Research On Anomaly Detection Of Water Distribution System Based On Multivariate Feature Learning

The water distribution system is a typical industrial control system,and the operational data of the equipment in the system has multidimensional and temporal characteristics.The anomaly detection method for water distribution systems refers to the method of constructing a normal behavior model of system operation based on multidimensional temporal state data,and detecting abnormal behavior from it.Existing anomaly detection methods can be divided into methods for detecting anomalies in static data and methods for detecting anomalies in concept drift data streams,based on the characteristics of changes in system state data.On the one hand,the current deep anomaly detection methods for static data suffer from insufficient exploration of the correlations between multidimensional time series,especially for the hidden nonlinear relationships between data with stronger dynamics and higher dimensions,making it difficult for the model to make accurate judgments on anomalies.On the other hand,the deep anomaly detection methods for concept drift data streams suffer from issues such as delayed model updates,lagging of the new model leading to efficiency decay,all new data being used for training leading to model suboptimality,and self-poisoning during self-updating of unsupervised anomaly detection models due to the lack of labels.To address the above issues,this thesis conducts research on multidimensional time series anomaly detection methods for both static data and concept drift data streams in water distribution systems.Two multidimensional time series anomaly detection methods are proposed: one is an anomaly detection method based on threedimensional residual convolution and Transformer,and the other is an anomaly detection method based on knowledge distillation.The main research work of this thesis includes the following aspects:(1)To address the issue of insufficient exploration of correlations among multidimensional time series,this thesis proposes an anomaly detection method based on three-dimensional residual convolution and Transformer,which improves the accuracy of multidimensional time series anomaly detection.This method uses a threedimensional convolutional network to capture the temporal dimension information and the interaction information between features of multidimensional time series,and integrates the captured information into the Transformer architecture to further learn global dependencies.It achieves the exploration of potential relationships from local to global,thereby improving the accuracy of multidimensional time series anomaly detection.(2)To balance the efficiency of model updates and detection performance,and reduce the problem of self-poisoning during the model’s self-updating process,this thesis proposes a knowledge distillation-based anomaly detection method to solve the concept drift problem faced by the water distribution system.This method pre-trains the teacher model on a large historical dataset and guides the student model to train on new data,taking advantage of the student model’s smaller parameter size and faster training speed to improve training efficiency.A dynamic algorithm for adjusting model parameters is designed to adjust the weights of the student model’s parameters through local inference calculations on new data,in order to improve the model’s responsiveness to new distribution data.In addition,a one-class support vector machine is used to eliminate outliers and reduce the impact of self-poisoning on the anomaly detection model.