Effective signal enhancement and interference separation method for ground penetrating radar underground disease data

Abstract

The invention discloses a ground penetrating radar underground disease signal enhancement and interference separation method, which is used for improving the detection precision of holes, void, looseness and leakage. The method comprises the following steps: performing denoising, amplitude normalization and time window processing on B-scan data to construct clean input; generating curvature guide features by using the curvature modeling network, and enhancing disease reflection; disease features are extracted and strengthened through a time domain-frequency domain double-domain attention network, and meanwhile earth surface and multi-path interference is restrained; dividing the features into background and disease abnormal components by using low-rank sparse decomposition, and highlighting disease signals; and finally, fusing each feature to generate an enhanced graph, and realizing automatic identification and distinguishing of diseases. The method does not need artificial feature engineering and end-to-end processing, is suitable for various underground disease detection scenes, and has high precision, robustness and good generalization.

Description

Technical Field

[0001] This invention belongs to the field of underground target detection and signal processing technology, specifically relating to a data enhancement method for underground defect detection by ground penetrating radar, particularly a deep learning signal enhancement and interference separation method for defects such as cavitation, voids, loosening, and leakage. Background Technology

[0002] Underground defects such as cavities, voids, loosening, and leakage are common in road, tunnel, and pipeline structures, making their effective identification crucial. Ground-penetrating radar (GPR) is widely used for detecting underground defects due to its non-destructive testing and rapid scanning capabilities. However, its B-scan echo data is susceptible to interference from surface reflections, metal edges, multipath scattering, and environmental noise, resulting in the mixing of background and anomalous features in the signal, making it difficult to effectively separate weak defect reflections.

[0003] Existing methods often rely on single features in the time or frequency domains for analysis, making it difficult to fully utilize multi-domain information. Traditional feature extraction methods underutilize the hypercurved reflection structures generated by subsurface defects and lack effective modeling of geometric features. Existing methods typically fail to effectively distinguish between stable geological backgrounds and local anomalous reflections, making detection results susceptible to interference. Therefore, it is necessary to propose a subsurface defect signal processing method that combines curvature feature modeling, time-frequency dual-domain feature enhancement, and low-rank sparse decomposition to achieve effective signal enhancement and interference separation, thereby improving the accuracy and stability of subsurface defect identification. Summary of the Invention

[0004] In view of this, the present invention provides an effective signal enhancement and interference separation method for ground penetrating radar underground disease data, realizing deep learning signal enhancement and interference separation for underground diseases such as cavities, voids, looseness, and leakage.

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

[0006] This invention provides a method for detecting underground defects based on the multi-frequency characteristics of ground-penetrating radar, comprising the following steps:

[0007] (1) B-scan data preprocessing (S1): Acquire and preprocess ground-penetrating radar B-scan data in underground areas, and perform amplitude standardization, time window calibration, DC drift removal and noise suppression on the echo data.

[0008] (2) Reflection curvature modeling network (S2): Construct a curvature guidance feature network based on reflection curvature modeling, estimate the curvature of hypercurve reflections generated by underground diseases (including cavities, voids, looseness and leakage) in B-scan data, generate reflection curvature guidance map, and enhance the curvature sensitive features related to diseases.

[0009] (3) Time-frequency dual-domain attention enhancement network (S3): Construct a time-frequency dual-domain attention feature extraction network, input the preprocessed B-scan data into the time-domain convolutional network and the frequency-domain transformation network respectively, and achieve the enhancement of disease reflection energy and the suppression of noise such as surface reflection, metal edge interference, and multipath scattering through the dual-domain attention mechanism.

[0010] (4) Low-rank sparse interference separation network (S4): A low-rank sparse interference separation network is introduced to decompose the feature map after dual-domain enhancement into a low-rank background component and a sparse disease component. The low-rank component is used to characterize the stable stratum background structure, and the sparse component is used to characterize the local abnormal reflections caused by cavities, voids, looseness and leakage.

[0011] (5) Cross-domain fusion module (S5): The curvature guidance feature, dual-domain attention output feature and sparse disease feature are fused to generate disease enhancement map, so as to achieve significant presentation and differentiation of underground diseases.

[0012] The present invention has the following beneficial effects:

[0013] 1. By modeling the reflection curvature and using a curvature-guided feature network, we can accurately estimate the curvature of hypercurve reflections formed by diseases, enhance disease features such as weak reflections and thin-layer anomalies that are difficult to identify by traditional methods, and make the diseased area present a clearer abnormal structure in the B-scan image.

[0014] 2. The dual-domain attention mechanism can simultaneously focus on temporal distribution anomalies and frequency band energy anomalies, effectively distinguishing between disease reflections and noises such as strong surface reflections, metal edge interference, and multipath scattering, thus significantly improving the signal-to-noise ratio of weak disease signals.

[0015] 3. The low-rank sparse interference separation network extracts the stable stratum background as a low-rank component and the local reflection of the disease as a sparse component. This makes the isolated strong reflection of cavities, the strong reflection of thin layers of voids, the weak diffusion reflection of loose soil and the wetting anomaly caused by seepage prominent in the form of sparse peaks, thus achieving high-precision disease enhancement. Attached Figure Description

[0016] 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.

[0017] Figure 1 The flowchart illustrates a data augmentation method for detecting underground defects using ground-penetrating radar, as provided by this invention. Detailed Implementation

[0018] 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.

[0019] The first aspect of this embodiment focuses on B-scan data from ground-penetrating radar beneath urban roads. First, the original signal undergoes amplitude normalization, time-window calibration, and DC component removal. Adaptive filtering is then used for noise suppression to obtain a smoother reflection profile. The preprocessed B-scan is input into a curvature-guided feature network. A reflection curvature map is obtained using a second-order gradient operator and local hypercurve fitting. Based on the curvature amplitude, areas related to geological defects are assigned higher weights to highlight potential defect reflections. The data is then input into a temporal convolutional network and a frequency-domain discrete wavelet transform network. A dual-domain attention module (temporal-frequency domain) is constructed to enhance defect reflection energy and suppress reflections from metal pipeline edges, strong surface reflections, and multipath scattering noise. The enhanced feature map is input into a low-rank sparse interference separation network. Matrix factorization is used to represent the layered structure of the formation as a low-rank component, separating cavities, voids, looseness, and seepage reflections into sparse components. By fusing curvature-guided features, dual-domain attention features, and sparse disease features through multiple channels, a disease enhancement map is obtained, which can clearly present the abnormal location and reflection pattern of the disease, providing high-quality input for subsequent disease identification and localization.

[0020] The second aspect of this embodiment addresses shallow road void scenarios. Ground-penetrating radar data is input into a curvature-guided feature network. Curvature estimates are obtained by calculating the second derivative of local reflection curves, and a local polynomial curve fitting method is used to finely model hypercurve reflections. High curvature regions typically correspond to the reflection interfaces at the upper and lower boundaries of the void; therefore, the network automatically increases the feature weights of these regions based on the curvature amplitude, making the originally weak and discontinuous reflections of the void clearer. This method significantly enhances the reflection of thin layers of shallow voids and maintains high reflection sensitivity even under conditions of complex surface waves and noise interference.

[0021] The third aspect of this embodiment employs a dual-domain attention feature extraction network to process seepage data in wet soil environments. First, in the time domain, multi-scale convolution is used to extract the time delay and waveform spread features caused by seepage. Then, in the frequency domain, short-time Fourier transform is used to extract the mid-to-low frequency energy anomalies caused by seepage. The dual-domain attention mechanism dynamically assigns weights based on the correlation between time-domain and frequency-domain features, simultaneously enhancing both the weak reflections from wetting in the frequency domain and the hysteretic reflections in the time domain, while effectively suppressing the strong reflection energy from the surface and random scattering noise. This embodiment verifies the enhancement capability of the dual-domain attention network for weak defects in humid and noisy environments.

[0022] The fourth aspect of this embodiment analyzes data from the complex background environment under the utility tunnel. A low-rank sparse interference separation network is used to perform matrix decomposition on the enhanced features of the dual domain, obtaining the low-rank background component and the sparse defect component. The low-rank component clearly preserves the continuous reflection of the strata and the morphology of the structural surfaces, while the sparse component extracts the isolated strong reflection of cavities, the thin hypercurved reflection of void boundaries, the weak diffusion energy of loose soil, and the local damp anomalies caused by seepage water as independent peak values, greatly improving the visibility of the target area. The separated sparse defect reflections have obvious spatial concentration and structural interpretability, making them easy to use directly in subsequent detection networks.

[0023] 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. An effective signal enhancement and interference separation method for ground-penetrating radar (GPR) data on underground defects, characterized in that, include: S1: Acquire underground ground-penetrating radar B-scan data, and perform amplitude normalization, time window calibration, DC drift removal and noise suppression processing on the data; S2: Construct a curvature guidance feature network to estimate the reflection curvature in the B-scan data and generate curvature guidance features; S3: Construct a dual-domain feature extraction network in the time and frequency domains, extract time-domain features and frequency-domain features from the preprocessed data respectively, and weight the dual-domain features through an attention mechanism; S4: Perform low-rank sparse decomposition on the dual-domain features to obtain low-rank background components and sparse anomaly components. S5: The curvature-guided features, dual-domain features, and sparse anomaly components are fused to obtain enhanced features.

2. The method according to claim 1, characterized in that, The curvature guidance features are obtained through second derivatives, gradient operators, or local curve fitting.

3. The method according to claim 1, characterized in that, The dual-domain feature extraction network achieves feature fusion through a time-domain and frequency-domain attention weighting mechanism.

4. The method according to claim 1, characterized in that, The low-rank sparse decomposition is achieved through matrix decomposition to separate background structure from anomalous features.

5. The method according to claim 1, characterized in that, The method is implemented based on a deep learning framework and trained using multi-source data.

Images