An architecture for a diagnostic/prognostic system with rough set feature selection and diagnostic decision fusion capabilities.

摘要

Fault detection and identification (FDI) and failure prediction (FP) are two crucial parts of industrial health monitoring systems. Therefore, various techniques have been developed for those purposes. Among them, approaches to discover the relationships between input conditions and the corresponding abnormalities have been gaining more attention than ever before. However, the methods reported over the past decade mainly focused on specific applications. A general framework for FDI and FP still has not been formulated. Therefore, this research aims, by applying data mining techniques, at providing a systematic framework to identify the most relevant input features from a set of predefined features that correspond to a specific abnormality. In addition, diagnostic rule generation methods are investigated within this framework.; Diagnostic decision making is another focus. Recent research has demonstrated the advantages of using multiple diagnostic sources instead of depending on a sole source. Combining the diagnostic information from multiple sources can enhance diagnostic decisions. Among such fusion methods, two frequently adopted techniques are weighting fusion and Dempster-Shafer evidential theory. Each method, however, has its own disadvantages and advantages. Therefore, an innovative and generalized combination method taking only the advantages of each technique is introduced.; Specifically, the major contributions of this research includes the followings: feature preparation methods to obtain potential features from raw data, rough set based feature selection methods for FDI and FP, diagnostic rule generation using rough set methods to provide the structure of a classifier, a classification tool named Arrangement Fuzzy Neural Network Classifier (AFNNC) to increase the flexibility of the diagnostic module design, and an innovative diagnostic decision method based on Dempster-Shafer evidential theory and weighting fusion technique to increase diagnostic accuracy.; To demonstrate the feasibility of the methodology in practical use, the proposed methods are applied to three different applications: a Navy chiller system, Process Demonstrator, and an automotive backlight inspection system. The first two examples show how the methodology could be applied to industrial processes, and the last one exemplifies the availability of the methodology in image-based inspection areas. Consequently, the application results demonstrate the feasibility of applying the proposed methods within the industrial arena.