Prediction of Performance Degradation Trends of Turnout Machine Based on Autoencoder and TCN Network

In response to the frequent failures of turnout machines, which make it difficult for on-site staff to predict these failures, this paper proposes a method for predicting the performance degradation trend of turnout machines based on Autoencoder (AE) and Temporal Convolutional Network (TCN). First, time-domain, frequency-domain, and time-frequency features of the turnout machine's action power data are extracted to construct a feature vector set that reflects the degradation process of the turnout machine. KPCA is utilized to remove redundant information present in multiple-domain features. A Healthy Indicator (HI) is constructed based on the AE network. Subsequently, a performance degradation prediction model based on TCN is built to achieve the prediction of the turnout machine's degradation trend. The effectiveness of the method is verified using on-site datasets. The outcomes demonstrate that KPCA successfully reduces the feature set from 22 dimensions to 2 dimensions, retaining 81.33% of the information from the original optimal feature set. Additionally, the approach of constructing a healthy indicator based on the AE network proves to be effective. The generated healthy indicator accurately reflects the degradation process of the turnout machine's performance over time. The model developed using the TCN algorithm accurately forecasts the performance degradation of the turnout machine. This model outperforms prediction models constructed based on LSTM networks and GRU networks in terms of prediction accuracy. This method is generally applicable and has practical guidance for on-site maintenance.