Anomaly Detection Of Multivariate Time Series Data Based On Representation Learning

Sensors in many real-world systems such as smart buildings,factories,power plants and data centers generate a large amount of multidimensional time series data.Time series anomaly detection prevents losses by continuously monitoring the real-time health status of these systems.Multivariate time series data has more complex nature compared to single variate time series,and detecting anomalies among multivariate time series is a challenging problem.1)Due to the dynamic complexity of multivariate time series data,traditional statistical-based anomaly detection methods can no longer meet the requirements.2)In addition to time-dependent relationships,relationships among multiple variables have to be considered.3)In addition,time series data is often sparsely labeled and the class imbalance of labels in the anomaly detection environment leads to difficulties in anomaly detection by supervised models such as deep classification models.Therefore,existing unsupervised learning methods perform anomaly detection by constructing predictive or reconstruction models on normal data.Reconstruction-based models have difficulty capturing small anomalies due to the nature of the structure,while existing prediction-based models have not yet fully utilized temporal dependencies and multivariate interrelationships for anomaly detection.Moreover,temporal relationships are often implicit and difficult to construct features by hand,so representational learning can be useful.The recent success of contrast learning for representation learning in computer vision has led to a series of work on contrastive learning-based temporal representation learning,however,existing contrastive learning-based time series representation learning focuses on learning inter-temporal correlations while ignoring temporal dependencies,and very little work is applicable to multidimensional temporal anomaly detection.To address the above challenges,this thesis proposes Clea Ano,a novel unsupervised multi-dimensional time series anomaly detection model based on contrastive representation learning.1)In this thesis,a multivariate time series anomaly detection model is constructed by combining the contrastive representation learning and time series prediction tasks by sharing the multi-task structure of the underlying encoder,and the model is jointly optimized by the contrast loss and the prediction loss,and the time series anomaly detection is performed based on the prediction error and threshold.2)In order to ensure that the input and output lengths are equal and can handle variable-length inputs,this thesis designs an encoder based on dilated causal convolution to initially extract representations.3)This thesis designs a contrastive representation learning framework based on the principle of context consistency to learn temporal dependencies and inter-temporal associations.In order to generate different contrastive views without introducing additional noise,we perform cropping and masking operations on the time series along the time dimension.4)In order to improve the model’s ability to model longterm temporal dependencies,this thesis introduces an attention mechanism and designs a prediction branch based on multi-head attention.5)Finally,this thesis conducts comparative experiments on six public real data sets.Compared with six multi-type baseline models,the Clea Ano model ranks first in the F1 score.The experimental results show that the Clea Ano model has excellent robustness.for other baseline models.And each component and important hyperparameters of the model are analyzed through ablation experiments and parameter sensitivity experiments,and the results further prove the effectiveness of the model design.