Bond Graph-CNN based hybrid fault diagnosis with minimum labeled data

Abstract

Fault Isolation is a critical step in any fault diagnosis method, and the difficulty increases with the complexity of the system. When there are not enough sensors in the system, the traditional model-based methods have trouble isolating the fault. The current data-driven FDI techniques generally emphasize accuracy and rarely draw attention to the lack of readily accessible labeled data in the industry. This study aims to develop a hybrid fault diagnosis method by combining the well-established graphical technique of Bond-Graph (BG) with the powerful pattern recognition ability of Convolutional Neural Network (CNN) to improve the overall fault isolation performance. A new formalism named BG-CNN method is proposed, which can utilize the residuals generated from the BG model in a CNN for improved fault isolation with a minimal number of labeled data. The single incipient faults as well as multiple simultaneous faults can be isolated by this method. The BG-CNN method demonstrates a high level of performance for the FDI of a Direct Current (DC)-motor with a relatively small number of labeled samples. In comparison, the traditional CNN method using raw sensor data requires a significantly larger number of labeled samples to achieve a similar level of performance.