Bearing Fault Diagnosis Using ACWGAN-GP Enhanced by Principal Component Analysis

Bin Chen, Chengfeng Tao, Jie Tao, Yuyan Jiang, Ping LiSustainability 2023 · DOI

TL;DR

This paper tackles the challenge of diagnosing bearing faults with limited data by proposing a novel data augmentation technique. They enhance an Auxiliary Classifier Generative Adversarial Network (ACGAN) with Principal Component Analysis (PCA) and Wasserstein distance with gradient penalty (WGAN-GP) to generate realistic synthetic fault data, significantly improving diagnostic accuracy on a standard benchmark dataset.

What Problem Does This Paper Address?

Rolling bearings are critical components in many rotating machinery systems, including wind turbines. Their failure can lead to significant financial losses and safety hazards. Early fault diagnosis is crucial for predictive maintenance. However, a major hurdle in developing accurate deep learning models for bearing fault diagnosis is the scarcity of labeled fault data. Collecting and labeling such data is expensive and time-consuming, leading to imbalanced datasets and low classification accuracy for existing models. This paper aims to address this "small sample" problem by generating high-quality synthetic fault data to augment limited real-world datasets.

Methods

The core of this work is an improved Generative Adversarial Network (GAN) architecture, specifically an ACWGAN-GP enhanced with Principal Component Analysis (PCA). Here's a breakdown:

  • ACGAN (Auxiliary Classifier GAN): This GAN variant incorporates class labels into both the generator and discriminator. This allows it to generate not only realistic data but also data belonging to specific fault categories, improving the quality and controllability of generated samples.
  • WGAN-GP (Wasserstein GAN with Gradient Penalty): To address the training instability issues common in GANs (like gradient explosion/disappearance), the authors adopt WGAN-GP. This uses Wasserstein distance for a more stable training objective and a gradient penalty term to enforce the Lipschitz constraint, leading to more reliable training.
  • PCA Enhancement: The novelty lies in how PCA is integrated. Instead of using random noise as the generator's input, the authors first transform the 1D time-domain vibration signals into 2D grayscale images. Then, PCA is applied to these images to reduce their dimensionality and extract key feature vectors. This PCA-derived feature vector, along with the fault label, serves as the input to the generator. This approach aims to preserve original data characteristics and improve the generator's controllability.
  • Two-Dimensional Convolutional Network (2D-CNN): For the fault classification task itself, a 2D-CNN is employed. This is chosen because the input data is converted into 2D grayscale images, making 2D convolutions suitable for feature extraction.

Datasets

The proposed model was trained and evaluated using the Case Western Reserve University (CWRU) bearing dataset. This dataset is a standard benchmark for bearing fault diagnosis research. It includes vibration data collected under various motor loads and contains multiple fault types (inner ring, outer ring, ball faults) and damage severities, along with a "normal" state. The authors converted the 1D time-series vibration signals into 2D grayscale images (64x64 pixels) for processing.

Key Insights and Ideas

  • PCA as a Feature Extractor for GAN Input: Using PCA-reduced feature vectors derived from 2D image representations of vibration signals as input to the GAN generator is a novel way to guide data generation. It leverages the dimensionality reduction capabilities of PCA to feed more informative, characteristic-rich data to the generator, moving beyond simple random noise.
  • Bridging 1D Signals and 2D Image Processing: The transformation of 1D vibration signals into 2D images is a crucial preprocessing step that enables the use of powerful 2D CNNs for both data generation (via GANs that can process image-like data) and classification.
  • Synergy of ACGAN, WGAN-GP, and PCA: The combination of these techniques creates a robust framework. ACGAN provides class-conditional generation, WGAN-GP ensures stable training, and PCA enhances the input representation for better sample generation.
  • Effectiveness in Low-Data Regimes: The paper strongly emphasizes that this approach is particularly effective when real fault data is scarce, demonstrating significant accuracy improvements even with reduced training set sizes.

Key Results

  • Improved Sample Similarity: The PCA-ACWGAN-GP model generated synthetic samples that were highly similar to real fault samples, as quantitatively measured by Euclidean Distance (ED) and Cosine Distance (CD). For example, for fault label 0, the ED was 0.45187 and CD was 0.97517, outperforming ACGAN and ACWGAN-GP.
  • Enhanced Classification Accuracy: When the generated data was used to augment a limited training set, the 2D-CNN classifier achieved significantly higher accuracy. The overall accuracy reached 99.2% with the PCA-ACWGAN-GP enhanced dataset, compared to 94.6% for ACGAN and 97.8% for ACWGAN-GP.
  • Robustness to Data Scarcity: Even when the training data was reduced to 40% of the original size, the PCA-ACWGAN-GP enhanced classifier maintained a high accuracy of 97.3%, demonstrating its effectiveness in low-data scenarios.

Limitations and Future Work

  • Computational Cost: While the paper doesn't explicitly detail computational costs, GAN training, especially with complex architectures and large datasets, can be computationally intensive.
  • Generalizability to Other Datasets: The study primarily focuses on the CWRU dataset. Further validation on diverse bearing datasets from different machinery types and operating conditions would strengthen the findings.
  • Future Directions: The authors suggest relaxing limitations on generated sample quality and the number of real samples. They also aim to optimize the model structure and improve computational efficiency. Exploring more advanced feature extraction techniques and different GAN architectures could be future avenues.

Summary updated

Abstract

Rolling bearings are one of the most widely used parts in all kinds of rotating machinery (including wind power equipment) and also one of the most easily damaged parts, which makes fault diagnosis of rolling bearings a promising research field. To this end, recent studies mainly focus on fault diagnosis cooperating with deep learning. However, in practical engineering, it is very challenging to collect massive fault data, resulting in low accuracy of bearing fault classification. To solve the problem, an auxiliary classifier optimized by a principal component analysis method is proposed to generate an adversarial network model in which Wasserstein distance and gradient penalty are used to improve the stability of the network training process in case of over-fitting and gradient disappearance during model training. Specifically, we implement the model system using two main components. First, the one-dimensional time domain signal is transformed into a two-dimensional grayscale image and the principal component analysis is employed to reduce the dimension of the original data; this is instead of random noise as the input of the generator thereby preserving the characteristics of the original data to a certain extent. Second, in a generative adversarial network, the label information of the fault data is inserted into the generator to achieve supervised learning, thereby improving the data generation performance and reducing the training time cost. The experimental results show that our model could produce high-quality samples that are similar to real samples and that it could significantly improve the classification accuracy of fault diagnosis in the case of insufficient fault samples.
BibTeX
@inproceedings{chen2023bearing,
  author = {Bin Chen and Chengfeng Tao and Jie Tao and Yuyan Jiang and Ping Li},
  title = {{Bearing Fault Diagnosis Using ACWGAN-GP Enhanced by Principal Component Analysis}},
  booktitle = {Sustainability},
  year = {2023},
  doi = {10.3390/su15107836},
}
Figure