Health status evaluation of responder transmission unit based on CNN-transformer encoder-BiLSTM model

Abstract The train control system is a critical piece of equipment in the field of high-speed railway signaling. Timely and comprehensive mastery of its health status, shifting daily maintenance from breakdown maintenance to condition-based maintenance, plays a vital role in guaranteeing safe train operation and improving equipment life. A CNN-Transformer encoder-BiLSTM serial-parallel neural network is proposed to predict and evaluate the health status of train control equipment. First, the collected multi-dimensional degradation status information is subjected to multi-scale local feature extraction by a Convolutional Neural Network (CNN). The improved Transformer encoder module performs feature extraction through parallel multi-head attention. By introducing three different attention mask mechanisms into the Transformer encoder, it focuses on data at different positions respectively to extract their correlations. Secondly, the Bidirectional Long Short-Term Memory (BiLSTM) module extracts, memorizes, and processes the degradation information with positional encoding input to the Transformer encoder to extract the correlation of historical sequences, thereby improving prediction accuracy. Meanwhile, an improved particle swarm optimization algorithm is introduced to establish dynamic nonlinear inertia weights to find the optimal hyperparameter combination of the model. Finally, the regression layer outputs the predicted value of the Health Index (HI) percentage. The model is capable of conducting real-time and accurate health status prediction within the entire service life of the equipment, which will assist the ground maintenance site in timely maintaining equipment triggering health warnings, thereby ensuring train operation safety while significantly reducing maintenance workload and financial costs. Experiments indicate that the prediction accuracy of the model reaches over 96%, providing a new perspective for the field of health status prediction based on deep learning models, and possessing important reference value for the safe maintenance of train control equipment and the assurance of train safe operation.