Anomaly Detection Method for Multivariate Time Series Data Based on BLTranAD

The text proposes a deep learning model, BLTranAD, for anomaly detection in time series data, which combines Bidirectional Long Short-Term Memory (BiLSTM) and Transformer to more effectively capture anomalous patterns in time series data, improving upon the mTranAD algorithm. BiLSTM allows the model to take both past and future context into account, which allows BiLSTM to pass a more comprehensive context to the Transformer model, a combination that allows the model to better adapt to patterns on different time scales and capture potentially anomalous patterns more efficiently. The tanh activation function used in the model limits the output range to [−1,1]. This constraint helps to smoothly normalize the output, particularly in the computation of anomaly scores, enabling a more balanced modeling of both anomalous and normal patterns. Experiments on four public datasets, compared against 12 advanced methods, demonstrate that BLTranAD achieves superior performance in time series anomaly detection.