A ground penetrating radar underground disease target detection method based on meta learning

Abstract

This invention discloses a ground-penetrating radar (GPR) method for detecting underground defects based on meta-learning. The method first denoises, normalizes, and enhances the GPR signal to construct multi-scale features; it then uses ResNet50 to extract spatial features and combines wavelet packet decomposition and channel attention mechanisms to strengthen weakly scattering defect features; a model-independent meta-learning algorithm is introduced to enable the model to quickly adapt to different regions and soil conditions under limited sample conditions; finally, the features are input into an improved YOLOv8 detection network for defect localization and classification. Through model fine-tuning and loss optimization strategies, the generalization ability and detection stability across environments are improved. This invention can achieve high-precision identification of underground defects such as cavities, voids, loosening, and seepage, and has good engineering application value.

Description

Technical Field

[0001] This invention relates to underground disease detection technology, and more particularly to a ground-penetrating radar method for detecting underground disease targets based on meta-learning, used to identify disease characteristics such as cavities, voids, looseness and leakage in underground soil. Background Technology

[0002] Underground defects such as cavities, voids, loosening, and seepage are widely present in engineering structures such as road, tunnel, and bridge foundations, making accurate detection crucial. Ground-penetrating radar (GPR) is widely used due to its non-destructive and rapid capabilities; however, its echo signals are easily affected by soil dielectric properties, moisture content, and burial depth variations, often suffering from strong noise interference, complex reflection characteristics, and indistinct signals for weak defects. This results in insufficient stability of traditional detection methods in complex environments.

[0003] While existing deep learning-based detection methods can automatically extract features, they typically rely on large amounts of labeled data, and their generalization ability is poor under different regions or soil conditions, making them difficult to adapt to cross-scenario application requirements. Single-scale feature extraction struggles to simultaneously capture large-scale structural information and small-scale anomaly features, and while wavelet packet decomposition can provide multi-scale information, its utilization of key subbands remains insufficient. Therefore, it is necessary to propose a subsurface disease detection method that combines multi-scale feature modeling with a meta-learning mechanism to improve recognition accuracy and adaptability in environments with limited samples and complex conditions. Summary of the Invention

[0004] In view of this, the present invention provides a ground-penetrating radar target detection method based on meta-learning. By integrating wavelet packet transform, multi-scale feature extraction, attention mechanism and Model-Agnostic Meta-Learning algorithm, it can achieve high-precision positioning and classification of defects such as cavitation, voids, loosening and leakage, and has the ability to adapt quickly across regions.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] This invention provides a method for detecting underground defects using ground-penetrating radar based on meta-learning, comprising the following steps:

[0007] (1) Data acquisition and preprocessing (S1): Radar reflection signals of underground soil are acquired by ground penetrating radar, and the original signals are denoised, normalized and amplitude enhanced. At the same time, the signals are segmented at multiple scales to construct a spatiotemporal feature map, providing high-quality input for subsequent feature extraction.

[0008] (2) Feature extraction network (S2): ResNet50 is used as the backbone network to extract the spatial features of radar signals through multi-layer convolution, and skip connections are added between different convolutional layers to retain low-level detailed information.

[0009] (3) Time-frequency enhancement module (S3): The extracted features are decomposed into multi-scale sub-band signals by wavelet packet transform, and the weights are adaptively allocated through the channel attention mechanism to enhance the weak reflection signal, reduce high-frequency noise interference, and improve the response capability of disease features.

[0010] (4) Meta-learning rapid adaptation module (S4): The Model-Agnostic Meta-Learning algorithm is adopted to enable the model to quickly optimize the parameters of the feature extraction and detection modules through a small number of labeled samples, so as to achieve rapid adaptation to the disease characteristics in new areas, different soil types or burial depths.

[0011] (5) Disease detection and classification (S5): The enhanced features are input into the YOLOv8 detector, and the PAN+FPN feature fusion structure is used. A new P2 layer is added to enhance the low-scale feature expression, so as to achieve accurate location and category prediction of cavitary, void, loose and seeping diseases. At the same time, the disease bounding box and prediction confidence are output.

[0012] (6) Adaptive optimization and update (S6): Based on the new sample data obtained in practice, the feature extraction network and YOLOv8 detector are fine-tuned. Through dynamic learning rate adjustment and loss function weighting mechanism, the generalization ability and robustness of the model in cross-regional and cross-soil environment are enhanced.

[0013] Compared with existing technologies, it has the following significant advantages:

[0014] 1. By introducing the Model-Agnostic Meta-Learning algorithm, the model can quickly adapt to new working conditions, new regions, or different soil conditions with a small number of samples, solving the problem of decreased accuracy of traditional deep learning methods in cross-regional detection.

[0015] 2. By employing wavelet packet multi-scale decomposition combined with channel attention mechanism, more discriminative disease features can be extracted in different frequency bands, and invalid high-frequency noise can be automatically suppressed, making weak reflective targets such as cavities, voids, looseness and leakage easier to identify, thereby improving detection sensitivity and depth adaptability.

[0016] 3. By adding a P2 layer to the PAN+FPN structure, the low-level feature expression is improved, enabling the detector to effectively identify shallow, small and weak-contrast disease targets, while maintaining high-precision positioning of medium and large disease targets. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0018] Figure 1 The flowchart illustrates a ground-penetrating radar method for detecting underground defects based on meta-learning, as provided in this invention.

[0019] Specific implementation method

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] In the first aspect of the embodiment, ground-penetrating radar is used to collect radar reflection signals from underground soil. After acquisition, the raw signals are preprocessed, including denoising, normalization, and amplitude enhancement, to improve the signal-to-noise ratio and the identifiability of weakly reflective targets. The preprocessed signals are then segmented at multiple scales to construct a spatiotemporal feature map, providing high-quality input for subsequent feature extraction by a deep network. For feature extraction, ResNet50 is used as the backbone network, extracting spatial features through multi-layer convolutions, and introducing skip connections between different convolutional layers to preserve low-level detail information. The extracted features are then decomposed into multi-scale sub-band signals using wavelet packet transform, and weights are automatically assigned using a channel attention mechanism to enhance weakly reflective disease signals while suppressing high-frequency noise interference. The Model-Agnostic Meta-Learning algorithm is employed, enabling the model to quickly optimize the parameters of the feature extraction and detection modules using a small number of labeled samples, achieving rapid adaptation to disease characteristics in new areas and different soil conditions. The enhanced YOLOv8 detector utilizes a PAN+FPN feature fusion structure and adds a P2 layer to strengthen low-scale feature representation, achieving accurate localization and classification of cavities, voids, loosening, and leakage defects. It also outputs defect bounding boxes and prediction confidence scores. In practical applications, the network is fine-tuned based on newly sampled data, and combined with dynamic learning rate adjustment and loss function weighting mechanisms, significantly improving the model's generalization ability and robustness across regions and soil types. Detection results are used to generate a visual report through non-maximum suppression and probability filtering, including defect location, category, and confidence score, providing an intuitive reference for underground facility maintenance and decision-making.

[0022] In the second aspect of the embodiment, to further improve the detection capability of small or low-contrast diseases, the wavelet packet transform module was optimized. A fixed tree structure was used for multi-level decomposition, and the most effective sub-band features were adaptively selected through a channel attention mechanism to ensure that weak reflection signals can fully express disease information while suppressing background noise interference. The feature extraction network still uses the ResNet50 backbone, but a multi-scale fusion module is added between different convolutional layers to preserve and enhance the structural information of medium and large-sized diseases. In the meta-learning module, the parameters of the feature extraction network and the YOLOv8 detector are initialized through the Model-Agnostic Meta-Learning algorithm, enabling the model to converge quickly with a small number of new samples and possess efficient adaptability. The YOLOv8 detector adds a P2 layer on the basis of PAN+FPN to enhance the detection of low-scale disease targets and improves the localization accuracy of medium and large-sized diseases through multi-scale fusion. This embodiment also introduces a cross-regional training strategy and regularization constraints, combined with a loss function weighting mechanism, to enable the model to maintain stable performance under different soil moisture, dielectric constant, burial depth, and complex environmental conditions. Experimental results show that the optimized model can not only accurately locate underground cavities, voids, looseness and leakage, but also provide clear and intuitive interpretation results in terms of confidence assessment and visualization.

[0023] In the third aspect of this embodiment, an adaptive optimization strategy is further introduced into the detection process. Newly acquired sample data is input into the model, and the feature extraction network and YOLOv8 detector are fine-tuned. A dynamic learning rate adjustment strategy is adopted, enabling the model to quickly adapt to new environments while maintaining its original performance. For scenarios with complex soil conditions and varying moisture levels, a loss function weighting mechanism is used to selectively optimize different types of diseases, thereby improving the model's sensitivity to a small number of samples or abnormal samples. In the time-frequency enhancement module, wavelet packet decomposition combined with a channel attention mechanism dynamically adjusts the weights of each sub-band, achieving automatic selection and enhancement of disease signals in different frequency bands. After processing by this embodiment, the model can not only quickly identify cavities, voids, loosening, and leakage diseases in different underground media, but also the visualization interface of the detection results can display the disease boundary box, category, and confidence level, facilitating intuitive analysis and decision support for engineers. Furthermore, this embodiment demonstrates the high adaptability and robustness of the method in practical engineering applications, providing an effective solution for underground disease detection in complex environments.

[0024] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A ground-penetrating radar method for detecting underground defects based on meta-learning, characterized in that, include: S1: Collect underground ground-penetrating radar signals, and perform noise reduction, normalization and amplitude enhancement processing on the signals, while performing multi-scale segmentation to construct feature maps; S2: ResNet50 is used as the backbone network to extract features, and multi-scale information is preserved through cross-layer connections; S3: The extracted features are decomposed into wavelet packets to obtain multi-scale sub-band features, and then weighted using a channel attention mechanism; S4: The model-independent meta-learning algorithm is used to initialize the parameters of the feature extraction module and the detection module to improve the model's adaptability to new environments; S5: Input the processed features into the YOLOv8 detection network for target detection. The detection network adopts a feature fusion structure and includes a low-scale feature layer to output the disease category and location parameters. S6: Fine-tune the model based on the new samples to improve detection performance in cross-regional environments.

2. The method according to claim 1, characterized in that, The wavelet packet decomposition adopts a fixed tree structure and combines an attention mechanism to select sub-band features.

3. The method according to claim 1, characterized in that, The feature fusion structure of the YOLOv8 detection network includes FPN and PAN structures, and a low-scale feature layer is set for small target detection.

4. The method according to any one of claims 1-3, characterized in that, The feature extraction module and the detection module are initialized based on the model-independent meta-learning algorithm.

5. The method according to any one of claims 1-4, characterized in that, Improve the generalization ability of the model by weighting the loss function and applying regularization constraints.

Images