Research On Anomaly Detection Method Of Time Series Based On Multi Modal Attention

Anomaly detection,also known as anomaly mining or outlier detection,aims to identify and detect data that are different from some features of normal data from a large number of data.Time series anomaly detection is to detect a small number of points with outliers,mutations and other anomalies in time series data.It is an important research direction in data mining field.The existing single-mode time series anomaly detection methods usually only analyze the time series features,ignoring the correlation and complementarity of different features of timing information in multimode space,resulting in an unsatisfactory anomaly detection results.This paper proposes a time series anomaly detection method based on multimodal attention mechanism.In order to use the features of multiple modes of timing information,this paper proposes a Multimode Joint Learning Model（MJLM）.Two encoder structures are used to learn timing and noise information of time series,and the low dimensional embedding of the two modes is fused as the input of decoder.The cross-modal interaction learning between timing and noise signals is realized during the reconstruction of timing mode and noise mode.Finally,the reconstruction error of original data and reconstructed data are reduced by multiple iterations,so that the reconstructed features resemble the original features to the maximum extent possible.Considering the feature information cannot be extracted effectively in the process of multimodal information fusion,this paper introduces Multi Modal Attention（MMA）for the fusion of two modes and effective extraction of fusion features.The attention mechanism can obtain valuable information for anomaly detection,and suppress worthless information.A Time Series Anomaly Detection Algorithm Based on Multi Modal Attention（TSADM）is proposed,and the algorithm can extract multiple correlation features of time series from multimodal perspective in an unsupervised manner,so as to obtain more excellent anomaly detection performance than traditional methods.The experimental results of five real data sets in time series data sets UCR show that the performance of TSADM in anomaly detection task is 9.27% and 2.82% higher than that of the traditional single mode anomaly detection method in terms of AUC and AP respectively.The results show that the proposed anomaly detection algorithm can achieve more effective anomaly detection than the traditional single mode anomaly detection algorithm.