Cable fault location with convolutional neural network and attention mechanism

Abstract

Traditional cable-fault locating relies on technicians to spot fault signals and tweak pulse delays by hand, making the process slow and prone to subjective errors. We, therefore, provide a technique for locating cable faults that relies on the clever identification of magnetic and acoustic waveforms. A set of convolutional neural network-based acoustic and magnetic waveform intelligent recognition algorithm is designed by analyzing the acoustic field, magnetic field distribution law, and the principle of acoustic and magnetic synchronous quorum location. The key differences between acoustic–magnetic signals and interference signals are portrayed in the time–frequency diagram generated by continuous wavelet transform (CWT), and a CNN-attention mechanism recognition model is constructed to correctly recognize the acoustic–magnetic signals. Additionally, an automatic calibration algorithm of acoustic–magnetic time difference based on an empirical modal decomposition-Teager energy operator (EMD-TEO) is designed for the problem of observation error in the manual calibration of acoustic–magnetic time difference. This algorithm uses EMD decomposition to filter out aliasing noise and uses TEO to automatically calibrate the waveform's starting point, quantify the acoustic–magnetic time difference of the acoustic–magnetic signals, and identify the precise location of the point of failure based on the dimension of the time difference. The experimental results prove that the CNN-attention mechanism model has an accuracy of 97.56%, a precision of 97.37%, a recall rate of 97.38%, an F1-value index of 97.29%, and a running time of 0.43 s. The automatic calibration results of the EMD-TEO are basically in agreement with the average value of the manual calibration, and the measurement error is within 5%.