Multi-fidelity sub-label-guided transfer network with physically interpretable synthetic datasets for rotor fault diagnosis

Abstract

The development of deep-learning-based diagnostic models for prognostics and health management relies on data from the target system. However, when the system does not exist (e.g., during the design phase), acquiring health data is impossible, making accurate fault diagnosis models difficult to develop. To resolve this challenge, this study proposes a fault diagnosis framework for rotating machinery that uses physically interpretable synthetic data generation and sub-label-guided transfer learning. An empirical model based on domain knowledge is used to generate large quantities of low-cost, low-fidelity data, although system configuration and expected operating condition information are incorporated into multi-body dynamics simulations to create a small amount of accurate, high-fidelity data. Low-fidelity data are labeled with a primary label that represents dominant fault modes and with sub-labels corresponding to detailed categories of faults, whereas high-fidelity data are assigned only a primary label for the dominant fault mode. Multi-fidelity data are preprocessed to create two-dimensional orbit plots and one-dimensional handcrafted features, which are then used as inputs for the transfer learning model. To effectively train the model, while considering the correlation between multi-fidelity data, this study designs a sub-label-guided transfer network (SlgTN) that combines sub-label guidance with transfer learning. In the pretraining process, feature learning is performed using low-fidelity data, with two parallel classification layers designed to effectively learn both the primary label and the sub-label. Center loss and cross entropy loss are incorporated. Experimental results from applying the method to a Bently-Nevada rotor kit demonstrate the effectiveness and validity of the proposed method. © 2025 Elsevier Ltd