Machinery fault diagnosis-oriented regularization for nonlinear system identification: Framework and applications

Abstract

In this article, we propose a novel framework for machinery fault diagnosis based on nonlinear system identification, called Identification for Fault Diagnosis (I4FD) The focus and necessity of the framework is that it can mitigate the effects of external environmental changes and enhance diagnostic accuracy. The framework integrates regularized data-driven modeling and frequency analysis. During the modeling process, prior physical knowledge about the diagnostic target is incorporated through a penalty parameter, leading to fault diagnosis-oriented regularization (FDoR). FDoR tailors the model specifically for fault diagnosis (FD) applications, offering new insights into FD-oriented system identification. The regularized NARX modeling in this paper does not end when a model is built by using information in a period of time, but uses the updated data for continuous dynamic modeling. After the model is identified, frequency analysis is then used to extract model-based features, which change significantly when faults occur. The effectiveness of the I4FD framework is demonstrated through simulations and real cases, highlighting its advantages over traditional methods and its industrial potential.