| Time series anomaly detection is a highly active research area,with a wide range of applications in finance,weather,wind power,and computer science,among others.However,there are still many challenges in detecting anomalies from large amounts of normal data,capturing complex real-world data distributions to detect abnormal data,and effectively detecting anomalies in the absence of labels.In current multi-dimensional time series anomaly detection tasks,there are still several shortcomings:1.Lack of effective methods to detect multiple types of anomalies in multi-dimensional time series ata;2.Lack of effective ways to capture complex real-world data distributions in order to detect abnormal data that does not conform to the distribution;3.How to solve the problem of not being able to detect subtle local anomalies when successfully capturing the data distribution.Based on these issues,this paper proposes a new multi-dimensional time series anomaly detection scheme,named TAGAN.In the anomaly detection method,based on reconstruction,the method is good at capturing the data distribution,while based on prediction,the method is better at detecting local subtle anomalies.Therefore,TAGAN combines the advantages of these two types of methods and uses a joint training approach of reconstruction and prediction to detect anomalies in multi-dimensional time series.In the reconstruction branch,TAGAN uses an adversarial training approach based on an autoencoder structure to learn the data distribution,and uses two different loss functions,the Wasserstein distance with a gradient penalty term and the Wasserstein divergence,during the adversarial training process.In the prediction branch,TAGAN uses an attention mechanism and a combination of Long Short-Term Memory network structures to predict time series data.Specifically,this paper’s research work mainly includes the following three aspects:1.An architecture based on the joint training of reconstruction and prediction networks is proposed for detecting time-dependent anomalies.The algorithm can effectively capture the dependencies between highdimensional data.2.Designing a new reconstruction network by combining autoencoder and adversarial training methods.Compared with six other anomaly detection algorithms,this new structure improves the search efficiency and overall performance of the generator(encoder+decoder).3.Introducing loss functions with gradient penalty terms for the Wasserstein distance and the Wasserstein divergence,which can alleviate pattern collapse in the generator and avoid gradient vanishing.In addition,the stability of the training process and the reliability of the algorithm are also enhanced to a certain extent.The performance of the TAGAN algorithm was tested on four realworld datasets(MSL,SMAP,SMD,and SWaT).The numerical experiments show that the proposed algorithm performs better than six baseline anomaly detection algorithms. |
