A tunnel multi-modal intelligent detection method based on multi-sensor cooperation

Abstract

The application discloses a kind of based on multi-sensor cooperation's tunnel multi-modal intelligent detection method, comprising: through the multi-dimensional original detection data set of tunnel structure surface, geological body inside and environment state related data of multiple sensing devices collected by tunnel global multi-sensing cooperation intelligent detection early warning platform deployment, call geological texture door control chart attention network extraction and enhance geological texture feature, adopt tunnel section deformation inversion algorithm to analyze section collection data to characterize section deformation state, utilize radar image texture reconstruction algorithm to process radar detection data, restore tunnel concealed part structure state.Processed multi-modal feature data is fused and associated, constructs global detection feature space, and outputs tunnel structure state detection result and abnormal signal identification after platform analysis and judgment.The application realizes the globalization, precision and intelligentization of tunnel detection, and provides reliable technical support for tunnel structure health monitoring.

Description

Technical Field

[0001] This invention relates to the field of intelligent tunnel detection technology, and in particular to a multimodal intelligent tunnel detection method based on multi-sensor collaboration. Background Technology

[0002] With the high-quality advancement of infrastructure construction, the application scale of tunnels in transportation networks continues to expand, making the need for safety monitoring during their construction and operation increasingly urgent. As semi-enclosed, narrow spaces, tunnels face complex challenges such as insufficient lighting, humid environments, and concealed structural defects. Traditional monitoring methods are no longer sufficient to meet the requirements of comprehensive, high-frequency, and precise monitoring. Simultaneously, the increasing number of aging tunnels and the resulting pressure from growing traffic flow highlight the growing importance of tunnel structural health monitoring, necessitating breakthroughs in intelligent and collaborative monitoring technologies to achieve comprehensive perception and timely early warning of tunnel structural conditions.

[0003] Existing tunnel inspection technologies are mainly divided into two categories: traditional manual inspection and single-function automated inspection. Manual inspection relies on workers walking around, assessing the tunnel's structural condition through visual inspection combined with simple instruments. Automated inspection, on the other hand, often uses single-type sensing devices, or integrates multiple modules but lacks an effective collaborative mechanism. It collects data on tunnel surface features and internal response signals separately, processes the data using independent data analysis methods, and then integrates the results manually or through a simple system. Some technologies attempt to introduce sensor networks, but data processing is often limited to single-dimensional analysis, lacking in-depth correlation and fusion analysis of multi-source data, making it difficult to form a comprehensive understanding of the tunnel's structural condition.

[0004] Existing technologies suffer from two main drawbacks: First, insufficient collaboration among multi-source data sources. Data generated by various detection devices is scattered across different systems, lacking a unified analysis and fusion architecture. This results in severe data silos, hindering the full utilization of the complementary advantages of multiple sensors. Consequently, the characterization of tunnel structure is not comprehensive or accurate, particularly in effectively reconstructing structural information from concealed areas. Second, the targeted nature of feature extraction and data processing is weak. There is a lack of intelligent processing mechanisms specifically adapted to tunnel geological textures, cross-sectional deformation, and radar imagery. The ability to capture subtle features in complex environments is limited, leading to issues such as lost details and missed detections. Furthermore, the automation and intelligence levels of data processing and analysis are low, affecting the reliability and timeliness of detection results. Summary of the Invention

[0005] To overcome the shortcomings and deficiencies of existing technologies, this invention provides a multi-sensor collaborative intelligent detection method for tunnels with multiple modes.

[0006] The technical solution adopted in this invention is a multi-sensor collaborative intelligent detection method for tunnels, comprising the following steps: S1, collecting surface feature data of tunnel structures, internal response signals of geological bodies, and environmental state parameters through multiple types of sensing devices deployed in a tunnel-wide multi-sensor collaborative intelligent detection and early warning platform, and constructing a multi-dimensional original detection dataset; S2, calling a geological texture gating graph attention network to extract and enhance features from the collected geological body-related data, dynamically adjusting the weight allocation of different geological texture features through a gating mechanism to strengthen the effective representation capability of geological information; S3, using a tunnel cross-section deformation inversion algorithm to analyze the tunnel cross-section deformation. S2 involves analyzing the collected data to establish a mapping relationship between cross-sectional deformation and sensor data, thus characterizing the cross-sectional deformation state; S3 involves processing the radar detection data using a radar image texture reconstruction algorithm to reconstruct the texture feature map of the tunnel's internal structure and restore the structural state of the tunnel's hidden parts; S4 involves fusing and associating the multimodal feature data processed from S2 to S4 based on a multi-sensor collaborative tunnel multimodal intelligent detection architecture to construct a full-domain detection feature space; and S6 involves analyzing and judging the fused feature data through a full-domain multi-sensor collaborative intelligent detection and early warning platform for the tunnel, outputting the tunnel structural state detection results and abnormal signal identifiers.

[0007] Furthermore, the expression for the geological texture gating map attention network is:

[0008]

[0009]

[0010] in, This is the attention weight matrix for the gating graph. The number of geological texture feature channels. It is the Sigmoid activation function. , For the network's learnable weight matrix, For the first Original geological texture features of the passage. For bias vectors, For the first The graph adjacency matrix corresponding to the channel features This is an element-wise multiplication operation. This is the aggregated original geological texture feature matrix. The enhanced geological texture feature matrix is ​​represented by ReLU and Softmax as activation functions.

[0011] Furthermore, the expression for the tunnel cross-section deformation inversion algorithm is as follows:

[0012]

[0013]

[0014] in, This is the cross-sectional deformation offset matrix. This refers to the number of sampling points for cross-sectional inspection. These are the deformation inversion coefficients. For the first Sensor measurement data at sampling points, For the first Gradient values ​​of the sampled data points The correlation matrix of sampling point locations, To obtain the true cross-sectional deformation data through inversion, To perform the matrix inversion operation, For deformation inversion regularization matrix, for The transpose of .

[0015] Furthermore, the expression for the radar image texture reconstruction algorithm is:

[0016]

[0017] in, The reconstructed radar image texture matrix, These represent the number of rows and columns of the radar image, respectively. For texture reconstruction weight coefficients, For radar imagery Line 1 The original pixel values ​​of the column. For Gabor filtering operators, For the first Line 1 The filtering direction angle of the column pixels, These are the filter wavelength parameters. This is the texture reconstruction correction matrix.

[0018] Furthermore, the multi-type sensing devices include a distributed fiber optic sensing module, a millimeter-wave radar detection module, a three-dimensional laser scanning module, an ultrasonic sensing array, and an infrared thermal imaging module. Each module establishes a communication connection with the tunnel full-domain multi-sensor collaborative intelligent detection and early warning platform through a 5G industrial gateway. The number of feature channels in the geological texture gating map attention network is set to 64 to 256, the number of neurons in the network hidden layer is 256 to 1024, and the number of iterations is 100 to 500. The sampling point interval of the tunnel cross-section deformation inversion algorithm is set to 0.1 to 0.5, and the regularization matrix is... The diagonal elements range from 0.01 to 0.1; the Gabor filter orientation angle of the radar image texture reconstruction algorithm. The value range is 0 to Filter wavelength The value ranges from 2 to 8.

[0019] Further, S2 includes the following sub-steps: S21, performing feature channel separation on the collected geological body-related data, dividing the data into different subsets according to the geological texture type, and establishing an association mapping between each subset and the geological structure attributes; S22, inputting each data subset into the feature encoding layer of the geological texture gating graph attention network, extracting local texture features through convolution operations, and generating an initial feature map; S23, using a gating mechanism to adjust the weights of the initial feature map, dynamically allocating attention resources according to the recognizability and relevance of the geological texture, and suppressing redundant information interference; S24, performing global association modeling on the weighted feature map through a graph neural network, fusing local and global feature information, and outputting enhanced geological texture feature data.

[0020] Further, step S3 includes the following sub-steps: S31, calibrating the sampling point coordinates of the tunnel cross-section data, establishing a global coordinate system for the cross-section, and determining the spatial position parameters of each sampling point; S32, extracting the sensor measurement data of each sampling point, calculating the statistical characteristics and gradient information of the data, and constructing a sampling point data matrix; S33, inputting the data matrix into the tunnel cross-section deformation inversion algorithm, substituting the preset inversion coefficients and correlation matrix, and calculating the cross-section deformation offset; S34, correcting the inversion results through matrix inversion and regularization processing to obtain the tunnel cross-section deformation characterization data.

[0021] Further, S4 includes the following sub-steps: S41, performing pixel-level alignment on the radar detection data to eliminate texture shifts caused by differences in detection viewing angles and constructing a uniform-scale original radar image; S42, extracting grayscale features and texture frequency information from the original radar image to determine the parameter range and reconstruction weight coefficients of the Gabor filter; S43, substituting the radar image texture reconstruction algorithm, and reconstructing the detailed texture of the radar image through filtering and weighted calculation; S44, performing consistency verification on the reconstructed texture map, correcting texture distortion areas, and outputting a complete texture feature map of the tunnel's internal structure.

[0022] Further, S5 includes the following sub-steps: S51, unifying the dimensions of the multimodal feature data output from S2 to S4, converting geological texture features, cross-sectional deformation data, and radar texture maps into feature vectors of the same dimension; S52, constructing a multimodal feature association matrix, calculating the similarity coefficient between different types of features, and determining the feature fusion weights; S53, using a weighted fusion strategy to perform aggregation operations on the multimodal feature vectors to generate preliminary fused features; S54, removing redundant information in the fused features through a feature filtering mechanism, retaining the calibrated feature components, and constructing a global detection feature space.

[0023] A multi-sensor collaborative intelligent tunnel multimodal detection method is proposed. This method is implemented through a multi-sensor collaborative intelligent tunnel multimodal system, comprising: a tunnel multimodal data acquisition and transmission unit, a geological texture feature intelligent extraction and enhancement unit, a tunnel cross-section deformation inversion calculation unit, a radar image texture adaptive reconstruction unit, a multimodal feature collaborative fusion processing unit, and a full-domain detection result analysis and labeling unit. The tunnel multimodal data acquisition and transmission unit establishes bidirectional data transmission mechanisms with the geological texture feature intelligent extraction and enhancement unit, the tunnel cross-section deformation inversion calculation unit, and the radar image texture adaptive reconstruction unit. According to the transmission link, the collected multi-dimensional data is synchronously distributed to each feature processing unit. The output ends of the geological texture feature intelligent extraction and enhancement unit, the tunnel cross-section deformation inversion calculation unit, and the radar image texture adaptive reconstruction unit are all connected to the input end of the multi-modal feature collaborative fusion processing unit to transmit their respective processed feature data. The output end of the multi-modal feature collaborative fusion processing unit is connected to the global detection result analysis and identification unit to transmit the fused global feature spatial data to this unit for analysis and judgment. Finally, the global detection result analysis and identification unit outputs the tunnel structure status detection results and abnormal signal identification.

[0024] Beneficial Effects: This invention proposes a multi-sensor collaborative intelligent tunnel detection method. It collaboratively collects multi-dimensional tunnel data using multiple types of sensors. Through intelligent algorithms and a collaborative architecture, it integrates data from various sensors, including distributed fiber optics and millimeter-wave radar, using a tunnel-wide multi-sensor collaborative intelligent detection and early warning platform. This breaks down data silos and constructs a full-domain detection feature space through multi-modal feature fusion, solving the problem of insufficient multi-source data collaboration in traditional technologies. It fully leverages the complementary advantages of each sensor to comprehensively and accurately depict the tunnel's structural state and clearly reconstruct information about hidden areas. Simultaneously, it specifically designs a geological texture gating map attention network, a tunnel cross-section deformation inversion algorithm, and a radar image texture reconstruction algorithm to selectively extract and process key tunnel features, enhance the ability to capture weak features, avoid detail loss and missed detections, and improve the automation and intelligence level of data processing. This effectively compensates for the shortcomings of existing technologies in feature extraction, which lacks specificity. The multi-unit collaborative architecture further ensures the reliability and timeliness of the detection results. Attached Figure Description

[0025] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention.

[0026] Figure 2 This is a flowchart of method step S2 of the present invention;

[0027] Figure 3 This is a flowchart of method step S3 of the present invention;

[0028] Figure 4This is a flowchart of method step S4 of the present invention;

[0029] Figure 5 This is a flowchart of step S5 of the method of the present invention. Detailed Implementation

[0030] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0031] like Figure 1 As shown, a multi-sensor collaborative intelligent detection method for tunnels includes the following steps: S1, collecting surface feature data of the tunnel structure, internal response signals of the geological body, and environmental state parameters through multiple types of sensing devices deployed in a tunnel-wide multi-sensor collaborative intelligent detection and early warning platform, and constructing a multi-dimensional original detection dataset; S2, calling a geological texture gating graph attention network to extract and enhance features from the collected geological body-related data, and dynamically adjusting the weight allocation of different geological texture features through a gating mechanism to enhance the ability to effectively represent geological information; S3, using a tunnel cross-section deformation inversion algorithm to process the collected tunnel cross-section data. The process involves: S4, analyzing and establishing a mapping relationship between cross-sectional deformation and sensor data to characterize the state of cross-sectional deformation; S5, using a radar image texture reconstruction algorithm to process radar detection data, reconstructing the texture feature map of the tunnel's internal structure, and restoring the structural state of the tunnel's hidden parts; S6, based on a multi-sensor collaborative tunnel multi-modal intelligent detection architecture, fusing and associating the multi-modal feature data processed from S2 to S4 to construct a full-domain detection feature space; and S7, analyzing and judging the fused feature data through a full-domain multi-sensor collaborative intelligent detection and early warning platform for the tunnel, and outputting the tunnel structural state detection results and abnormal signal identifiers.

[0032] Step S1 is implemented through a multi-sensor collaborative intelligent detection and early warning platform covering the entire tunnel. Distributed fiber optic sensing modules, millimeter-wave radar detection modules, 3D laser scanning modules, ultrasonic sensor arrays, and infrared thermal imaging modules are deployed at preset intervals inside the tunnel. Each module establishes a stable communication connection with the platform via a 5G industrial gateway to ensure real-time data transmission. During the detection process, the distributed fiber optic sensing modules collect strain-related characteristic data of the tunnel structure surface in a continuous sampling mode. The millimeter-wave radar detection module focuses on collecting response signals inside the geological body. The infrared thermal imaging module simultaneously captures environmental parameters such as temperature. The 3D laser scanning module and ultrasonic sensor array supplement data related to structural geometric features and internal defects, respectively. The sampling frequency and data accuracy of each sensor are strictly controlled during the data acquisition process to ensure the synchronization of timestamps for different types of data. Finally, all collected tunnel structure surface feature data, geological body internal response signals, and environmental parameters are summarized, categorized and labeled according to data type and collection location, constructing a multi-dimensional, multi-source raw detection dataset. This provides a comprehensive, complete, and high-quality data foundation for subsequent data analysis and processing steps, ensuring the accuracy and effectiveness of the entire detection process.

[0033] In step S2, the geological texture gating attention network is first invoked. This network has 64 to 256 feature channels, 256 to 1024 hidden layer neurons, and undergoes 100 to 500 iterations of training to ensure sufficient feature extraction and processing capabilities. After inputting the geological body-related data collected in step S1 into the network, the data undergoes preliminary processing via the network's built-in feature channel separation mechanism. Different data subsets are divided according to geological texture type, and a mapping relationship is established between these subsets and geological structural attributes. Subsequently, the network dynamically adjusts the weight allocation of different geological texture features through a gating mechanism. Based on the discriminability and relevance of each texture feature, attention resources are prioritized for effective geological information that significantly impacts the detection results, while suppressing redundant information and interference signals. During weight adjustment, the gating mechanism dynamically adjusts the allocation ratio based on real-time data features, ensuring that effective geological information is sufficiently enhanced. Through feature extraction and enhancement processing by the network, the representational ability of geological information is significantly improved, providing accurate and effective geological texture feature data for subsequent tunnel structural state analysis. This ensures that subsequent detection steps can conduct in-depth analysis based on high-quality feature data.

[0034] In step S3, the tunnel cross-section data collected in step S1 is preprocessed. A global coordinate system for the tunnel cross-section is established using coordinate calibration technology, clarifying the spatial location parameters of each sampling point. The sampling point interval is strictly set to 0.1 to 0.5 to ensure uniform distribution and coverage of the entire cross-section area. Subsequently, a tunnel cross-section deformation inversion algorithm is used to analyze the calibrated cross-section data. In this algorithm, the diagonal elements of the regularization matrix range from 0.01 to 0.1. By appropriately setting the regularization parameters, overfitting during the inversion process is avoided. During algorithm implementation, a mapping model between cross-section deformation and sensor data is first established. This model is used to process the sensor data at each sampling point. Combined with preset deformation inversion coefficients and the correlation information between sampling point locations, the deformation of each part of the tunnel cross-section is accurately calculated. Through multiple rounds of iterative computation and error correction of the algorithm, the accuracy of the mapping relationship model is gradually optimized, and finally the accurate characterization of the deformation state of the tunnel cross section is achieved. Deformation data of each region of the cross section is obtained, providing core data support for judging whether there is deformation anomaly in the tunnel structure and the degree of anomaly, and ensuring that potential structural safety hazards can be detected in a timely manner.

[0035] In step S4, the radar detection data collected in step S1 is first pixel-level aligned to eliminate texture shift caused by differences in detection viewing angles, constructing a uniform-scale original radar image to lay the foundation for subsequent texture reconstruction. Then, a radar image texture reconstruction algorithm is used to perform depth processing on the original radar image. In this algorithm, the Gabor filter direction angle ranges from 0 to π, and the filter wavelength ranges from 2 to 8. By appropriately setting the filter parameters, the algorithm's ability to extract texture features from the radar image is enhanced. During the algorithm implementation, the grayscale features and texture frequency information of the original radar image are first extracted to determine the texture reconstruction weight coefficients. Then, through filtering and weighted calculations, the detailed textures in the radar image are reconstructed, focusing on restoring the texture features of the tunnel's internal structure. During the reconstruction process, multi-dimensional texture information is integrated using quadrature operations to ensure that the reconstructed texture feature map accurately reflects the tunnel's internal structural state. Especially for concealed parts of the tunnel, the algorithm's precise processing clearly restores their structural details, compensating for the shortcomings of traditional detection techniques in detecting concealed parts, and providing detailed and accurate texture data support for a comprehensive understanding of the tunnel's internal structural state.

[0036] Step S5 is implemented based on a multi-sensor collaborative tunnel multimodal intelligent detection architecture. First, the geological texture feature data, tunnel cross-section deformation data, and radar image texture feature maps processed in steps S2 to S4 are subjected to dimensional unification processing. This converts multimodal feature data of different types and dimensions into feature vectors of the same dimension, ensuring data format consistency for fusion and association. Then, a multimodal feature association matrix is ​​constructed. By calculating the similarity coefficients between different types of features, the weight allocation ratio of each type of feature in the fusion process is determined. The weight allocation is strictly set according to the degree of influence of the features on the detection results. A weighted fusion strategy is used to aggregate the multimodal feature vectors after unification, generating preliminary fused features. Based on this, a feature filtering mechanism is used to remove redundant information from the fused features, retaining key feature components and avoiding interference from redundant data on the detection results. The entire fusion process strictly follows the logic of multimodal data collaborative fusion, fully explores the inherent correlation between different types of feature data, and finally constructs a full-domain detection feature space. This feature space includes key information on the surface, interior and environment of the tunnel structure, providing comprehensive and systematic feature data support for the analysis and judgment in step S6, and ensuring the comprehensiveness and accuracy of the detection results.

[0037] Step S6 is implemented through the tunnel's full-domain multi-sensor collaborative intelligent detection and early warning platform. After receiving the full-domain detection feature space data constructed in step S5, the platform activates its built-in intelligent analysis and judgment system to perform in-depth analysis of the data. During the analysis, the system verifies and comprehensively evaluates various feature data in the full-domain detection feature space according to preset detection standards and judgment logic. It focuses on analyzing the tunnel structural status reflected by geological texture features, cross-sectional deformation data, and radar image texture features to identify the presence of abnormal signals and structural defects. The system strictly adheres to relevant detection specifications during the analysis process, combining the collaborative correlation information of multi-modal feature data to accurately determine the health status of the tunnel structure. Abnormal signals detected during the detection process are classified and labeled according to parameters such as signal strength and impact range. Finally, the platform outputs a detection report including the overall tunnel structural status assessment results, detailed detection data for each part, and the specific location and characteristics of abnormal signals. The abnormal signal identification is clear and unambiguous, providing tunnel maintenance personnel with accurate and detailed detection results, facilitating timely and targeted maintenance measures, ensuring tunnel operational safety, and providing data reference for subsequent detection scheme optimization.

[0038] Preferably, the expression for the geological texture gating map attention network is:

[0039]

[0040]

[0041] in, This is the attention weight matrix for the gating graph. The number of geological texture feature channels. It is the Sigmoid activation function. , For the network's learnable weight matrix, For the first Original geological texture features of the passage. For bias vectors, For the first The graph adjacency matrix corresponding to the channel features This is an element-wise multiplication operation. This is the aggregated original geological texture feature matrix. The enhanced geological texture feature matrix is ​​represented by ReLU and Softmax as activation functions.

[0042] Specifically, the geological texture gated graph attention network, based on the multi-channel correlation of geological texture features and the need for effective information enhancement, first analyzes the feature distribution patterns of multi-channel original geological texture data to determine an architecture combining gating and graph attention mechanisms. First, linear transformations and activation processes are performed on the features of each channel. Multiplication operations are introduced to integrate multi-channel weights to highlight the contribution of key channels. Then, element-wise multiplication is used to achieve coordinated adjustment of gating and attention weights. Based on the attention weights obtained in the first step, a feature enhancement model is constructed by combining the residual connections of the original feature matrix with adaptive adjustment of the gating weights to avoid gradient vanishing and enhance effective features. The number of feature channels is set to 64 to 256, the number of neurons in the hidden layer of the network is 256 to 1024, the number of training iterations is 100 to 500, and the bias vector is determined through initialization training optimization. The implementation of this network requires first separating geological body-related data by channel and then inputting it into the feature encoding layer. Local features are extracted through convolution operations, and attention resources are dynamically allocated through a gating mechanism to suppress redundant information. Finally, enhanced features are output through global correlation modeling of graph neural networks, which accurately captures key information of geological texture, improves the accuracy and distinguishability of geological feature representation, provides high-quality geological texture data support for subsequent tunnel structure status analysis, and ensures effective identification of geological structural defects during the detection process.

[0043] Preferably, the expression for the tunnel cross-section deformation inversion algorithm is:

[0044]

[0045]

[0046] in, This is the cross-sectional deformation offset matrix. This refers to the number of sampling points for cross-sectional inspection. These are the deformation inversion coefficients. For the first Sensor measurement data at sampling points, For the first Gradient values ​​of the sampled data points The correlation matrix of sampling point locations, To obtain the true cross-sectional deformation data through inversion, To perform the matrix inversion operation, For deformation inversion regularization matrix, for The transpose of .

[0047] Specifically, the tunnel cross-section deformation inversion algorithm is based on the modeling requirement of the mapping relationship between cross-section deformation and sensor data. Combining the spatial correlation of sampling points and data error correction logic, it integrates the contributions of all cross-section sampling points by weighted summation of the measurement data and gradient information of each sampling point, and then multiplies this summation with the positional correlation matrix to establish a preliminary correlation between deformation offset and sensor data. Based on the obtained deformation offset, overfitting in the inversion process is suppressed through a regularization matrix. The inverse solution of the mapping relationship is achieved by combining matrix transpose and inversion operations, thereby restoring the true cross-section deformation data. The sampling point interval is set to 0.1 to 0.5, the diagonal elements of the regularization matrix range from 0.01 to 0.1, and the deformation inversion coefficients are trained and calibrated using a large amount of measured data. The implementation of this algorithm requires first calibrating the coordinates of the tunnel cross-section data, constructing a sampling point data matrix, inputting it into the algorithm, and then calculating the cross-sectional deformation offset by substituting preset parameters. The result is then corrected through regularization and matrix inversion to establish a quantitative relationship between the sensor data and the cross-sectional deformation. This overcomes the problem of insufficient inversion accuracy in traditional algorithms, enables accurate characterization of the tunnel cross-sectional deformation state, provides core data support for judging whether there are deformation anomalies in the tunnel structure, and ensures the timely detection of structural safety hazards.

[0048] Preferably, the expression for the radar image texture reconstruction algorithm is:

[0049]

[0050] in, The reconstructed radar image texture matrix, These represent the number of rows and columns of the radar image, respectively. For texture reconstruction weight coefficients, For radar imagery Line number The original pixel values ​​of the column. For Gabor filtering operators, For the first Line number The filtering direction angle of the column pixels, For the filter wavelength parameter, This is the texture reconstruction correction matrix.

[0051] Specifically, the radar image texture reconstruction algorithm is based on the multi-pixel correlation characteristics and detail restoration requirements of radar image textures. By analyzing the mapping rules between original radar pixel values ​​and texture features, it introduces a Gabor filter operator to capture texture information from different directions and wavelengths. A dual quadrature operation is used to integrate the weight contributions of row and column pixels. Combined with a texture reconstruction correction matrix, it corrects deviations generated during filtering and weighting, constructing a complete texture reconstruction model. The Gabor filter direction angle ranges from 0 to π, and the filter wavelength ranges from 2 to 8. The texture reconstruction weight coefficients are determined through image texture similarity analysis. The implementation of this algorithm requires first aligning the radar detection data at the pixel level to construct a uniform-scale original radar image. Image grayscale features and texture frequency information are extracted to determine the parameter range. These are then substituted into the algorithm for filtering and weighting operations. Finally, consistency checks correct texture distortion areas, restoring the detailed texture of the tunnel's internal structure in the radar image, especially for concealed structural information. This overcomes the defects of texture blurring and detail loss in traditional radar image processing, providing clear and accurate texture data for a comprehensive understanding of the tunnel's internal structural state, supporting the accuracy of subsequent detection and analysis.

[0052] Preferably, the multi-type sensing devices include a distributed fiber optic sensing module, a millimeter-wave radar detection module, a three-dimensional laser scanning module, an ultrasonic sensing array, and an infrared thermal imaging module. Each module establishes a communication connection with the tunnel full-domain multi-sensor collaborative intelligent detection and early warning platform through a 5G industrial gateway. The number of feature channels in the geological texture gating map attention network is set to 64 to 256, the number of neurons in the hidden layer is 256 to 1024, and the number of iterations is 100 to 500. The sampling point interval of the tunnel cross-section deformation inversion algorithm is set to 0.1 to 0.5, and the regularization matrix is... The diagonal elements range from 0.01 to 0.1; the Gabor filter orientation angle of the radar image texture reconstruction algorithm. The value range is 0 to Filter wavelength The value ranges from 2 to 8.

[0053] Preferred, such as Figure 2As shown, S2 includes the following sub-steps: S21, performing feature channel separation on the collected geological body-related data, dividing the data into different subsets according to the geological texture type, and establishing the association mapping between each subset and the geological structure attributes; S22, inputting each data subset into the feature encoding layer of the geological texture gating graph attention network, extracting local texture features through convolution operations, and generating an initial feature map; S23, using a gating mechanism to adjust the weights of the initial feature map, dynamically allocating attention resources according to the recognizability and relevance of the geological texture, and suppressing redundant information interference; S24, performing global association modeling on the weighted feature map through a graph neural network, fusing local and global feature information, and outputting enhanced geological texture feature data.

[0054] Specifically, step S2 achieves precise extraction and enhancement of geological texture features through four sub-steps. S21 first performs feature channel separation on the geological body-related data collected in step S1, dividing the data into subsets based on different types of geological textures such as lithological texture and fracture texture. Simultaneously, it establishes a mapping relationship between each subset and attributes such as tunnel surrounding rock structure and geological structure, ensuring a high degree of consistency between data classification and actual geology. S22 inputs each data subset into the feature encoding layer of a geological texture gating map attention network. This network has 64 to 256 feature channels and extracts local texture features through 3×3 convolutional kernels, generating an initial feature map that includes basic texture information, ensuring no loss of local details. S23 uses a gating mechanism to adjust the weights of the initial feature map. The number of neurons in the network's hidden layer is 256 to 1024. Attention resources are dynamically allocated based on the geological texture's recognition score and structural correlation coefficient, adaptively adjusting the weight values ​​within the range of 0.1 to 0.9, focusing on enhancing effective information and suppressing redundant interference. S24 uses a graph neural network to perform global correlation modeling on the weighted feature map, with 100 to 500 iterations of training. It integrates local texture details with global geological correlation information and outputs enhanced geological texture feature data. This process improves the accuracy of geological feature representation through step-by-step parameter control and process optimization, providing high-quality data support for subsequent detection and analysis.

[0055] Preferred, such as Figure 3 As shown, step S3 includes the following sub-steps: S31, calibrating the sampling point coordinates of the tunnel cross-section data, establishing a global coordinate system for the cross-section, and determining the spatial position parameters of each sampling point; S32, extracting the sensor measurement data of each sampling point, calculating the statistical characteristics and gradient information of the data, and constructing a sampling point data matrix; S33, inputting the data matrix into the tunnel cross-section deformation inversion algorithm, substituting the preset inversion coefficients and correlation matrix, and calculating the cross-section deformation offset; S34, correcting the inversion results through matrix inversion and regularization processing to obtain the tunnel cross-section deformation characterization data.

[0056] Specifically, step S3 achieves the inversion of tunnel cross-sectional deformation through four sub-steps. S31 first calibrates the sampling point coordinates of the tunnel cross-sectional data collected in step S1. A global coordinate system is established based on the tunnel centerline and cross-sectional contour. The sampling point interval is strictly set to 0.1 to 0.5, covering the entire cross-section according to the principle of equidistant distribution, ensuring that the number of sampling points meets the detection accuracy requirements and accurately determining the three-dimensional spatial position parameters of each sampling point. S32 extracts the sensor measurement data from each sampling point, calculates the mean, variance, and other statistical characteristics of each sampling point's data, and simultaneously solves for the first and second-order gradient information of the data, integrating them to form a sampling point data matrix including multi-dimensional information. The matrix dimension is consistent with the number of sampling points. S33 inputs the data matrix into the tunnel cross-sectional deformation inversion algorithm, substitutes the deformation inversion coefficients calibrated by the measured data, and constructs a computational model based on the sampling point position correlation matrix. The tunnel cross-sectional deformation offset is calculated through matrix multiplication, with the offset calculation accuracy controlled within a preset threshold range. S34 performs matrix inversion and regularization on the deformation offset. The diagonal elements of the regularization matrix range from 0.01 to 0.1, which effectively suppresses overfitting during the inversion process. After correcting the error, accurate tunnel cross-section deformation characterization data is obtained. A complete closed loop is formed from data preprocessing to result optimization to ensure accurate characterization of cross-section deformation.

[0057] Preferred, such as Figure 4 As shown, S4 includes the following sub-steps: S41, performing pixel-level alignment on the radar detection data to eliminate texture shifts caused by differences in detection viewing angles and constructing a uniform-scale original radar image; S42, extracting grayscale features and texture frequency information from the original radar image to determine the parameter range and reconstruction weight coefficients of the Gabor filter; S43, substituting the radar image texture reconstruction algorithm, and reconstructing the detailed texture of the radar image through filtering and weighted calculation; S44, performing consistency verification on the reconstructed texture map, correcting texture distortion areas, and outputting a complete texture feature map of the tunnel's internal structure.

[0058] Specifically, step S4 performs high-quality reconstruction of radar image texture. S41 first performs pixel-level alignment processing on the radar detection data acquired in step S1, using a coordinate transformation algorithm to eliminate texture shifts caused by different detection perspectives, unifying all radar data to a standard scale of 1024×1024, constructing a scale-consistent, shift-free original radar image, laying the foundation for subsequent processing. S42 extracts the grayscale histogram features and texture frequency spectrum information of the original radar image, determining the parameter range of the Gabor filter through frequency analysis. The filter direction angle ranges from 0 to π, divided into 16 equal parts, and the filter wavelength ranges from 2 to 8, adaptively selected according to texture density. Simultaneously, texture reconstruction weight coefficients are calculated, with coefficient values ​​distributed between 0.2 and 0.8 based on texture importance. S43 substitutes the parameters and coefficients into the radar image texture reconstruction algorithm, enhancing the original pixel values ​​through the Gabor filter operator. Combining weighted operations and double quadrature operations, row and column pixel information is integrated to reconstruct the detailed texture of the radar image, strengthening the structural information of concealed parts. S44 performs consistency verification on the reconstructed texture map, identifies distorted regions through adjacent pixel texture similarity analysis and texture continuity detection, and adjusts the distorted regions using an interpolation correction algorithm. Finally, it outputs a complete and clear texture feature map of the tunnel's internal structure, ensuring that the reconstructed texture highly matches the actual structure.

[0059] Preferred, such as Figure 5 As shown, step S5 includes the following sub-steps: S51, unifying the dimensions of the multimodal feature data output from S2 to S4, converting geological texture features, cross-sectional deformation data, and radar texture maps into feature vectors of the same dimension; S52, constructing a multimodal feature association matrix, calculating the similarity coefficient between different types of features, and determining the feature fusion weights; S53, using a weighted fusion strategy to perform aggregation operations on the multimodal feature vectors to generate preliminary fused features; S54, removing redundant information in the fused features through a feature filtering mechanism, retaining the calibrated feature components, and constructing a global detection feature space.

[0060] Specifically, step S5 performs deep fusion of multimodal features. S51 first unifies the dimensions of the three types of feature data output from steps S2 to S4, converting geological texture feature data, tunnel cross-section deformation data, and radar image texture maps into 512-dimensional feature vectors. A feature scaling algorithm is used to ensure that the numerical range of all vectors remains consistent, eliminating the influence of dimensional differences on the fusion result. S52 constructs a multimodal feature association matrix, with the matrix dimension being the cube of the feature vector dimension. A cosine similarity algorithm is used to calculate the similarity coefficients between different types of features, with coefficient values ​​ranging from -1 to 1. The fusion weights for each type of feature are determined based on the coefficient values, with a total weight of 1. Important features have a weight ratio of no less than 0.3. S53 employs a weighted fusion strategy to aggregate the unified-dimensional multimodal feature vectors, generating preliminary fusion features through vector multiplication and addition operations. During the fusion process, the dimension of the feature vectors remains unchanged to ensure information integrity. S54 uses a threshold filtering mechanism to remove redundant information from the fused features. The filtering threshold is set to 0.05, retaining key feature components whose absolute value is greater than the threshold and removing invalid and redundant data. Finally, a full-domain detection feature space including core information is constructed. The dimension of the feature space is consistent with the dimension of the fused feature vector, providing comprehensive and efficient feature support for analysis and judgment.

[0061] 10. A multi-sensor collaborative intelligent tunnel detection method according to any one of claims 1-9, characterized in that the method is implemented through a multi-sensor collaborative intelligent tunnel system, comprising: a tunnel multi-modal data acquisition and transmission unit, a geological texture feature intelligent extraction and enhancement unit, a tunnel cross-section deformation inversion calculation unit, a radar image texture adaptive reconstruction unit, a multi-modal feature collaborative fusion processing unit, and a full-domain detection result analysis and labeling unit; the tunnel multi-modal data acquisition and transmission unit, the geological texture feature intelligent extraction and enhancement unit, the tunnel cross-section deformation inversion calculation unit, and the radar image texture adaptive reconstruction unit are integrated into the intelligent system. Each reconstruction unit establishes a bidirectional data transmission link to synchronously distribute the collected multi-dimensional data to each feature processing unit. The outputs of the intelligent extraction and enhancement unit of geological texture features, the tunnel cross-section deformation inversion calculation unit, and the adaptive reconstruction unit of radar image texture are all connected to the input of the multimodal feature collaborative fusion processing unit to transmit their respective processed feature data. The output of the multimodal feature collaborative fusion processing unit is connected to the global detection result analysis and identification unit to transmit the fused global feature spatial data to this unit for analysis and judgment. Finally, the global detection result analysis and identification unit outputs the tunnel structure status detection results and abnormal signal identification.

[0062] A multi-sensor collaborative intelligent tunnel detection method is proposed, constructing a detection system that deeply integrates multi-sensor collaboration with dedicated intelligent algorithms to achieve comprehensive, precise, and intelligent tunnel detection. By integrating multiple professional sensing modules and building a data transmission and processing architecture through a unified intelligent detection and early warning platform, the method breaks the traditional situation of scattered and isolated multi-source data, allowing detection data from different dimensions to form complementary linkages, comprehensively covering the monitoring needs of tunnel surface, interior, and environmental conditions. Simultaneously, specifically designed algorithms are adapted for geological texture feature extraction, cross-sectional deformation inversion, and radar image reconstruction, enhancing the ability to capture complex structural information and subtle defect features, and improving the analytical accuracy and effectiveness of detection data.

[0063] This method addresses the problem of insufficient data collaboration with a multi-unit collaborative architecture. Through a specialized multimodal feature fusion processing flow, it performs unified dimensional transformation, correlation calculation, and redundancy removal on data collected from various sensors, constructing a full-domain detection feature space. This fully explores the inherent correlations within the data, comprehensively restores the true state of the tunnel structure, and particularly accurately presents information about hidden parts. By using algorithms and detailed processing steps, it overcomes the weakness of insufficient targeting in feature extraction. Through a multi-stage refined data processing flow, combined with algorithmic adaptation to the unique structural features of tunnels, it effectively captures weak defect signals in complex environments, avoiding the loss of details and missed detections. At the same time, the automated processing architecture improves detection efficiency, enhances result reliability and response timeliness, and changes the limitations of traditional detection methods.

[0064] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," "link," and "fix" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0065] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various equivalent changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-sensor collaborative intelligent detection method for tunnels with multiple modes, characterized in that, Includes the following steps: S1. A multi-sensor collaborative intelligent detection and early warning platform for the entire tunnel uses various types of sensors to collect surface feature data of the tunnel structure, internal response signals of the geological body, and environmental state parameters, constructing a multi-dimensional raw detection dataset. S2. A geological texture gating map attention network is used to extract and enhance features from the collected geological body-related data. A gating mechanism dynamically adjusts the weight allocation of different geological texture features, strengthening the ability to effectively represent geological information. S3. A tunnel cross-section deformation inversion algorithm is used to analyze the collected tunnel cross-section data, establishing a mapping relationship between cross-section deformation and sensor data to represent the cross-section deformation state. S4. A radar image texture reconstruction algorithm is used to process radar detection data, reconstructing the texture feature map of the tunnel's internal structure and restoring the structural state of concealed parts of the tunnel. S5. Based on a multi-sensor collaborative tunnel multi-modal intelligent detection architecture, the multi-modal feature data processed in S2 to S4 are fused and correlated to construct a full-domain detection feature space. S6. The tunnel full-domain multi-sensor collaborative intelligent detection and early warning platform analyzes and judges the fused feature data, outputting tunnel structure state detection results and abnormal signal identifiers.

2. The tunnel multimodal intelligent detection method based on multi-sensor collaboration according to claim 1, characterized in that, The expression for the geological texture gating map attention network is: ； ； in, This is the attention weight matrix for the gating graph. The number of geological texture feature channels. It is the Sigmoid activation function. , For the network's learnable weight matrix, For the first Original geological texture features of the passage. For bias vectors, For the first The graph adjacency matrix corresponding to the channel features For element-wise multiplication, This is the aggregated original geological texture feature matrix. The enhanced geological texture feature matrix is ​​represented by ReLU and Softmax as activation functions.

3. The tunnel multimodal intelligent detection method based on multi-sensor collaboration according to claim 1, characterized in that, The expression for the tunnel cross-section deformation inversion algorithm is as follows: ； ； in, This is the cross-sectional deformation offset matrix. This refers to the number of sampling points for cross-sectional inspection. These are the deformation inversion coefficients. For the first Sensor measurement data at sampling points, For the first Gradient values ​​of the sampled point measurement data The correlation matrix of sampling point locations, To obtain the true cross-sectional deformation data through inversion, To perform the matrix inversion operation, For deformation inversion regularization matrix, for The transpose of .

4. The tunnel multimodal intelligent detection method based on multi-sensor collaboration according to claim 1, characterized in that, The expression for the radar image texture reconstruction algorithm is: ； in, The reconstructed radar image texture matrix, These represent the number of rows and columns of the radar image, respectively. For texture reconstruction weight coefficients, For radar imagery Line number The original pixel values ​​of the column. For Gabor filtering operators, For the first Line number The filtering direction angle of the column pixels, These are the filter wavelength parameters. This is the texture reconstruction correction matrix.

5. The tunnel multimodal intelligent detection method based on multi-sensor collaboration according to claim 1, characterized in that, The multi-type sensing devices include a distributed fiber optic sensing module, a millimeter-wave radar detection module, a three-dimensional laser scanning module, an ultrasonic sensing array, and an infrared thermal imaging module. Each module establishes a communication connection with the tunnel's full-domain multi-sensor collaborative intelligent detection and early warning platform through a 5G industrial gateway. The number of feature channels in the geological texture gating map attention network is set to 64 to 256, the number of neurons in the network's hidden layer is 256 to 1024, and the number of iterations is 100 to 500. The sampling point interval of the tunnel cross-section deformation inversion algorithm is set to 0.1 to 0.5, and the regularization matrix... The diagonal elements range from 0.01 to 0.1; Gabor filter orientation angle in radar image texture reconstruction algorithm The value range is 0 to Filter wavelength The value ranges from 2 to 8.

6. The tunnel multimodal intelligent detection method based on multi-sensor collaboration according to claim 1, characterized in that, S2 includes the following sub-steps: S21, performing feature channel separation on the collected geological body-related data, dividing the data into different subsets according to the geological texture type, and establishing the association mapping between each subset and the geological structure attributes; S22, inputting each data subset into the feature encoding layer of the geological texture gating graph attention network, extracting local texture features through convolution operations, and generating an initial feature map; S23, using a gating mechanism to adjust the weights of the initial feature map, dynamically allocating attention resources according to the recognizability and relevance of the geological texture, and suppressing redundant information interference; S24, performing global association modeling on the weighted feature map through a graph neural network, fusing local and global feature information, and outputting enhanced geological texture feature data.

7. The tunnel multimodal intelligent detection method based on multi-sensor collaboration according to claim 1, characterized in that, S3 includes the following sub-steps: S31, calibrating the sampling point coordinates of the tunnel cross-section data, establishing a global coordinate system for the cross-section, and determining the spatial position parameters of each sampling point; S32, extracting the sensor measurement data of each sampling point, calculating the statistical characteristics and gradient information of the data, and constructing a sampling point data matrix; S33, inputting the data matrix into the tunnel cross-section deformation inversion algorithm, substituting the preset inversion coefficients and correlation matrix, and calculating the cross-section deformation offset; S34. By performing matrix inversion and regularization, the inversion results are corrected to obtain tunnel cross-sectional deformation characterization data.

8. The tunnel multimodal intelligent detection method based on multi-sensor collaboration according to claim 1, characterized in that, S4 includes the following sub-steps: S41, performing pixel-level alignment on the radar detection data to eliminate texture shifts caused by differences in detection viewpoints and constructing a uniform-scale original radar image; S42, extracting grayscale features and texture frequency information from the original radar image to determine the parameter range and reconstruction weight coefficients of the Gabor filter; S43, substituting the radar image texture reconstruction algorithm, and reconstructing the detailed texture of the radar image through filtering and weighted calculation; S44, performing consistency verification on the reconstructed texture map, correcting texture distortion areas, and outputting a complete texture feature map of the tunnel's internal structure.

9. The tunnel multimodal intelligent detection method based on multi-sensor collaboration according to claim 1, characterized in that, S5 includes the following sub-steps: S51, unifying the dimensions of the multimodal feature data output from S2 to S4, converting geological texture features, cross-sectional deformation data, and radar texture maps into feature vectors of the same dimension; S52, constructing a multimodal feature association matrix, calculating the similarity coefficient between different types of features, and determining the feature fusion weights; S53, using a weighted fusion strategy to perform aggregation operations on the multimodal feature vectors to generate preliminary fused features; S54, removing redundant information in the fused features through a feature filtering mechanism, retaining the calibrated feature components, and constructing a global detection feature space.

10. A multi-sensor collaborative intelligent tunnel detection method according to any one of claims 1-9, characterized in that, This method is implemented through a multi-sensor collaborative tunnel multimodal intelligent system, comprising: a tunnel multimodal data acquisition and transmission unit, a geological texture feature intelligent extraction and enhancement unit, a tunnel cross-section deformation inversion calculation unit, a radar image texture adaptive reconstruction unit, a multimodal feature collaborative fusion processing unit, and a global detection result analysis and labeling unit. The tunnel multimodal data acquisition and transmission unit establishes bidirectional data transmission links with the geological texture feature intelligent extraction and enhancement unit, the tunnel cross-section deformation inversion calculation unit, and the radar image texture adaptive reconstruction unit, synchronously distributing the acquired multi-dimensional data to each feature processing unit. The outputs of the geological texture feature intelligent extraction and enhancement unit, the tunnel cross-section deformation inversion calculation unit, and the radar image texture adaptive reconstruction unit are all connected to the input of the multimodal feature collaborative fusion processing unit, transmitting their respective processed feature data. The output of the multimodal feature collaborative fusion processing unit is connected to the global detection result analysis and labeling unit, transmitting the fused global feature spatial data to this unit for analysis and judgment. Finally, the global detection result analysis and labeling unit outputs the tunnel structure status detection results and abnormal signal labels.

Images