A Hybrid Method Applied to Improve the Efficiency of Full-Waveform Inversion for Pavement Characterization

Jingwei Zhang, Shengbo Ye, Li Yi, Yuquan Lin, Hai Liu, Guangyou FangSensors 2018 · DOI

TL;DR

This paper introduces a novel hybrid approach combining Multilayer Perceptrons (MLPs) with a local optimization algorithm to significantly improve the efficiency of full-waveform inversion for Ground Penetrating Radar (GPR) data. By using MLPs to provide accurate initial estimates for pavement layer properties, the method drastically reduces the computational cost of inversion, achieving comparable accuracy to traditional methods with far fewer objective function calculations.

What Problem Does This Paper Address?

Ground Penetrating Radar (GPR) is a valuable non-destructive technique for characterizing pavement layers, specifically estimating their thickness and dielectric permittivity. However, a significant challenge arises when pavement layers are "thin" relative to the GPR pulse width. In such cases, the reflected signals from the top and bottom of a layer can overlap, making it difficult to accurately determine the layer's properties using standard echo detection methods. This "thin layer problem" necessitates more sophisticated inversion techniques. Traditional full-waveform inversion (FWI) methods, while capable of handling these complexities, often suffer from computational inefficiency due to the ill-posed nature of the inverse problem, which involves navigating a landscape of numerous local minima.

Methods

The proposed solution is a two-stage hybrid method:

  1. MLP-based Initial Estimation: The first stage utilizes a trained Multilayer Perceptron (MLP) network. This network is trained on simulated GPR data generated using a physically-based GPR forward model. The MLP takes a GPR signal as input and outputs a classification corresponding to specific intervals of pavement layer thickness and permittivity. The midpoint values of these predicted intervals are then used as initial estimates for the layer parameters.
  2. Local Optimization for Refinement: In the second stage, these MLP-generated initial estimates are fed into a local optimization algorithm, specifically the Levenberg-Marquardt (LM) algorithm. This algorithm then performs a more precise inversion to refine the parameter estimates and achieve the final optimal solution.

This hybrid approach leverages the pattern recognition capabilities of MLPs to overcome the initial "guess" problem in FWI, thereby guiding the local optimizer towards the global optimum more efficiently.

Datasets

  • Simulated GPR Data: The primary dataset for training the MLPs was generated using a GPR forward model based on Lambot's far-field approximation. This simulation involved two pavement models: a two-layer model and a three-layer model, with the first layer always being "thin." The parameters (thickness and permittivity) were varied within specified ranges and divided into intervals to create a multi-class classification problem for the MLPs. A total of 27,000 training data points were generated.
  • Measured GPR Data: A field experiment was conducted on TIAN BEI road in Beijing, China, using a mono-static GPR system. This measured data was used to validate the performance of the hybrid method in a real-world scenario.

Key Insights and Ideas

  • MLPs as Intelligent Initial Guess Generators: The core innovation is using MLPs not just for classification but specifically to provide highly accurate initial solutions for FWI. This bypasses the need for random initial guesses or computationally expensive global search methods.
  • Bridging the Gap Between Classification and Regression: The method effectively bridges the gap between classification (MLPs) and regression (FWI) by translating classification labels into parameter intervals and then using their midpoints as initial solutions for the regression problem.
  • Efficiency Gains are Substantial: The paper demonstrates that by providing a good initial guess, the computationally intensive FWI can be significantly accelerated, requiring far fewer iterations to converge.
  • Robustness to Parameter Range: The performance of the hybrid method is shown to be dependent on the ranges of parameters used to train the MLPs. If the actual pavement parameters fall outside these trained ranges, the MLP's classification accuracy decreases, impacting the initial solution and, consequently, the inversion accuracy. However, adjusting the training ranges can effectively address this.

Key Results

  • Computational Efficiency: The hybrid method reduced the number of objective function calculations from 450 (for a conventional global optimization algorithm) to just 30 for simulated data, while achieving the same inversion precision. For the field experiment, the DE algorithm required 500 calculations, while the hybrid method needed only 70.
  • Accuracy: The hybrid method achieved high accuracy in both simulated and measured data. For a measured GPR signal, the thickness estimation error was as low as 1.2 mm.
  • Classification Accuracy: The trained MLPs achieved a maximum classification accuracy of 98.6% on a test set, indicating their ability to accurately categorize GPR signals based on pavement properties.
  • Performance with Out-of-Range Parameters: Even when parameters were slightly outside the trained ranges (e.g., 5% deviation), the hybrid method showed better performance than the DE algorithm, with significantly smaller estimation errors.

Limitations and Future Work

  • Dependence on Training Data Ranges: The paper explicitly states that the minimum layer thickness and permittivity detectable by the hybrid method depend on the ranges of training parameters used for the MLPs. If actual pavement parameters fall significantly outside these ranges, the MLP's classification accuracy can decrease, impacting the initial solution and overall inversion performance.
  • Future Work: The authors suggest extending this hybrid method to other situations by incorporating corresponding training data. Future research will focus on developing more efficient methods for providing initial solutions for full-waveform inversion.

Summary updated

Abstract

Ground penetrating radar (GPR), as a nondestructive testing tool, is suitable for estimating the thickness and permittivity of layers within the pavement. However, it would become problematic when the layer is thin with respect to the probing pulse width, in which case overlapping between the reflected pulses occurs. In order to deal with this problem, a hybrid method based on multilayer perceptrons (MLPs) and a local optimization algorithm is proposed. This method can be divided into two stages. In the first stage, the MLPs roughly estimate the thickness and the permittivity of the GPR signal. In the second stage, these roughly estimated values are used as the initial solution of the full-waveform inversion algorithm. The hybrid method and the conventional global optimization algorithm are respectively used to perform the full-waveform inversion of the simulated GPR data. Under the same inversion precision, the objective function needs to be calculated for 450 times and 30 times for the conventional method and the hybrid method, respectively. The hybrid method is also applied to a measured data, and the thickness estimation error is 1.2 mm. The results show the high efficiency and accuracy of such hybrid method to resolve the problem of estimating the thickness and permittivity of a "thin layer".
BibTeX
@inproceedings{zhang2018hybrid,
  author = {Jingwei Zhang and Shengbo Ye and Li Yi and Yuquan Lin and Hai Liu and Guangyou Fang},
  title = {{A Hybrid Method Applied to Improve the Efficiency of Full-Waveform Inversion for Pavement Characterization}},
  booktitle = {Sensors},
  year = {2018},
  doi = {10.3390/s18092916},
}
Figure