Gear Target Detection and Fault Diagnosis System Based on Hierarchical Annotation Training

Haojie Huang, Qianhong Liang, Rui Wu, Dan Yang, Jia Wang, Rong Zheng, Zhezhuang XuMachines 2025 · DOI

TL;DR

This paper introduces a novel hierarchical annotation and training approach for gear fault diagnosis, specifically addressing challenges posed by complex industrial environments and imbalanced datasets. By first identifying the entire gear and then pinpointing specific fault regions, their method achieves robust gear target detection and fault diagnosis, outperforming traditional approaches on self-constructed, realistic datasets.

What Problem Does This Paper Address?

Gears are critical components in industrial machinery, and their health directly impacts system safety and efficiency. Detecting faults like missing or broken teeth early is crucial to prevent costly downtime and accidents. However, existing methods often struggle with:

  • Complex Industrial Environments: Real-world gear images are often captured under varying lighting conditions, different shooting angles, and with cluttered backgrounds, which can confuse standard detection models.
  • Imbalanced Datasets: In industrial settings, healthy gears are far more common than faulty ones. This data imbalance makes it difficult for machine learning models to learn and accurately identify rare fault types, often leading to false positives or negatives.
  • Lack of Interpretability: Traditional end-to-end deep learning models can be "black boxes," making it hard to understand why a particular diagnosis was made.

This paper aims to tackle these challenges by developing a more robust and interpretable system for gear fault detection and diagnosis.

Methods

The core of this work lies in a two-stage, hierarchical approach:

  1. Hierarchical Annotation: Instead of directly annotating fault regions, the process is split into two steps:
  2. Stage 1: Gear Target Detection: The first stage focuses on accurately segmenting the entire gear from complex backgrounds, regardless of its condition. This is achieved using a standard object detection format (YOLO).
  3. Stage 2: Fault Region Localization: Once the gear is isolated, the second stage focuses on precisely identifying and labeling the specific faulty areas (missing or broken teeth). This hierarchical approach helps to disentangle the problem of finding the gear from the problem of identifying the fault.
  4. Semi-Supervised Learning Framework: To address the data imbalance, the authors propose a semi-supervised learning framework that combines Convolutional Neural Networks (CNNs) with the K-means clustering algorithm.
  5. CNN Feature Extraction: A YOLOv8 backbone is used to extract feature matrices from the gear images.
  6. K-means Clustering: These features are then clustered using K-means to identify distinct groups representing healthy, missing, and broken teeth.
  7. Hybrid Decision Making: A novel decision-making process combines the predictions from the CNN and K-means. If their predictions are consistent, the result is accepted. If not, a distance metric and a threshold (M) are used to decide which prediction is more reliable, leveraging the strengths of both methods.

Datasets

The paper introduces self-constructed imbalanced datasets for training and evaluation. These datasets were created using images captured under various challenging conditions (different angles, lighting, backgrounds) to simulate real-world industrial scenarios. They specifically generated three degrees of imbalance: 2:1:1, 5:1:1, and 10:1:1 (Healthy:Missing:Broken).

They also mention using a public dataset (from Roboflow) for comparison, noting that this dataset features simpler, pure white backgrounds, which leads to poorer generalization in real-world tasks.

Key Insights and Ideas

  • Hierarchical Annotation is Key: Breaking down the problem into "find the gear" then "find the fault" significantly improves robustness, especially in cluttered environments. This is a crucial departure from traditional end-to-end methods.
  • Semi-Supervised Learning for Imbalance: Combining CNNs with K-means offers a powerful way to leverage unlabeled data and mitigate the issues caused by a scarcity of fault samples.
  • Hybrid Decision-Making for Robustness: The fusion of CNN and K-means predictions, with a carefully chosen threshold, allows the system to benefit from both supervised learning's accuracy and unsupervised clustering's ability to find patterns in imbalanced data.
  • Real-world Data Matters: Models trained on idealized public datasets perform poorly when applied to realistic, complex industrial images, highlighting the importance of data that reflects actual operating conditions.

Key Results

  • Improved Accuracy: The proposed hierarchical method, especially when combined with the semi-supervised framework, achieved high accuracy across various imbalance ratios. For instance, with a 2:1:1 imbalance, the accuracy reached 0.95 (supervised) and 0.94 (semi-supervised).
  • Robustness to Imbalance: While accuracy slightly decreased with increasing imbalance in the supervised approach, the semi-supervised method showed improved accuracy at higher imbalance ratios (e.g., 0.90 for 10:1:1 imbalance).
  • Effective Fault Localization: The hierarchical approach demonstrated strong performance in both gear detection and fault localization, with metrics like precision and recall generally remaining high.
  • G-mean Superiority: When compared against models trained on a public dataset, the self-constructed datasets with the proposed method yielded significantly higher G-mean values, indicating better overall performance and generalization.

Limitations and Future Work

  • Small Fault Areas: The authors acknowledge that the fault areas identified in the second stage are very small, posing a challenge for precise localization and potentially impacting accuracy.
  • Threshold Setting: The fixed threshold (M) used in the semi-supervised decision-making process is noted as a limitation. They suggest that adaptive threshold adjustment based on specific task requirements would be beneficial.
  • Future Directions:
    • Data Augmentation: Employing more advanced data augmentation techniques.
  • More Powerful Networks: Exploring deeper or more complex deep learning architectures.
  • Adaptive Thresholds: Developing methods for dynamically adjusting the decision threshold in the semi-supervised framework.
  • Further Research: Continued research into improving the positioning accuracy of fault areas.

Summary updated

Abstract

Gears are the core components of transmission systems, and their health status is critical to the safety and stability of the entire system. In order to efficiently identify the typical fault types such as missing teeth and broken teeth in gears, this paper collects a rich sample under complex backgrounds from different shooting angles and lighting conditions. Then a hierarchical approach is used to describe gear faults on the image. The gear samples are first segmented for image extraction and then finely labeled for gear fault regions. In addition, imbalanced datasets are produced to simulate the environment with fewer fault samples in the actual industrial process. Finally, a semi-supervised learning framework is trained based on the above method and applied in actual environment. The experimental results show that the model performs well in gear target detection and fault diagnosis, demonstrating the effectiveness of the proposed method.
BibTeX
@inproceedings{huang2025gear,
  author = {Haojie Huang and Qianhong Liang and Rui Wu and Dan Yang and Jia Wang and Rong Zheng and Zhezhuang Xu},
  title = {{Gear Target Detection and Fault Diagnosis System Based on Hierarchical Annotation Training}},
  booktitle = {Machines},
  year = {2025},
  doi = {10.3390/machines13100893},
}
Figure