Time series anomaly detection using LSTM and attention

Time series data anomaly detection is to identify observations or patterns in some chronologically ordered points which are significantly inconsistent with expected patterns or normal behavior. These patterns may indicate unexpected events, unusual behavior, failures, or other unusual conditions in the system. Time series anomaly detection has important applications in many fields, including industry, finance, medical care, etc. We designed a time series anomaly detection algorithm using LSTM(Long Short-Term Memory) and attention. By adding an attention layer after the LSTM network, the network can pay more attention to more relevant features in the input multivariate time series, thereby improving accuracy and recall. In the data preprocessing process, the decision tree algorithm is used on the original data set to remove features that have little impact on the anomaly detection results, so as to reduce the computational complexity of anomaly detection, and improve the efficiency of anomaly detection. Experiments on the industrial time series data set SWaT data set show that the LSTM-Attention model proposed in this article is better to the basic LSTM network in terms of precision, recall, F1score and other indicators, and achieves good time series data anomaly detection results.