In order to solve the problem that the unlabeled data of the temperature part of the motor controller cannot be utilized in the motor troubleshooting of new energy vehicles, LSTM-autoencoding density spatial clustering (LSTM-AE-DBSCAN) motor controller temperature anomaly label recognition method is proposed. Firstly, the text outlines a process whereby the neurons in the hidden layer are replaced with LSTM neurons using the autoencoder (AE) algorithm to extract time series data features while preserving the key information of the data structure. Secondly, the motor controller monitoring data is divided by using the sliding window approach to obtain optimal LSTM-AE model performance. Then, the temporal features of the new energy vehicle motor controller no-abnormality monitoring data are extracted and reconstructed from the input data. Besides, the improved DBSCAN algorithm trains the difference between the reconstructed value and the actual value in an anomalous manner. The EPS and Min-Samples parameters are generated by chaotic sequences, and the silhouette coefficients are used as fitness values. These values are then optimized by the chaotic game optimization (CGO) algorithm to obtain the optimal parameters of the DBSCAN algorithm. This approach reduces human involvement and minimizes clustering errors. Finally, the time series data from the motor controller during the operation of the real vehicle is input into the model described above. Comparative experiments have been conducted and the proposed enhanced method has a higher precision of 0.985, a recall of 1, and a recognition time of 1.47 s compared to K-means, LOF, and single DBSCAN anomaly recognition methods, proving the effectiveness of the proposed model.
Zhou et al. (Tue,) studied this question.