Prediction of railway track irregularity based on TCN-BiLSTMs-Attention model

Track irregularities can significantly reduce the comfort and safety of train operation. If the development trend of track irregularities can be predicted, the railway management department can issue early warnings to ensure the safe and smooth operation of trains. The sequence data associated with track irregularities exhibit complex nonlinear and non-stationary characteristics, and predictions based solely on the track quality index (TQI) do not take into account the influence of other factors. To address this issue, this paper proposes a model based on TCN-BiLSTMs-Attention for predicting the development trends of track irregularities. First, the Time Convolutional Network (TCN) is employed to extract features from the preprocessed multivariate input data of track irregularities, so as to capture the impact of various indicators on TQI values and provide richer inputs for subsequent model stages. The data is then fed into a multi-layer bidirectional long short-term memory network (BiLSTM) for training. The stacked BiLSTM layers utilize the non-linear changes and long-term dependencies of TQI sequence data that LSTM networks can learn, as well as the bidirectional LSTM, to simultaneously process past and future information. This increases the depth of the model, further improving its learning ability. Subsequently, the attention mechanism module is used for information extraction, enhancing the model's sensitivity to key information and its ability to capture long-distance dependencies. Finally, the prediction results are output through a fully connected layer. Comparative experiments with multiple models have shown that the proposed TCN-BiLSTMs-Attention model yields more accurate predictions and has strong robustness.