Research On Fine-Grained And Fast Multivariate Time Series Anomaly Detection Algorithm

Nowadays,the amount of data processed by large-scale Internet online service systems and IoT systems are becoming more and more complex.Multiple monitoring data in these systems constitute a multivariate time series,also known as multivariate time series.When there are loopholes and threats in the system,the multivariate time series will show an abnormal state.If these abnormal states are detected,measures can be taken to avoid the loss of resources and property effectively.Therefore,multivariate time series anomaly detection(MTSAD)become more and more important.However,the high dimensionality and complexity of multivariate time series data make data maintenance and anomaly detection difficult.In practical applications,MTSAD still has two problems.First,the current main work is to detect abnormal timestamps,but it is actually necessary to determine in detail which indicator on which timestamp is abnormal,which is called fine-grained anomaly detection(FGAD).There is currently less work in this area,and although some work mentions that it can be further extended based on the results of timestamp anomaly detection,none of them perform the quantitative evaluation.Second,most of the existing MTSADs use methods based on deep learning and machine learning,and the training time is long,and after the environment changes,new models need to be retrained for new data,which will greatly increase the time.Therefore,this paper studies the problems of fine-grained detection in multivariate time series,and the detection time is too long and the adaptability to environmental changes is not enough.Aiming at the problem of fine-grained detection in multivariate time series,this paper first verifies that extended methods for timestamp anomaly detection cannot work well in finegrained detection tasks due to the lack of spatial dimension anomaly features.Then,a Multivariate Time Series Fine-Grained Anomaly Detection(MFGAD)framework is proposed,which can accurately locate the fine-grained anomaly by detecting timestamps and index anomalies by performing temporal and spatial feature extraction respectively..Based on this framework,this paper designs an index detection algorithm based on graph attention neural network timestamp and attention convolutional long short-term memory network,and combines the two results for fine-grained anomaly detection.Extensive experiments on the SWaT(Safe Water Treatment)real dataset show that our algorithm can achieve higher F1 scores.Aiming at the problems of long detection time and insufficient adaptability to environmental changes in multivariate time series,this paper proposes a fast multivariate time series anomaly detection(FMAD)algorithm based on matrix filling.The input matrix data is processed by clustering and anti-abnormal sampling method,and then the sampled matrix is filled and restored by the matrix filling method to reconstruct the input matrix,and the error after reconstruction can be judged to identify anomalies.The reconstructed multivariate time series implicitly inherits the normal behavior without involving any explicit learning.The algorithm effectively improves the detection speed of the anomaly detection model for multivariate time series and has good performance on multiple service system datasets.