Dominant Feature Identification for Industrial Fault Detection and Isolation Applications

Fault Detection and Isolation (FDI) is crucial to reduce production costs and down-time in industrial machines. In this paper, we show how to find a reduced feature subset which is optimal in both estimation and clustering least square errors using a new Dominant Feature Identification (DFI) method. It is shown how to apply DFI to fault detection by two methods that seek to identify the important features in a given set of faults. Then, based on the determined reduced feature set, a Neural Network (NN) is used for online fault classification. The DFI technique reduces the number of features and hence potentially the number of sensors required, and the NN allows reduction in the required signal processing for multiple fault prediction in the proposed two-stage framework. Our experimental results on an industrial machine fault simulator show the effectiveness in fault diagnosis and classification. Accuracy of 99.4% for fault identification is observed when using proposed new DFI followed by NN classification, reducing the number of required features from 120 to 13 and the number of sensors from 8 to 4. This translates to significant cost savings and prerequisites for next generation of intelligent diagnosis and prognosis systems.