Data-driven Based Multi-mode Anomaly Detection Methods

Due to the influences of factors, such as properties of raw materials, equipment abrasion and process loading, the complex industrial system has multiple stable operation modes, and the transition processes among the stable modes have obvious dynamic features. Most of the conventional multi-mode monitoring methods have ignored monitoring of the transition processes; under the condition of taking the transition processes into account, as effective feature extraction of the transition processes hasn’t been carried out, neither mode division nor multi-mode anomaly detection achieves satisfying effects. Besides, industrial process data mostly have the multi-scale features, and more studies on multi-mode anomaly detection under the multi-scale framework are required. This thesis adopts the data driving based method to study the above problems in order to facilitate safe and efficient operation of the multi-mode system. The main tasks are as follows:(1) By extracting the differential geometry features of the process data and describing the process dynamic features, this thesis put forward a differential geometry clustering algorithm based mode division method on the basis of augmented data.(2) Starting with the dynamic gradient feature of the transition mode data, this thesis studies the dynamic feature extraction technology and put forward the anomaly detection model of the transition modes on the basis of time-varying rolling balls to effectively conduct the transition mode anomaly detection.(3) Aiming at the multi-mode process provided with the multi-scale features, this thesis launches the multi-mode anomaly detection study under the multi-scale framework, establishes the stable mode multi-scale principal element analysis model and the transition mode multi-scale differential geometry model respectively, and ultimately realizes the multi-mode anomaly detection under the multi-scale framework.