Deep Generative Model-based Time Series Anomaly Detection

Time series anomaly detection is referred as detecting anomalies from normal values giving observed time series.Time series anomaly detection,e.g.anomaly detection for time series data,is of great significance both in academia and industry.Time series anomaly detection can automatically detect anomalies from a large number of Internet company business or industrial machinery monitoring data,and raise alarms to avoid unexpected losses.Existing time series anomaly detection algorithms mainly include feature extraction and end-to-end methods,but they have serious problems:(1)For feature-based methods,a large number of features are extracted.This causes the problem of extremely high feature dimensions;(2)For end-to-end anomaly detection,autoencoders are used to detect anomalies.However,existed methods ignore missing values in the data.Plus,they cannot utilize the existed prior labels in real-world applications;(3)The reconstruction task is ineffective for local and contextual anomalies.In addition,due to the powerful capabilities of neural networks,it is easy to reconstruct anomalous samples as well.In order to tackle the above issues,this paper focuses on unsupervised time series anomaly detection.First,for problem(1),we propose an anomaly detection framework based on knowledge distillation.Our framework consists of a teacher network and several student networks.They can all be regarded as encoders and have the same structure.The judgment of anomaly is all performed in the space after dimensionality reduction,which avoids the problem of too high feature dimensionality.We first use the self-supervised objective function to pre-train the teacher network,and then train the student network to approach the output of the teacher network(that is,use the knowledge learned by the teacher network to "teach" students).Since the student network has not been trained on anomalous samples,for anomalous samples,not only the output of teachers and students are very different,but the output between different students will also be very different.Therefore,the abnormality can be judged based on these two angles.Secondly,for problem(2),we build a model based on the adversarial autoencoder,and propose a contrast reconstruction loss and training filling mechanism to solve the problem of using a small number of prior labels and missing values,respectively.The contrast reconstruction loss can reduce the correct reconstruction of abnormal samples compared to the original reconstruction loss.The training filling mechanism will use the strong generation ability of the confrontation generation network to continuously supplement the missing values during the training process.Finally,for problem(3),we use additional prediction modules and memory modules to solve the problem of insufficient detection of local anomalies in reconstruction tasks and easy overfitting.The newly added prediction module uses an additional prediction task to predict the value of the next time step by using existing observations,thereby achieving a better detection effect for local anomalies.In addition,the memory module explicitly models the normal mode to ensure that the decoder only reconstructs the vector in the memory module,so that the probability of good reconstruction of abnormal samples is greatly reduced.In order to verify the effectiveness of the model,we conducted extensive experiments,and the experimental results show the effectiveness of our method in time series anomaly detection.