Research And Application Of Anti-noise Anomaly Detection Method For Time Series Data

Data analysis and mining is the process of analyzing information and knowledge that is difficult to discover directly from enormous,mixed and noisy data.Anomaly detection is one of the research areas of data analysis and mining.With the rapid development of modern industry,intelligent detection and monitoring technology have been widely used in industrial systems,resulting in an enormous amount of time-series data.Anomaly detection based on deep learning plays an important role in the safety and stability of application systems.Noise on the attributes and labels of the data is inevitable due to the capability limitations of data collection equipment,the difficulty of labeling massive amounts of data and other factors,which brings great challenges to the task of anomaly detection based on deep learning.Based on deep learning theory,it is of great theoretical significance and practical value to study robust anti-noise anomaly detection method and build an anomaly detection system for real-world time-series detection data.This paper conducts a deep study on the anti-noise anomaly detection of time-series data.Firstly,it conducts a study on the robust learning method for the noisy label problem in the data,then conducts a study on the anti-noise anomaly detection for the data attribute noise problem,and finally develops an anomaly detection system for dynamic inspection data by integrating the research achievements on noisy label and attribute noise for the characteristics and anomaly detection needs of railway dynamic inspection data.The main research content of this paper are as follows:(1)For noisy label issue,a novel robust learning paradigm called joint training by combining consistency and diversity(Jo Ca D)is proposed in this paper.The Jo Ca D is devoted to maximize the prediction consistency of the networks while keeping enough diversity on their representation learning.Meanwhile,aiming to reconcile the relationship between consistency and diversity,an effective implementation is proposed which dynamically adjusts joint loss.Comparative experiments are conducted on the publicly available datasets MINIST,CIFAR-10 and CIFAR-100 under four noise conditions,as well as on the publicly available real-world dataset Clothing1 M and the time-series dataset UCR.The results demonstrate that the proposed method achieves the best classification accuracy compared to all the comparative methods,and achieves a smoother classification performance especially in the presence of elevated noise.(2)In order to address the issue that time-series data in industrial scenarios usually contain noisy on attributes and sparse labels,an unsupervised anomaly detection algorithm based on dual graph deviation networks(DGDN-USAD)is proposed in this paper.The method is divided into two main phases,which are rough anomaly detection and label correction respectively.Firstly,a dual graph deviation network is trained using a joint training method to capture association relationships between attributes of normal time-series data to discriminate anomalies,and then label correction is implemented in a dual adversarial dynamic Shapelet network.Comparative experiments on two real-world datasets,SWa T and WADI,show that DGDN-USAD achieves the optimal recall compared to the comparative method,which validates the practicality of this method.(3)Railway dynamic inspection data is a special kind of time-series data,with multiple attributes,data containing noise,few labels and difficulties in annotation,etc.Based on the previous studies,this paper designs and develops an anomaly detection system for dynamic inspection data,taking full consideration of the actual needs of dynamic inspection data anomaly detection.The system embeds dual graph deviation networks with joint consistency and diversity(DGDN/Jo Ca D)and DGDN-USAD,with functional features such as model management,anomaly detection and data visualization,providing technical support for the intelligent analysis and application of dynamic inspection data.