental results showed that the proposed SPWVD-C method could obtain more accurate and clearer 2-D TFR than other common-used time-frequency transform methods such as Ensemble Empirical Mode Decomposition (EEMD), Short-time Fourier Transform (STFT), and Synchronous Wavelet Transforms (SSWT). The formed S-DCNN was more simple and effective than VGG architectures (such as VGG11, VGG16 and VGG19), and had significant advantages as far as recognition accuracy and time cost. The propose pipeline was suitable for non-destructive testing and other engineering applications where it was difficult to obtain a large number of training samples.