Cable Broken Wire Signal Recognition Based on Convolutional Neural Network

Abstract

Due to the long-term exposure of bridge ties to complex environments, their internal steel wires are prone to corrosion damage, which may lead to tie breakage accidents if not detected in time. Although existing advanced monitoring methods can be used to obtain the broken wire signal, they either still need the damage to be identified manually or are limited by the training data set. To address this problem, a model combination consisting of a classification model and three regression models was built based on convolutional neural networks to predict the location of broken wires after first classifying them based on features. We developed software-containing data set generation and model performance testing functions, in which we used original algorithms to expand the broken wire data set for training based on the measured data obtained from FBG sensors with a sampling frequency of 100 Hz, thus generating more than 22,000 types of data. The performance test results showed that the model combination successfully detected 11,972 broken wires among 12,000 test data points generated by the algorithm, with a recognition success rate of 99.77% and an average time of 0.0076 s between the predicted location and the actual broken wire location, with an error rate of 0.38%. In the test of 118 real broken wires, the model detected all the abnormalities, and the average time between the predicted location and the actual broken wire location was 0.0695 s, with an error of 3.48%. This verified the feasibility of using artificial intelligence to accurately identify broken wire signals and can provide a reference for the subsequent intelligent identification of tie abnormalities.