Intelligent Debonding Detection in GFRP Rock Bolts via Piezoelectric Time Reversal and CNN-SVM Model

Zhenyu Zhang, Yang Liu, Yixuan Bai, Jianfeng Si, Zhaolong Zhang, Shengwu TuSensors 2025 · DOI

TL;DR

This paper introduces a novel method for detecting debonding damage in GFRP rock bolts using piezoelectric sensors and a combined time-reversal signal processing technique with a CNN-SVM machine learning model. The approach achieves high accuracy (up to 100%) in identifying debonding defects based on their length, location, and number, offering a promising solution for ensuring structural integrity in underground engineering.

What Problem Does This Paper Address?

GFRP (glass-fiber-reinforced polymer) rock bolts are crucial for tunnel and slope stabilization due to their strength and corrosion resistance. However, these bolts are susceptible to degradation, particularly interfacial debonding between the bolt and grout, or grout and rock. This debonding can severely compromise the anchoring system's integrity, potentially leading to catastrophic failures like slope collapse. Existing non-destructive testing methods often struggle with low signal-to-noise ratios, insensitivity to minor defects, and difficulty in characterizing complex damage scenarios. This paper aims to address these limitations by developing a more robust and accurate method for detecting and evaluating GFRP rock bolt debonding.

Methods

The study employs a two-pronged approach:

  1. Piezoelectric Time Reversal (TR): This technique uses piezoelectric transducers to generate and receive ultrasonic waves. The core idea of time reversal is to exploit the inherent spatiotemporal focusing properties of wave propagation. By recording a signal, reversing its time domain, and re-emitting it, the method can amplify subtle defect-induced signal distortions and suppress noise, effectively enhancing the detection of small flaws. This is particularly useful for identifying debonding by analyzing how it alters the wave propagation.
  2. CNN-SVM Hybrid Model: To interpret the complex signals captured by the TR method, a Convolutional Neural Network (CNN) combined with a Support Vector Machine (SVM) is utilized. The CNN acts as a powerful feature extractor, learning discriminative patterns from time-frequency representations of the received signals. The SVM then takes these extracted features and performs the classification of debonding states (e.g., length, location, number of defects). This hybrid approach leverages the CNN's ability to handle complex data and the SVM's strength in robust classification, especially with limited data.

The study also incorporates COMSOL Multiphysics simulations to model the GFRP rock bolt system and analyze wave propagation under various debonding conditions before experimental validation.

Datasets

The study utilizes two primary sources of data:

  • Numerical Simulation Data: Generated using COMSOL 6.1, simulating GFRP rock bolt specimens with controlled debonding defects of varying lengths (0-1200 mm), locations, and quantities.
  • Experimental Data: Collected from laboratory-fabricated GFRP rock bolt specimens with artificially induced debonding defects. These specimens were designed to mimic real-world conditions, and signals were acquired using a piezoelectric sensing system.

The experimental data was partitioned into an 80% training set and a 20% test set for model evaluation, with a focus on maintaining class balance through stratified randomization.

Key Insights and Ideas

  • Debonding as a Stiffness Reducer: The core principle is that debonding defects reduce the structural stiffness of the GFRP rock bolt system. This reduction in stiffness directly impacts the propagation of ultrasonic waves, leading to measurable changes in signal characteristics.
  • Time Reversal for Signal Enhancement: The time-reversal technique is crucial for amplifying weak signals and suppressing noise, making it possible to detect subtle changes caused by debonding that might otherwise be missed.
  • Time-Frequency Maps as Rich Features: Wavelet time-frequency maps of the experimental signals provide a rich representation of the wave's behavior over time and frequency, which is highly effective for the CNN-SVM model to learn from.
  • CNN-SVM Synergy: Combining CNN for feature extraction and SVM for classification proves highly effective, allowing for accurate identification of debonding even in complex scenarios.

Key Results

The study reports significant findings from both simulations and experiments:

  • Debonding Length: Increasing debonding length generally leads to:
  • Increased signal amplitude (10.96% in simulations, 54.9% in experiments).
  • A shift in the main peak frequency (decrease in simulations, slight increase in experiments).
  • Greater energy concentration.
  • Debonding Location: Location has a minimal impact on amplitude but influences waveform clarity and can cause frequency shifts.
  • Number of Debonding Defects: Multiple defects lead to:
  • Decreased signal amplitude (16.68% in simulations, 3.0% in experiments).
  • Elevated main peak frequency.
  • More severe wave packet overlap.
  • CNN-SVM Model Performance: The developed CNN-SVM model achieved exceptional classification accuracy:
  • 99% for debonding length.
  • 100% for debonding position.
  • 100% for the number of debonding defects. All these results surpassed the 95% reliability threshold.

Limitations and Future Work

The authors acknowledge that this study focused on isolated variables (debonding length, location, and quantity). Real-world degradation often involves coupled factors, such as simultaneous mechanical loads and environmental effects. Future work will aim to address these interactions. Additionally, the study did not investigate different types of defects beyond interfacial debonding, which could be a focus for future research. The paper also notes that numerical simulations sometimes overestimate amplitude reduction compared to experimental results, suggesting the need for damping correction factors in future modeling.

Summary updated

Abstract

To address the challenge of detecting debonding damage in glass-fiber-reinforced polymer (GFRP) rock bolt anchorage structures, this study proposes a time reversal detection method based on piezoelectric sensing and a Convolutional Neural Network-Support Vector Machine (CNN-SVM) model. Through COMSOL 6.1 numerical simulations and laboratory experiments, the influence of debonding length, location, and quantity on the characteristics of detection signals was investigated. The results indicate that an increase in debonding length leads to a rise in the amplitude of the focused signal, a reduction in the main peak frequency, and greater energy concentration around the main peak. Specifically, the amplitude increased by 10.96% (simulations) and 54.9% (experiments) for lengths from 0 to 1200 mm, while the peak frequency decreased by 3.43% (simulations) or increased slightly (experiments). When the debonding location changes, the amplitude remains stable, while the main peak frequency increases by 4.94% in simulations and shifts to higher frequencies experimentally, and the energy exhibits an increasing trend. An increase in the number of debonding points results in decreased amplitude, elevated main peak frequency, and more severe wave packet overlap. Multi-defect configurations reduced the amplitude by 16.68% (simulations) and 3% (experiments), with peak frequency increases of up to 3.35%. Based on these characteristics, a CNN-SVM evaluation model was constructed, using the wavelet time-frequency maps of experimental signals as input and the debonding state as output. The model achieved evaluation accuracy rates of 99%, 100%, and 100% under varying debonding lengths from 10 to 100 mm, different debonding positions, and increasing numbers of debonding defects, all exceeding 95%, thereby validating the reliability and high precision of the proposed method.
BibTeX
@inproceedings{zhang2025intelligent,
  author = {Zhenyu Zhang and Yang Liu and Yixuan Bai and Jianfeng Si and Zhaolong Zhang and Shengwu Tu},
  title = {{Intelligent Debonding Detection in GFRP Rock Bolts via Piezoelectric Time Reversal and CNN-SVM Model}},
  booktitle = {Sensors},
  year = {2025},
  doi = {10.3390/s25237208},
}
Figure