Municipal road cavity detection method based on artificial intelligence

Abstract

The invention provides a municipal road cavity detection method based on artificial intelligence, and belongs to the technical field of road cavity detection.The municipal road cavity detection method comprises the steps that a sound shock excitation device and a seismic wave sensor array are arranged in a detection area to synchronously collect double-source detection data, and signal preprocessing is conducted through band-pass filtering and a wavelet denoising algorithm; a sparse representation algorithm is adopted to convert into a sparse matrix, a condition number is optimized through a conjugate gradient method, space block splitting and global optimal registration fusion based on convex optimization are carried out on optimized data, principal component analysis dimension reduction is applied to extract a fusion feature matrix, a pre-trained convolutional neural network model is input to obtain a preliminary detection result, and a final detection result is obtained. And when the confidence coefficient is insufficient, calculating and optimizing a parameter vector through a sound shock parameter adjustment function and a seismic parameter adjustment function according to the detection environment parameters, and reconfiguring equipment parameters to carry out secondary fine detection, so that the technical problem of low road cavity detection accuracy is solved.

Description

Technical Field

[0001] This invention belongs to the field of road cavity detection technology, and more specifically, relates to an artificial intelligence-based method for detecting cavities in municipal roads. Background Technology

[0002] Cavity detection in municipal roads is a crucial technical means for the safe maintenance of urban infrastructure. Traditional detection methods mainly include ground-penetrating radar (GPR), seismic wave detection, and acoustic wave detection. GPR technology identifies underground cavity structures by emitting electromagnetic waves and receiving reflected signals, and is widely used in the safety inspection of roads, bridges, and underground pipelines. Seismic wave detection utilizes the differences in the propagation characteristics of artificially excited seismic waves in different media to detect abnormal underground structures. Acoustic wave detection identifies cavity locations by analyzing the reflection and refraction phenomena of sound waves during underground propagation. However, in practical engineering applications, single detection technologies are often affected by factors such as the complexity of geological conditions, environmental noise interference, and limitations in equipment accuracy, leading to high false positive and false negative rates. Especially under complex geological conditions or in the presence of multiple underground interference sources, the reliability of traditional detection methods is significantly reduced. The most significant technical challenge in traditional technologies lies in how to effectively integrate multiple detection data sources to improve overall detection accuracy. Since the data obtained by different detection technologies have significant differences in spatiotemporal reference, signal characteristics, and noise distribution, simple data superposition or linear fusion often cannot fully utilize the advantages of various detection technologies, resulting in limited improvement in the accuracy of the fused detection results or even the problem of fusion error accumulation. In other words, existing technologies have the technical problem of insufficient accuracy in multi-source detection data fusion, leading to low accuracy in hole detection. Summary of the Invention

[0003] In view of this, the present invention provides an artificial intelligence-based method for detecting cavities in municipal roads, which can solve the technical problem of low accuracy in cavity detection due to insufficient fusion accuracy of multi-source detection data in the prior art.

[0004] This invention is implemented as follows: It provides an artificial intelligence-based method for detecting cavities in municipal roads, comprising: collecting and preprocessing acoustic and seismic detection datasets of the road data; converting the processed data into acoustic and seismic sparse matrices using a sparse representation algorithm; employing a conjugate gradient method as a preprocessing strategy for the acoustic and seismic sparse matrices, improving the condition number of the sparse matrices through preconditioner transformation; partitioning the optimized acoustic and seismic sparse matrices into blocks according to spatial location, obtaining sets of acoustic and seismic submatrices; aligning the corresponding submatrices in the acoustic and seismic submatrices spatially and temporally, performing registration using a global optimal solution mechanism based on convex optimization, and utilizing the convexity of the objective function to ensure the finding of registration parameters. The global optimal solution is then fused into a unified feature submatrix set. This unified feature submatrix set is then subjected to principal component analysis (PCA) for dimensionality reduction and eigenvector extraction to obtain a fused feature matrix. This fused feature matrix is ​​input into a pre-trained convolutional neural network hole detection model for computation, yielding preliminary hole detection results. The output includes hole location coordinates and a confidence score. A confidence threshold analysis is performed on the preliminary hole detection results. When the confidence score is below the threshold, supplementary detection is conducted to obtain a confidence evaluation value. This confidence evaluation value is input into acoustic and seismic parameter adjustment functions to calculate acoustic and seismic parameter vectors and seismic parameter vectors. Based on these vectors, the detection equipment parameters are reconfigured, and secondary fine-tuning is performed on areas with confidence evaluation values ​​below the threshold to obtain the final hole detection results.

[0005] The specific method for collecting acoustic and seismic detection datasets and seismic detection datasets for roads is as follows: an acoustic excitation device and a seismic wave sensor array are arranged on the surface of the road detection area. The acoustic excitation device uses low-frequency acoustic waves for excitation, and the seismic wave sensor array uses a grid layout to collect underground structure response signals. The acoustic excitation device is activated to emit acoustic pulse signals, and the reflected echo data received by the seismic wave sensor array is collected simultaneously.

[0006] The sparse representation algorithm is used to convert high-dimensional detection data into a low-dimensional sparse matrix representation, which retains key feature information while reducing computational complexity. The bandpass filter is used to filter out high-frequency noise and low-frequency interference in the detection signal and retain the signal components within the effective frequency band. The wavelet denoising algorithm is used to remove random noise in the signal through wavelet decomposition and reconstruction, thereby improving the signal-to-noise ratio.

[0007] The preprocessing strategy of the conjugate gradient method is used to improve the numerical properties of the original sparse matrix by constructing a precondition submatrix, thereby accelerating the convergence speed of subsequent calculations. The precondition subtransformation is used to convert the ill-conditioned matrix into a well-conditioned matrix with a smaller condition number, thereby improving the stability of numerical calculations.

[0008] The condition number represents the ratio of the maximum singular value to the minimum singular value of the matrix, reflecting the stability of the matrix numerical calculation. The global optimal solution mechanism based on convex optimization is used to find the globally optimal registration parameters during the data registration process, avoiding local optima.

[0009] The convexity of the objective function refers to the mathematical property that the registration error function has a single extreme point, ensuring that the optimization algorithm converges to the global optimal solution. The spatial coordinate alignment is used to unify the data acquired by different sensors into the same coordinate system. The timing synchronization is used to eliminate the time delay differences between different sensors.

[0010] The principal component analysis is used to extract the main direction of change in the data, thereby achieving feature dimensionality reduction and key information retention. The feature vector extraction is used to extract vector representations that characterize the underground structure features from the dimensionality-reduced data.

[0011] The acoustic vibration parameter adjustment function is used to optimize the operating parameters of the acoustic vibration equipment based on the detection environment and preliminary results. The inputs include confidence assessment value, detection depth, soil type coefficient, temperature compensation factor and humidity influence factor, and the output is an acoustic vibration parameter vector.

[0012] The seismic parameter adjustment function is used to adjust the parameters of the seismic wave detection equipment according to geological conditions and detection results. The inputs include confidence assessment value, formation density, propagation velocity, attenuation coefficient and interference intensity, and the output is a seismic parameter vector.

[0013] The confidence threshold analysis is used to evaluate the reliability of the detection results and identify areas that need to be supplemented for detection. The confidence score is generated by the Sigmoid activation function of the output layer of the convolutional neural network hole detection model, and the value ranges from 0 to 1. The confidence evaluation value is obtained by weighted averaging of the confidence scores of all detection points in the preliminary hole detection results.

[0014] The acoustic excitation device is used to generate low-frequency acoustic signals that propagate underground and receive reflected echoes. It detects underground cavity structures by measuring the difference in the propagation speed of acoustic waves in different media. The seismic wave sensor array is used to receive ground vibration signals and identify underground abnormal structures by analyzing the amplitude and phase changes of the vibration waveform.

[0015] The convolutional neural network hole detection model employs an 8-layer convolutional structure. Layers 1 to 3 have a 7×7 kernel size, a stride of 2, padding of 3, and a ReLU activation function. Layers 4 to 6 have a 5×5 kernel size, a stride of 1, padding of 2, and a ReLU activation function. Layers 7 and 8 have a 3×3 kernel size, a stride of 1, padding of 1, and a Sigmoid activation function. Each convolutional layer is followed by a batch normalization layer and a max-pooling layer with a 2×2 pooling kernel. The final output layer is a fully connected layer, outputting the hole location coordinates and confidence score.

[0016] The acoustic vibration parameter vector includes acoustic vibration excitation frequency, acoustic vibration excitation intensity, and acoustic vibration pulse interval. The acoustic vibration excitation frequency is obtained by adjusting the acoustic vibration parameter adjustment function according to the detection depth and soil type coefficient, and is used to reconfigure the working frequency of the acoustic vibration excitation device.

[0017] The acoustic excitation intensity is obtained by adjusting the acoustic parameter adjustment function according to the temperature compensation factor and the humidity influence factor, and is used to reconfigure the output power of the acoustic excitation device. The acoustic pulse interval is obtained by adjusting the acoustic parameter adjustment function according to the confidence evaluation value, and is used to reconfigure the pulse emission interval time of the acoustic excitation device.

[0018] The seismic parameter vector includes seismic wave frequency, sensor gain, and sampling frequency. The seismic wave frequency is obtained by adjusting the seismic parameter adjustment function according to the formation density and propagation velocity, and is used to reconfigure the receiving frequency range of the seismic wave sensor array.

[0019] The sensor gain is obtained by adjusting the seismic parameter adjustment function according to the attenuation coefficient and interference intensity, and is used to reconfigure the signal amplification factor of the seismic wave sensor array. The sampling frequency is obtained by adjusting the seismic parameter adjustment function according to the confidence evaluation value, and is used to reconfigure the data acquisition frequency of the seismic wave sensor array.

[0020] This invention addresses the technical problem of insufficient accuracy in multi-source detection data fusion by constructing a dual-source data fusion system for acoustic and seismic detection and employing a global optimal solution mechanism based on convex optimization for data registration and feature extraction. The invention utilizes a sparse representation algorithm to convert high-dimensional data obtained from different detection technologies into a structured sparse matrix representation. A preprocessing strategy using the conjugate gradient method optimizes the matrix condition number, ensuring numerical stability during data preprocessing. Simultaneously, spatial coordinate alignment and temporal synchronization techniques eliminate benchmark differences between different sensors, laying a solid foundation for subsequent data fusion. In the feature extraction stage, principal component analysis is used to extract key information from the fused feature matrix, and a pre-trained convolutional neural network model enables intelligent hole identification. When detection confidence is insufficient, the system can adaptively adjust the operating parameters of the acoustic and seismic detection equipment based on environmental parameters, performing secondary fine-tuning to improve detection accuracy. In summary, this invention, through intelligent fusion of multi-source data and adaptive parameter optimization, solves the technical problem of low hole detection accuracy due to insufficient multi-source detection data fusion accuracy in existing technologies. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method of the present invention.

[0022] Figure 2 This is a convergence curve of the training loss function of the convolutional neural network in Example 2.

[0023] Figure 3 This is a comparison chart of acoustic and seismic parameters before and after optimization in Example 2. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0025] like Figure 1 The diagram shown is a flowchart of an artificial intelligence-based method for detecting cavities in municipal roads provided by this invention. This method includes the following steps:

[0026] S01. An acoustic excitation device and a seismic wave sensor array are arranged on the surface of the detection area. The acoustic excitation device uses low-frequency acoustic waves with a frequency f∈[20, 100]Hz for excitation. The seismic wave sensor array uses a grid layout with a spacing d∈[0.5, 2.0]m to collect underground structure response signals.

[0027] S02. Start the acoustic vibration excitation device to emit acoustic pulse signals, and simultaneously collect the reflected echo data received by the seismic wave sensor array to obtain acoustic vibration detection dataset and seismic detection dataset. The acquisition time t∈[60,120]s;

[0028] S03. Preprocess the acoustic seismic detection dataset and the seismic detection dataset. Eliminate background noise by bandpass filtering and wavelet denoising algorithm. Use sparse representation algorithm to convert the processed data into acoustic seismic sparse matrix and seismic sparse matrix.

[0029] S04. A preprocessing strategy using the conjugate gradient method is adopted for the acoustic-seismic sparse matrix and the seismic sparse matrix. The condition number of the sparse matrix is ​​improved by the preconditioner transformation. The optimization range of the condition number is κ∈[10,100].

[0030] S05. The optimized acoustic-seismic sparse matrix and seismic sparse matrix are divided into blocks according to spatial location, with each block being a 5×5 grid cell, to obtain the acoustic-seismic submatrix set and the seismic submatrix set.

[0031] S06. Spatial coordinate alignment and temporal synchronization are performed on the corresponding sub-matrices in the acoustic-seismic sub-matrix set and the seismic sub-matrix set. A global optimal solution mechanism based on convex optimization is used for registration. The convexity of the objective function is used to ensure that the global optimal solution of the registration parameters is found and then fused into a unified feature sub-matrix set.

[0032] S07. The unified feature submatrix set is reduced in dimensionality by principal component analysis and eigenvector extraction to obtain the fused feature matrix. The dimensionality compression ratio of the fused feature matrix is ​​α∈[0.3, 0.5].

[0033] S08. Input the fused feature matrix into the pre-trained convolutional neural network hole detection model for calculation to obtain preliminary hole detection results. The output includes hole location coordinates and confidence scores.

[0034] S09. Perform confidence threshold analysis on the preliminary cavity detection results. When the confidence score β < 0.75, perform supplementary detection to obtain the confidence evaluation value. The confidence evaluation value is used as the input of the acoustic parameter adjustment function.

[0035] S10. Input the confidence assessment value, detection depth, soil type coefficient, temperature compensation factor and humidity influence factor into the acoustic parameter adjustment function to calculate the optimized acoustic parameter vector composed of acoustic excitation frequency, acoustic excitation intensity and acoustic pulse interval.

[0036] S11. Input the confidence assessment value, formation density, propagation velocity, attenuation coefficient and interference intensity into the seismic parameter adjustment function to calculate the optimized seismic parameter vector composed of seismic wave frequency, sensor gain and sampling frequency;

[0037] S12. Based on the acoustic and seismic parameter vectors and the seismic parameter vectors, the detection equipment parameters are reconfigured, and secondary fine detection is performed on areas with confidence assessment values ​​below the threshold to obtain the final cavity detection results.

[0038] The acoustic excitation device generates low-frequency acoustic signals that propagate underground and receive reflected echoes, detecting underground cavities by analyzing the differences in the propagation speed of sound waves in different media. The seismic wave sensor array receives ground vibration signals and identifies underground anomalous structures by analyzing the amplitude and phase changes of the vibration waveforms.

[0039] The sparse representation algorithm is used to convert high-dimensional detection data into a low-dimensional sparse matrix representation, preserving key feature information while reducing computational complexity. The bandpass filtering removes high-frequency noise and low-frequency interference from the detection signal, retaining signal components within the effective frequency band. The wavelet denoising algorithm removes random noise from the signal through wavelet decomposition and reconstruction, improving the signal-to-noise ratio. The preprocessing strategy of the conjugate gradient method improves the numerical properties of the original sparse matrix by constructing a precondition submatrix, accelerating the convergence speed of subsequent calculations. The precondition subtransformation converts ill-conditioned matrices into well-conditioned matrices with smaller condition numbers, improving numerical computation stability. The condition number represents the ratio of the maximum to the minimum singular value of the matrix, reflecting the stability of the matrix numerical computation. The global optimal solution mechanism based on convex optimization finds the globally optimal registration parameters during data registration, avoiding local optima. The convexity of the objective function refers to the mathematical property that the registration error function has a single extreme point, ensuring that the optimization algorithm converges to the global optimum.

[0040] The spatial coordinate alignment is used to unify data acquired by different sensors into the same coordinate system. The temporal synchronization is used to eliminate time delay differences between different sensors. The principal component analysis is used to extract the main directions of change in the data, achieving feature reduction and preservation of key information. The feature vector extraction is used to extract vector representations characterizing the underground structure features from the dimensionality-reduced data.

[0041] The convolutional neural network hole detection model employs an 8-layer convolutional structure. Layers 1 to 3 have a 7×7 kernel size, a stride of 2, padding of 3, and ReLU activation. Layers 4 to 6 have a 5×5 kernel size, a stride of 1, padding of 2, and ReLU activation. Layers 7 and 8 have a 3×3 kernel size, a stride of 1, padding of 1, and Sigmoid activation. Each convolutional layer is followed by a batch normalization layer and a max-pooling layer with a 2×2 kernel size. The final output layer is a fully connected layer, outputting the hole location coordinates and confidence score. The training dataset for the convolutional neural network hole detection model was obtained by manually constructing different types of holes in experimental road sections of length L∈[100, 500]m. The hole types include circular holes with a diameter D. ∈[0.2, 2.0]m, major axis D of the elliptical cavity ∈[0.5, 3.0]m minor axis D ∈[0.2, 1.5]m, length D of the rectangular hole Width D ∈ [0.8, 4.0]m ∈[0.3, 2.0]m, equivalent diameter D of irregular cavity Hole depth H ∈ [0.4, 2.5] m, cavity volume V ∈ [0.5, 5.0] m, cavity volume V ∈ [0.1, 50.0] Data was collected by mining 156 hole combinations within the above parameter range in the experimental road section. The training steps of the convolutional neural network hole detection model include: a data preprocessing stage where 12,480 sets of collected training samples are divided into training and validation sets in an 8:2 ratio; a model initialization stage where the Xavier initialization method is used to set the network weights; a training parameter setting stage where the learning rate lr∈[0.001, 0.01], batch size ∈[16, 64], and training epochs ∈[100, 500] are set; a forward propagation stage where the loss function value is calculated; a backpropagation stage where the network parameters are updated; and a validation evaluation stage where the accuracy and recall of the model on the validation set are calculated. Training is stopped when the accuracy on the validation set does not improve for 20 consecutive epochs.

[0042] The confidence threshold analysis is used to assess the reliability of the detection results and identify areas requiring supplementary detection. The confidence score is generated by the sigmoid activation function of the output layer of the convolutional neural network hole detection model, with a value range of [0, 1]. The confidence evaluation value is obtained by weighted averaging of the confidence scores of all detection points in the preliminary hole detection results. The detection depth is calculated from the waveform propagation time measured by the seismic wave sensor array and the known wave velocity.

[0043] The soil type coefficients are pre-determined based on the physical properties of the soil in the detection area: 1.2 for sandy soil, 0.8 for clay, and 1.5 for rock and soil. The temperature compensation factor is determined based on the influence of ambient temperature T on the speed of sound wave propagation, calculated as 1 + 0.002 × (T - 20). The humidity influence factor is determined based on the influence of soil moisture content W on signal attenuation, calculated as 1 - 0.01 × W. The stratigraphic density is determined using soil density data obtained from geological exploration.

[0044] The propagation velocity is determined by testing the propagation velocity of seismic waves in different geological strata. The attenuation coefficient is calculated by analyzing the energy loss of the seismic wave signal during propagation. The interference intensity is quantified by measuring the influence of underground pipelines and reinforced concrete structures on the detection signal.

[0045] The acoustic-vibration parameter adjustment function is used to optimize the operating parameters of the acoustic-vibration equipment based on the detection environment and preliminary results. The inputs include confidence assessment value, detection depth, soil type coefficient, temperature compensation factor and humidity influence factor, and the output is an acoustic-vibration parameter vector.

[0046] The seismic parameter adjustment function is used to adjust the parameters of the seismic wave detection equipment according to geological conditions and detection results. The inputs include confidence assessment value, formation density, propagation velocity, attenuation coefficient and interference intensity, and the output is a seismic parameter vector.

[0047] The acoustic excitation frequency is adjusted by the acoustic parameter adjustment function based on the detection depth and soil type coefficient, and is used to reconfigure the operating frequency of the acoustic excitation device. The acoustic excitation intensity is adjusted by the acoustic parameter adjustment function based on the temperature compensation factor and humidity influence factor, and is used to reconfigure the output power of the acoustic excitation device. The acoustic pulse interval is adjusted by the acoustic parameter adjustment function based on the confidence assessment value, and is used to reconfigure the pulse emission interval time of the acoustic excitation device. The seismic wave frequency is adjusted by the seismic parameter adjustment function based on the formation density and propagation velocity, and is used to reconfigure the receiving frequency range of the seismic wave sensor array. The sensor gain is adjusted by the seismic parameter adjustment function based on the attenuation coefficient and interference intensity, and is used to reconfigure the signal amplification factor of the seismic wave sensor array. The sampling frequency is adjusted by the seismic parameter adjustment function based on the confidence assessment value, and is used to reconfigure the data acquisition frequency of the seismic wave sensor array. The grid layout is used to ensure the uniform distribution of the seismic wave sensor array in the detection area, providing sufficient spatial sampling density. The dimensionality compression ratio represents the ratio of the fused feature matrix dimension to the original data dimension.

[0048] The specific implementation methods of the above steps are described in detail below.

[0049] The specific implementation of step S01 involves first determining the location of the acoustic vibration excitation device based on the area and geological conditions of the detection zone. The device uses an electromagnetic low-frequency generator to produce acoustic signals with a frequency range of 20 to 100 Hz. This frequency range is capable of effectively penetrating road surface materials and propagating in the underground medium. The seismic wave sensor array uses accelerometer-type sensors, arranged with a grid spacing of 0.5 to 2.0 m. The grid spacing is selected based on the spatial Nyquist sampling theorem to ensure the capture of the spatial frequency characteristics of the smallest target cavity. The sensor array covers the entire detection area, and each sensor node has triaxial detection capability, enabling simultaneous recording of vertical and horizontal ground vibration signals.

[0050] The specific implementation of step S02 involves activating the acoustic vibration excitation device to emit acoustic signals according to a preset pulse sequence. The pulse width is set to 50 to 100 milliseconds, the pulse interval to 200 to 500 milliseconds, and the acoustic intensity is adjusted according to the required detection depth. The seismic wave sensor array simultaneously begins data acquisition, with a sampling frequency set to 1000 to 2000 Hz and an acquisition time window set to 60 to 120 seconds. The data acquisition system uses GPS clock synchronization technology to ensure the consistency of the time reference of each sensor node, achieving a time synchronization accuracy at the microsecond level. The acoustic vibration detection dataset records the amplitude, phase, and arrival time information of the reflected echoes, while the seismic detection dataset records the triaxial acceleration time series of ground vibrations.

[0051] The specific implementation of step S03 involves digital signal processing of the original detection data. First, a fourth-order Butterworth bandpass filter is used to remove noise interference outside the frequency range, with the passband range set to 5 to 200 Hz and the stopband attenuation reaching over 40 dB. The wavelet denoising algorithm uses Dobesie wavelet basis functions for multi-scale decomposition, with a decomposition level of 6. A soft thresholding function is used to remove noise components from the wavelet coefficients, with the threshold value determined based on the square root of the signal variance. The sparse representation algorithm constructs an overcomplete dictionary based on a dictionary learning method, with the number of dictionary atoms set to 2 to 3 times the dimension of the original data. An orthogonal matching pursuit algorithm is used to solve for the sparse coefficients, with the sparsity controlled within the range of 10% to 20%.

[0052] The specific implementation of step S04 addresses the ill-conditioned nature of acoustic-seismic sparse matrices and seismic sparse matrices by constructing a preconditioning submatrix using incomplete LU decomposition. The fill factor for the decomposition is set to 0.01 to 0.05 to balance computational complexity and preprocessing effectiveness. During the iterative solution process using the conjugate gradient method, the convergence criterion is set as the L2 norm of the residual vector being less than... The maximum number of iterations is limited to 1000. Condition number optimization is evaluated through eigenvalue analysis, with the target condition number controlled between 10 and 100. When the condition number exceeds this range, the precondition sub-parameters are adjusted and the condition is reconstructed.

[0053] The specific implementation of step S05 involves spatially dividing the optimized sparse matrix into blocks based on the spatial distribution of the sensor array. Each block corresponds to data from 5×5 sensor nodes, and the block size is chosen to balance computational efficiency and spatial resolution requirements. The acoustic-seismic submatrix set and the seismic submatrix set maintain the same spatial index relationship, and each submatrix contains data from all time samples within its corresponding region. The block division process employs a sliding window method, maintaining a 50% overlap between adjacent blocks to ensure that void features in the boundary regions are not lost during segmentation.

[0054] The specific implementation of step S06 involves data registration and fusion of the acoustic-seismic sub-matrix and the seismic sub-matrix. Spatial coordinate alignment uses an affine transformation model to compensate for sensor installation errors, with transformation parameters including translation, rotation, and scaling components. Timing synchronization employs a cross-correlation function to calculate the time delay, achieving a delay compensation accuracy of one-tenth of the sampling interval. Convex optimization is achieved using the interior-point method, with the objective function designed as a weighted least squares form of the registration error, and weighting coefficients adaptively adjusted based on signal quality. The convergence criterion for the global optimal solution is set as the relative change in the objective function value being less than [a certain value]. This ensures the stability of the registration parameters.

[0055] The specific implementation of step S07 involves applying principal component analysis (PCA) to the unified feature submatrix set for feature dimensionality reduction. First, the eigenvalues ​​and eigenvectors of the data covariance matrix are calculated and sorted according to the eigenvalue size. The dimensionality compression ratio is set to 0.3 to 0.5, retaining 30% to 50% of the original dimensionality. The corresponding principal components are selected as the basis vectors for dimensionality reduction. During feature vector extraction, singular value decomposition (SVD) is used to improve numerical stability, and the threshold for truncating singular values ​​is set to 0.01 times the maximum singular value. Each element of the fused feature matrix represents the comprehensive underground structural features of the corresponding spatial location.

[0056] The specific implementation of step S08 involves inputting the fused feature matrix into the trained convolutional neural network model for hole detection. The model's input layer receives standardized feature matrix data, with zero-mean unit variance standardization. The forward propagation process extracts spatial pattern features of holes through multi-layer convolution operations, while pooling operations reduce feature dimensionality and enhance the model's translation invariance. The output layer generates hole location coordinates and confidence scores through fully connected layers and a sigmoid activation function. The coordinate accuracy reaches one-tenth of the grid cell, and the confidence score ranges from 0 to 1.

[0057] The specific implementation of step S09 involves performing a reliability assessment on the preliminary detection results output by the convolutional neural network. A confidence threshold is set to 0.75. When the confidence score of a detection point is lower than this threshold, it is marked as an area requiring further detection. The confidence assessment value is calculated by spatially weighted averaging the confidence scores of all detection points within the detection area. The weighting function uses a Gaussian kernel function, and the kernel width parameter is determined based on the sensor spacing. The assessment process simultaneously considers the spatial continuity of the detection results and the statistical characteristics of the confidence distribution to identify potential false positives and false negatives.

[0058] The specific implementation of step S10 involves constructing an acoustic vibration parameter adjustment function to achieve adaptive optimization of the detection parameters. Input parameters include confidence assessment values, detection depth, soil type coefficient, temperature compensation factor, and humidity influence factor. The acoustic vibration excitation frequency is adjusted based on the propagation characteristics of sound waves in different media; the greater the depth, the lower the frequency to enhance penetration. The soil type coefficient adjusts the frequency response characteristics. The acoustic vibration excitation intensity is adjusted according to environmental compensation factors. The reference value for the temperature compensation factor is 1 plus 0.002 times the difference between the temperature and 20 degrees Celsius, and the reference value for the humidity influence factor is 1 minus 0.01 times the soil moisture content. The acoustic vibration pulse interval is adaptively adjusted according to the confidence assessment value; the lower the confidence level, the shorter the interval to increase detection density.

[0059] The specific implementation of step S11 involves constructing a seismic parameter adjustment function to optimize the parameters of the seismic wave detection equipment. Input parameters include confidence assessment values, formation density, propagation velocity, attenuation coefficient, and interference intensity. The optimal detection frequency range for the seismic wave is determined based on formation density and propagation velocity; higher frequencies are used for high-density formations to improve resolution. The sensor gain is adaptively adjusted based on the attenuation coefficient and interference intensity; the gain is increased in areas with severe attenuation to compensate for signal attenuation, while the gain is appropriately reduced in areas with high interference intensity to avoid signal saturation. The sampling frequency is dynamically adjusted based on the confidence assessment value; the sampling frequency is increased in low-confidence areas to obtain more detailed time-domain information.

[0060] The specific implementation of step S12 involves reconfiguring the detection equipment based on the optimized acoustic-seismic parameter vector and seismic parameter vector, and performing secondary fine-tuning detection on areas marked as low-confidence. Equipment reconfiguration includes adjusting the operating frequency, output power, and pulse interval of the acoustic-seismic excitation device, and adjusting the receiving frequency range, signal amplification factor, and sampling frequency of the seismic wave sensor array. The secondary detection employs a higher spatial sampling density and a longer data acquisition time, reducing the sensor spacing to 50% of the original and extending the acquisition time to 1.5 times the original. The final detection result is obtained by fusing the results of the initial and secondary detections, and a confidence-weighted fusion method is used to determine the final cavity location and geometric parameters.

[0061] The convolutional neural network hole detection model employs a deep learning architecture to intelligently identify underground cavities. The model structure includes eight convolutional layers, batch normalization layers, pooling layers, and fully connected layers. Layers 1 to 3 use 7×7 convolutional kernels with a stride of 2 and padding of 3, employing a rectified linear function (RCF) as the activation function. This layer primarily extracts low-level spatial features from the input feature matrix. Layers 4 to 6 use 5×5 convolutional kernels with a stride of 1 and padding of 2, also employing an RCF as the activation function. This layer further abstracts spatial features and enhances the non-linear expressive power of the features. Layers 7 and 8 use 3×3 convolutional kernels with a stride of 1 and padding of 1, employing a sigmoid function as the activation function. This layer is responsible for the final feature mapping and hole probability calculation. Each convolutional layer is followed by a batch normalization layer to standardize features, accelerating model training convergence and improving generalization ability. The max-pooling layer uses a 2×2 pooling kernel for downsampling, reducing feature dimensionality and enhancing the model's robustness to spatial displacement. The fully connected layer maps convolutional features to the output space, generating hole location coordinates and confidence scores.

[0062] The training dataset was created by artificially constructing cavities of different types and sizes in experimental road sections ranging from 100 to 500 meters in length to obtain real-world labeled data. Cavity types included circular cavities with diameters ranging from 0.2 to 2.0 meters, elliptical cavities with major axes of 0.5 to 3.0 meters and minor axes of 0.2 to 1.5 meters, rectangular cavities with lengths of 0.8 to 4.0 meters and widths of 0.3 to 2.0 meters, and irregular cavities with equivalent diameters of 0.4 to 2.5 meters. Cavity depths ranged from 0.5 to 5.0 meters, and volumes ranged from 0.1 to... A diverse range of training samples was constructed using 156 combinations of dilation parameters. The data acquisition process involved 80 measurements under each dilation configuration, resulting in 12,480 training samples, each containing a fused feature matrix and corresponding dilation label information. During data preprocessing, the training samples were randomly divided into training and validation sets at an 8:2 ratio. The training set was used for model parameter learning, while the validation set was used for model performance evaluation and hyperparameter tuning. Model initialization employed the Xavier initialization method to set network weights. This method adaptively determines the variance of the weight distribution based on the input and output dimensions of the layers, effectively avoiding gradient explosion and vanishing gradient problems. Training parameters included a learning rate of 0.001 to 0.01, batch size of 16 to 64, and the number of training epochs of 100 to 500. The Adam gradient descent algorithm was used for optimization. During training, the cross-entropy loss function was used to measure the difference between the predicted results and the true labels, and the network parameters were updated using the backpropagation algorithm. During the validation and evaluation phase, performance metrics such as accuracy, recall, and F1 score of the model on the validation set are calculated. If the accuracy on the validation set does not improve for 20 consecutive training rounds, an early stopping strategy is adopted to terminate training and prevent the model from overfitting.

[0063] It should be noted that the key technical concepts of this invention include four aspects: multi-physics field fusion detection, adaptive parameter adjustment, deep learning intelligent recognition, and secondary fine detection. The multi-physics field fusion detection technology combines acoustic excitation and seismic wave sensing, utilizing the different propagation characteristics of acoustic and seismic waves in the underground medium to achieve complementary detection. Acoustic excitation has strong penetration and directionality, while seismic wave sensing has high spatial resolution and sensitivity. The fusion of these two physical fields significantly improves the accuracy and reliability of cavity detection, overcoming the limitations of single detection methods. The adaptive parameter adjustment technology constructs acoustic and seismic parameter adjustment functions, automatically optimizing equipment operating parameters based on geological conditions, climatic factors, and preliminary detection results. This technology enables the detection system to adapt to different working conditions, improving the stability of detection accuracy and the intelligence level of the equipment, avoiding the tedious process of manual parameter adjustment required in traditional methods. Deep learning-based intelligent recognition technology employs convolutional neural networks to construct a cavity detection model. Through multi-layer convolutional operations, it automatically extracts spatial pattern features of underground structures. Compared to traditional signal processing methods, it possesses stronger feature representation and pattern recognition capabilities, and can handle complex underground structures and various interference factors, significantly improving the intelligence and accuracy of cavity identification. Secondary fine-tuning detection technology supplements low-reliability areas based on confidence assessment results. By optimizing detection parameters and increasing sampling density, it achieves refined detection, effectively reducing false positives and false negatives, and improving the overall integrity and reliability of the detection system.

[0064] The synergistic effect of the above four key technical approaches constitutes a complete intelligent detection system. Multi-physics field fusion provides rich detection information, adaptive parameter adjustment ensures the stability of the detection system in various environments, deep learning intelligent recognition achieves high-precision automated recognition, and secondary fine detection ensures the integrity of the detection results. The four technical approaches support and promote each other, forming a more advanced, reliable, and intelligent technical solution than traditional detection methods, with significant advantages in detection accuracy, adaptability, automation, and reliability.

[0065] It should be noted that in actual municipal road inspections, the underground environment often presents a variety of complex factors, including different types of soil conditions, varying temperature and humidity, interference from underground pipelines, and differences in soil density. These factors significantly affect the propagation characteristics of the detection signal and the reliability of the detection results. Traditional detection methods typically use fixed equipment parameter configurations, which cannot be dynamically adjusted according to specific environmental conditions, leading to a significant decrease in detection effectiveness under complex environments. This invention, by establishing acoustic and seismic parameter adjustment functions, can intelligently calculate the optimal equipment operating parameters based on multi-dimensional environmental parameters such as detection depth, soil type coefficient, temperature compensation factor, humidity influence factor, soil density, propagation velocity, attenuation coefficient, and interference intensity. These parameters include acoustic excitation frequency, excitation intensity, pulse interval, as well as seismic wave frequency, sensor gain, and sampling frequency. This enables the detection system to adaptively optimize for complex environmental conditions, significantly improving the system's detection stability and accuracy under various geological conditions.

[0066] Traditional detection methods often rely on manual judgment or simple threshold comparisons to assess the reliability of detection results after the test is completed, lacking real-time reliability assessment and quality control mechanisms during the detection process. This lagging assessment not only affects detection efficiency but may also lead to the omission of low-quality detection results. This invention establishes a real-time detection quality assessment system based on a confidence threshold by outputting a confidence score through a convolutional neural network model. When the confidence score of the detection result falls below the set threshold, the system can immediately identify areas requiring supplementary detection and automatically trigger a secondary fine-tuning process. This real-time quality control mechanism ensures that each detection point meets the preset reliability requirements, avoiding the impact of low-quality detection results on the overall detection accuracy. Furthermore, by using the confidence score as input to a parameter adjustment function, a closed-loop feedback control of detection quality and equipment parameter optimization is achieved, further improving the overall performance and reliability of the detection system.

[0067] Specifically, the principle of this invention is as follows: The core principle that enables this invention to solve the problem of insufficient accuracy in multi-source detection data fusion lies in establishing a complete unified processing and intelligent fusion framework for multi-source heterogeneous data. First, by simultaneously deploying an acoustic excitation device and a seismic wave sensor array, this invention obtains dual-source detection data with complementary characteristics. Acoustic detection mainly utilizes the difference in the propagation speed of sound waves in different media to identify cavity boundaries, while seismic wave detection detects underground structural anomalies by analyzing the amplitude and phase changes of the vibration waveform. The differences in the physical mechanisms of these two detection technologies provide rich information redundancy for data fusion. Second, this invention employs a sparse representation algorithm to convert the original high-dimensional detection signal into a low-dimensional sparse matrix representation. This conversion not only preserves the key feature information in the signal but also reduces computational complexity, creating favorable conditions for subsequent data processing and fusion. Simultaneously, the preprocessing strategy of the conjugate gradient method improves the numerical characteristics of the sparse matrix, ensuring computational stability and convergence during data processing. In the data fusion stage, the most critical innovation of this invention lies in employing a globally optimal solution mechanism based on convex optimization for data registration. Due to the convexity of the objective function, the optimization algorithm guarantees finding the globally optimal registration parameters, avoiding the problem of getting trapped in local optima in traditional registration methods. This globally optimal registration ensures precise alignment of different detection data in both spatial and temporal dimensions, laying the foundation for high-quality data fusion. By extracting the main directions of change in the fused data using principal component analysis, this invention effectively compresses data dimensions while retaining the most important feature information. The fused feature matrix contains complementary information from the two detection technologies, significantly improving the accuracy of subsequent hole identification. Finally, the adaptive parameter adjustment mechanism established in this invention intelligently adjusts the operating parameters of the detection equipment based on the confidence level of the preliminary detection results and the detection environment parameters, achieving targeted and refined detection. This closed-loop feedback mechanism ensures that the detection system maintains high detection accuracy and reliability under various complex environmental conditions.

[0068] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0069] In this embodiment, the specific implementation methods of steps S01 and S02 are the same as those described above, and will not be repeated in detail here.

[0070] The specific implementation of step S03 involves preprocessing the acoustic seismic detection dataset and the seismic detection dataset. The transfer function of the bandpass filter is expressed as:

[0071] ;

[0072] In the formula, This is the frequency domain transfer function of a fourth-order Butterworth bandpass filter; Angular frequency; This is the filter order, with a value of 4; The filter coefficients are determined based on the passband range of 5 to 200 Hz. The wavelet transform of the wavelet denoising algorithm is expressed as:

[0073] ;

[0074] In the formula, These are the wavelet transform coefficients; This is a scale parameter, with a value ranging from 1 to 64; These are displacement parameters; For input signals; It is the conjugate of the Dobermann wavelet basis functions. The optimization objective function of the sparse representation algorithm is:

[0075] ;

[0076] In the formula, The input signal vector; For an overcomplete dictionary matrix, the dimension is... ,in For signal dimension, This represents the number of atoms in the dictionary. It is a sparse coefficient vector; This is the sparsity regularization parameter, with a value ranging from 0.01 to 0.1; It is a 2-norm; It is a zero norm.

[0077] The specific implementation of step S04 involves preprocessing the sparse matrix using the conjugate gradient method. The iterative formula for the conjugate gradient method is as follows:

[0078] ;

[0079] ;

[0080] ;

[0081] ;

[0082] In the formula, For the first The solution vector for the next iteration; For the first The step size of the next iteration; For the first The search direction for the next iteration; For the first The residual vector of the next iteration; It is a coefficient matrix; is the conjugate coefficient. The condition number is calculated using the following formula:

[0083] ;

[0084] In the formula, For matrix condition number; For matrix The maximum singular value; For matrix The smallest singular value.

[0085] The specific implementation method of step S05 is the same as described above, and will not be repeated in detail here.

[0086] The specific implementation of step S06 involves registering and fusing the acoustic-seismic submatrix and the seismic submatrix. The objective function for convex optimization is expressed as:

[0087] ;

[0088] In the formula, The registration parameter vector includes translation, rotation, and scaling parameters; It is the affine transformation matrix; For the first A set of acoustic seismic sub-matrices; For the first Each seismic submatrix; For the first The weight coefficients of each submatrix; This represents the total number of sub-matrices. The formula for calculating the unified feature matrix after fusion is:

[0089] ;

[0090] In the formula, For the first A fusion feature submatrix; The registered acoustic seismic sub-matrix; The registered seismic submatrix; The fusion weighting coefficient has a value range of 0.3 to 0.7.

[0091] The specific implementation of step S07 is to use principal component analysis for feature dimensionality reduction. The formula for calculating the covariance matrix is:

[0092] ;

[0093] In the formula, It is the covariance matrix; The number of samples; For the first A fusion feature submatrix; Let be the mean of the eigenvectors. The eigenvalue decomposition formula is:

[0094] ;

[0095] In the formula, The eigenvector matrix; Let be the eigenvalue diagonal matrix. The formula for calculating the fused feature matrix after dimensionality reduction is:

[0096] ;

[0097] In the formula, The feature matrix after dimensionality reduction; For the front The eigenvector matrix corresponding to each principal component; This is the original fused feature matrix; The number of principal components to be retained.

[0098] The specific implementation method of step S08 is the same as described above, and will not be repeated in detail here.

[0099] The specific implementation of step S09 involves performing a confidence analysis on the preliminary cavity detection results. The formula for calculating the detection depth is as follows:

[0100] ;

[0101] In the formula, For detection depth; The propagation velocity of seismic waves is obtained through geological exploration and wave velocity testing. The waveform propagation time is calculated from the arrival time of the first wave recorded by the seismic wave sensor array. The soil type coefficient is calculated using the following formula:

[0102] ;

[0103] In the formula, This is the soil type coefficient, determined based on geological exploration results. The confidence level assessment value is calculated using the following formula:

[0104] ;

[0105] In the formula, This is the confidence level assessment value; For position Confidence score at the location; Spatial weights for the corresponding positions; and These represent the number of rows and columns in the detection region, respectively. Spatial weights are calculated using a Gaussian kernel function.

[0106] ;

[0107] In the formula, The center position of the weight; The Gaussian kernel width parameter is set to 0.5 times the sensor spacing.

[0108] The specific implementation of step S10 is to construct an acoustic vibration parameter adjustment function, and the acoustic vibration excitation frequency adjustment formula is as follows:

[0109] ;

[0110] In the formula, The optimized acoustic excitation frequency; The reference frequency is 60Hz. For detection depth; The maximum detection depth is set to 5m. This represents the soil type coefficient. The formula for adjusting the acoustic excitation intensity is:

[0111] ;

[0112] In the formula, The optimized acoustic excitation intensity; As the reference strength; This is the temperature compensation factor; Humidity is a factor influencing humidity. The formula for calculating the temperature compensation factor is:

[0113] ;

[0114] In the formula, The ambient temperature is expressed in degrees Celsius. The formula for calculating the humidity influence factor is:

[0115] ;

[0116] In the formula, Soil moisture content, expressed as a percentage. The formula for adjusting the acoustic pulse interval is:

[0117] ;

[0118] In the formula, The optimized pulse interval; The reference pulse interval is 300 milliseconds.

[0119] The specific implementation of step S11 involves constructing a seismic parameter adjustment function, with the seismic wave frequency adjustment formula as follows:

[0120] ;

[0121] In the formula, The optimized seismic wave frequency; The reference seismic wave frequency is set at 50Hz. Formation density; Let be the propagation velocity of seismic waves in the Earth's strata. The sensor gain adjustment formula is:

[0122] ;

[0123] In the formula, For the optimized sensor gain; Reference gain; The attenuation coefficient; The interference intensity is given. The sampling frequency adjustment formula is:

[0124] ;

[0125] In the formula, The optimized sampling frequency; The reference sampling frequency is 1500Hz.

[0126] The specific implementation of step S12 involves performing a secondary detection based on the optimized parameters, and the final detection result fusion formula is as follows:

[0127] ;

[0128] In the formula, This is the final test result; This is the initial test result; This is the result of a second test; The formula for calculating the fusion weights is as follows: ,in This is the confidence level assessment value for the secondary detection.

[0129] The principles and effects of each formula and mathematical model are as follows: The bandpass filter transfer function is based on the Butterworth filter design principle, and is constructed by... The frequency domain response characteristics effectively remove high-frequency noise and low-frequency interference from the detected signal. Compared to simple frequency domain truncation methods, this transfer function ensures a flat response in the passband and rapid attenuation in the stopband, significantly improving the signal-to-noise ratio and signal quality. The wavelet transform formula adopts continuous wavelet transform theory, using the scale parameter... and displacement parameters The modulation enables multi-scale analysis of the signal;

[0130] ;

[0131] Compared to the traditional Fourier transform, the wavelet transform possesses superior time-frequency localization properties, effectively extracting transient features and local anomalies from signals, thus providing richer feature information for subsequent hole identification. The sparse representation optimization objective function combines minimizing reconstruction error with sparsity constraints, through... Norm measures the accuracy of reconstruction, through Norms control sparsity;

[0132] ;

[0133] Compared to traditional signal representation methods, sparse representation can efficiently represent complex signals with a small number of dictionary atoms, reducing data dimensionality while preserving key features, thus improving computational efficiency and noise resistance. The iterative formula of the conjugate gradient method is based on the principle of conjugate direction search, avoiding the jagged effect of gradient descent by constructing conjugate search directions.

[0134] ;

[0135] Compared to conventional iterative methods, the conjugate gradient method exhibits faster convergence and better numerical stability, making it particularly suitable for solving large sparse matrix equation systems. The condition number calculation formula evaluates the numerical stability of a matrix by using the ratio of the maximum singular value to the minimum singular value.

[0136] ;

[0137] Compared to direct solution methods, condition number optimization ensures the accuracy and reliability of numerical computation, avoiding computational errors caused by ill-conditioned matrices. The convex optimization objective function adopts a weighted least squares form, using weight coefficients... Adaptively adjust the registration importance of different regions, where the affine transformation matrix has the following specific form:

[0138] ;

[0139] In the formula, This is the scaling factor; The rotation angle; The translation parameter is used. Compared to traditional rigid registration methods, the convexity of this objective function guarantees the existence and uniqueness of the global optimum, avoids local optimum traps, and improves registration accuracy and stability. The formula for calculating the fusion feature matrix achieves effective fusion of acoustic and seismic data through linear weighting.

[0140] ;

[0141] Compared to single data sources, multi-physics data fusion fully leverages the complementary advantages of different detection mechanisms, improving the accuracy and reliability of detection results. Principal component analysis (PCA) covariance matrix and eigenvalue decomposition formulas are based on statistical principles, achieving dimensionality reduction by extracting the main directions of data variation. The general form of eigenvalue decomposition is:

[0142] ;

[0143] In the formula, For the first 1 eigenvector; These are the corresponding eigenvalues. Compared to simple feature selection methods, principal component analysis preserves the maximum variance information in the data, maintaining the integrity of the original information to the greatest extent while reducing computational complexity. The confidence score calculation formula uses a spatially weighted average method, assigning different weights to different positions through a Gaussian kernel function. The general form of the Gaussian kernel function is:

[0144] ;

[0145] In the formula, The Gaussian kernel function; The spatial distance is used. Compared to a simple arithmetic mean, this method considers spatial correlation and local consistency, providing a more accurate reliability assessment and a reliable basis for adaptive parameter adjustment. The series of formulas for acoustic parameter adjustment functions comprehensively consider the detection depth. Soil type Ambient temperature ,humidity and confidence level Multiple influencing factors, including seismic parameters, were considered, and adaptive optimization of parameters was achieved through mathematical modeling. Compared with fixed parameter settings, this adaptive adjustment mechanism significantly improved the adaptability of the detection system to different working conditions and the stability of detection accuracy. The series of formulas for seismic parameter adjustment functions are based on seismic wave propagation theory and signal processing principles, and utilize factors such as formation density. speed of transmission attenuation coefficient By comprehensively considering geological parameters, intelligent optimization of seismic detection parameters is achieved. Compared with empirical parameter settings, this theoretical parameter adjustment method improves the scientific validity and accuracy of the detection results. The final detection result fusion formula achieves optimal fusion of the initial and secondary detection results through confidence-weighted calculation.

[0146] ;

[0147] Compared to the simple method of taking the maximum or average value, this weighted fusion strategy makes full use of the advantages of different detection stages, improving the completeness and reliability of the final detection results.

[0148] To better understand and implement this invention, the following is a specific application scenario of this invention, Example 2:

[0149] A technical team needed to conduct cavity detection on a 300m-long urban main road. This section of road is located on soft soil and has experienced frequent road surface collapses in recent years. The soil in the detection area is mainly clay, with dense underground pipelines. The ambient temperature was 25℃, and the soil moisture content was 15%. The technical team used an AI-based municipal road cavity detection method to conduct a comprehensive inspection of this road section.

[0150] First, the technical team deployed acoustic vibration excitation devices and seismic wave sensor arrays within a 300m × 8m detection area with a grid spacing of 1.2m. The acoustic vibration excitation devices were set to a low-frequency acoustic wave excitation frequency of 65Hz, and the seismic wave sensor array consisted of 420 sensor nodes, forming a 25×17 grid array. The sensor array adopted a waterproof and sealed design, enabling it to adapt to complex road surface environments.

[0151] After the detection system is activated, the acoustic vibration excitation device begins emitting acoustic pulse signals, with a data acquisition time set to 90 seconds. During this period, the seismic wave sensor array simultaneously receives reflected echo signals and ground vibration signals. The system automatically records the acoustic vibration detection dataset, containing 42,000 data points, while the seismic detection dataset contains 63,000 sampling points. Throughout the data acquisition process, the system monitors signal quality in real time to ensure data integrity and accuracy.

[0152] In the data preprocessing stage, the technical team filtered and denoised the collected raw data. A bandpass filter with a frequency band of 20-100Hz was used to remove high-frequency noise and low-frequency interference, and then a wavelet denoising algorithm was used to further improve the signal-to-noise ratio. The processed data was then converted into an acoustic-seismic sparse matrix (420×315 dimensions) and a seismic sparse matrix (420×472 dimensions) using a sparse representation algorithm. The sparse representation process removed redundant information from the original high-dimensional data, retaining key feature components.

[0153] To improve the numerical properties of the matrices, the technical team employed a preprocessing strategy using the conjugate gradient method for the two sparse matrices. By constructing a preconditioning submatrix, the condition number of the original matrix was reduced from 156 to 35, significantly improving the numerical stability of the matrix. After the preconditioning submatrix transformation, the convergence speed of subsequent calculations was improved by approximately 60%.

[0154] Next, matrix partitioning is performed, dividing the optimized acoustic-seismic sparse matrix and seismic sparse matrix into 5×5 grid sub-matrices based on spatial location. The acoustic-seismic sparse matrix is ​​decomposed into 1260 sub-matrices, and the seismic sparse matrix is ​​decomposed into 1418 sub-matrices. Each sub-matrix corresponds to a different spatial location within the detection area, facilitating subsequent spatial registration and feature fusion.

[0155] In the data fusion phase, a globally optimal solution mechanism based on convex optimization was used for registration. The technical team performed spatial coordinate alignment and temporal synchronization on corresponding submatrices in the acoustic-seismic and seismic submatrices. During registration, the convexity of the objective function ensured that the optimization algorithm could find the globally optimal registration parameters. Ultimately, 1260 unified feature submatrices were obtained, each containing fused acoustic-seismic and seismic feature information.

[0156] During feature extraction, the unified feature submatrix set is dimensionality reduced using principal component analysis (PCA). A dimensionality compression ratio of 0.4 is set, compressing the original feature matrix from 1260×25 dimensions to a fused feature matrix of 1260×10. PCA extracts the main directions of change in the data, retaining 99.2% of the feature information while significantly reducing computational complexity.

[0157] like Figure 2 As shown, the technical team inputs the fused feature matrix into a pre-trained 8-layer convolutional neural network hole detection model. This model was trained on 12,480 training samples of 156 different types of holes, achieving a validation set accuracy of 94.3%. The first three layers of the model use 7×7 convolutional kernels, the middle three layers use 5×5 convolutional kernels, and the last two layers use 3×3 convolutional kernels, followed by batch normalization and max pooling operations after each layer. The model outputs preliminary hole detection results, identifying 18 potential hole locations. See Table 1 for details.

[0158] Table 1. Statistics of Preliminary Cavity Detection Results

[0159] During the confidence threshold analysis phase, the technical team set the confidence threshold at 0.75. The analysis results showed that 6 out of 18 detection points had confidence scores below the threshold, requiring supplementary testing. The weighted average confidence assessment value was 0.73, indicating significant uncertainty in the detection area.

[0160] like Figure 3As shown, during the parameter optimization process, the technical team input the confidence assessment value (0.73), average detection depth (2.35m), clay coefficient (0.8), temperature compensation factor (1.01), and humidity influence factor (0.85) into the acoustic-seismic parameter adjustment function. The function outputs the optimized acoustic-seismic parameter vector: the acoustic excitation frequency is adjusted to 58Hz, the acoustic excitation intensity is adjusted to 85W, and the acoustic pulse interval is adjusted to 0.8s. Simultaneously, the confidence assessment value and formation density (1850) are also used. With a propagation speed of 420 m / s, an attenuation coefficient of 0.12, and an interference intensity of 0.25 input into the seismic parameter adjustment function, the seismic parameter vector is obtained: the seismic wave frequency is adjusted to 30-80 Hz, the sensor gain is adjusted to 45 dB, and the sampling frequency is adjusted to 250 Hz.

[0161] In the second detection phase, the technical team reconfigured the detection equipment based on optimized parameters, performing detailed detection on six areas with confidence levels below the threshold. The adjusted equipment parameters significantly improved detection accuracy, with the signal-to-noise ratio of the newly acquired data increasing by 28%. The detailed detection process lasted 45 seconds, yielding higher-quality detection data. The final detection results confirmed the existence of 12 real cavities, including 3 large cavities with diameters exceeding 1.5m and 9 medium-sized cavities with diameters ranging from 0.5 to 1.2m. The cavity depths ranged from 1.2 to 3.8m, with a total volume of approximately [missing information]. The detection accuracy improved by 15% compared to the preliminary results, effectively reducing missed detections and false detections.

[0162] The technical team also discovered that underground pipelines within the inspection area interfered with the test results, mainly concentrated in the 80-120m section. Through adaptive parameter adjustment, the impact of pipeline interference on the inspection accuracy was successfully reduced to below 5%. The entire inspection process took approximately 4 hours, representing an efficiency improvement of about 8 times compared to traditional borehole inspection methods.

[0163] As shown in Table 2, the key technical parameters in the detection process are as follows:

[0164] Table 2 Statistics of Key Technical Parameters

[0165] Compared to traditional detection methods, this invention significantly improves detection accuracy and reliability through multi-source data fusion. Acoustic seismic detection can effectively identify larger cavities, while seismic wave detection is more sensitive to small cavities; the combination of these two technologies enables accurate detection of cavities across all scales. A convex optimization-based registration mechanism ensures the global optimality of data fusion, avoiding the local optima problem common in traditional methods. The application of a convolutional neural network model makes the cavity identification process more intelligent, enabling it to automatically learn complex underground structural features. The adaptive parameter adjustment function dynamically optimizes the equipment's operating state according to the actual detection environment, improving the method's adaptability to different geological conditions. The application of sparse representation and principal component analysis significantly reduces computational complexity while maintaining detection accuracy, making large-scale road detection possible.

[0166] It should be noted that the variables involved in this invention are explained in detail in Tables 3 and 4.

[0167] Table 3. Variable Explanation Table (Part 1)

[0168] Table 4. Variable Explanation Table (Part Two)

[0169] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting cavities in municipal roads based on artificial intelligence, characterized in that, include: The acoustic seismic detection dataset and earthquake detection dataset of the road were collected and preprocessed. The sparse representation algorithm was used to convert the processed data into acoustic seismic sparse matrices and earthquake sparse matrices. A preprocessing strategy using the conjugate gradient method is adopted for the acoustic-seismic sparse matrix and the seismic sparse matrix, and the condition number of the sparse matrix is ​​improved by preconditioner transformation. The optimized acoustic-seismic sparse matrix and the seismic sparse matrix are then partitioned into blocks according to spatial location to obtain the acoustic-seismic submatrix set and the seismic submatrix set. Spatial coordinate alignment and temporal synchronization are performed on corresponding submatrices in the acoustic-seismic submatrix set and the seismic submatrix set. A global optimal solution mechanism based on convex optimization is used for registration. The convexity of the objective function ensures that the globally optimal solution for the registration parameters is found, and the submatrices are then fused into a unified feature submatrix set. The unified feature submatrix set is then subjected to principal component analysis for dimensionality reduction and eigenvector extraction to obtain the fused feature matrix. The fused feature matrix is ​​input into a pre-trained convolutional neural network hole detection model for calculation, yielding preliminary hole detection results. The output includes hole location coordinates and confidence scores. A confidence threshold analysis is performed on the preliminary hole detection results. When the confidence score is below the threshold, supplementary detection is conducted to obtain a confidence evaluation value. The confidence evaluation value is input into acoustic and seismic parameter adjustment functions to calculate acoustic and seismic parameter vectors and seismic parameter vectors. Based on the acoustic and seismic parameter vectors and the seismic parameter vectors, the detection equipment parameters are reconfigured, and secondary fine detection is performed on areas with confidence evaluation values ​​below the threshold to obtain the final hole detection results.

2. The method for detecting cavities in municipal roads based on artificial intelligence according to claim 1, characterized in that, The specific method for collecting acoustic and seismic detection datasets and seismic detection datasets for roads is as follows: an acoustic excitation device and a seismic wave sensor array are arranged on the surface of the road detection area. The acoustic excitation device uses low-frequency acoustic waves for excitation, and the seismic wave sensor array uses a grid layout to collect underground structure response signals. The acoustic excitation device is activated to emit acoustic pulse signals, and the reflected echo data received by the seismic wave sensor array is acquired simultaneously.

3. The method for detecting cavities in municipal roads based on artificial intelligence according to claim 2, characterized in that, The sparse representation algorithm is used to convert high-dimensional detection data into low-dimensional sparse matrix representation, retaining key feature information while reducing computational complexity. The bandpass filter is used to filter out high-frequency noise and low-frequency interference in the detection signal, retaining signal components within the effective frequency band. The wavelet denoising algorithm is used to remove random noise in the signal through wavelet decomposition and reconstruction, thereby improving the signal-to-noise ratio.

4. The method for detecting cavities in municipal roads based on artificial intelligence according to claim 3, characterized in that, The preprocessing strategy of the conjugate gradient method is used to improve the numerical properties of the original sparse matrix by constructing a precondition submatrix, thereby accelerating the convergence speed of subsequent calculations. The precondition subtransformation is used to convert the ill-conditioned matrix into a well-conditioned matrix with a smaller condition number, thereby improving the stability of numerical calculations.

5. The method for detecting cavities in municipal roads based on artificial intelligence according to claim 4, characterized in that, The condition number represents the ratio of the maximum singular value to the minimum singular value of the matrix, reflecting the stability of the matrix numerical calculation. The global optimal solution mechanism based on convex optimization is used to find the globally optimal registration parameters during the data registration process, avoiding local optima.

6. The method for detecting cavities in municipal roads based on artificial intelligence according to claim 5, characterized in that, The convexity of the objective function refers to the mathematical property that the registration error function has a single extreme point, ensuring that the optimization algorithm converges to the global optimal solution. The spatial coordinate alignment is used to unify the data acquired by different sensors into the same coordinate system. The timing synchronization is used to eliminate the time delay differences between different sensors.

7. The method for detecting cavities in municipal roads based on artificial intelligence according to claim 6, characterized in that, The principal component analysis is used to extract the main direction of change in the data, thereby achieving feature dimensionality reduction and key information retention. The feature vector extraction is used to extract vector representations of underground structural features from the dimensionality-reduced data.

8. The method for detecting cavities in municipal roads based on artificial intelligence according to claim 7, characterized in that, The acoustic-vibration parameter adjustment function is used to optimize the operating parameters of the acoustic-vibration equipment based on the detection environment and preliminary results. The inputs include confidence assessment value, detection depth, soil type coefficient, temperature compensation factor and humidity influence factor, and the output is an acoustic-vibration parameter vector.

9. The method for detecting cavities in municipal roads based on artificial intelligence according to claim 8, characterized in that, The seismic parameter adjustment function is used to adjust the parameters of the seismic wave detection equipment according to geological conditions and detection results. The inputs include confidence assessment value, formation density, propagation velocity, attenuation coefficient and interference intensity, and the output is a seismic parameter vector.

10. The method for detecting cavities in municipal roads based on artificial intelligence according to claim 9, characterized in that, The confidence threshold analysis is used to evaluate the reliability of the detection results and identify areas that need to be supplemented with detection. The confidence score is generated by the Sigmoid activation function of the output layer of the convolutional neural network hole detection model, and the value ranges from 0 to 1. The confidence evaluation value is obtained by weighted averaging of the confidence scores of all detection points in the preliminary hole detection results.

Images