SSFLNet: A Novel Fault Diagnosis Method for Double Shield TBM Tool System

Peng Zhou, Chang Liu, Jiacan Xu, Dazhong Ma, Zinan Wang, Enguang HeSensors 2024 · DOI

TL;DR

This paper introduces SSFLNet, a novel deep learning approach for diagnosing faults in the tool system of Tunnel Boring Machines (TBMs). By integrating dynamic simulations with a Stacked Autoencoder and Support Vector Data Description (SAE-SVDD), SSFLNet achieves high accuracy in pinpointing wear locations on disk cutters, crucial for maintaining TBM efficiency and safety.

What Problem Does This Paper Address?

Tunnel Boring Machines (TBMs) are essential for large-scale excavation, but their tooling systems, particularly disk cutters, are prone to wear. This wear significantly reduces boring efficiency, increases maintenance costs, and can even lead to safety hazards. Diagnosing and localizing these failures, especially subtle wear, is challenging during operation. Current methods often rely on manual inspections during costly shutdowns, leading to decreased efficiency. The paper aims to address this by developing an automated, accurate, and real-time fault diagnosis method for TBM disk cutter wear.

Methods

The authors propose a novel method called SSFLNet, which combines several techniques:

  • 3D Modeling and Dynamic Simulation: They created a 3D model of the TBM hydraulic thrust and tool system using SolidWorks and performed dynamic simulations in Adams. This allowed them to analyze how wear on disk cutters affects the load on the hydraulic propulsion cylinder.
  • Hydraulic System Simulation: The hydraulic propulsion system was modeled and simulated using AMESIM. Pressure signals from the hydraulic cylinder chambers and flow signals were captured under various wear conditions.
  • Feature Extraction (SAE): A Stacked Autoencoder (SAE) is used to extract features from the multi-channel pressure signals. The SAE learns to reconstruct the input signals, and the hidden layer representations capture the essential characteristics related to tool wear.
  • Fault Classification (SVDD): Support Vector Data Description (SVDD) is employed to build a hypersphere that defines the boundary between normal and faulty states. This method is used to classify the extracted features into different failure regions.
  • SSFLNet Architecture: The proposed SSFLNet integrates the SAE for feature extraction and SVDD for classification. The network's hyperparameters, such as the number of layers, neural units, epoch value, batch size, and learning rate, were systematically tuned to optimize performance.

Datasets

The paper does not explicitly mention using pre-existing public datasets. Instead, it describes the creation of a custom dataset through experimental simulations and lab bench tests:

  • Simulated Data: Data was generated from dynamic simulations of the TBM hydraulic propulsion system using Adams and AMESIM software, capturing pressure signals under different disk cutter wear scenarios (no wear, 5mm wear, 15mm wear).
  • Experimental Data: A lab bench setup was used to collect data from a TBM hydraulic propulsion system. The experiment involved simulating failures in four distinct regions (upper partition, lower partition, left partition, right partition) by adjusting spring mechanisms. A total of 3000 feature data points were collected, with 2700 used for training and 300 for testing.

Key Insights and Ideas

  • Hydraulic Pressure as a Key Indicator: The study highlights that subtle changes in hydraulic cylinder pressure signals, particularly in the rodless cavity, are highly indicative of disk cutter wear, even when other signals might appear stable.
  • Synergy of Simulation and ML: Combining physics-based dynamic simulations (Adams, AMESIM) with machine learning (SAE-SVDD) provides a powerful approach to understanding complex system behavior and diagnosing faults.
  • SAE-SVDD for Robust Feature Learning: The SAE effectively learns discriminative features from noisy hydraulic signals, while SVDD provides a robust classification mechanism for identifying failure regions, even with limited anomalous data.
  • Optimal Network Depth: The research demonstrates that there's an optimal number of layers for the SAE (around 5 layers in this case) beyond which performance gains diminish, and training time increases significantly.

Key Results

  • High Accuracy: The SSFLNet model achieved an impressive 90% accuracy in determining the failure area of the cutter head.
  • Superior Performance: SSFLNet outperformed traditional SVDD and SAE-SVM models in accuracy. For instance, the SSFL model achieved 95% accuracy compared to SVDD's 66.7% and SAE-SVM's 88.3% for the "None" (no failure) category. For the "Left partition" failure, SSFLNet reached 91.7% accuracy, significantly better than SVDD (8.3%) and SAE-SVM (75%).
  • Efficient Training: Despite its complexity, SSFLNet required significantly less training time (323s) compared to SAE-SVM (459s), while offering higher accuracy.
  • Effective Feature Separation: Visualizations (t-SNE) showed that SSFLNet effectively separated features from different failure categories, minimizing overlap and improving classification.

Limitations and Future Work

  • Normal Wear Focus: The paper primarily focuses on diagnosing failures caused by normal wear of disk cutters, acknowledging that abnormal wear (chipping, dislodgement) is less frequent and not the main subject of this study.
  • Specific TBM Model: The simulations and experiments were conducted on a specific double-shield TBM model. Generalizability to other TBM types might require further validation.
  • Data Availability: The dataset used is custom-generated. Future work could involve validating the SSFLNet on larger, diverse datasets from real-world TBM operations.
  • Real-time Implementation: While the method shows promise for real-time diagnosis, further research into optimizing the computational efficiency for on-board deployment would be beneficial.
  • Other Failure Modes: Expanding the diagnostic capabilities to include other TBM tool system failure modes beyond disk cutter wear would enhance its practical utility.

Summary updated

BibTeX
@article{zhou2024ssflnet,
  author = {Peng Zhou and Chang Liu and Jiacan Xu and Dazhong Ma and Zinan Wang and Enguang He},
  title = {{SSFLNet: A Novel Fault Diagnosis Method for Double Shield TBM Tool System}},
  journal = {Sensors},
  year = {2024},
  doi = {10.3390/s24082631},
}
Figure