Bearing Fault Diagnosis Under Small Data Set Condition: A Bayesian Network Method With Transfer Learning for Parameter Estimation

Yongyan Hou, Ao Yang, Wenqiang Guo, Zheng En-rang, Qinkun Xiao, Zhigao Guo, Zixuan HuangIEEE Access 2022 · DOI

TL;DR

This paper introduces Varying Coefficient Transfer Learning (VCTL), a novel algorithm that enables accurate bearing fault diagnosis even when labeled training data is scarce. By dynamically aggregating knowledge from related "resource domains" (source tasks) and weighting them based on their relevance to the target task, VCTL significantly outperforms traditional statistical methods and existing transfer learning techniques in small-data scenarios.

What Problem Does This Paper Address?

In the world of the Industrial Internet of Things (IIoT), bearings are critical components whose failure can lead to catastrophic downtime. While AI-driven fault diagnosis has made huge strides, most state-of-the-art models (like deep neural networks) are "data-hungry," requiring massive amounts of labeled vibration data to function reliably.

In real-world engineering, collecting these labels is often expensive, dangerous, or technically difficult. Bayesian Networks (BNs) are excellent tools for modeling such systems because they handle uncertainty well, but they face a major hurdle: parameter learning. When you only have a "small data set," standard methods like Maximum Likelihood Estimation (MLE) produce highly inaccurate probability tables, leading to unreliable diagnoses.

Methods

The authors propose the VCTL algorithm to bridge the gap between data scarcity and model accuracy. The core idea is to "borrow" knowledge from existing, related datasets (resource domains) to help train the target model.

Key innovations include:

  • Dynamic Balancing Weight ($\beta$): Instead of a fixed ratio, VCTL uses a function that calculates a "Sample Size Factor" (SSF). If the target data is extremely small, the model relies heavily on resource knowledge. As more target data is collected, the model automatically shifts its weight toward the target samples.
  • Relevance Weight Factors ($\omega_{SL}$): Not all source data is helpful. VCTL uses the Dirichlet Compound Multinomial (DCM) similarity measure to quantitatively determine how relevant a resource domain is to the target task. This prevents "negative transfer," where irrelevant data actually hurts model performance.
  • Parameter Aggregation: The final model parameters are a weighted combination of initial target estimates (via MAP or MLE) and the fused knowledge from the most relevant resource domains.

Datasets

The researchers validated their approach using two distinct types of data:

  1. Synthetic Benchmarks: Three classic Bayesian Networks from the BN repository—Weather, Asia, and Alarm—were used to test the algorithm's mathematical robustness across small, medium, and large parameter spaces.
  2. Real-World Industrial Data: The famous Case Western Reserve University (CWRU) Bearing Data Center dataset. They specifically looked at 12k driving end bearing fault data with different fault diameters (7 mil, 14 mil, and 21 mil) to simulate resource and target domains.

Key Insights and Ideas

  • The "Trigger" Mechanism: One of the most practical insights is the use of a threshold ($Tr$) to decide when to even bother with transfer learning. If the target dataset is large enough to stand on its own, the system bypasses the resource domains entirely, ensuring computational efficiency.
  • Similarity Matters: The paper demonstrates that simply "averaging" source data (as some older methods do) is insufficient. By using DCM similarity, VCTL can identify which specific "neighboring" machine or condition provides the most useful diagnostic clues.
  • Robustness to Noise: VCTL proved surprisingly resilient to "noisy" resource data. Even when the source datasets contained errors or irrelevant information, the relevance weighting successfully filtered out the "garbage," maintaining higher accuracy than the state-of-the-art Local Linear Pooling (LoLP) method.

Key Results

  • Superior Accuracy in Small Sets: In experiments with target sample sizes as low as 15 to 100, VCTL consistently achieved the lowest Kullback-Leibler Divergence (KLD)—a measure of how much the learned model differs from the "ground truth."
  • Diagnostic Boost: In the real-world bearing case study (14 mil fault diagnosis), VCTL achieved an average diagnosis accuracy of 90.30% with a small dataset (size 60), outperforming the LoLP method by 6.71%.
  • Convergence: As the dataset size increased to 300 (sufficient data), VCTL's accuracy rose to 91.96%, proving that it converges toward classical optimal performance as data becomes plentiful.
  • Efficiency: Despite the extra steps for similarity calculation, the computational complexity remains nearly identical to standard MAP estimation, making it viable for real-time industrial monitoring.

Limitations and Future Work

While VCTL handles noisy resource data well, it still relies on the existence of some structurally similar resource domains. If the target system is entirely unique with no related historical data, the benefits of transfer learning vanish.

The authors suggest that future research could explore structure transfer learning—not just learning the parameters (probabilities) but also "borrowing" the actual architecture of the Bayesian Network from related tasks. Additionally, applying VCTL to more complex, multi-state industrial systems beyond bearings remains an open and promising frontier.

Summary updated

Abstract

Bearings are broadly applied in various types of industrial systems. Fault diagnosis, as a promising way for reliability of modern industrial internet of thing applications, has attracted increasing attention from both academia and industry fields. Being ideal modeling and inference tool in uncertainty situations, Bayesian network (BN) is becoming increasingly popular in many systems. However, in practical uncertain and complicated engineering surroundings, it’s difficult or expensive to collect massive labeled fault data for the sake of fault diagnosis model learning. To address the issue of BN parameter learning under small data set conditions, this paper proposes a Varying Coefficient Transfer Learning (VCTL) algorithm based on aggregation and transfer learning, that considers both knowledge from the resource domain and the resource relevance contributions. The balancing weight function is designed to determine whether the learning task in the resource domain is activated. Relevance weight factors are proposed to measure the relevance of resource and target parameters quantitatively, by combing parameter information from resource domains with those obtained from the target domain, using maximum a posterior (MAP) or maximum likelihood estimation (MLE). Finally, the target parameters are aggregated with both the target initial parameters and the parameter knowledge from the resource domain. Based on VCTL, a bearing fault diagnosis approach is proposed and verified. The experimental results show that, under the condition of the small data set, learning accuracy of VCTL algorithm with varying coefficient aggregation is better than MLE algorithm, MAP algorithm or state-of-the-art parameter transfer method, local linear pooling transfer learning (LoLP) algorithm. Under the condition of sufficient data set, learning accuracy of VCTL algorithm approaches the classical MLE or MAP, and the correctness of the proposed algorithm is verified. Moreover, we illustrate the successful application to real-world bearing fault diagnosis case with VCTL, where we had access to expert-provided resource knowledge and real fault diagnosis data.
BibTeX
@inproceedings{hou2022bearing,
  author = {Yongyan Hou and Ao Yang and Wenqiang Guo and Zheng En-rang and Qinkun Xiao and Zhigao Guo and Zixuan Huang},
  title = {{Bearing Fault Diagnosis Under Small Data Set Condition: A Bayesian Network Method With Transfer Learning for Parameter Estimation}},
  booktitle = {IEEE Access},
  year = {2022},
  doi = {10.1109/access.2022.3151240},
}
Figure