Fault Detection And Diagnosis Based On Kernel Slow Feature Analysis

With the development of computer science and information technology, and also with the expanding production scale, the modern industrial control system has become increasingly complicated, integrated and intelligent. A large number of produced sensor data is stored during the industrial processes. It is of great significance to effectively extract the underlying important information about the current operation status of the system from the data, as it can be utilized for process monitoring and to improve the production efficiency and quality. In recent years, the data-driven based fault detection and diagnosis technology has been highlighted widely and is the important topic in the field of process monitoring.Slow feature analysis(SFA) is a new method for feature extraction and dimensionality reduction. It draws the most slowly varying components from temporal input sequences, which are called slow features. In essence, SFA projects the input signal into the high dimensional feature space through a nonlinear expansion and then searches for the optimal linear combination in the feature space, thus producing the slow features. Slow features are the high-level representation of the system information abstracted from original input signal and can characterize the inherent properties of the system. Therefore, SFA has great potential to dig the true patterns of the industrial processes. In this article, attempts are made to apply SFA to fault detection and diagnosis.Specially, the main contents of this article are as follows:(1) SFA, as a novel method for feature extraction, is introduced into the field of fault detection and diagnosis in order to explore its application value and potential in process monitoring. SFA learns the invariances from temporal sequence signals. The learning of invariances is very useful for data analysis and pattern recognition. As for temporal sequences, invariance means the extracted most slowly varying components, which reveal the inherent properties of the system.(2) The nonlinear expansion stage of SFA is realized with the help of kernel method, resulting in KSFA. The complete fault detection model based on KSFA is proposed. The 2S statistic and SPE statistic are constructed. The KSFA-based fault detection model is proved to be valid through the simulation experiments on the platform of TE process.(3) The fault diagnosis model based on KSFA-SVM is proposed. This model combines the advantage of KSFA and SVM. In other words, it can on the hand catch the intrinsic characteristic of the system, and on the other hand inherits the excellent classification performance of SVM. Simulation experiments executed on the TE process platform verify the good performance of KSFA-SVM diagnosis model under condition of single fault and multiple faults respectively.