An underground disease intelligent identification method based on multi-physical field feature fusion

Abstract

The application discloses an underground disease intelligent identification method based on multi-physical field feature fusion. The method uses multi-source information such as ground penetrating radar B-scan data, dielectric constant distribution and conductivity distribution, extracts time domain, frequency domain and medium parameter features through a learnable wavelet transform and a multi-network structure, realizes cross-domain feature alignment and importance weighting in an adaptive fusion module, and forms a comprehensive underground target representation. Via a detection network, the positioning and classification of diseases such as cavities, voids, loose and leakage are output. The model is trained by using a joint loss containing classification, position regression, frequency domain consistency and electromagnetic propagation constraints. The method can effectively improve the accuracy and physical credibility of underground disease identification.

Description

Technical Field

[0001] This invention relates to the field of ground-penetrating radar signal processing and intelligent identification technology, and in particular to an intelligent identification method for underground diseases based on multi-physics field feature fusion, which is applicable to the automatic detection and identification of underground diseases such as cavities, voids, looseness and leakage. Background Technology

[0002] Underground defects such as cavities, voids, loosening, and leakage are widespread in infrastructure such as roads, bridges, tunnels, and pipelines. If they are not detected and treated in a timely manner, they can easily lead to structural instability and other safety problems. Therefore, efficient and accurate identification methods are urgently needed. Existing detection technologies mainly include ground-penetrating radar and measurements of dielectric constant and conductivity.

[0003] While ground-penetrating radar (GPR) offers advantages such as being non-destructive and rapid, it is susceptible to the influence of underground media, and its reliance on single-time-domain analysis makes it difficult to fully utilize frequency-domain information, thus limiting identification accuracy. Dielectric constant and conductivity can reflect changes in media properties, but lack spatial structural information, making it difficult to independently locate defects. Existing methods still suffer from problems such as large differences in feature distribution, susceptibility to noise, and lack of physical constraints during multi-source data fusion, resulting in insufficient stability and reliability of identification results. Therefore, it is necessary to propose a method for identifying underground defects that integrates multi-physics information and introduces physical constraints. Summary of the Invention

[0004] In view of this, the present invention provides an intelligent identification method for underground defects based on multi-physics feature fusion. By fusing and modeling multi-source data such as ground-penetrating radar B-scan data, dielectric constant, and conductivity, and combining the frequency domain analysis capability of wavelet transform with the adaptive feature extraction capability of deep learning, the method achieves accurate identification of defects such as cavitation, voids, loosening, and leakage, improving the detection performance and physical reliability of weakly reflective defects. To achieve the above objectives, the present invention adopts the following technical solution:

[0005] This invention provides an intelligent identification method for underground diseases based on multi-physics field feature fusion, comprising the following steps:

[0006] S1: Acquire multi-source data for underground disease analysis, including underground B-scan radar data collected by ground penetrating radar, dielectric constant distribution data and conductivity distribution data related to underground media, wherein the underground diseases include cavitation, voids, loosening and leakage.

[0007] S2: Perform two-dimensional wavelet transform, wavelet packet decomposition, or deformable wavelet transform based on learning strategy on the B-scan radar data to obtain frequency domain sub-band features of different scales, directions, and frequency bands. The wavelet basis, decomposition path, or frequency band weight is adaptively determined through attention mechanism or structure search strategy to enhance the salience of the disease reflection area.

[0008] S3: Input the dielectric constant distribution data, conductivity distribution data and original B-scan time domain data into the corresponding feature extraction network to generate: a first feature reflecting the spatial variation of dielectric parameters, a second feature reflecting the electromagnetic dissipation mechanism, a third feature reflecting the radar reflection structure mode, and a fourth feature generated from the frequency domain sub-band, wherein the feature extraction network is a convolutional neural network, a residual network, a converter network, a time-frequency joint network or a physical guidance network;

[0009] S4: Input the first feature, second feature, third feature and fourth feature into the adaptive feature fusion module for fusion. The adaptive feature fusion module includes a weight prediction branch, a spatial attention module and a cross-domain feature alignment module. The weight prediction branch is used to generate learnable weights for each feature. The spatial attention module is used to enhance the signal of the underground target area and suppress background noise. The cross-domain feature alignment module is used to eliminate the feature scale differences between the dielectric parameter domain, frequency domain and time domain, thereby obtaining the fused feature representation.

[0010] S5: Input the fused features into the detection head network and output the location parameters and category labels of the underground target. The location parameters include lateral location, depth or bounding box information, and the category labels include cavitation, void, looseness and leakage.

[0011] S6: The model is trained using a joint loss function, which includes classification loss, location regression loss and frequency domain consistency loss, wherein the frequency domain consistency loss is used to constrain the consistency of the response of the diseased area in the time domain and the frequency domain.

[0012] In a preferred embodiment, the wavelet transform in step S2 employs a learnable wavelet basis, wavelet packet optimal basis search, or adaptive subband selection strategy to enhance the frequency domain response of the diseased area.

[0013] In a preferred embodiment, the adaptive feature fusion module in step S4 further includes a mutual information maximization unit for improving the correlation between different physical field features.

[0014] In a preferred embodiment, the joint loss function in step S6 further includes an electromagnetic propagation constraint loss, used to make the fused features conform to the laws of electromagnetic wave propagation and attenuation.

[0015] The present invention has the following significant beneficial effects:

[0016] 1. This invention utilizes the spatial distribution information of dielectric constant and conductivity to characterize the properties of underground media from two aspects: electromagnetic wave propagation speed and energy dissipation. This enables the model to more accurately distinguish different types of defects such as cavities, voids, looseness, and leakage, thereby improving the accuracy and stability of identification.

[0017] 2. By introducing learnable wavelet bases, wavelet packet optimal basis search, and subband attention mechanism, the frequency domain representation can automatically focus on the frequency band most sensitive to disease reflection, suppress irrelevant noise, and solve the problem that traditional fixed tree structure wavelet packets cannot take into account different scenarios.

[0018] 3. The adaptive feature fusion module proposed in this invention can automatically perform scale alignment and distribution correction between the time domain, frequency domain and dielectric parameter domain, and enhance the complementarity between multi-source features through mutual information maximization units, thereby generating more physically consistent fused features.

[0019] 4. The frequency domain consistency loss and electromagnetic propagation constraint loss introduced in model training can effectively restrict the model output to follow the physical laws of electromagnetic wave attenuation and reflection, making the detection results more reliable and interpretable, and reducing false alarms and false detections. Attached Figure Description

[0020] 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.

[0021] Figure 1 The flowchart shows a method for intelligent identification of underground diseases based on multi-physics field feature fusion provided by the present invention.

[0022] Specific implementation method

[0023] 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.

[0024] This embodiment provides a method for intelligent identification of underground diseases based on multi-physics feature fusion. Ground-penetrating radar (GPR) B-scan data of underground areas is collected, and the dielectric constant and conductivity distributions describing the properties of the underground medium are acquired simultaneously. The obtained B-scan data is input into a two-dimensional wavelet transform and learnable wavelet packet decomposition module, which automatically selects the optimal sub-band that highlights the reflection characteristics of the disease through an attention mechanism. Based on this, a multi-path feature extraction model composed of a convolutional network and a physical guidance network is established to extract dielectric constant features, conductivity features, time-domain radar features, and frequency-domain wavelet features, which are then fed into an adaptive feature fusion module. In the fusion module, multi-source features are uniformly modeled through weighted prediction branches, spatial attention, and cross-domain feature alignment mechanisms to generate a fused feature representation. Finally, the fused features are input into a detection head network, which outputs the location parameters and category labels of the underground diseases. The overall model is trained using a joint loss of classification loss, location regression loss, and frequency domain consistency loss, thereby achieving intelligent identification of cavitation, voids, looseness, and seepage diseases.

[0025] The second aspect of this embodiment employs deformable wavelet transform for frequency domain feature extraction. Addressing the limitation of traditional wavelet packet decomposition's fixed tree structure in adapting to the frequency differences of various disease types, this invention introduces a learnable wavelet basis and an adaptive subband selection module, enabling the system to automatically determine the optimal wavelet decomposition path and subband weights during the training phase. The system first performs initial wavelet decomposition on the B-scan data, then further subdivides or merges a subband through structure search-based control branch selection, prioritizing the retention of the most significant frequency bands reflected by the disease. Subsequently, the optimized wavelet subbands are jointly input into the feature fusion module along with dielectric constant, conductivity, and the original B-scan time-domain features. The fusion module utilizes mutual information maximization units to enhance the correlation between time-frequency-medium parameters and spatial attention to enhance the signal in the disease area, ensuring the final detection model maintains high detection accuracy even in scenarios with weak reflection and deeply buried diseases.

[0026] The third aspect of this embodiment introduces electromagnetic propagation constraint loss to enhance the physical consistency of the detection model. The collected B-scan data, along with the acquired dielectric constant and conductivity distributions, are preprocessed. A physical guidance network is then constructed to map the dielectric constant and conductivity features to the corresponding electromagnetic propagation model parameters, constraining the spatial attenuation trend of radar reflection intensity. During model training, a frequency domain consistency loss is constructed by calculating the response differences of the affected area in the time and frequency domains. This is combined with electromagnetic attenuation constraints derived from Maxwell's equations, ensuring that the feature learning process conforms to the laws of energy dissipation and medium reflection. Finally, the model outputs the location and type of underground defects via the detection head, achieving high-reliability identification of complex defects such as cavities, voids, looseness, and leakage. This implementation significantly reduces false positives due to noise and improves the model's cross-regional generalization ability.

[0027] 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 method for intelligent identification of underground diseases based on multi-physics field feature fusion, characterized in that, include: S1: Acquire multi-source data for underground disease analysis, including underground B-scan radar data collected by ground penetrating radar, dielectric constant distribution data and conductivity distribution data related to underground media, wherein the underground diseases include cavitation, voids, loosening and leakage. S2: Perform two-dimensional wavelet transform, wavelet packet decomposition, or deformable wavelet transform based on learning strategy on the B-scan radar data to obtain frequency domain sub-band features of different scales, directions, and frequency bands. The wavelet basis, decomposition path, or frequency band weight is adaptively determined through attention mechanism or structure search strategy to enhance the salience of the disease reflection area. S3: Input the dielectric constant distribution data, conductivity distribution data and original B-scan time domain data into the corresponding feature extraction network to generate: a first feature reflecting the spatial variation of dielectric parameters, a second feature reflecting the electromagnetic dissipation mechanism, a third feature reflecting the radar reflection structure mode, and a fourth feature generated from the frequency domain sub-band, wherein the feature extraction network is a convolutional neural network, a residual network, a converter network, a time-frequency joint network or a physical guidance network; S4: Input the first feature, second feature, third feature and fourth feature into the adaptive feature fusion module for fusion. The adaptive feature fusion module includes a weight prediction branch, a spatial attention module and a cross-domain feature alignment module. The weight prediction branch is used to generate learnable weights for each feature. The spatial attention module is used to enhance the signal of the underground target area and suppress background noise. The cross-domain feature alignment module is used to eliminate the feature scale differences between the dielectric parameter domain, frequency domain and time domain, thereby obtaining the fused feature representation. S5: Input the fused features into the detection head network and output the location parameters and category labels of the underground target. The location parameters include lateral location, depth or bounding box information, and the category labels include cavitation, void, looseness and leakage. S6: The model is trained using a joint loss function, which includes classification loss, location regression loss and frequency domain consistency loss, wherein the frequency domain consistency loss is used to constrain the consistency of the response of the diseased area in the time domain and the frequency domain.

2. The method according to claim 1, characterized in that, The wavelet transform employs learnable wavelet bases, wavelet packet optimal base search, or adaptive subband selection strategies to enhance the frequency domain response of the diseased area.

3. The method according to claim 1, characterized in that, The adaptive feature fusion module also includes a mutual information maximization unit, which is used to improve the correlation between different physical field features.

4. The method according to claim 1, characterized in that, The joint loss function also includes electromagnetic propagation constraint loss, which is used to make the fused features conform to the laws of electromagnetic wave propagation and attenuation.

Images