Study On Methods For Fault Diagnosis Based On Fisher Discriminant Analysis

As the automation level in modern industry improves, the productive process is becoming more sophisticated. The traditional fault detection and diagnosis methods are no longer applicable. In order to timely and precisely diagnose the fault and avoid losses in lives and properties and protect environment away from contamination, it is urgent to propose new and effective theories and methods for fault diagnosis. Fault diagnosis methods base on data- driven are effective in large-scale industrial process, and they do not require precise math model. Only through analyzing process data by multivariate statistical theory can it provide useful information for fault detection and diagnosis. This paper combines Fisher Discriminate Analysis (FDA) with Kernel Function and ICA respectively and proposes two improved FDA methods for fault diagnosis.Contents and acheivements of the paper are as follows:1. Concepts and common methods for fault detection and diagnosis are illustrated, and the development status of multivariate statistics based fault detection and diagnosis is also summarized.2. Basic theories of multivariate statistics based fault detection and diagnosis are illustrated. PCA and FDA are discussed in detail, and their applications on fault detection and diagnosis are also briefly introduced.3. In order to lower the operation amount in FDA, an improved Kernel based FDA is proposed. Feature Sample Selection works well in reducing computation of FDA. Cosine Kernel Function is applied to enhance the performance of original Polynomial Kernel Function and Nearest Feature Line Classifier is used to improve the classifying accuracy of model.4. For the assumption in FDA that the data follow Gaussian Distribution limits FDA's application, a mixed model fault detection method is presented. First, Independent Component Analysis is used to extract independent components from observed variables, then Feature Scaling Kernel FDA is applied to diagnose faults. This method effectively reduces mistakes in fault classification and improves fault diagnosis performance.