A fissure automatic positioning method fusing ultrasonic coda wave inversion and deep learning

Abstract

The application discloses a kind of fissure automatic positioning methods of fusing ultrasonic coda wave inversion and deep learning, it is related to coal mine deep mining roof disaster prevention and control technical field, including: along the direction of roof deployment active ultrasonic sensor array, acquisition reference datum full waveform data of initial state and dynamic monitoring full waveform data of dynamic state;According to the sliding window cross-correlation analysis of acquisition data, the relative travel-time offset of each sensor in different coda time window is extracted;Spatial sensitivity kernel function is constructed, and the mapping relationship between relative travel-time offset and roof three-dimensional space wave velocity disturbance is established;Inversion solves three-dimensional wave velocity disturbance distribution in roof, obtains three-dimensional space damage field;The image sequence of three-dimensional space damage field and the stress data synchronously collected are input into deep learning model, and the fissure positioning result is output.The application realizes that coal roadway roof fissure is from one-dimensional signal to three-dimensional space sub-meter precision automatic positioning, and has physical reliability and engineering practicability.

Description

Technical Field

[0001] This invention belongs to the fields of roof disaster prevention and control in deep coal mining, non-destructive testing in geological engineering, and intelligent rock mechanics, and particularly relates to an automatic fracture location method that integrates ultrasonic wake wave inversion and deep learning. Background Technology

[0002] As mining depth increases, deep coal roadways face extreme environments characterized by high ground stress, high ground temperature, high karst water pressure, and strong mining disturbances. Under these conditions, roof failure manifests as a disaster pattern where micro-cracks quietly emerge, expand, and connect internally, ultimately leading to a sudden collapse. This pattern is highly concealed and unpredictable. Traditional manual inspections and surface displacement monitoring cannot detect early-closing or semi-closing latent micro-cracks. Therefore, achieving non-destructive, high-precision three-dimensional positioning and dynamic tracking of these micro-cracks is a fundamental technical problem that urgently needs to be solved.

[0003] Currently, there are three main technologies for detecting the internal structure of surrounding rock in deep tunnels, but all of them have inherent limitations: One method is borehole optical imaging, which obtains images of cracks in the borehole wall through drilling. It is intuitive but is a point-based destructive detection method with a large blind zone between holes. It is also a static observation method that cannot track the dynamic evolution of cracks. Dense drilling is extremely costly.

[0004] Second, traditional ultrasonic and acoustic emission monitoring: Traditional ultrasonic relies on first arrival parameters and is extremely insensitive to early microcrack disturbances at the micrometer to millimeter level, and the signal is easily submerged by noise; Acoustic emission is a passive receiving technology that can only capture events of rapid release of strain energy, has limited ability to sense the crack initiation stage, and cannot achieve precise three-dimensional positioning of crack geometry.

[0005] Third, seismic / electromagnetic tomography has low resolution (decimeter level) in tunnel-scale detection, making it difficult to distinguish millimeter to centimeter-level cracks; moreover, the inversion problem based on ray tracing is highly ill-conditioned, relies on the initial model, and is prone to positioning distortion. The equipment is expensive, the process is cumbersome, and it is difficult to achieve continuous automated monitoring.

[0006] In recent years, ultrasonic wake wave technology has gained popularity due to its ability to detect changes in the elastic parameters of a medium with an accuracy of up to 10. -6 While attracting significant attention due to its scale, deep learning faces a fundamental bottleneck: "extremely sensitive perception but ambiguous spatial positioning." It typically only outputs a scalar value representing the overall damage level, failing to map damage information to specific locations in three-dimensional space. On the other hand, although deep learning has significant advantages in pattern recognition, its direct application faces two major challenges: purely data-driven "black box" models lack constraints from the laws of rock fracture mechanics (such as the Mohr-Coulomb criterion), potentially outputting physically impossible positioning results; and the scarcity of truly labeled data on deep coal mine fractures easily leads to model overfitting and poor generalization ability.

[0007] In summary, existing technologies suffer from a core technical contradiction: ultrasonic wake waves are sensitive but cannot pinpoint locations, while deep learning has recognition capabilities but lacks physical constraints. A single technical approach cannot simultaneously meet the comprehensive requirements of high sensitivity, high precision, localization, interpretability, and robustness for locating roof fractures in deep coal roadways. How to overcome the bottlenecks of point-based, passive monitoring and achieve precise three-dimensional positioning of fractures throughout their entire lifecycle is a critical technical challenge that urgently needs to be addressed in this field. Summary of the Invention

[0008] To solve the above-mentioned technical problems, this invention provides an automatic crack localization method that integrates ultrasonic wake wave inversion and deep learning. By deeply integrating and organically combining active ultrasonic wake wave full-field sensing technology, three-dimensional damage field high-resolution inversion imaging technology based on diffusion approximation and sparse regularization, and a data-physical dual-drive deep learning model (TCN-LSTM-AM) with embedded rock fracture physical mechanism constraints, a major technical breakthrough is achieved in three dimensions: (1) achieving non-destructive, highly sensitive, and accurate three-dimensional automatic localization of microcracks inside the roof of coal roadways; (2) solving the core theoretical bottleneck of ultrasonic wake wave sensing being sensitive but with ambiguous localization, and converting one-dimensional time-series wake wave signals into three-dimensional spatial crack network images; (3) overcoming the inherent defects of pure data-driven deep learning models being black boxes and unreliable, and ensuring the interpretability and engineering reliability of the localization results by introducing physical mechanism constraints.

[0009] This invention provides an automatic crack localization method integrating ultrasonic wake wave inversion and deep learning, comprising the following steps: An active ultrasonic sensor array is deployed along the direction of the top plate inside the borehole in the top plate to acquire reference full waveform data of the initial state and dynamic monitoring full waveform data of the dynamic state in full matrix capture mode. Based on the sliding window cross-correlation analysis of the reference full waveform data and the dynamic monitoring full waveform data, the relative travel time offset of each sensor in different wake time windows is extracted. A spatial sensitivity kernel function is constructed based on the diffusion approximation theory, and a mapping relationship between relative travel time offset and three-dimensional spatial wave velocity disturbance of the top plate is established based on the spatial sensitivity kernel function. Based on the mapping relationship, the three-dimensional wave velocity disturbance distribution inside the top plate is solved by sparse regularization inversion to obtain the three-dimensional spatial damage field; The image sequence of the three-dimensional spatial damage field at multiple consecutive monitoring times, the synchronously acquired stress data, and the original ultrasonic coma full waveform data are input together into a deep learning model with embedded physical constraints that integrates time-memory-attention fusion, and outputs the fracture location results and the four-level early warning probability distribution of fracture development. The output fracture location results and the four-level early warning probability distribution of fracture development are stored in a structured format and visualized.

[0010] Optionally, the active ultrasonic sensor array is arranged in a multi-row, multi-column non-uniform staggered grid layout. The sensor is coupled to the surrounding rock of the borehole wall in the top plate through high-strength epoxy resin. Each sensor unit has a built-in moisture-proof sealing protective layer and a temperature sensor to achieve automatic compensation for temperature drift.

[0011] Optionally, the process of extracting the relative travel time offset of each sensor within different wake time windows based on the sliding window cross-correlation analysis of the reference full waveform data and the dynamic monitoring full waveform data includes: The wake wave start time is determined based on a preset multiple of the arrival time of the direct wave, and the wake wave cutoff time is determined based on the signal-to-noise ratio threshold. Multiple consecutive wake wave time windows are defined between the start time and the cutoff time. For each wake wave time window, a sliding window cross-correlation function is used to compare the current dynamic monitoring waveform with the reference waveform segment by segment, and the cross-correlation coefficient corresponding to each segment is calculated. The time delay corresponding to the peak value of the cross-correlation coefficient is used as the relative travel time offset of the corresponding sensor pair within the current wake wave time window.

[0012] Optionally, the process of constructing a spatial sensitivity kernel function based on the diffusion approximation theory, and establishing a mapping relationship between the relative travel time migration and the three-dimensional spatial wave velocity disturbance of the roof based on the spatial sensitivity kernel function, includes: Based on the diffusion approximation theory, the Green's function is used to calculate the sensitivity coefficients of each sensor pair to wave velocity disturbances at various points in space within the wake time window, so as to construct a spatial sensitivity kernel function. The relative travel time offsets of each sensor pair are arranged into observation data vectors, and the wave velocity disturbances after discretization of the three-dimensional space of the top plate are arranged into parameter vectors to be determined. The spatial sensitivity kernel function is used as the coefficient matrix to establish a linear mapping relationship between the observation data vectors and the parameter vectors to be determined.

[0013] Optionally, the process of obtaining the three-dimensional spatial damage field by solving the three-dimensional wave velocity perturbation distribution inside the top plate through sparse regularization inversion based on the mapping relationship includes: Based on the mapping relationship, an objective function is constructed, including a data fitting term, a first-norm sparse regularization term, and a total variation regularization term. The objective function is iteratively solved using a fast iterative shrinkage threshold algorithm until the convergence criterion is met. The obtained wave velocity relative disturbance distribution is used as the three-dimensional spatial damage field.

[0014] Optionally, the deep learning model with embedded physical constraints and temporal-memory-attention fusion includes: a temporal convolutional network module, a long short-term memory network module, and an attention mechanism module; The temporal convolutional network module includes stacked multi-layer dilated causal convolutional layers, with the dilation coefficient of each layer increasing exponentially; the long short-term memory network module employs a multi-layer bidirectional long short-term memory network; the attention mechanism module is used to apply spatial-temporal dual attention weights; the loss function of the deep learning model includes a data-driven loss term and a physical constraint penalty term based on the Moore-Coulomb fracture criterion and the strain energy release rate criterion.

[0015] Optionally, the process of training a deep learning model includes: A synthetic dataset covering multi-dimensional fracture geometry and rock mass physical property changes was generated through finite element numerical simulation. This synthetic dataset was then used to pre-train a deep learning model. The pre-trained model was further fine-tuned using real-world labeled data validated in the field. An adaptive moment estimation optimizer, combined with a learning rate decay strategy, was employed for end-to-end training of the deep learning model, while an early stopping mechanism was used to prevent overfitting during training. Optionally, the process of inputting the image sequence of the three-dimensional spatial damage field at multiple consecutive monitoring times, the synchronously acquired stress data, and the original ultrasonic coma waveform data into a time-series-memory-attention fusion deep learning model with embedded physical constraints, and outputting the fracture location result and the four-level early warning probability distribution of fracture development includes: The system uses a sequence of images of the three-dimensional spatial damage field from multiple consecutive monitoring times, synchronously acquired stress data, and raw ultrasonic coma waveform data as joint inputs. Multi-scale spatiotemporal features are extracted using a temporal convolutional network module, and the temporal evolution dependence of fractures is captured using a long short-term memory network module. A spatial-temporal dual attention mechanism adaptively focuses on the core region of densely developed fractures in the damage field and the moment when fracture geometric features change most significantly. The output includes fracture location results and a four-level early warning probability distribution for fracture development. The fracture location results include the three-dimensional spatial coordinates of the fracture, fracture geometric parameters, damage probability, and fracture type label. The four-level early warning probability distribution includes probability outputs for four levels: stable, attention, early warning, and near-slip.

[0016] Optionally, the process of visualizing the output fracture location results includes: The fracture location results are integrated into the three-dimensional digital model of the tunnel and displayed using a combination of voxel rendering and ellipsoid primitives.

[0017] Optionally, the method also includes verifying the output fracture location results using borehole television verification and core sampling verification.

[0018] Compared with the prior art, the present invention has the following advantages and technical effects: (1) The present invention adopts a strategy of focusing the sensor array in the borehole of the top plate, which concentrates all the detection energy in the key area of ​​the top plate. Compared with the full-section layout scheme, the number of sensors is greatly reduced, the system hardware cost and on-site installation workload are reduced, and the ray coverage density and inversion resolution in the target area are improved. It has significant engineering economy and practicality.

[0019] (2) All sensors are arranged inside the borehole, eliminating the need for large-scale invasive operations on the roof and enabling non-destructive continuous detection. They can track the three-dimensional spatial evolution trajectory of cracks from initiation to penetration over a long period of time, providing reliable technical support for early warning and intelligent prevention and control of roof disasters in coal mines.

[0020] (3) This invention is based on the deployment of an active ultrasonic sensor array in the borehole of the roof and the acquisition of full waveform data in the full matrix acquisition mode. The tail wave travel time offset is extracted by cross-correlation of the sliding window. The spatial sensitivity kernel function is constructed by using the diffusion approximation theory and the mapping relationship between travel time offset and wave velocity disturbance is established. Then, the three-dimensional spatial damage field is reconstructed by sparse regularization inversion. The high sensitivity of ultrasonic tail wave to changes in the elastic parameters of the medium is transformed into a three-dimensional damage image with a clear spatial distribution. This overcomes the fundamental defect of existing tail wave technology that is sensitive but has ambiguous spatial positioning. It realizes high-resolution transparent perception of the internal fracture network of the coal roadway roof from one-dimensional time sequence signal to three-dimensional spatial position.

[0021] (4) This invention inputs the three-dimensional damage field image sequence and stress data from multiple consecutive monitoring times into a time-memory-attention fusion deep learning model with embedded physical constraints. This allows the model to utilize both time-series evolution information and stress field information. Furthermore, the rock fracture mechanics criterion is introduced into the loss function as a constraint to ensure that the output fracture location results conform to the Mohr-Coulomb fracture criterion and the strain energy release rate criterion. This solves the problem of physical unreliability caused by the black box of pure data-driven deep learning models, achieves sub-meter level three-dimensional location accuracy, and significantly improves the accuracy and engineering credibility of fracture location.

[0022] (5) The present invention realizes an end-to-end automatic crack location process: from the acquisition of raw waveform data to the extraction of wake travel time offset, damage field inversion and reconstruction, deep learning model inference output of crack location parameters, and then to structured storage and visualization display, the whole process does not require manual intervention, replacing the traditional crack picking and interpretation steps that rely on manual experience, and the single location response time meets the real-time requirements of dynamic monitoring in coal mines. Attached Figure Description

[0023] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic diagram of the spatial layout of the active ultrasonic wake wave monitoring system for the roof of a coal roadway according to an embodiment of the present invention; Figure 2 This is a schematic diagram of a three-dimensional reconstruction of the fracture network of the roof strata in a deep mining roadway according to an embodiment of the present invention. Figure 3 This is a diagram illustrating the internal architecture and topology of a deep learning model according to an embodiment of the present invention. Figure 4 This is a flowchart illustrating the automatic location algorithm for cracks in the roof of deep roadways according to an embodiment of the present invention. Detailed Implementation

[0024] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0025] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0026] Example 1 This embodiment provides an automatic crack localization method that integrates ultrasonic wake wave inversion and deep learning, including the following steps: An active ultrasonic sensor array is deployed along the direction of the top plate inside the borehole in the top plate to acquire reference full waveform data of the initial state and dynamic monitoring full waveform data of the dynamic state in full matrix capture mode. Based on the sliding window cross-correlation analysis of the reference full waveform data and the dynamic monitoring full waveform data, the relative travel time offset of each sensor in different wake time windows is extracted. A spatial sensitivity kernel function is constructed based on the diffusion approximation theory, and a mapping relationship between relative travel time offset and three-dimensional spatial wave velocity disturbance of the top plate is established based on the spatial sensitivity kernel function. Based on the mapping relationship, the three-dimensional wave velocity disturbance distribution inside the top plate is solved by sparse regularization inversion to obtain the three-dimensional spatial damage field; The image sequence of the three-dimensional spatial damage field at multiple consecutive monitoring times, the synchronously acquired stress data, and the original ultrasonic coma full waveform data are input together into a deep learning model with embedded physical constraints that integrates time-memory-attention fusion, and outputs the fracture location results and the four-level early warning probability distribution of fracture development. The output fracture location results and the four-level early warning probability distribution of fracture development are stored in a structured format and visualized.

[0027] As a specific implementation method, the automatic crack localization method proposed in this embodiment includes the following four core steps that are interconnected and progressive: Step S1—Optimized layout of active ultrasonic wake wave sensor array focusing on the roof and high-quality full waveform data acquisition; Step S2—High-resolution inversion and reconstruction of the three-dimensional spatial damage field of the roof based on wake wave interferometry (CWI), diffusion approximation theory and sparse regularization algorithm; Step S3—Automatic and accurate crack localization and geometric parameter extraction based on a deep learning model (TCN-LSTM-AM) with embedded physical constraints and time-series-memory-attention fusion; Step S4—Output, visualization and verification of crack localization results.

[0028] Step S1: Optimization and full waveform data acquisition of the active ultrasonic wake array on the focusing top plate: S1.1 Engineering Geological Condition Analysis and Monitoring Area Delineation: Before deployment, a detailed engineering geological condition survey of the coal roadway section to be monitored is conducted, focusing on collecting the following key parameters: roadway cross-sectional morphology (width, height, arch shape), roof strata sequence (thickness, elastic modulus, Poisson's ratio, P-wave velocity, S-wave velocity, etc. of each layer, including sandy mudstone, fine sandstone, and coal seam), magnitude and direction of the geostress field (horizontal maximum principal stress, minimum principal stress, and vertical stress), historical data on working face advance speed and mining disturbance intensity, and the location, scale, and development characteristics of previous roof accidents. Based on the above geological information, numerical simulation software (such as FLAC3D and ANSYS) is used to predict high-risk sections of roof stress field and fracture development, thereby scientifically delineating the scope and boundaries of the key monitoring area.

[0029] S1.2 Optimized Layout of Sensor Array Focusing on the Roof: Addressing the complex characteristics of deep roadways with high elevation, high temperature, high humidity, and high turbulence, as well as the specific prevention and control needs for roof collapse disasters, this embodiment abandons the traditional practice of blindly deploying sensors across the entire roadway cross-section (including the sidewalls and floor). Instead, it adopts a dedicated array design strategy focused on the roof. In the designated key monitoring section (typically 30m to 80m in length), based on the anisotropic characteristics of the rock strata and the analysis results of sensor coverage capabilities, an optimized non-uniform grid design scheme is adopted. Integrated transmitting / receiving active ultrasonic piezoelectric sensors (PZT-5H type, center operating frequency 50 to 200kHz, specific frequency selected according to roof lithology) are deployed in pre-constructed inclined boreholes in the roof. The number of sensors ranges from 30 to 60, with a typical layout of 5 rows × 8 columns (40 sensors in total), a row spacing of 2.0 to 3.0m, and a column spacing of 1.5 to 2.0m, arranged in an alternating pattern. All sensors are fully and tightly coupled to the borehole wall rock using high-strength epoxy resin (Araldite 2011) coupling agent, ensuring an acoustic coupling efficiency of no less than 85%. Each sensor unit incorporates a moisture-proof sealing protective layer and a high-precision temperature sensor, enabling long-term stable operation and automatic temperature drift compensation in high-humidity and high-temperature environments downhole. The sensors are connected to a multi-channel ultrasonic acquisition edge computing master control station installed in an explosion-proof box on the sidewall of the tunnel via highly shielded armored signal transmission cables.

[0030] S1.3 Full Matrix Capture (FMC) Mode Data Acquisition: The sensor array operates in Full Matrix Capture (FMC) mode, with the following specific protocol: The edge computing master station sequentially controls each sensor node in the array as an active excitation source, transmitting N-cycle (N=3 to 10, typically 5 cycles) sinusoidal pulse ultrasonic waves modulated by the Hanning window into the roof rock mass. The peak excitation voltage is 150 to 300V, and the transmission frequency matches the sensor's center operating frequency. Simultaneously with the excitation of a single node, all other sensor nodes in the array synchronously enter receiving mode, recording all waveform data completely at a sampling frequency of no less than 10MHz and an analog-to-digital converter (ADC) accuracy of 16 bits or higher. This includes direct waves, first and multiple reflected waves, and the long-duration ultrasonic tailband signal formed after complex multiple scattering in the fractured roof rock mass. After traversing all N excitation nodes, a complete N×N dimensional ultrasonic full matrix waveform dataset is obtained.

[0031] S1.4 Baseline Dataset Acquisition and Dynamic Monitoring Data Acquisition: During the initial service phase when the roof is not affected by strong mining (typically when the working face advances to more than 30m ahead of the monitoring section), several sets of FMC full matrix waveform data are continuously collected at a set frequency (1 to 2 times per day) to establish a high-quality reference baseline dataset (Background Dataset) of the initial state of the roof rock mass. As the working face gradually approaches the monitoring section, the acquisition frequency is automatically increased (once every 2 to 4 hours), entering dynamic monitoring mode. All raw waveform data are transmitted in real time to the downhole edge computing master station via 100 Mbps Ethernet fiber optic cable for real-time digital filtering, signal-to-noise ratio optimization, temperature drift compensation correction, and other preprocessing operations. The processed data is then synchronously uploaded to the ground data center server.

[0032] As an additional implementation method, in step S1, the layout scheme of the ultrasonic sensor array can be adaptively optimized according to the following factors: (1) The sensor position is optimized using compressed sensing theory to minimize the number of sensors required while ensuring the inversion resolution of the target detection area, so as to reduce the engineering implementation cost; (2) The sensor layout is appropriately densified in the section where roof accidents occur frequently in the past to achieve differentiated monitoring accuracy configuration; (3) A dedicated high-power excitation sensor and a high-sensitivity receiving sensor with separate acoustic wave transmission and reception functions are introduced to further improve the signal transmission intensity and reception sensitivity, and solve the problem of insufficient tail wave signal-to-noise ratio under deep strong attenuation rock mass conditions.

[0033] Step S2: Inversion and reconstruction of the three-dimensional spatial damage field based on the wake interferometry (CWI), diffusion approximation theory, and sparse regularization algorithm: S2.1 Fine Segmentation and Travel Time Migration Extraction of Ultrasonic Wake Wave Signals: For the acquired full-matrix waveform data, fine segmentation of the wake wave time period is first performed. Based on the elastic wave propagation and diffusion theory, the start time of the wake wave segment... The cutoff time is typically defined as 2.0 to 5.0 times the arrival time of the direct P-wave (ensuring entry into the diffusion propagation region dominated by multiple scattering). The signal-to-noise ratio (SNR) threshold is used for judgment (typically 6 dB). The moving window cross-correlation function in the wake interferometry (CWI) method is employed to compare the current wake signal u(t) with the reference state wake signal. (t) Perform segment-by-segment fine comparison, and find the time delay corresponding to the peak value of the cross-correlation coefficient in each moving window. The sensor can accurately calculate its relative travel time offset within different wake time windows. Its accuracy can reach The magnitude is far greater than that of traditional initial arrival picking methods.

[0034] S2.2 Construction of a 3D Spatial Sensitivity Kernel Function Based on Diffusion Approximation Theory: To address the core challenge of sensitive sensing but ambiguous positioning, this embodiment establishes a spatial sensitivity kernel function based on the diffusion approximation theory. This function quantitatively describes the contribution weight of elastic parameter perturbation at each location point in the three-dimensional space of the roof to the specific contribution of each sensor to the wake travel time offset. Under the assumption of an isotropic homogeneous scattering medium, the diffusion approximation equation equates the energy propagation process of sound waves in a complex scattering medium to a diffusion process, and its Green's function... The propagation law of energy from the source point to the field point is described, where D is the effective diffusion coefficient (determined by the rock mass scattering coefficient and absorption coefficient, which needs to be obtained through laboratory calibration), and t is the time taken for the ultrasonic wake wave to propagate from the excitation source to a certain point in space, i.e., the propagation time or diffusion time. For the propagation path from the excitation node s to the receiving node r, during the wake wave time period [ , Within, it disturbs the wave velocity at spatial coordinate point x. The travel time offset sensitivity kernel function of (x) / c (x) can be derived from the convolution integral of the Green's function in the diffusion equation. For actual deep coal and rock masses, an anisotropic correction factor and a frequency-dependent attenuation term are further introduced to improve the calculation accuracy of the sensitive core under complex rock mass conditions. Finally, a complete system sensitivity matrix K is constructed, with dimensions M (number of sensor pairs) × P (number of three-dimensional mesh voxels).

[0035] S2.3 High-resolution inversion of 3D damage field based on sparse regularization: The following 3D inversion optimization objective function based on sparse regularization is established: { + + } Where d is the observed data vector (composed of the wake travel offsets of all sensor pairs). (Composition); m is the vector of inversion parameters to be determined, representing the relative perturbation of wave velocity at each grid point in the three-dimensional space of the top plate. (x) / c, whose spatial distribution constitutes the mathematical expression of the three-dimensional damage field; K is the system sensitivity matrix; The L1 norm sparsity regularization parameter utilizes the prior knowledge of the sparsity of local clustering of fractures in the rock mass to preserve the sharp boundary characteristics of the fracture region and suppress false diffusion. The isotropic total variation (TV) regularization parameter is used to suppress block artifacts in the inversion results; Represents the L1 norm; Let L1 norm represent the gradient vector. The Fast Iterative Shrinkage Thresholding (FISTA) algorithm is used to efficiently solve the above objective function, typically requiring 50 to 200 iterations. The convergence criterion is that the relative change in the residual between two consecutive iterations is less than 1 / 2. After the inversion is completed, a three-dimensional wave velocity disturbance distribution image inside the roof is output, which is the visualization result of the three-dimensional spatial damage field. The area with a large relative decrease in wave velocity corresponds to the spatial region with highly developed cracks or damage, thereby realizing the fundamental transformation from one-dimensional time-series wake wave signal to three-dimensional spatial crack network image, providing high-quality input for the accurate automatic localization of subsequent deep learning models.

[0036] As an additional implementation method, in step S2, when constructing the diffusion approximation sensitive core, the following corrections can be further introduced to improve the inversion accuracy: (1) Anisotropic diffusion correction: for the natural anisotropy of layered sedimentary rock bodies (the difference in P-wave velocity in different directions can reach 10% to 30%), the anisotropic diffusion equation is used to replace the isotropic diffusion equation to construct a sensitive core that is more consistent with the real physical properties of the strata; (2) Inelastic attenuation correction: the frequency-dependent quality factor Q is introduced to correct the attenuation of the wake energy with the propagation distance, so as to eliminate the systematic deviation of the attenuation factor on the inversion results; (3) The pore elastic wave theory constraint is introduced to consider the influence of the pore fluid (groundwater) of the rock mass on the propagation characteristics of the ultrasonic wake wave, and improve the detection and location capabilities of the water-bearing fracture area.

[0037] As an alternative implementation, in step S2, the extraction of the wake travel time offset is further performed using a combination of the stretching method and the moving window cross-correlation method (MWCC). Quality control is performed by comparing the consistency of the results of the two methods. Data points whose deviation between the two methods exceeds a set threshold (10%) are automatically marked as substandard and removed to ensure the high quality and reliability of the inversion input data.

[0038] Step S3: Automatic and accurate crack localization and geometric parameter extraction based on a deep learning model (TCN-LSTM-AM) with embedded physical constraints and temporal-memory-attention fusion: S3.1 Systematic Construction of Multi-Source Heterogeneous Training Datasets: To address the scarcity of fracture annotation data in deep coal mines, this embodiment employs a transfer learning strategy that combines large-scale generation via numerical simulation with fine-tuning using a small amount of on-site calibration data. In the numerical simulation phase, a detailed three-dimensional acoustic elastic numerical model is established based on the actual geological parameters of the tunnel using finite element software (COMSOL Multiphysics) or finite difference software. A large-scale synthetic dataset covering multi-dimensional parameter variations is systematically generated. The parameters include: fracture geometric parameters (full coverage of location coordinates, uniform sampling of dip angles from 0 to 90 degrees, uniform sampling of strike angles from 0 to 360 degrees, equivalent radius from 0.05m to 1.5m, and opening degree from closed to fully open), fracture network parameters (random combination of 3 to 8 single fractures, double fractures, and multiple fractures, and fracture density field distribution at different development stages), rock mass background parameters (P-wave velocity from 2500 to 5500 m / s, Poisson's ratio from 0.15 to 0.35, weak / medium / strong scattering intensity, and anisotropy factor), and load and stress state (horizontal stress ratio from 0.5 to 1.5, and mining load coefficient from 1.0 to 3.5). Each simulation generates corresponding full-matrix ultrasonic waveform data and a 3D damage field inversion image, while simultaneously recording precise 3D parameters of the fracture as training labels. The total size of the synthesized dataset is no less than 20,000 sets. Based on this, a small amount (200 to 500 sets) of real-world labeled data verified by borehole television or core sampling is used to fine-tune the pre-trained model, effectively bridging the domain gap between simulation and reality.

[0039] S3.2 as Figure 3 As shown, the architecture design of the TCN-LSTM-AM fusion neural network model is as follows: In this embodiment, a multi-module cascaded deep learning network with embedded physical constraints is constructed. Its input is the three-dimensional spatial damage field image matrix obtained by step S2 reconstruction, the synchronously acquired regional three-dimensional stress tensor data, and the original ultrasonic coma waveform data (or the coma parameters after feature extraction, such as the coma energy envelope attenuation rate, the spectral centroid offset, the cross-correlation coefficient time series, etc.). Its output is the high-precision three-dimensional spatial positioning coordinates (X,Y,Z) of the fracture, geometric parameters (equivalent radius R, tilt angle α, direction β), damage probability p, and the four-level early warning probability distribution of fracture development (including the probability values ​​of four levels: stable, attention, early warning, and near slip). The network is composed of the following three functional modules cascaded: (1) Temporal convolutional network module (TCN Module): composed of L layers (L=4 to 8) of dilated causal convolutional layers stacked together, with the dilation coefficient of each layer increasing exponentially ( = ), enabling the network to obtain an exponentially increased receptive field without increasing the number of parameters, and to efficiently capture local high-frequency features (such as the wave velocity change signal in the stress concentration area at the crack tip) and macroscopic structural features (such as the spatial topological relationship of multiple crack penetration zones) across different spatial scales in the damage field image. Each convolutional layer is followed by a batch normalization (BatchNorm), ReLU activation function and Dropout (rate=0.2) layer, and residual connections ensure the effective propagation of gradients in the deep network. (2) Long Short-Term Memory Network Module (LSTM Module): It consists of 2 to 3 layers of bidirectional LSTM (BiLSTM) stacked together, with 256 to 512 hidden units in each layer. The LSTM module receives the multi-scale spatiotemporal feature sequence extracted by TCN, and uses its forget gate, input gate and output gate gate gate control memory mechanism to capture the long-term historical dependence law of the damage field from the initiation of microcracks to the gradual expansion in multiple continuous monitoring times, so as to realize the accurate modeling of the temporal evolution of crack spatial location. (3) Attention Mechanism Module (AM): A spatial-temporal dual attention mechanism is applied to the temporal features of the LSTM output. Spatial attention adaptively focuses on the core region of densely developed cracks in the damage field through a learnable attention weight matrix; temporal attention enables the network to automatically identify and focus on the moment when the geometric features of cracks change most significantly, effectively filtering out background noise and interference from irrelevant temporal information, and significantly improving the model's localization accuracy and noise robustness.

[0040] S3.3 Design and Embedding of the Physics-Informed Constraint Layer: During the backpropagation training of the TCN-LSTM-AM model, a physical constraint penalty term based on the principles of rock fracture mechanics is forcibly introduced into the total loss function. The total loss function is defined as follows: = + × .in, The data-driven loss term (using mean squared error MSE, which measures the deviation between the model output crack location coordinates and the predicted values ​​of geometric parameters and the actual labeled values) is used. The physical constraint loss term is composed of a weighted superposition of the following two sub-physical constraint terms: First, the Mohr-Coulomb Criterion Constraint, which ensures that the crack shear stress and normal stress predicted by the model satisfy the Mohr-Coulomb yield condition (where c is the cohesion). (The internal friction angle is calibrated by in-situ rock mechanics experiments); second, the strain energy release rate constraint ensures the strain energy released during predicted fracture propagation. Not exceeding the critical energy release rate corresponding to the rock fracture toughness (Obtained from laboratory fracture toughness tests). Physical constraint weighting coefficients ( =0.1 to 1.0 (determined by hyperparameter search). By hard-coding the above mechanical constraints into the optimization objective of the model, the black-box effect is completely broken, so that the final output crack location result has strict physical consistency and engineering interpretability.

[0041] S3.4 Model Training Strategy and Localization Performance Evaluation: An adaptive moment estimation optimizer, namely the Adam optimizer, is used (initial learning rate lr=0.001). =0.9, =0.999), combined with a cosine annealing learning rate scheduling strategy, the TCN-LSTM-AM model is trained end-to-end with 200 to 500 training epochs and batch sizes of 16 to 64. An early stopping mechanism (patience=30) is used to prevent overfitting. The dataset is divided into training, validation, and test sets in a 7:1.5:1.5 ratio. On the independent test set, the method of this embodiment has a mean error (MAE) of no more than 0.15m for the three-dimensional localization of the roof crack center point, no more than 0.08m for the equivalent radius prediction, and no more than 5 degrees for the dip angle prediction. The overall localization accuracy and robustness are significantly better than existing technologies.

[0042] As an additional implementation, in step S3, the input features of the TCN-LSTM-AM network can be further extended to multi-time-step fusion input: the three-dimensional damage field image sequence obtained by continuous monitoring at the current time and the previous N-1 times (N=5 to 20) is used as joint input, so that the deep learning model can directly learn the dynamic expansion pattern of the crack from the continuous temporal evolution image, thereby further improving the model's early identification ability of new cracks and the prediction accuracy of the evolution trend of crack geometric parameters.

[0043] As an alternative implementation, in step S3, the input of the TCN-LSTM-AM model can also be fused with the frequency domain features (tailwave energy envelope attenuation rate, spectral centroid offset, cross-correlation coefficient time series) of the original full-matrix ultrasound waveform data after feature extraction, and together with the three-dimensional damage field image, constitute a multimodal fusion input, further improving the model's ability to distinguish, identify and accurately locate different types of cracks (tension cracks, shear cracks).

[0044] Step S4: Output, visualization, and accuracy verification of the three-dimensional fracture localization results: S4.1 Structured output of three-dimensional fracture location results: After the TCN-LSTM-AM model completes inference, the system automatically outputs the following set of structured location parameters for each identified fracture: (1) Three-dimensional spatial coordinates (X,Y,Z), with the starting point of the roadway monitoring section as the origin, X as the roadway direction, Y as the roof width direction, and Z as the vertical depth from the roof surface; (2) Fracture geometric parameters - equivalent radius R (representing the scale of fracture development), dip angle alpha (the angle with the horizontal plane, ranging from 0 to 90 degrees), strike be ta (horizontal angle with north, ranging from 0 to 360 degrees); (3) damage probability p (confidence of the model in the development of cracks at this location, ranging from 0 to 1); (4) crack type label (tensional crack / shear crack / mixed crack, judged jointly by stress state and crack geometric characteristics); (5) four-level early warning probability distribution, including the probability output of four levels: stable period (no significant crack activity), attention period (micro-crack initiation), early warning period (accelerated crack propagation), and near-slip period (crack penetration, high-risk collapse), for engineers to respond and make decisions in a graded manner. All positioning results are stored in real time in standardized JSON format and pushed to the host computer for subsequent engineering applications.

[0045] S4.2 3D Visualization of Location Results: The structured fracture location parameters output from step S4.1 are integrated into the 3D digital model of the coal roadway. A combination of voxel rendering and ellipsoidal geometry is used for 3D visualization: the damage field voxel map encodes fracture density in grayscale or pseudo-color (darker / warmer colors represent higher damage); each identified fracture is represented by a semi-transparent ellipsoid, with the ellipsoid's major axis corresponding to the fracture orientation, its major axis radius corresponding to the equivalent radius R, and its transparency corresponding to the damage probability p. The spatial topological relationships of multiple fractures (whether they converge into surfaces, cut each other, etc.) are clearly visible in the 3D view. The visualization interface supports arbitrary angle rotation, profile cutting, and fracture parameter pop-up display, providing engineers with an intuitive panoramic view of the 3D spatial distribution of roof fractures.

[0046] S4.3 Engineering Verification Methods for Positioning Accuracy: To verify the positioning accuracy of the method in this embodiment, the following two engineering verification methods are provided: (1) Borehole Camera Verification Method: A small number of verification boreholes are constructed at the coordinates of the fracture-dense area output by the system. The predicted position is directly observed using a borehole camera. The actual development depth of the fractures revealed by the borehole imaging is compared with the Z coordinate predicted by the model to evaluate the vertical positioning error; (2) Core Sampling Verification Method: Core samples are taken from the verification boreholes. The actual spacing, dip angle, and filling characteristics of the fractures on the core are statistically analyzed. The fracture geometric parameters output by the model are systematically compared to evaluate the positioning accuracy and the accuracy of the geometric parameter prediction. The above verification methods are periodic accuracy verification tools and do not affect the system's online positioning and continuous monitoring capabilities.

[0047] As an additional implementation method, in step S4, the positioning results can be fused and analyzed with the data from the mine pressure monitoring system (roof and floor approach distance, support working resistance data) and the distributed optical fiber strain monitoring system (BOTDR / BOTDA) to cross-validate and comprehensively evaluate the positioning results, improve the reliability of the fracture positioning conclusion, and provide engineers with a more comprehensive basis for evaluating the development status of roof fractures.

[0048] In this embodiment, the above technical solution achieves the following significant effects: (i) By combining the high sensitivity of ultrasonic wake interferometry with the spatial sensitive kernel derivation based on the diffusion approximation theory and sparse regularized three-dimensional inversion, the damage information contained in the one-dimensional time-series wake signal was successfully restored into a three-dimensional spatial crack network image inside the roof. This scheme solves the problem that existing wake technology is sensitive but has ambiguous positioning, and improves the detection capability of cracks inside the roof from an average damage index with no spatial resolution to three-dimensional spatial positioning with sub-meter accuracy.

[0049] (ii) The Mohr-Coulomb fracture criterion and the strain energy release rate fracture mechanics criterion were embedded as physical penalty terms into the optimization objective function of the deep learning model, thus constructing a data-driven and physical mechanism-constrained fusion learning framework. This framework ensures that the fracture location coordinates and geometric morphology output by the model strictly follow the basic laws of rock mechanics, avoiding the unreliable physical predictions that may arise from a purely data-driven model. This results in high engineering interpretability and field credibility of the location results, with the fracture center point location accuracy reaching sub-meter level.

[0050] (iii) The use of a time-memory-attention fusion deep learning network replaces the traditional crack picking process that relies on human experience, enabling automatic identification of crack features from three-dimensional wave velocity perturbation images and output of quantitative positioning parameters. This greatly improves positioning efficiency and objectivity, with a short single complete inference response time, and can support near real-time tracking and positioning of dynamic crack development processes.

[0051] (iv) By adopting a sensor layout strategy that focuses on the top plate, all detection energy is concentrated in the key area of ​​the top plate. Compared with the full-section layout scheme, the number of sensors used is significantly reduced, the system hardware cost and on-site installation workload are reduced. At the same time, the ray density and spatial sampling rate in the target monitoring area are improved, thereby obtaining better inversion resolution, which is conducive to high-precision crack location.

[0052] (v) All sensors are placed inside the boreholes in the roof, eliminating the need for large-scale invasive operations on the roof and enabling non-destructive continuous detection. The system can operate for extended periods, dynamically recording the three-dimensional spatial location of cracks from initiation to expansion and penetration as a complete evolution trajectory. The obtained data can be used by the engineering team for subsequent analysis and application, meeting the needs of intelligent and continuous detection in coal mines.

[0053] Example 2 This embodiment takes a three-dimensional automatic positioning project for roof fractures in a deep fully mechanized mining roadway as an example: Background: A longwall mining face (approximately 1050m deep) in a coal mine has a roadway with a net width of 5.2m and a net height of 3.8m, featuring a semi-circular arched roof. The roof strata, from bottom to top, consist of: immediate roof mudstone (approximately 3m thick, P-wave velocity approximately 2800m / s), underlying fine sandstone (approximately 8m thick, P-wave velocity approximately 4200m / s), and interbedded sandy mudstone. The working face advances at a rate of approximately 3 to 5 m / d. Historically, this working face has experienced several periodic roof collapses and localized cave-ins during mining operations, making roof fracture location extremely difficult. This implementation example fully covers the implementation details of the four steps, verifying the effectiveness of the method proposed in Example 1 for fracture location in engineering practice.

[0054] Step 1: Engineering geological analysis and optimized deployment of the focused ultrasonic wake array on the roof: like Figure 1 The diagram shows the spatial layout of the active ultrasonic wake wave monitoring system for the coal roadway roof in this embodiment. It illustrates the array configuration in which transmitting sensors (Tx) and receiving sensors (Rx) are arranged alternately in the boreholes of the roadway roof. The acoustic rays cross and cover the entire area of ​​the roof. The edge computing master station is connected to the array through transmission cables to form a full matrix capture (FMC) network.

[0055] Based on the detailed collection of engineering geological data (borehole columnar sections, laboratory rock mechanics parameters, and geostress test reports) for this roadway section, FLAC3D numerical simulation was used to predict high-risk sections of roof stress concentration and fracture development. It was determined that a monitoring section should be set up in the roadway roof within 50m ahead of the working face, with a monitoring volume of approximately 50m (roadway direction) × 5m (roof width) × 12m (roof depth).

[0056] Sensor deployment scheme: Forty broadband PZT-5H piezoelectric ultrasonic sensors (25mm diameter, 100kHz center frequency, measured -6dB bandwidth of 65 to 150kHz) are arranged in a 5-row × 8-column staggered grid (row spacing 2.5m, column spacing 1.8m). The sensor housing is made of nickel-plated brass (IP68 dustproof and waterproof rating). The drilled holes are fully filled with epoxy resin (Araldite 2011, acoustic impedance of approximately 2.8MRayl after curing) to ensure a high-quality acoustic coupling interface between the sensor and the surrounding rock. The coupling efficiency has been verified to be no less than 87%. The coaxial shielded signal cables from each sensor run along the roof and converge into a 40-channel ultrasonic acquisition edge computing main control station (sampling rate 12MHz, ADC resolution 16-bit, built-in FPGA real-time digital filter) fixed in an explosion-proof enclosure on the left side of the roadway. The overall power consumption of the system is no more than 150W, meeting the intrinsically safe explosion-proof requirements for mining.

[0057] Baseline data acquisition: When the working face is about 80m in front of the monitoring section, the system is started and run in FMC mode. A complete baseline dataset is collected once at 8:00 and 20:00 every day (each acquisition takes about 3 minutes). After 7 days of continuous acquisition, a total of 14 sets of high-quality baseline FMC full matrix data (40×40=1600 path waveforms, each waveform duration is 1ms, and the number of sampling points is 12000) are obtained, and a high-quality reference baseline database is established.

[0058] Step 2: Ultrasonic wake signal processing and three-dimensional damage field inversion and reconstruction: like Figure 2 The image shown is a schematic diagram of the three-dimensional reconstruction effect of the fracture network (damage field) of the roof rock strata in the deep mining roadway in this embodiment. The fracture network in the 20m×20m×20m three-dimensional space in the image is reconstructed and visualized from the ultrasonic wake signal by the sparse regularization inversion algorithm. The densely developed fracture area is clearly presented in the image with high-density lines, which reflects the fundamental transformation from one-dimensional time series signal to three-dimensional spatial fracture image.

[0059] When the working face advances to about 35m from the monitoring section (the impact of mining disturbance begins to be significant), the dynamic monitoring mode is triggered, and the sampling frequency is automatically increased to once every 3 hours.

[0060] Wakewave signal extraction and travel time offset calculation: For each newly acquired FMC full matrix data, the system automatically extracts the waveform from 120 microseconds to 520 microseconds after the arrival of the P-wave as the wakewave analysis period. During this period, the signal-to-noise ratio (SNR) meets the quality control standard of no less than 8 dB. A sliding cross-correlation function with a window length of 30 microseconds and a step size of 5 microseconds is used to perform cross-correlation analysis between the current wakewave signal and the corresponding reference signal for each path, automatically extracting the travel time offset. Statistical analysis shows that during the initial mining phase (when the working face is approximately 35m from the monitoring area), the average travel time deviation for each path is approximately 0.08 microseconds, corresponding to a relative change in wave velocity of approximately 1.6 × 10⁻⁶. This has far exceeded the system's minimum detectable accuracy (approximately 5×). The signal characteristics are obvious.

[0061] Three-dimensional damage field inversion: based on the laboratory-calibrated effective diffusion coefficient D=0.018 / s (after anisotropic correction), calculate the 3D spatial sensitivity kernel function corresponding to the path of all 1600 sensor pairs at 300 microseconds, and construct the system sensitivity matrix K (dimension 1600×125000, corresponding to a 50×50×50 3D grid within the monitoring volume of the top plate, grid size 0.4m×0.4m×0.4m). Set the sparsity regularization parameter lambda=0.02 and the TV regularization parameter mu=0.005, and use the FISTA algorithm for iterative solution (a total of 120 iterations, the objective function converged, and the relative change of the residuals <8× The inversion results show that a significant low wave velocity anomaly zone appears at the junction of the immediate roof and the main roof (approximately 2.5m deep from the roadway roof), with a relative decrease in wave velocity. / c reaches a maximum of 12.3%, and the abnormal area has a planar projection area of ​​approximately 2.8 square meters. The three-dimensional damage field image clearly reveals the spatial distribution of internal fracture development.

[0062] Step 3: Automatic crack localization using the TCN-LSTM-AM deep learning model: The three-dimensional damage field image sequence (dimension 5×50×50×50) at five consecutive monitoring times (t-4h, t-3h, t-2h, t-1h, t) was combined with the synchronously acquired triaxial stress data. =28.5MPa, =22.3MPa, =26.0MPa, =3.2MPa, =1.8MPa, =2.5MPa) as joint input, and imported into the pre-trained and fine-tuned TCN-LSTM-AM model. Model configuration: TCN uses 6 layers of dilated convolution (dilation coefficient 1 / 2 / 4 / 8 / 16 / 32, number of channels 64 / 128 / 256 / 256 / 256 / 256); BiLSTM uses 2 layers (512 hidden units); a spatial + temporal dual attention mechanism is applied; the fully connected output layer outputs a 7-dimensional vector (X,Y,Z,R, tilt angle, orientation,p).

[0063] Model location results (response time approximately 2.3 seconds): The system identified the primary hazardous fracture cluster C1, with center coordinates (X=23.4m, Y=11.8m, Z=2.6m) (originating from the monitoring area's starting point, where Z is the depth from the roadway roof), equivalent radius R=0.62m, dip angle 67 degrees, strike 248 degrees, and damage probability p=0.48. Simultaneously, a secondary fracture cluster C2 was identified, with center coordinates (X=31.2m, Y=14.3m, Z=3.1m), equivalent radius R=0.35m, and damage probability p=0.27. The system stores all fracture location results in structured JSON format in real time and pushes them to the underground safety monitoring display screen and the surface data center for engineers to view and analyze.

[0064] Step 4: Output, visualization, and accuracy verification of crack location results: like Figure 4 The diagram shown is the overall flowchart of the automatic coal roadway roof crack location algorithm in this embodiment, demonstrating a complete four-stage technical route: the first stage is physical layer data acquisition (full waveform acquisition of roof ultrasonic wake waves) → the second stage is analytical layer feature extraction and three-dimensional imaging (CWI + diffusion approximation + sparse regularized high-resolution inversion) → the third stage is intelligent layer data and object dual-drive positioning (TCN-LSTM-AM + physical constraints) → the fourth stage is output layer result output and verification (structured output of crack three-dimensional coordinates + visualization + accuracy verification), and the system integrates the full-process automated execution capability.

[0065] Visualization of Location Results: The 3D location coordinates and geometric parameters of fracture clusters C1 and C2 are imported into the 3D digital model of the tunnel. The spatial morphology of the fractures is rendered using semi-transparent ellipsoidal primitives, and the wave velocity disturbance intensity distribution is presented using pseudo-color damage field voxel maps. In the 3D visualization interface, fracture cluster C1 (R=0.62m, p=0.48) is highlighted with a high-transparency, high-saturation ellipsoid, while fracture cluster C2 (R=0.35m, p=0.27) is displayed with a low-transparency ellipsoid. The relative positions and spatial relationships of the two are clear at a glance, providing engineers with an intuitive 3D panorama of the roof fracture development status.

[0066] Positioning accuracy verification: After the monitoring period ended, two verification boreholes (3.5m deep) were drilled above the fracture-dense area C1 indicated by the system, and borehole television observation was conducted. The borehole television images clearly revealed a significant fracture development zone near the predicted location (Z=2.6m) (measured depth range of 2.4m to 2.8m). The fracture dip angle revealed by the borehole imaging was approximately 65 to 70 degrees, which highly matched the model's predicted dip angle of 67 degrees. The vertical positioning error did not exceed 0.2m, verifying that the positioning accuracy of this method meets the requirements for engineering applications.

[0067] The above embodiments fully demonstrate that the automatic positioning method proposed in Embodiment 1 has a strong ability to detect cracks, accurate three-dimensional spatial positioning accuracy, and objective and reliable positioning results under deep and complex geological conditions. It can effectively provide technical support for the refined detection and engineering assessment of cracks in the roof of deep coal mines and has broad prospects for promotion and application.

[0068] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for automatic crack localization integrating ultrasonic wake wave inversion and deep learning, characterized in that, Includes the following steps: An active ultrasonic sensor array is deployed along the direction of the top plate inside the borehole in the top plate to acquire reference full waveform data of the initial state and dynamic monitoring full waveform data of the dynamic state in full matrix capture mode. Based on the sliding window cross-correlation analysis of the reference full waveform data and the dynamic monitoring full waveform data, the relative travel time offset of each sensor in different wake time windows is extracted. A spatial sensitivity kernel function is constructed based on the diffusion approximation theory, and a mapping relationship between relative travel time offset and three-dimensional spatial wave velocity disturbance of the top plate is established based on the spatial sensitivity kernel function. Based on the mapping relationship, the three-dimensional wave velocity disturbance distribution inside the top plate is solved by sparse regularization inversion to obtain the three-dimensional spatial damage field; The image sequence of the three-dimensional spatial damage field at multiple consecutive monitoring times, the synchronously acquired stress data, and the original ultrasonic coma full waveform data are input together into a deep learning model with embedded physical constraints that integrates time-memory-attention fusion, and outputs the fracture location results and the four-level early warning probability distribution of fracture development. The output fracture location results and the four-level early warning probability distribution of fracture development are stored in a structured format and visualized. The deep learning model that integrates temporal-memory-attention with embedded physical constraints includes: a temporal convolutional network module, a long short-term memory network module, and an attention mechanism module; The temporal convolutional network module includes stacked multi-layer dilated causal convolutional layers, with the dilation coefficient of each layer increasing exponentially; the long short-term memory network module employs a multi-layer bidirectional long short-term memory network; the attention mechanism module is used to apply spatial-temporal dual attention weights; the loss function of the deep learning model includes a data-driven loss term and a physical constraint penalty term based on the Moore-Coulomb fracture criterion and the strain energy release rate criterion.

2. The method according to claim 1, characterized in that, The active ultrasonic sensor array is deployed in a multi-row, multi-column non-uniform staggered grid layout. The sensors are coupled to the surrounding rock of the borehole wall in the top plate through high-strength epoxy resin. Each sensor unit has a built-in moisture-proof sealing protective layer and a temperature sensor to achieve automatic compensation for temperature drift.

3. The method according to claim 1, characterized in that, Based on the sliding window cross-correlation analysis of the reference full waveform data and the dynamic monitoring full waveform data, the process of extracting the relative travel time offset of each sensor within different wake time windows includes: The wake wave start time is determined based on a preset multiple of the arrival time of the direct wave, and the wake wave cutoff time is determined based on the signal-to-noise ratio threshold. Multiple consecutive wake wave time windows are defined between the start time and the cutoff time. For each wake wave time window, a sliding window cross-correlation function is used to compare the current dynamic monitoring waveform with the reference waveform segment by segment, and the cross-correlation coefficient corresponding to each segment is calculated. The time delay corresponding to the peak value of the cross-correlation coefficient is used as the relative travel time offset of the corresponding sensor pair within the current wake wave time window.

4. The method according to claim 1, characterized in that, The process of constructing a spatial sensitivity kernel function based on the diffusion approximation theory, and establishing a mapping relationship between relative travel time migration and three-dimensional spatial wave velocity disturbance of the roof based on the spatial sensitivity kernel function, includes: Based on the diffusion approximation theory, the Green's function is used to calculate the sensitivity coefficients of each sensor pair to wave velocity disturbances at various points in space within the wake time window, so as to construct a spatial sensitivity kernel function. The relative travel time offsets of each sensor pair are arranged into observation data vectors, and the wave velocity disturbances after discretization of the three-dimensional space of the top plate are arranged into parameter vectors to be determined. The spatial sensitivity kernel function is used as the coefficient matrix to establish a linear mapping relationship between the observation data vectors and the parameter vectors to be determined.

5. The method according to claim 4, characterized in that, Based on the aforementioned mapping relationship, the process of obtaining the three-dimensional spatial damage field by solving the three-dimensional wave velocity perturbation distribution inside the top plate through sparse regularization inversion includes: Based on the mapping relationship, an objective function is constructed, including a data fitting term, a first-norm sparse regularization term, and a total variation regularization term. The objective function is iteratively solved using a fast iterative shrinkage threshold algorithm until the convergence criterion is met. The obtained wave velocity relative disturbance distribution is used as the three-dimensional spatial damage field.

6. The method according to claim 1, characterized in that, The process of training a deep learning model includes: A synthetic dataset covering the changes in multi-dimensional fracture geometric parameters and rock mass physical parameters is generated through finite element numerical simulation. The deep learning model is pre-trained using the synthetic dataset. The pre-trained deep learning model is then fine-tuned using real labeled data verified in the field. An adaptive moment estimation optimizer is used, combined with a learning rate decay strategy, to perform end-to-end training of the deep learning model. At the same time, an early stopping mechanism is used to prevent overfitting during the training process.

7. The method according to claim 1, characterized in that, The process of inputting image sequences of the three-dimensional spatial damage field from multiple consecutive monitoring times, synchronously acquired stress data, and raw ultrasonic coma waveform data into a time-series-memory-attention fusion deep learning model with embedded physical constraints, and outputting fracture location results and a four-level early warning probability distribution for fracture development, includes: The system uses a sequence of images of the three-dimensional spatial damage field from multiple consecutive monitoring times, synchronously acquired stress data, and raw ultrasonic coma waveform data as joint inputs. Multi-scale spatiotemporal features are extracted using a temporal convolutional network module, and the temporal evolution dependence of fractures is captured using a long short-term memory network module. A spatial-temporal dual attention mechanism adaptively focuses on the core region of densely developed fractures in the damage field and the moment when fracture geometric features change most significantly. The output includes fracture location results and a four-level early warning probability distribution for fracture development. The fracture location results include the three-dimensional spatial coordinates of the fracture, fracture geometric parameters, damage probability, and fracture type label. The four-level early warning probability distribution includes probability outputs for four levels: stable, attention, early warning, and near-slip.

8. The method according to claim 1, characterized in that, The process of visualizing the output fracture location results includes: The fracture location results are integrated into the three-dimensional digital model of the tunnel and displayed using a combination of voxel rendering and ellipsoid primitives.

9. The method according to claim 1, characterized in that, It also includes using borehole television verification and core sampling verification methods to verify the output fracture location results.

Images