Online anomaly detection method integrating LSTM and MGD for suspension system of maglev trains

In view of online anomaly detection of suspension system of maglev trains, a detection method based on long short-time memory (LSTM) neural network and multivariate Gaussian distribution (MGD) was proposed. The LSTM time series prediction model for learning the normal operation of the suspension system was established, and the prediction errors under normal conditions were obtained. Based on the prediction errors of the gap, current, and acceleration, the MGD model reflecting the distribution characteristics of the prediction errors under normal conditions was built. The log probability density was used as the detection indicator, the online detection logic was designed, and the threshold was set using F₁ score as the evaluation criterion of the detection effect. To verify the effectiveness of the proposed method, the operation line data of the maglev train were used to simulate online data, and the proposed method was applied to detect and analyze track crossing joint, rail smashing, and acceleration signal anomalies. Research results show that the F₁ scores of the proposed method for detecting the above three anomalies are 100.00%, 97.85%, and 83.33%, respectively. The detection indicators of the proposed method are significantly different between normal and abnormal conditions, reflecting the specific time period from anomaly occurrence of the suspension system to normal adjustment, and the algorithm takes about 2 s on average. Compared with the method based on the Gaussian distribution of hyperspheres, the proposed method achieves an average improvement of 1.9% in detection rate. Specifically, it achieves a 9.4% increase in detecting short-duration track crossing joint anomaly. Therefore, the proposed method can achieve online anomaly detection of the suspension system's state data.