A physics-informed feature weighting method for bearing fault diagnostics

Abstract

Intelligent bearing diagnostics has gained popularity over the last few years. However, most of the diagnostic methods are developed under the assumption that training and test data sets are collected under the same working conditions. This assumption is rare in practical scenarios because rotating machinery usually works under wide ranges of rotational speeds and loads. As bearings work under complex and time-varying operating conditions, the test data might come from a data distribution outside the training distribution. Purely data-driven diagnostic models often cannot provide reliable classifications for out-of-distribution test data. To tackle this challenge, this paper proposes a physics-informed feature weighting method for bearing diagnostics. First, a signal processing step is proposed that leverages physical knowledge of bearing faults to extract discriminative features that are robust to bearing speed variation. Then, a novel physics-informed feature weighting layer is developed to assign higher weights for features located closer to bearing fault characteristic frequencies. The feature weighting layer enhances the model’s sensitivity towards the fault-related features among the speed invariant features. Through experiments on three bearing datasets, the effectiveness of the proposed method is validated and shown to have promise for bearing fault diagnostics under different operating conditions. This study also details the deployment of a physics-informed convolutional neural network model on an Industrial Internet of Things (IIoT) device, where edge computing gives users a real-time evaluation of bearing health.