Research On Sequential Anomaly Detection Based On Deep Neural Network

Anomaly detection has been a widely studied issue in applications such as system health management,intrusion detection,healthcare,bioinformatics,fraud detection and mechanical fault detection.In real life,there is a large amount of data with sequence characteristics.Due to the diversity of application scenarios of sequence data,the problem of anomaly detection of sequence data is a very high value application.Anomaly detection of sequence data is a very challenging research question for two reasons.First of all,the anomaly data is difficult to obtain,under normal circumstances we can only get a relatively small number of anomalous data instances.Second,the sequence characteristics of the data are hard to be learned by the model.Deep neural network is a general term for neural networks with many hidden layers.In recent years,it has made significant breakthroughs in applications such as computer vision and speech recognition.Among them,the Autoencoder has been widely used in a variety of dimensionality reduction and denoising because of its good non-linear feature extraction ability.Another kind of similar structure,as a representative of the deep generation model,Variational Autoencoder,achieved good results in text generation,image style migration and other fields in unsupervised learning and semi-supervised learning scene,and the recurrent neural network is widely used in various types of sequence data modeling.Combined with the above analysis,we can solve the problem of sequential anomaly detection by constructing a model based on deep neural network.The main work of this paper includes the following points: Analyze the basic neural network architectures of Autoencoder,Variational Autoencoder,Recurrent neural network,Long short-term memory,and the latest proposed model named Variational Recurrent neural network;An Autoencoder Recurrent neural network model was established based on Autoencoder and Long short-term memory.Based on the above,a new algorithm of sequential anomaly detection was proposed.By using the semi-supervised learning method and Variational Recurrent neural network model,A new algorithm of sequential anomaly detection was proposed.In the proposed algorithm,a KL loss annealing method and sequence anomaly score integration method are proposed to improve the anomaly detection effect.In this paper,we use some real-world datasets that have been used in the study of sequential anomaly detection.We conduct a large number of parameter sensitivity tests on the proposed algorithm and select some baseline anomaly detection algorithms for experimental comparison.We find that the proposed algorithm is based on some data sets can exceed the detection results of these baseline anomaly detection algorithms,The new sequential anomaly detection algorithm proposed in this paper is used to solve the feasibility of the universal sequence anomaly detection problem,and through the experimental analysis,the advantages and limitations of the algorithm are analyzed.