CNN-LSTM Hybrid Model to Promote Signal Processing of Ultrasonic Guided Lamb Waves for Damage Detection in Metallic Pipelines

Li Shang, Zi Zhang, Fujian Tang, Qi Cao, Hong Pan, Zhibin Lin2023 · DOI

TL;DR

This paper introduces a novel CNN-LSTM hybrid model for detecting damage in metallic pipelines using ultrasonic guided lamb waves. The model achieves a remarkable 94.8% accuracy, outperforming individual CNN and LSTM models, and demonstrates the significant benefit of incorporating pre-defined time, frequency, and time-frequency domain features for improved robustness, even in noisy conditions.

What Problem Does This Paper Address?

Onshore oil and gas pipelines are critical infrastructure, but they are susceptible to various forms of damage like corrosion and mechanical defects. Detecting these issues early is crucial for preventing failures and ensuring safety. Ultrasonic guided lamb waves offer a promising non-destructive testing (NDT) method for monitoring these pipelines. However, processing the complex signals generated by these waves, especially under challenging operational conditions and in the presence of noise, is a significant challenge. Traditional signal processing methods can struggle with this complexity, and while deep learning approaches like CNNs and LSTMs have shown promise, their effectiveness in this specific application, particularly with carefully selected features, needed further investigation.

Methods

The core of this research is a CNN-LSTM hybrid model. This architecture combines the strengths of Convolutional Neural Networks (CNNs) for extracting spatial and spectral features from the ultrasonic signals (across time, frequency, and time-frequency domains) and Long Short-Term Memory (LSTM) networks for capturing the temporal dependencies within these features.

Key methodological aspects include:

  • Wavelet Threshold Denoising: A pre-processing step to reduce noise in the raw ultrasonic signals, aiming to improve the signal-to-noise ratio.
  • Feature Engineering: Instead of relying solely on end-to-end deep learning, the study extracts a comprehensive set of twenty-nine features. These features are categorized into:
  • Ten dimensional time-domain indicators.
  • Six dimensionless time-domain indicators.
  • Thirteen frequency-domain indicators. This approach aims to leverage "physics-informed" feature extraction to enhance the deep learning model's performance.

  • Data Dimension Reduction: Principal Component Analysis (PCA) and Kernel Principal Component Analysis (KPCA) were explored to reduce the dimensionality of the extracted features, though their impact on accuracy was also evaluated.

  • Model Comparison: The proposed CNN-LSTM model was benchmarked against standalone CNN and LSTM models to demonstrate its superiority.
  • Noise Robustness Testing: The study systematically investigated the impact of varying levels of white Gaussian noise (from 3 dB to 15 dB SNR) on the classification performance of the models.

Datasets

The study utilized a dataset collected from field tests involving six types of pipeline samples with different simulated damage scenarios (notches, weldments, epoxy coating). A total of 336 groups of samples were collected, with 240 used for training and 96 for testing. Each sample contained 3000 sampling points. The dataset represented six distinct classes of pipeline conditions. No new benchmark datasets were introduced; rather, the existing experimental data was curated and processed.

Key Insights and Ideas

  • Hybrid Power: The CNN-LSTM hybrid model significantly outperforms individual CNN and LSTM models, highlighting the synergistic benefit of combining convolutional feature extraction with recurrent temporal modeling for this type of signal.
  • Feature Engineering Matters: Contrary to the "black box" perception of deep learning, the inclusion of carefully selected, physics-informed features (time, frequency, time-frequency domains) demonstrably improves the robustness and accuracy of the deep learning model. This suggests a powerful avenue for enhancing deep learning performance in specialized domains.
  • Noise Resilience: While deep learning models are generally sensitive to noise, the CNN-LSTM model shows remarkable resilience, maintaining high accuracy even at moderate noise levels (SNR=9 dB and above). This is crucial for real-world pipeline monitoring where environmental noise is common.
  • Denoising Benefits: Pre-processing the signals with wavelet threshold denoising significantly improves the classification accuracy of the CNN-LSTM model, underscoring the importance of signal conditioning.
  • Dimension Reduction Nuance: While PCA and KPCA did not significantly boost classification accuracy in this study, they could potentially improve computational efficiency with larger datasets.

Key Results

  • Highest Accuracy: The CNN-LSTM hybrid model achieved a peak classification accuracy of 94.8% using the twenty-nine feature parameters.
  • Benchmarking: This significantly surpassed the accuracy of the CNN model (85.4%) and the LSTM model (86.5%).
  • Denoising Impact: Denoising the data improved the CNN-LSTM model's accuracy from 77.1% (original data) to 87.5%.
  • Feature Impact: Using the twenty-nine feature parameters boosted accuracy to 94.8%, a substantial improvement over original data (77.1%) and denoised data (87.5%).
  • Noise Performance: At a high SNR (15 dB), the CNN-LSTM model achieved 93.8% accuracy. Even at a moderate SNR of 9 dB, accuracy remained high at 83.3%. The model's performance degraded significantly only at very low SNRs (e.g., 3 dB, 33.3% accuracy).

Limitations and Future Work

  • Dataset Size and Complexity: The study acknowledges that the dataset was limited to six common defect types and may not cover the full spectrum of real-world pipeline issues. More complex scenarios involving signal propagation, reflection, and scattering were not extensively explored.
  • Generalizability: The effectiveness of the proposed method on a wider range of pipeline materials, geometries, and more complex defect interactions needs further validation.
  • Computational Cost: While the CNN-LSTM model is effective, it is computationally more intensive than standalone CNN or LSTM models, especially for training.
  • Future Directions: The authors suggest expanding the dataset to include more defect types and complex scenarios. Investigating the impact of different feature selection methods and exploring more advanced deep learning architectures could further enhance performance and efficiency. The potential of dimension reduction techniques for improving computational efficiency with larger datasets also warrants further exploration.

Summary updated

BibTeX
@misc{shang2023cnn,
  author = {Li Shang and Zi Zhang and Fujian Tang and Qi Cao and Hong Pan and Zhibin Lin},
  title = {{CNN-LSTM Hybrid Model to Promote Signal Processing of Ultrasonic Guided Lamb Waves for Damage Detection in Metallic Pipelines}},
  year = {2023},
  doi = {10.3390/s23167059},
}
Figure