Condition monitoring of train air-conditioning systems using an LSTM-MSET framework with attention and adaptive memory

With the rapid development of urban rail transit systems, the workload associated with train inspection and maintenance has increased significantly. Traditional periodic maintenance approaches are increasingly unable to meet growing operational demands, leading to a shift toward more scientific and efficient condition-based maintenance strategies. To address the challenges in monitoring the operational state of train air-conditioning systems, such as the scarcity of abnormal data, low sensitivity in anomaly detection, and a high false alarm rate, this paper proposes a novel method that integrates long short-term memory (LSTM) networks with the multivariate state estimation technique (MSET). The method combines the strength of LSTM in time-series modeling with the suitability of MSET for multivariate state estimation and anomaly detection, it further incorporates temporal and feature-level attention mechanisms. In the offline phase, a memory matrix is constructed by selecting representative healthy samples from historical data using the Minibatch-KMeans clustering algorithm. In the online monitoring phase, an adaptive dynamic memory bank update mechanism is introduced to improve the model’s adaptability to varying operating conditions. Experiments were conducted using data collected from the train control and monitoring system (TCMS) of a metro train. The results show that the proposed model achieves excellent prediction performance under healthy conditions, with RMSE of 0.384 and MAE of 0.208, outperforming baseline models. In fault detection, it provides early warnings 62 hours in advance with a 0% false alarm rate, demonstrating strong sensitivity and robustness.