A Threshold-Control Generative Adversarial Network Method for Intelligent Fault Diagnosis

Xinyu Li, Sican Cao, Liang Gao, Long WenComplex Syst. Model. Simul 2021 · DOI

Abstract

In modern industrial environments, intelligent fault diagnosis is essential for maintaining system reliability. However, deep learning models often require vast amounts of labeled data, which is difficult to obtain in practice. This paper proposes the Threshold-Control Generative Adversarial Network (TCGAN), a framework that converts 1D vibration signals into 2D images and introduces a dynamic threshold-control mechanism to stabilize GAN training. By regulating the training frequency of the generator and discriminator based on real-time classification accuracy, TCGAN achieves superior performance with limited labeled data, reaching up to 99.96% accuracy on benchmark datasets.

Introduction

The rapid advancement of smart manufacturing has made the early identification of mechanical faults—such as bearing or impeller wear—critical to preventing catastrophic failures. While Deep Learning (DL) has automated feature extraction, its reliance on large, expert-labeled datasets remains a significant bottleneck. In industrial settings, labeled data is scarce and expensive to produce.

Generative Adversarial Networks (GANs) offer a potential solution by generating pseudo-data to augment small datasets. However, standard GANs are notoriously unstable; the discriminator often learns too quickly, leading to a training imbalance that causes the model to fail. This paper addresses these challenges through TCGAN, which stabilizes adversarial training for high-accuracy fault diagnosis under data-constrained conditions.

Methodology

The TCGAN approach consists of three integrated stages: signal preprocessing, architectural design, and a novel training control mechanism.

Signal-to-Image Conversion

To leverage the spatial feature extraction capabilities of Convolutional Neural Networks (CNNs), 1D vibration signals are reshaped into 2D gray-pixel images. This conversion allows the model to capture periodic patterns and visual textures inherent in mechanical vibrations, which are more effectively processed by 2D architectures than traditional 1D signal processing methods.

TCGAN Architecture

The framework employs a Generator and a Discriminator. The Generator uses deconvolutional layers to map random noise into pseudo-vibration images. The Discriminator is a CNN based on the LeNet-5 architecture, modified with a final softmax layer. This allows the discriminator to perform multi-class fault classification in addition to distinguishing between real and synthetic data.

Threshold-Control Mechanism

The core innovation of this work is the dynamic threshold-control mechanism. Unlike standard "open-loop" GANs, TCGAN uses a "closed-loop" feedback system. The training process monitors the testing accuracy (TestAcc) against a predefined threshold.

Flowchart of the TCGAN training process with the dynamic threshold-control mechanism.

As shown in the flowchart, if the accuracy is below the threshold, the discriminator is trained more frequently (e.g., twice as often as the generator) to accelerate convergence. Once the threshold is reached, the training ratio is adjusted to maintain balance, preventing the discriminator from overpowering the generator and ensuring stable feature learning.

Experimental Results

The TCGAN model was validated using the Case Western Reserve University (CWRU) bearing dataset and the Self-Priming Centrifugal Pump (SPCP) dataset. To simulate real-world data scarcity, the model was trained on small subsets of 250, 500, and 1000 labeled samples.

Performance and Stability

The threshold-control mechanism significantly improved both the speed and stability of the training process. On the CWRU dataset, TCGAN achieved 99.96% accuracy with 1000 samples. Even with only 250 samples, it maintained 98.67% accuracy, compared to 94.87% for a standard GAN without threshold control.

Comparison of training accuracy over iterations for TCGAN with and without the threshold-control mechanism.

The results indicate that the threshold-control logic prevents the training collapse often seen in traditional GANs, allowing the model to consistently reach higher accuracy levels more quickly.

Comparison with Other Methods

In head-to-head comparisons, TCGAN outperformed established methods such as Support Vector Machines (SVM), Deep CNNs (DCNN), and Sparse Auto-encoders. Notably, TCGAN using limited data often exceeded the performance of these competitors even when they were trained on the full available dataset.

Conclusion

TCGAN provides a robust solution for industrial fault diagnosis where labeled data is limited. By combining 2D signal representation with a regulated adversarial training strategy, the method achieves high accuracy and stability. Future research will focus on automating threshold selection and extending the framework to other types of industrial sensor data, such as thermal or acoustic signals.

Summary updated

BibTeX
@article{li2021threshold,
  author = {Xinyu Li and Sican Cao and Liang Gao and Long Wen},
  title = {{A Threshold-Control Generative Adversarial Network Method for Intelligent Fault Diagnosis}},
  journal = {Complex Syst. Model. Simul.},
  year = {2021},
  doi = {10.23919/csms.2021.0006},
}
Figure