| Time series data usually contains some anomalies,which do not conform to expected patterns.Accurate anomaly detection with low time delay is helpful in extracting important information and handling anomalies timely.Anomaly detection for time series data is one of the important research issues in data mining,which has been widely used in network traffic anomaly detection,industrial fault detection,transaction fraud detection,etc.Anomaly detection methods for time series can be roughly divided into two categories:traditional proximity based methods and deep learning based methods.Aiming to detect time series anomalies,traditional proximity based methods generally capture temporal anomalies using the probability density ratio to measure the proximity between subsequences.Most of the deep learning based methods adopt recurrent neural network or long short term memory to capture temporal dependency,and use Auto Encoder to reconstruct data.Although these methods are effective to detect anomalies,they still have high time delay and low computational efficiency.To solve the above problems,this paper studies these two types of anomaly detection methods respectively.Our main contributions are as follows:（1）Traditional proximity based detection methods generally have high delay and computational complexity.To address these issues,this paper proposes an improved method（named as TVRSE-AD）based on total variation ratio separation distance,in which total variation is adopted to extract sequence fluctuation features.Due to the fact that the calculation avoids complex parameter estimation,TVRSE-AD achieves the higher computational efficiency and lower time delay.For reducing noise interference and improving the detection accuracy,the proposed method is further combined with the relative total variation.The experimental results show that TVRSE-AD performs well in terms of detection accuracy,low delay and computational efficiency.（2）The network of deep learning based methods contains many parameters resulting in slow computation speed.In order to improve the speed,SEAE-AD method based on serial Auto Encoders is proposed,which contains two Auto Encoders（AE1and AE2）with simple structure.The simple structure speeds up the training and inference.In addition,the output of AE1 is fed into AE2 to improve the decoding ability of the decoder of AE2.The way of serial training makes SEAE-AD achieve better detection accuracy.Experiment results show that the proposed method has better precision,recall,F1 score and faster inference than several state-of-art anomaly detection methods.TVRSE-AD achieves efficient anomaly detection by calculating the proximity between subsequences.With the increase of data dimension and volume,it becomes more difficult to describe the data distribution,resulting in lower anomaly detection rate and computational efficiency.Unlike this,SEAE-AD improves the performance effectively with the help of the strong feature extraction and data representation capabilities of the neural network.This paper provides two effective methods for anomaly detection,and provides some possible directions for future work. |
