A Study on Fault Diagnosis Incrementally Integrating Field Data from the Data-Limited Design Stage

Abstract

Rotating machinery plays a critical role across various industrial sectors, including power plants, manufacturing facilities, and process plants. However, due to their structural complexity and harsh operating conditions, faults can lead to the shutdown of entire systems. To prevent such failures, vibration-based fault diagnosis techniques have been actively studied. Nevertheless, ensuring the effectiveness of these techniques requires data covering diverse operating conditions and fault types. A key challenge is that, in the early stages of system design, experimental data are difficult to obtain, and the performance of existing diagnostic models may degrade as new operational data are collected during system operation. This study proposes an integrated fault diagnosis framework. It starts from the early design stage where data are unavailable and gradually incorporates field data collected during system operation. This enables continuous improvement in diagnostic performance. The study consists of three main research themes. First, a physics-based data generation method was developed using domain knowledge, geometry, and operating conditions. Low-fidelity data were generated using empirical equations. High-fidelity data were produced using rigid-body multibody dynamics (MBD) models. Additionally, a physical model was constructed using the finite element method (FEM), enabling the generation of high-quality datasets. Testbed experiments were used to calibrate and validate the models, ensuring the physical reliability of the generated data. Second, a transfer learning-based diagnostic model, termed Sub-label-guided Transfer Network (SlgTN), was proposed to effectively utilize the multi-fidelity data. The proposed SlgTN combines 1D handcrafted features and 2D orbit plots to simultaneously learn fault types and subtypes. To effectively utilize prior knowledge and a small amount of high-fidelity data, the model was fine-tuned using transfer learning. This tuning process significantly enhanced the accuracy of fault diagnosis. Third, an incremental learning framework was established to update the diagnostic model by incorporating newly collected data during system operation. It combines three techniques: domain incremental learning for handling domain shifts, class incremental learning for detecting new fault types, and novelty detection to guide model updates based on distributional changes. The effectiveness of the framework was demonstrated through experiments using a journal bearing testbed, where streaming data were incrementally incorporated. This study addresses the problem of data scarcity in the early design stage by combining physics-based models and data generation techniques. It also implements a diagnostic framework capable of flexibly incorporating field data. As a result, the effectiveness and scalability of fault diagnosis technology have been significantly improved. This framework is expected to make a practical contribution to various industrial applications, including real-time maintenance systems and digital twin-based prognostics.