Unsupervised Time Series Anomaly Detection Based On Generative Adversarial Networks

Time series anomaly detection can provide early warning for equipment failure,which is an important part of Prognostics and Health Management(PHM).Time series data obtained in practical applications are usually unsupervised.The traditional methods for anomaly detection cannot meet the increasing accuracy requirements,and the deep learning model is widely used.Generative Adversarial Networks(GAN)is a deep learning model which has been successfully applied to unsupervised time series anomaly detection.However,the existing GAN-based anomaly detection algorithms fail to take full advantage of the GAN model,resulting in insufficient detection accuracy.This thesis studies the accuracy improvement of GAN-based unsupervised time series anomaly detection algorithm.Two fusion anomaly detection models of optimized and improved GAN and Long Short-Term Memory Network(LSTM)are designed and verified by experiments.On this basis,the proposed anomaly detection algorithm is used for bearing fault early warning.The main achievements are as follows:(1)The GAN model uses Kullback-Leibler divergence to derive the objective function,which leads to the problem of mode collapse and vanishing gradient.Mode collapse leads to insufficient model diversity and vanishing gradients leads to under-fitting,which ultimately leads to decline in anomaly detection accuracy.Using Wasserstein distance to optimize the GAN objective function and improve it for anomaly detection can effectively reduce the above problems.In this way,not only the anomaly detection accuracy is improved,but also the difficulty of training and parameter tuning is reduced.(2)The existing GAN anomaly detection models usually use encoders or similar structures to establish reverse mapping to achieve anomaly detection,but such structures have a high loss of encoding information,resulting in insufficient accuracy.Therefore,a double GAN cyclic structure is proposed to replace the GAN with encoder structure,which effectively improves the accuracy of anomaly detection.(3)The original GAN anomaly detection model does not establish reverse mapping,but uses gradient sampling to achieve anomaly detection.This method has high accuracy but low efficiency.A method of clip search is proposed to improve the sampling process,which can effectively improve the efficiency of anomaly detection while retaining high accuracy.(4)LSTM can effectively capture strong time-dependent trends in time series.The fusion of LSTM and GAN can improve the efficiency of the model adapting to time trends,thus improving the accuracy of anomaly detection.NAB is a group of datasets used to test the time series anomaly detection algorithm.Five different datasets in NAB are used for experiments.The results show that the F1 score of the algorithm in this thesis has an improvement of at least 7.4% and an average of 12.6%compared with the other GAN-based and several classical anomaly detection algorithms.Furthermore,an anomaly analysis strategy is designed using confidence intervals and bearing vibration criteria.A bearing fault early warning method is constructed and verified on the XJTU-SY bearing dataset combined with the proposed algorithm.The work of this thesis provides a feasible and effective model for the problem of anomaly detection in unsupervised time series,and provides an efficient and accurate solution and idea for the problem of fault early warning.