Time Series Anomaly Detection And Diagnosis Method For Industrial Internet Of Things

The advancement of the Industrial Internet of Things,big data,and artificial intelligence has led to the digitization and intelligent upgrade of traditional industries.As a result,a large amount of time series data has been generated.Analyzing this data can provide accurate and reliable information services and serve as a decision-making basis,facilitating the transformation and upgrade of the manufacturing industry.Anomaly detection and diagnosis have become essential tools due to the complexity of the industrial Io T infrastructure,ensuring its effectiveness.However,as the amount of data increases,multivariate time series can exhibit varying correlations among their dimensions.This requires a deeper exploration of the interactions and relationships among the variables in the data.Additionally,due to the significant time covered by the data,identifying long-term contextual dependencies has become a critical factor in temporal modeling.Furthermore,few models account for the reasons behind the occurrence of anomalies.Therefore,this thesis aims to address some of the shortcomings of existing models for multivariate time series anomaly detection and diagnosis through the following research work.(1)To address the challenges of capturing variable interrelationships and longterm timing dependence in multivariate time series data,this thesis proposes a novel multivariate time series anomaly detection model based on an attention mechanism.The model comprises three modules: the multivariate interaction module,the time series modeling and training module,and the anomaly scoring and detection module.The multivariate interaction module utilizes a graphical attention network to model the feature correlation and temporal correlation among multivariate.The time series modeling and training module employs a Transformer fusion generating adversarial network to capture the long-term correlation in the sequence data.Finally,the anomaly scoring and detection module leverages an anomaly score to detect the occurrence of anomalies.(2)In response to the issue of analyzing the causes of multivariate time series anomalies,this thesis aims to identify specific dimensions where anomalies occur to locate the causes of anomalies.A dynamic anomaly diagnosis algorithm based on causal relationships has been developed,which uses an anomaly detection model as its foundation and consists of three sub-modules: the dimensional score sub-module,the threshold diagnosis sub-module,and the anomaly diagnosis sub-module.Firstly,the anomaly degree score is calculated by combining multi-dimensional scores.Then,an automated threshold-setting method is used to obtain the threshold for anomaly determination,instead of relying on threshold enumeration.Finally,by ranking the dimension-level anomaly scores over the anomaly duration based on causal relationships,the top-ranked dimensions are returned.(3)In response to the application scenarios of multi-sensor fusion in industrial Io T,a time-series anomaly detection and diagnosis system was designed and implemented.This system integrates multi-variable time-series anomaly detection and diagnosis models,which enable real-time monitoring of sensor equipment status through a data acquisition module.Initially,the system pre-processes monitoring data,subsequently employing preprocessed training data to train the anomaly detection and diagnosis model.Ultimately,the system detects anomalies and presents the detection results,providing vital support to management and maintenance personnel.