A Transformer Oil Temperature Anomaly Detection Method Based on LSTM-Attention

In response to the traditional detection methods have been difficult to effectively detect a large number of time-series oil temperature data and other problems. In this paper propose an unsupervised anomaly detection method that combines a long and short-term memory network (LSTM) with an Attention mechanism. Firstly, multiple LSTM units are formed into an LSTM network to collaborate with each other in order to learn the long-term dependencies of the data in the time series sequence. Subsequently, the Attention mechanism was introduced to automatically assign attention weights to the outputs of the LSTM units for each data point in a time series thereby performing weighted aggregation to obtain the predicted values. Finally, the oil temperature was evaluated whether it was anomalous or not based on the prediction error for each data in all the time series, and the experimental results showed that in comparison with the similar methods, both of them showed much better DBI (0.4477) and DVI (11.4516).