Research On Multivariate Time Series Anomaly Detection Method Based On Deep Neural Networks

With the development of sensors and other hardware technologies,more and more reliable time-series data can be collected,and time-series anomaly detection is an important task to find problems in time and avoid risks.It is not easy to establish a multidimensional time series anomaly detection system because the collected data not only has information of different dimensions and characteristics but also has horizontal and vertical connections among the data.Besides,the observation series may be asynchronous due to the different frequency of data collected by each sensor.Due to equipment failure or other reasons,the data also has the problem of missing observation values.In addition,the anomaly has no clear boundary and there is no unified definition of a multivariate time series anomaly.Focusing on the research of anomaly detection of sensor multivariate time series,this paper proposed missing value interpolation algorithm and anomaly detection algorithm of multivariate time series based on deep neural network respectively and compare the performance of multiple groups of experiments with other algorithms.Experimental results show the advantages of the proposed algorithm.Finally,a realtime anomaly monitoring and alarming system is designed to demonstrate the practicability of the proposed algorithm,and the system can better meet the needs of users.In recent years,deep learning has shown outstanding advantages in dealing with multidimensional time series.This paper proposes a definition of point anomalies in multivariate time series and an unsupervised deep learning method,the multilayer convolutional recurrent autoencoded anomaly detector(MCRAAD),which is used to detect anomalies in multivariate time series.The model obtains correlation through a feature matrix based on a sliding window,extracts information of the feature matrix sequence with a multi-layer convolution encoder,obtains spatial-temporal information of the feature matrix from the Conv LSTM cell,reconstructs feature matrix sequence with convolution decoder,and predicts its self-feature matrix.In addition,a threshold setting method is proposed to assist in anomaly determination.In this paper,the model is tested on a synthetic dataset and a real-world dataset,and the results show that the proposed method is superior to the comparison model in both detection capability and robustness.This model also provides an effective method for anomaly detection of multivariate time series in the real world.Imputing appropriate values according to the missing pattern of a dataset is a technique for dealing with missing data.In this paper,a method named Multivariate Time Series Missing Value Imputation End-to-End Method Based on GAN(MIEGAN)is proposed,which uses the antagonistic training of the generator and discriminator to train the optimal generator.The generator consists of a denoising autoencoder containing improved GRU units.The autoencoder can generate simulated fake data following the actual distribution.This improved GRU unit adds a decaying gate to extract spatial-temporal features and time-missing pattern features from multivariate time series with missing values.This model uses a synthetic dataset and a complete real-world dataset to verify our model's imputed missing data performance under various missing rates and asynchronous conditions.Experimental results show that the proposed method outperforms the comparison model in imputing missing data of multivariate time series.In this paper,a real-time anomaly monitoring and alarm web system is designed by integrating the missing value processing algorithm and anomaly detection algorithm of multivariate time series.The system can display the changes of sensor data in real-time and send a reminder to users in time when the sensor system is abnormal and provide abnormal analysis results to help users find anomalies quickly.