Research on Diesel Engine Fault Status Identification Method Based on Synchro Squeezing S-Transform and Vision Transformer

Siyu Li, Zichang Liu, Yunbin Yan, Rongcai Wang, Enzhi Dong, Zhonghua Cheng2023 · DOI

TL;DR

This paper introduces a novel method for identifying diesel engine faults by combining Synchro Squeezing S-Transform (SSST) for enhanced time-frequency analysis of vibration signals with Vision Transformers (ViT) for robust image-based classification. This approach achieves high accuracy in fault detection, outperforming traditional methods by effectively capturing complex signal features.

What Problem Does This Paper Address?

Diesel engines are critical components in heavy machinery, but their reliability decreases with usage, leading to frequent failures that can disrupt operations. Traditional methods for identifying these faults often struggle with the complex, non-linear, and non-smooth nature of vibration signals, especially when dealing with subtle or multiple concurrent faults. Existing time-frequency analysis techniques can suffer from low resolution and poor energy aggregation, while standard deep learning models might not fully capture the intricate temporal and spectral characteristics of these signals. This paper aims to bridge this gap by developing a more accurate and reliable fault identification system for diesel engines.

Methods

The core of this research lies in a two-pronged approach:

  • Synchro Squeezing S-Transform (SSST): This technique is employed to convert the raw one-dimensional vibration signals into two-dimensional time-frequency maps. SSST is an advancement over traditional S-transform and wavelet transforms, offering improved time-frequency resolution and better energy aggregation. This is crucial for highlighting subtle fault-related features that might be obscured in less refined time-frequency representations.
  • Vision Transformer (ViT): The generated time-frequency maps are then treated as images and fed into a ViT model. ViT, a powerful deep learning architecture originally designed for image classification, excels at capturing long-range dependencies and complex spatial features. By applying ViT to these time-frequency images, the method can effectively learn and distinguish between different fault states based on their unique visual patterns.

The paper also compares this SSST-ViT approach against several baseline methods: ST-ViT (using standard S-transform with ViT), SSST-2DCNN (using SSST with a 2D Convolutional Neural Network), and FFT spectrum-1DCNN (using Fast Fourier Transform with a 1D CNN).

Datasets

The study utilizes two primary sources of data:

  • Publicly Available CWRU Bearing Dataset: This dataset, commonly used for bearing fault diagnosis, was employed to initially verify the feasibility of the SSST-ViT method. It contains vibration data from bearings with various fault types and depths.
  • Actual Laboratory Data: The researchers conducted pre-set fault experiments on a CA6DF3-20E3 diesel engine using a high-pressure common rail diesel engine experimental bench. This involved simulating specific fault modes (e.g., cylinder misfires, clogged air filter) to collect real-world diesel engine vibration data.

Key Insights and Ideas

  • Bridging Time-Frequency and Image Recognition: The most significant insight is the successful integration of advanced time-frequency analysis (SSST) with a state-of-the-art image classification model (ViT). This synergy allows for the extraction of rich, discriminative features from complex vibration signals that are then effectively interpreted by a powerful visual recognition architecture.
  • SSST's Advantage: The paper highlights that SSST significantly improves upon traditional time-frequency methods by enhancing resolution and energy aggregation, making subtle fault signatures more discernible.
  • ViT's Suitability for Complex Signals: The research demonstrates that ViT, despite being an image-focused model, is highly effective in classifying the time-frequency representations of vibration signals, outperforming CNN-based approaches in this context.
  • Distinguishing Subtle Faults: The method shows a remarkable ability to differentiate between fault types that are difficult to distinguish even with visual inspection of time-domain waveforms or simpler time-frequency plots.

Key Results

The proposed SSST-ViT method demonstrated superior performance across both datasets:

  • Public Dataset (CWRU): Achieved an overall fault status identification accuracy of 98.31% on the test set.
  • Actual Laboratory Data: Achieved an overall fault status identification accuracy of 95.67% on the test set.

In comparative analysis, the SSST-ViT model consistently outperformed ST-ViT, SSST-2DCNN, and FFT spectrum-1DCNN in terms of accuracy, convergence speed, and stability on both training and validation sets. For instance, on the CWRU test set, SSST-ViT achieved 98.31% accuracy, while ST-ViT reached 95.27%, SSST-2DCNN 92.33%, and FFT spectrum-1DCNN 88.50%.

Limitations and Future Work

The authors acknowledge limitations due to the constraints of their experimental setup in creating pre-set fault conditions. For future work, they suggest expanding research to include experimental studies on diesel engines with multiple, concurrent faults, moving beyond single-fault simulations to better reflect real-world scenarios. They also propose investigating the combination of SSST-ViT with multi-part synthesis for a more comprehensive fault diagnosis.

Summary updated

BibTeX
@misc{li2023research,
  author = {Siyu Li and Zichang Liu and Yunbin Yan and Rongcai Wang and Enzhi Dong and Zhonghua Cheng},
  title = {{Research on Diesel Engine Fault Status Identification Method Based on Synchro Squeezing S-Transform and Vision Transformer}},
  year = {2023},
  doi = {10.3390/s23146447},
}
Figure