Time Series Anomaly Detection Based On Active Learning And Variational Auto-encoder

Anomaly detection is designed to find instances that do not match data patterns because anomalous data often contains important information.For traditional time series anomaly detection algorithms,for the data has high dimensional characteristic,part of the information often lost during the dimensionality reduction,causing detection errors.For the more popular anomaly detection methods based on deep learning in recent years,th ese semi-supervised methods require a training set containing only normal samples.However,the real data is a mixture of normal and abnormal with the anomalies have few and different characteristics.In the process of constructing the training set,manual labelin g is costly,and it is easy to mix noise samples,which reduces the accuracy of the model.In view of the above problems,this thesis proposes an active anomaly detection framework based on Variational Auto-Encoder(AL-BLVE).The framework is based on the pool-based batch learning strategy.In the mixed sample pool,a batch of minimum entropy samples is found through the improved uncertainty-based sampling strategy based on trend representation.And then anomaly time series and subsequences are found based on the BLVE classifier.Aiming at the above problems,this thesis proposes an active anomaly detection framework based on Variational Auto-Encoder(AL-BLVE).The framework is based on the pool-based batch learning strategy.In the mixed sample pool,without expert,a batch of minimum entropy samples is found through the improved uncertainty-based sampling strategy based on trend representation.At the same time,samples are generated by the improved Variational Auto-Encoder to find the abnormal time series and sub-sequences.For the sampling strategy module,this thesis proposes a similarity sampling strategy based on trend representation.The strategy uses a binary string to record the relative trend and shape characteristics of the local time and replaces the original PAA distance with the mean value.It is proved that the similarity method is the strict lower bound of the segmentation aggregation method PAA.In order to adapt to the anomaly detection scenario,according to the principle of minimum entropy,this method selects a batch of uncertain samples classified as high normal confidence,which can effectively reduce the number of training samples.As for the base classifier module,under the structure of the Variational Auto-Encoder,the BILSTM-VAE-ENCODER(BLVE)model is proposed in order to adapt to the time series.In order to adapt the model to time series,a bidirectional long-term and short-term memory network was added to the codec.In order to reduce the reconstruction errors and ensure the distribution of latent space,the reconstructed encoder is used to re-encode the generated samples to obtain a new latent space,and constrain it to the original difference between spaces.In the calculation of the abnormal score,in order to avoid the influence of noise on the pixel-level distance calculation,this paper uses the mean square error of the latent space.In this thesis,experiments are carried out on the UCR and UCI datasets.The results show that with the AUC as the evaluation index,the accurac y of similarity sample selection is increased by about 3% on average compared with the baseline model,and it is not sensitive to the size of the time window.At the same time,compared with the baseline method,the variation detection method based on the Variational Auto-Encoder(BLVE)has a significant performance improvement of about 9%.By mixing the samples for detection,the AL-BLVE method can achieve the results of BLVE model.