Subway station escalator data prediction based on MA-Transformer

The escalator system, which experiences data fluctuations during continuous daily operation, plays a vital role in subway stations. To accurately assess the health status of the escalator, it is crucial to predict its future condition. To mitigate the effects of significant fluctuations in data and leverage the correlations among cross-variables, the Multi-scale Average Pool Transformer (MA-Transformer) is proposed. The model pre-processes the data using average pooling layers with different kernel sizes, segregates the intricate escalator data into trend and residual components, and predicts them separately with two independent networks. This approach enhances the essential data representations and mitigates the adverse effects of frequent data fluctuations. Given the substantial number of variable types in escalator data, this paper introduces two modules that focus on the self-attention mechanism to model the inter-relationships between multiple variable types. Additionally, the trends and residuals output from different average pooling layers are fed into the self-attention layer, which provides a more comprehensive representation of features at the same time point. This enables the model to better capture the overall trend changes and the distribution of anomalies that deviate from the trend. This study collected data from 19 classes from 2 escalators at the engineering site for training and testing. Experimental results demonstrate that the MA-Transformer achieves superior detection accuracy, outperforming evaluated time-series prediction baseline models in terms of both Mean Absolute Error (MAE) and Root Mean Square Error (RMSE).