Bearing Health Status Evaluation Method Based on Multi-Scale Hybrid Features and Inception-Block Attention Bidirectional Physics-Informed Domain Adaptation Network

Yuan Xu, Xiao Wang, Yang Xiao, Qingfeng WangResearch Square 2025 · DOI

This paper introduces a sophisticated deep learning framework designed to solve one of the most persistent headaches in industrial maintenance: accurately predicting when a bearing is going to fail when the operating conditions (speed and load) are constantly changing.

TL;DR

The authors propose IBA-BPDAN, a neural network that combines multi-scale feature extraction with physics-informed constraints to evaluate bearing health. By bridging the "domain gap" between different machines and operating conditions, the model achieves superior accuracy in cross-device health monitoring compared to traditional deep learning methods.

What Problem Does This Paper Address?

In a laboratory, predicting bearing failure is relatively straightforward. However, in a factory, bearings operate under fluctuating speeds and heavy loads, creating "noisy" vibration signals that are difficult to interpret.

The core challenge is Domain Adaptation. A model trained on one specific bearing (the source domain) often fails when applied to a different machine or even the same machine under different loads (the target domain). This "distribution discrepancy" means that features indicating "wear and tear" on a test rig might look completely different on a centrifugal pump in the field. Current models often struggle with feature extraction and slow convergence when faced with these real-world variations.

Methods

The authors developed a multi-layered approach called the Inception-Block Attention Bidirectional Physics-Informed Domain Adaptation Network (IBA-BPDAN). Key components include:

  1. Multi-Scale Hybrid Feature Extraction: Instead of looking at raw data, they use Empirical Wavelet Transform (EWT) to decompose signals into various scales across time, frequency, and time-frequency domains.
  2. Physics-Informed Neural Network (PINN) Layers:
    • Mean Constraint Layer: This constrains input data within specific ranges based on physical expectations, preventing the model from being "tricked" by simple amplitude differences between machines.
  3. Inverse Physical Information Layer: This uses physical priors to adaptively tune the model’s hyperparameters, making the training process more efficient and grounded in reality.
  4. Advanced Architecture: The network uses Inception modules (parallel convolution kernels of different sizes) to capture both local and global features, paired with a Convolutional Block Attention Module (CBAM) to help the model focus on the most "informative" parts of the signal.
  5. Adversarial Domain Adaptation: The model uses Wasserstein distance to align the feature distributions of the source and target domains, essentially "teaching" the model to ignore machine-specific noise and focus on universal degradation patterns.
  6. New Evaluation Metric: They introduce the HSDTW-M index, which combines Dynamic Time Warping (DTW) and Maximized Mutual Information (MIC) to quantitatively measure how closely the model's output tracks the actual physical degradation trend.

Datasets

The model was validated using three distinct data sources to prove its "cross-device" capabilities:

  • XJTU-SY Bearing Dataset: A standard benchmark from Xi’an Jiaotong University.
  • PHM2012 Challenge Dataset: Data from the PRONOSTIA platform, widely used in the Prognostics and Health Management community.
  • Engineering Case Data: Real-world vibration data from a P3409A centrifugal pump at a petrochemical plant in China.

Key Insights and Ideas

  • Physics as a Guardrail: Purely data-driven models can "overfit" to the specific noise of a training machine. By embedding physical constraints (like mean deviation loss) directly into the network layers, the authors ensure the model's predictions remain "physically plausible."
  • Bidirectional Information Flow: The use of both forward physical constraints and "inverse" hyperparameter optimization allows the model to adapt to new environments much faster than standard transfer learning.
  • Geometric Correlation: The HSDTW-M index is a clever way to move beyond simple "pass/fail" classification. It treats health evaluation as a geometric trend-matching problem, which is much more useful for maintenance engineers who need to see the progression of a fault.

Key Results

The IBA-BPDAN model significantly outperformed standard Convolutional Neural Networks (Ret-CNN) and standard Domain Adaptation Networks (DAN):

  • Lower Error Metrics: In the HSDTW-M tests, the proposed model consistently scored lower (indicating a better match to the degradation trend) across almost all tasks. For example, on the P3409A pump, it achieved an index of 0.065, compared to 0.074 for Ret-CNN.
  • Clear Health Grading: The model successfully categorized bearing life into four distinct stages: Healthy, Good, Deterioration, and Failure, with clear "jumps" in the acceleration effective value curves that matched the predicted health transitions.
  • Robustness: The model maintained high accuracy even when trained on laboratory data and tested on real-world industrial pump data, proving its practical utility.

Limitations and Future Work

While the paper demonstrates excellent results for single-bearing degradation, it does not deeply explore mixed fault diagnosis (e.g., a bearing with both inner-race and outer-race damage occurring simultaneously). The authors suggest that future work could involve refining the virtual sample generation algorithms to better simulate these complex, multi-fault scenarios. Additionally, while the physics-informed layers are powerful, they require some level of domain expertise to set the initial constraints, which might limit "out-of-the-box" use for non-experts.

Summary updated

Abstract

<title>Abstract</title> Bearing health state recognition is often affected by variable speeds and heavy load working conditions, making fault signal features difficult to identify and resulting in challenges in health condition recognition methods, including feature extraction difficulties and cross-device recognition challenges. This paper proposes the Bearing Health Status Evaluation Method Based on Multi-Scale Hybrid Features and Inception-Block Attention Bidirectional Physics-Informed Domain Adaptation Network. A multi-scale feature extraction method is developed, and a Multi-Scale Hybrid Features and Inception-Block Attention Bidirectional Physics-Informed Domain Adaptation Network is introduced. This network uses physical information layers and inverse physical information layers to constrain multi-scale features and adaptively adjust model hyperparameters, incorporating Inception multi-scale convolution and convolutional self-attention mechanisms to enhance feature recognition capabilities. To validate the effectiveness of the model, this paper constructs a Health Status Dynamic Time Warping-Mic Index and uses the Xi’an Jiao tong University bearing degradation dataset, PHM2012 challenge dataset, and centrifugal pump engineering data for model validation. The results demonstrate that the model performs well in recognizing the health status of equipment under cross-operating conditions and cross-device scenarios.
BibTeX
@inproceedings{xu2025bearing,
  author = {Yuan Xu and Xiao Wang and Yang Xiao and Qingfeng Wang},
  title = {{Bearing Health Status Evaluation Method Based on Multi-Scale Hybrid Features and Inception-Block Attention Bidirectional Physics-Informed Domain Adaptation Network}},
  booktitle = {Research Square},
  year = {2025},
  doi = {10.21203/rs.3.rs-6303869/v1},
}
Figure