Research On Adaptive Time Series Anomaly Detection Algorithm Based On Multimodal Adversarial Learning

Anomaly detection,also known as outlier detection,is a detection process to find objects whose behavior differs greatly from the expected behavior.Time series anomaly detection is one of the important research directions in the field of machine learning.However,most of the existing methods for timing anomaly detection are single-mode learning,ignoring the correlation and complementarity of different feature distributions of timing information in multimodal space,and failing to make full use of the existing information for effective pattern mining,resulting in poor detection effect and other problems.For this reason,this paper proposes a multimodal anomaly detection method for time series to realize the joint learning of temporal and frequency domain characteristic distribution of time series information.To take advantage of the characteristics correlation of temporal information on the different space,this paper proposes a multimodal adversarial anomaly detection framework(MATGAN).It can excavate the relevance of the distribution of time series information in different feature spaces through multi-modal method,and adapting the best time series information to different feature dimensions through multimodal adaptive time series encoding framework(MATE).Finally,through the generative adversarial network(GAN)jointly learn the distribution of the time series in the two feature spaces of the time domain and the frequency domain,the multi-modal feature representation learning is carried out.A multimodal adversarial learning adaptive time series anomaly detection algorithm is proposed.The MATGAN algorithm is an anomaly detection method based on reconstruction errors.It can use the excellent feature representation and reconstruction ability of the multimodal feature representation framework to measure the anomaly patterns of features from two modal spaces,so as to obtain more excellent anomaly detection performance.The experimental results of six real data sets in the time series data sets UCR and MIT-BIH show that the performance of MATGAN in anomaly detection task is 10.21%and 15.87% higher than that of the traditional monomodal anomaly detection method in terms of AUC and AP,respectively,which proves the effectiveness of the proposed method.