Application Of Deep Learning In Time Series Anomaly Detection

Time series is a series of numbers determined by the research object in chronological order under the same standard,which reflects the regularity of the research object about time.In real life,time series data can be seen everywhere,and its research and application has developed and gradually become a subject.In many research directions of time series,anomaly detection is very hot in recent years.Due to a variety of uncontrollable factors,the observations of time series data may have various types of anomalies.The occurrence of anomalies will not only affect the data analysis,but also cause major harm to production,life and property.Therefore,how to detect anomalies in time series efficiently and accurately is a major challenge at present.In previous studies,many classical models and algorithms have been successfully used in time series anomaly detection.However,with the rapid development of big data and informatization,these classical models still have some limitations in the demand for high-performance computing and large amount of data.New technologies was born with the rapid development,and the deep learning technology of artificial intelligence provides new ideas for time series research.The study mainly studies the application of deep learning method in time series anomaly detection.Firstly,for solving the information loss of the original sequence model LSTM(long short term memory network)in time series modeling,original LSTM is appropriately improved to make it more suitable for time series.Then,the generation model VAE(variational auto encoder)is integrated into the sequence model LSTM,and a anomaly detection algorithm(lstmvae)combining sequence modeling and sequence generation is proposed based on the idea of sequence reconstruction.Through the experimental analysis of simulated data and real data,the practicability and effectiveness of the two algorithms are tested.The final experimental results show that S-LSTM and LSTM-VAE not only have excellent detection performance,but are suitable for large-scale data.