Ice rock mass crack monitoring and early warning method and system based on artificial intelligence

Abstract

The invention provides an artificial intelligence-based ice rock mass crack monitoring and early warning method and system, and the method comprises the steps: collecting real-time geological fluctuation data including sound wave reflection, stress distribution and temperature change signals through a multi-mode sensor array disposed on the surface of an ice rock mass, inputting the real-time geological fluctuation data into a pre-trained multi-dimensional feature fusion model, and carrying out the real-time geological fluctuation data fusion; and extracting a crack feature vector containing parameters such as crack space coordinates and the like. And calling a spatio-temporal evolution prediction model to carry out time sequence analysis on the feature vector, generating a probability distribution diagram containing a crack propagation direction and a critical fracture risk level in a future preset time period, and when the critical fracture risk level in the probability distribution diagram exceeds a preset threshold value, triggering an early warning signal and generating an emergency support scheme. According to the scheme, support point coordinates, support strength parameters and construction time sequence instructions are covered, effective monitoring and early warning of ice rock mass cracks are achieved, and targeted emergency measures are provided.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an ice-rock crack monitoring and early warning method and system based on artificial intelligence. Background Art

[0002] In the field of ice-rock engineering, crack monitoring and early warning technologies are crucial for ensuring the safety of infrastructure in high-altitude and cold regions. Traditional monitoring methods mainly rely on single-point sensors (such as strain gauges and inclinometers) or regular manual inspections, which have significant technical limitations. Single-point sensors can only obtain discrete data in local areas, making it difficult to capture the three-dimensional spatial distribution and dynamic evolution of cracks within ice-rock masses; manual inspections are limited by harsh natural environments and high-risk operating conditions, and cannot meet real-time and continuous monitoring needs. In addition, existing methods are mostly based on static threshold judgments or simple linear prediction models, which cannot adapt to the nonlinear deformation characteristics of ice-rock masses under the coupling of complex geological stress fields and temperature fields, resulting in low early warning accuracy and poor timeliness.

[0003] At the data processing and analysis level, relevant technologies lack the ability to integrate and analyze multi-physics field information, making it difficult to reveal the inherent connections between crack initiation, expansion, and multi-field coupling. For predicting crack evolution paths, existing methods often use deterministic models or empirical formulas, which are unable to quantify the uncertainty of the prediction results and make it difficult for decision makers to conduct risk probability assessments and optimize the allocation of emergency resources. In terms of early warning and response mechanisms, traditional solutions can only provide simple alarm information and lack intelligent linkage with support design and construction plans, resulting in delayed or insufficiently targeted emergency measures. Summary of the Invention

[0004] In view of the above-mentioned problems, in combination with the first aspect of the present invention, an embodiment of the present invention provides an ice and rock crack monitoring and early warning method based on artificial intelligence, the method comprising: A multimodal sensor array deployed on the surface of the ice and rock mass collects real-time geological fluctuation data, including acoustic wave reflection signals, stress distribution signals, and temperature change signals; Inputting the geological fluctuation data into a pre-trained multi-dimensional feature fusion model to extract the crack feature vector inside the ice rock body, wherein the crack feature vector includes the crack spatial coordinates, crack propagation rate and dynamic perception parameters of the stress concentration area; Calling a spatiotemporal evolution prediction model to perform a time series analysis on the crack characteristic vector to generate a probability distribution map of the crack evolution path, wherein the probability distribution map includes the crack propagation direction and critical fracture risk level within a preset future time period; According to the area where the critical fracture risk level in the probability distribution map exceeds a preset threshold, an early warning signal is triggered and a corresponding emergency support plan is generated. The emergency support plan includes support point coordinates, support strength parameters and construction timing instructions.

[0005] On the other hand, an embodiment of the present invention also provides an ice and rock crack monitoring and early warning system based on artificial intelligence, including a processor and a machine-readable storage medium, wherein the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0006] Based on the above aspects, the embodiment of the present application realizes multi-physical field coupled monitoring of micro-damage and macro-deformation inside the ice and rock mass by deploying a multimodal sensor array on the surface of the ice and rock mass to collect sound wave reflection, stress distribution and temperature change signals in real time. The pre-trained multi-dimensional feature fusion model extracts the dynamic perception parameters of the crack spatial coordinates, expansion rate and stress concentration area, which not only depicts the geometric morphology and mechanical evolution characteristics of the crack, but also establishes a probability distribution map of the crack expansion path through the spatiotemporal evolution prediction model. The probability distribution map can quantify the uncertainty of the crack expansion direction and the spatial distribution of the critical fracture risk in the future preset time period. When the critical fracture risk level exceeds the preset threshold, it can synchronously trigger an early warning signal and generate an emergency support plan including support points, strength parameters and construction sequence, significantly improving the safety and reliability of ice and rock mass projects in extreme environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1 It is a schematic diagram of the execution flow of the ice-rock crack monitoring and early warning method based on artificial intelligence provided by an embodiment of the present invention.

[0008] Figure 2 Schematic diagram of exemplary hardware and software components of an artificial intelligence-based ice and rock crack monitoring and early warning system provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0009] The present invention will be described in detail below with reference to the accompanying drawings. Figure 1 This is a flow chart of an artificial intelligence-based ice and rock crack monitoring and early warning method provided by an embodiment of the present invention. The artificial intelligence-based ice and rock crack monitoring and early warning method is introduced in detail below.

[0010] Step S110 , real-time geological fluctuation data is collected by a multimodal sensor array deployed on the surface of the ice and rock mass, wherein the geological fluctuation data includes an acoustic wave reflection signal, a stress distribution signal, and a temperature change signal.

[0011] In this embodiment, an engineering scenario inside an ice-rock mixed mountain is considered, for example, an underground facility construction project is carried out inside an ice-rock mountain, and a multimodal sensor array is deployed on the surface of the ice-rock body.

[0012] Specifically, to collect acoustic reflection signals, the sensor transmits acoustic waves of a specific frequency and then receives acoustic waves reflected from various structures within the ice and rock mass. For example, when an acoustic wave encounters a crack or an area of varying density within the ice and rock, the reflected wave's frequency, intensity, and phase characteristics change. The sensor continuously records information about these reflected waves, forming an acoustic reflection signal. For example, at a certain moment, the sensor transmits an acoustic wave with a frequency of 1000 Hz. Some time later, an acoustic wave is received that is reflected from a suspected crack 5 meters away from the sensor. The frequency of the sound wave changes to 950 Hz, and the intensity decreases. All of this data is accurately recorded.

[0013] Stress distribution signals are collected using stress sensors. Due to factors such as gravity, external pressure, and internal structural changes within the ice and rock mass, stress distribution is uneven. Stress sensors accurately measure the stress magnitude at each sampling point. For example, near the mountain's surface, stress may be relatively low due to temperature fluctuations and atmospheric pressure. Deep within the mountain, stress may be higher due to the weight of the ice and rock above. The sensors collect stress data at regular intervals, such as every 10 minutes, to generate stress distribution signals. For example, at a sampling point at a depth of 10 meters, the initial measured stress value is 1000 Pascals. Over time, due to creep of the ice and rock or external disturbances, the stress value may change, for example, to 1050 Pascals. These changes are recorded by the stress sensors.

[0014] The temperature of an ice and rock mass significantly affects its structural stability. Temperature sensors can monitor temperature changes at the surface and shallow depths of the mass in real time. Due to the varying thermal conductivity of ice and rock, temperature changes vary at different depths. For example, in winter, the temperature on the surface of a mountain may drop rapidly to -10°C. However, at a depth of two meters within the mountain, the temperature drops more slowly, starting at -2°C and dropping to -3°C after a day. Temperature sensors record these temperature changes hourly, generating a temperature change signal. The data collected by these different types of sensors collectively constitutes geological fluctuation data.

[0015] Step S120 , inputting the geological fluctuation data into a pre-trained multi-dimensional feature fusion model to extract the crack feature vector inside the ice rock body, wherein the crack feature vector includes the crack spatial coordinates, crack propagation rate, and dynamic perception parameters of the stress concentration area.

[0016] In this embodiment, after obtaining the above-mentioned geological fluctuation data, it can be input into a pre-trained multi-dimensional feature fusion model, which has been pre-trained with a large amount of sample geological fluctuation data and corresponding sample fracture annotation data.

[0017] Taking the aforementioned underground facility construction scenario as an example, the acoustic reflection signal data received by the multidimensional feature fusion model is first processed. Assuming the multidimensional feature fusion model has determined the relationship between acoustic reflection signals within a certain frequency range and cracks, such as that anomalies in reflection waves between 900 and 1000 Hz may be associated with cracks, the multidimensional feature fusion model can analyze the collected acoustic reflection signals and predict the reflection wave portions that fall within this frequency range and have specific intensity and phase characteristics, thereby determining the approximate area where cracks may exist.

[0018] The multidimensional feature fusion model can analyze stress distribution signals in conjunction with the mechanical properties of ice and rock masses. When a significant gradient in stress distribution occurs, such as a sudden increase from 1000 Pascals to 1500 Pascals in a specific area that doesn't conform to normal stress distribution patterns, the multidimensional feature fusion model can indicate the presence of stress concentration in that area, which is often associated with the presence and growth of cracks. Based on the stress distribution, the model calculates dynamic perception parameters such as the extent and degree of stress concentration.

[0019] Temperature change signals are also comprehensively considered by the model. In ice and rock, the temperature around cracks may differ from that of surrounding areas due to factors such as air or water flow. If the temperature change rate in one area is significantly higher than in other areas, the multidimensional feature fusion model can analyze this as a feature. For example, if the temperature in one area drops by 5°C in a short period of time, while the surrounding area only drops by 1°C, this suggests the presence of cracks in that area, as cracks may affect heat conduction.

[0020] By comprehensively analyzing these signals, the multidimensional feature fusion model can accurately extract the characteristic vectors of cracks within the ice-rock body. Suppose the multidimensional feature fusion model ultimately determines the presence of a crack 3 meters horizontally and 5 meters vertically from the sensor array, with a crack growth rate of 0.1 meters per month. There is also a stress concentration area surrounding the crack with a stress gradient of 100 Pascals per meter. These parameters constitute the crack characteristic vector.

[0021] Step S130 , calling a spatiotemporal evolution prediction model to perform time series analysis on the crack feature vector, and generating a probability distribution diagram of the crack evolution path, wherein the probability distribution diagram includes the crack propagation direction and critical fracture risk level within a future preset time period.

[0022] After obtaining the crack feature vector, the spatiotemporal evolution prediction model can be used for analysis. Continuing with the example of underground facility construction, this spatiotemporal evolution prediction model builds a hybrid neural network architecture that includes a long-short-term memory network and a three-dimensional convolution kernel.

[0023] The LSTM network begins to capture the time-dependent characteristics of the crack growth rate from the crack feature vector. For example, the spatiotemporal evolution prediction model can track changes in the crack growth rate over the past few months. If the crack growth rate has been 0.08, 0.09, and 0.1 meters per month over the past three months, respectively, showing a gradual upward trend, the spatiotemporal evolution prediction model can record this trend as a time-dependent feature.

[0024] The 3D convolution kernel extracts topological features of the crack's spatial morphology from the crack's feature vector. For example, based on the crack's spatial coordinates and the shape of the stress concentration area, the spatiotemporal evolution prediction model can construct the crack's approximate 3D shape, including topological features such as the crack's length, width, direction, and spatial relationship to surrounding stress concentration areas.

[0025] Then, the time-dependent features are dynamically fused with the topological features through a gating mechanism to generate dynamic fusion features that contain time-space correlations. Assume that based on the previous analysis, the spatiotemporal evolution prediction model predicts that the crack is likely to continue to expand along the current direction within the next three months, and due to the influence of stress concentration areas, the probability of crack expansion will be higher in certain directions.

[0026] Next, the dynamic fusion features are input into the discriminator network. The spatiotemporal evolution prediction model can obtain the spatial coordinate point sets of actual rupture events corresponding to the crack feature vectors. These spatial coordinate point sets may come from a database of past similar ice-rock rupture events or small-scale rupture events monitored in the field. These spatial coordinate point sets are then converted into a binary rupture mask map with the same resolution as the spatial probability distribution map output by the spatiotemporal evolution prediction model. For example, if a coordinate point was identified as a rupture point in a past rupture event, the corresponding spatial cell in the binary rupture mask map is marked as 1; otherwise, it is marked as 0.

[0027] Next, the cross-entropy loss is calculated between the probability value of each spatial unit in the spatial probability distribution map and the label value of the corresponding spatial unit in the binary fracture mask map to generate the initial adversarial loss component. Simultaneously, a set of spatial units in the spatial probability distribution map whose probability values exceed a preset probability threshold (e.g., 0.5) is extracted. The spatial intersection-over-union (IoU) of this set of spatial units with the fracture region in the binary fracture mask map is calculated to generate an IoU penalty term. The initial adversarial loss component and the IoU penalty term are linearly combined with the preset weight coefficient to generate the final adversarial training loss. In this way, the model continuously optimizes the prediction results, ultimately generating a probability distribution map of the fracture evolution path. This probability distribution map shows that within a preset time period in the future (e.g., the next three months), the probability of the fracture propagating in one direction is 0.6, while the probability of another direction is 0.3. It also indicates the critical fracture risk level for different regions, such as high in one region and medium or low in others.

[0028] Step S140: triggering an early warning signal and generating a corresponding emergency support plan based on the area where the critical fracture risk level in the probability distribution diagram exceeds a preset threshold, wherein the emergency support plan includes support point coordinates, support strength parameters, and construction timing instructions.

[0029] Assuming a preset threshold is the critical value for a high risk level, if the critical fracture risk level in a certain area of the probability distribution map exceeds the preset threshold, an early warning signal is triggered. For example, in the aforementioned underground facility construction scenario, if the critical fracture risk level is found to be high in the area 5 meters from the top of the underground facility and 8 meters horizontally, it means that the ice-rock structure in this area is extremely unstable and may fracture at any time, threatening the safety of the underground facility.

[0030] First, all spatial cells in the probability distribution map are traversed to identify a set of spatial cells with a critical fracture risk level exceeding a preset threshold. These cells are marked as the initial danger zone. A spatial connectivity analysis is performed on this initial danger zone. For example, if several adjacent spatial cells are found, although their individual risk levels are slightly below the threshold, the spatial distance between them is less than a preset safety radius (e.g., 1 meter). These adjacent spatial cells are then merged to form a continuous joint danger zone.

[0031] Next, the geological structural characteristics of the joint hazard zone were extracted. Previously collected geological fluctuation data and analysis results revealed that the maximum crack growth rate in this joint hazard zone was 0.12 meters per month, the average stress concentration factor was 1200 Pascals per square meter, and the temperature gradient distribution pattern showed a rapid decrease in temperature from the surface to the interior.

[0032] These geological structural characteristics are then matched against the characteristics of support cases in a pre-existing support solution library. The support solution library is constructed by acquiring a historical support case dataset, which contains support solutions for a variety of geological structures. Assume that the geological structural characteristics of some support cases in the support solution library are similar to those of the current joint hazard area. For example, the crack propagation rate, stress concentration factor, and temperature gradient distribution pattern in the historical cases are similar to those in the current area. The support case set with a matching degree exceeding a similarity threshold (e.g., 0.8) is selected.

[0033] An initial support plan template is generated based on the support strength parameters and construction sequence instructions from the support case collection. For example, one case in the support case collection recommends setting up support points every 2 meters around the danger zone. The support strength parameter is set to be able to withstand a pressure of 15,000 Newtons. The construction sequence instructions are to first set up support points at the front end of the crack propagation direction and then set up support points on both sides.

[0034] Next, the real-time monitoring of rock creep rate and ambient temperature changes is obtained. Assume that the real-time monitoring indicates an increase in the rock creep rate and an accelerated drop in ambient temperature. Based on these changes, the dynamic attenuation coefficient of the support strength parameter in the initial support scheme template and the time offset of the construction sequence instructions are calculated. For example, due to the increase in rock creep rate, the support strength parameter needs to be increased by 10%, from 15,000 Newtons to 16,500 Newtons; due to the accelerated drop in ambient temperature, the construction sequence instructions need to be advanced; support work originally scheduled to begin in 10 days needs to start in 5 days.

[0035] The support strength parameters of the initial support plan template were adjusted based on the dynamic attenuation coefficient, and the priority order of the construction sequence instructions was revised based on the time offset. This generated an emergency support plan that included updated support point coordinates, optimized support strength parameters, and revised construction sequence instructions. For example, the final emergency support plan determined that support points would be set every 1.8 meters around the joint hazard zone (updated support point coordinates), with a support strength parameter of 16,500 Newtons (optimized support strength parameters). The construction sequence instructions were to first set support points at the front end of the crack propagation direction, and then complete the installation of all support points within three days (revised construction sequence instructions).

[0036] Finally, the emergency support scheme is topologically mapped to the spatial coordinates of the joint hazard zone to verify the consistency of the support point coordinates with the spatial coverage of the crack propagation direction. If uncovered areas are found, such as a small area on one side of the crack propagation direction that is not covered by the support points, the support case set is re-iterated until all joint hazard zones are covered by the support scheme, thus ensuring the safety of the underground facilities.

[0037] Based on the above steps, the embodiment of the present application realizes multi-physical field coupled monitoring of micro-damage and macro-deformation inside the ice and rock mass by deploying a multimodal sensor array on the surface of the ice and rock mass to collect sound wave reflection, stress distribution and temperature change signals in real time. The pre-trained multi-dimensional feature fusion model extracts the dynamic perception parameters of the crack spatial coordinates, expansion rate and stress concentration area, which not only depicts the geometric morphology and mechanical evolution characteristics of the crack, but also establishes a probability distribution map of the crack expansion path through the spatiotemporal evolution prediction model. The probability distribution map can quantify the uncertainty of the crack expansion direction and the spatial distribution of the critical fracture risk in the future preset time period. When the critical fracture risk level exceeds the preset threshold, it can synchronously trigger an early warning signal and generate an emergency support plan including support points, strength parameters and construction sequence, significantly improving the safety and reliability of ice and rock mass projects in extreme environments.

[0038] In one possible implementation, the pre-training process of the multi-dimensional feature fusion model includes: Step S210, obtaining a sample geological fluctuation data set and corresponding sample fracture annotation data, wherein the sample geological fluctuation data includes a sample acoustic wave reflection signal, a sample stress distribution signal, and a sample temperature change signal, and the sample fracture annotation data includes the sample fracture spatial coordinates, the sample fracture expansion rate, and the dynamic perception parameters of the sample stress concentration area.

[0039] For example, during the early exploration phase, sensors are installed at various locations within the ice and rock mass. Over a period of time, these sensors monitor the intensity and duration of sound waves of varying frequencies reflected from within the mass. Sample stress distribution signals are obtained by using stress sensors to capture stress values at each sampling point, reflecting the stress distribution within the mass. Sample temperature change signals are provided by temperature sensors, such as temperature values recorded at different depths and time periods. The corresponding sample fracture annotation data includes the spatial coordinates of the sample fractures, the sample fracture growth rate, and dynamic sensing parameters of the sample stress concentration areas. This annotation data may be obtained through more precise measurement equipment or manual analysis, such as using high-precision imaging equipment to determine the spatial coordinates of the fractures, long-term tracking and monitoring to determine the fracture growth rate, and detailed analysis of the stress concentration areas to obtain dynamic sensing parameters.

[0040] Step S220 , performing time-frequency transformation on the sample geological fluctuation data to generate a joint feature matrix including sample acoustic wave spectrum features, sample stress gradient features, and sample temperature change rate.

[0041] In a possible implementation, step S220 includes: Step S221 , performing short-time Fourier transform on the sample sound wave reflection signal to generate a time-frequency spectrum, and extracting the energy distribution peak of a preset frequency band in the time-frequency spectrum as the sample sound wave spectrum feature.

[0042] Step S222 , performing spatial gradient calculation on the sample stress distribution signal to obtain the stress gradient amplitude and direction angle of each sampling point, and taking the stress gradient amplitude difference between adjacent sampling points as the sample stress gradient feature.

[0043] Step S223 , performing time window sliding average processing on the sample temperature change signal, and calculating the standard deviation of the temperature change in each time window as the sample temperature change rate.

[0044] In step S224 , the sample acoustic wave spectrum features, the sample stress gradient features, and the sample temperature change rate are aligned according to timestamps, and are normalized and then spliced into the joint feature matrix in the form of a three-dimensional tensor.

[0045] During this process, the acoustic wave reflection signal is processed within a specific time window to generate a time-frequency spectrum. For example, if the time window is set to 1 second, the acoustic wave reflection signal within this 1-second period is Fourier transformed to obtain a time-frequency spectrum. The energy distribution peak within a preset frequency band is then extracted from the time-frequency spectrum as the sample acoustic wave spectrum feature. Assuming the preset frequency band is 1000-1500 Hz, the energy distribution of the time-frequency spectrum within this frequency band is analyzed to find the value corresponding to the peak with the highest energy. This value is the sample acoustic wave spectrum feature. For the sample stress distribution signal, a spatial gradient calculation is performed. For each sampling point, the stress gradient amplitude and azimuth are calculated. For example, on a two-dimensional sampling plane, with a sampling point as the center, the stress changes at the surrounding adjacent sampling points are calculated to obtain the stress gradient amplitude. The azimuth is determined based on the direction of stress change. For example, if the stress increases from left to right, the azimuth is horizontal to the right. The difference in stress gradient amplitude between adjacent sampling points is then used as the sample stress gradient feature. For example, if the stress gradient amplitudes of a sampling point and its adjacent right-hand point are 1000 Pa and 1200 Pa, respectively, then the difference of 200 Pa is part of the sample stress gradient feature. A sliding average of the sample temperature change signal is performed over a time window, set to a 1-hour window. The standard deviation of the temperature change within each time window is calculated as the sample temperature change rate. For example, within a 1-hour window, the recorded temperatures are -2°C, -1.5°C, -1°C, and so on. The difference between each temperature value and the average temperature within that time window is calculated, and the square sum of these differences is calculated. Finally, the square root of the average of the square sum is taken to obtain the sample temperature change rate. Finally, the sample acoustic spectrum features, sample stress gradient features, and sample temperature change rate are aligned by timestamp and concatenated into a joint feature matrix in the form of a three-dimensional tensor after normalization. Assume that the value of the sample acoustic wave spectrum feature is 0.5, the value of the sample stress gradient feature is 0.3, and the value of the sample temperature change rate is 0.2. Through normalization, their value ranges are adjusted to the appropriate range, and then spliced into a joint feature matrix in the form of a three-dimensional tensor in the set order.

[0046] In step S230, the initialized deep cross-attention network is used to calculate the sample weight distribution coefficient of each modal data in the joint feature matrix through a multi-head attention mechanism, with the sample sound wave spectrum feature as the query vector, the sample stress gradient feature as the key vector, and the sample temperature change rate as the value vector.

[0047] In a deep cross-attention network, the multi-head attention mechanism can comprehensively consider the relationships between modal data from different representation subspaces. For example, for a query vector (sample acoustic spectrum features), when interacting with a key vector (sample stress gradient features), and a value vector (sample temperature change rate), weights can be calculated separately across multiple heads (assuming three heads). For the first head, a weight coefficient, such as 0.3, might be calculated based on the correlation between certain frequency components in the sample acoustic spectrum features and the stress change trend in the sample stress gradient features. For the second head, another weight coefficient, such as 0.2, is calculated based on the relationship between the sample acoustic spectrum features and the sample temperature change rate. For the third head, a weight coefficient of 0.1 is calculated, taking into account the more complex relationship between the three features. These weight coefficients calculated across the different heads are then integrated to obtain the final sample weight distribution coefficients for each modal data.

[0048] Step S240, dynamically weighted fusion is performed on the joint feature matrix based on the sample weight distribution coefficient, the fused sample crack feature vector is output, and the parameters of the deep cross-attention network are optimized through the contrast loss function to minimize the spatial distribution error between the sample crack feature vector and the sample crack annotation data.

[0049] For example, the joint feature matrix contains multiple elements. Based on the previously calculated weight coefficient of 0.3 of the sample acoustic spectrum feature, the weight coefficient of 0.3 of the sample stress gradient feature, and the weight coefficient of 0.4 of the sample temperature change rate, each element in the joint feature matrix can be multiplied by the corresponding weight coefficient, and then the weighted elements are added together to obtain the fused sample crack feature vector. The sample crack feature vector contains crack-related features after comprehensively considering the relationship between each modal data.

[0050] In a possible implementation, step S240 includes: Step S241 : Divide the fused sample crack feature vectors and the sample crack annotation data into spatial grids to generate corresponding sample crack feature matrices and sample annotation matrices.

[0051] Step S242: According to the spatial coordinates of the cracks in the sample annotation matrix, positive sample pairs and negative sample pairs are allocated in the sample crack feature matrix. The positive sample pairs are grid cells in the same sample crack feature matrix whose spatial distance is less than a preset threshold, and the negative sample pairs are grid cells in different sample crack feature matrices whose spatial distance is greater than the preset threshold.

[0052] Step S243 , calculating the cosine similarity of the grid cells in the positive sample pair, and extracting the Euclidean distance difference of the grid cells in the negative sample pair.

[0053] Step S244 : constructing a contrast loss function based on the cosine similarity and the Euclidean distance difference, wherein a weighted sum of a similarity maximization constraint term of the positive sample pair and a difference minimization constraint term of the negative sample pair is performed.

[0054] Step S245: Use the gradient descent algorithm to backpropagate the partial derivatives of the contrast loss function to update the attention weight coefficients and fully connected layer parameters in the deep cross-attention network until the spatial distribution error between the sample crack feature matrix and the sample annotation matrix converges to a preset tolerance range.

[0055] In this embodiment, it is assumed that the study area of the ice rock body is divided into 10×10 spatial grids, and each spatial grid corresponds to a portion of the sample crack feature vector and the sample crack annotation data. According to the spatial coordinates of the cracks in the sample annotation matrix, positive sample pairs and negative sample pairs can be allocated in the sample crack feature matrix. For example, if the preset threshold is set to a distance of 3 grid units, in the same sample crack feature matrix, if the spatial distance between two grid units is less than 3 grid units, they are regarded as positive sample pairs, such as the two grid units with coordinates of (2,3) and (3,3); and in different sample crack feature matrices, if the spatial distance between two grid units is greater than 3 grid units, they are regarded as negative sample pairs, such as the (1,1) grid unit in one sample crack feature matrix and the (5,5) grid unit in another sample crack feature matrix. Calculate the cosine similarity of the grid cells in the positive sample pairs. Assume that the vectors of the grid cells of the two positive sample pairs are [0.5, 0.3] and [0.4, 0.4] respectively. First calculate their dot product as 0.5×0.4+0.3×0.4=0.32. Then calculate the modulus of the two vectors respectively. The modulus of the first vector is √(0.5²+0.3²)=0.5831, and the modulus of the second vector is √(0.4²+0.4²)=0.5657. Then divide the dot product by the product of the two moduli to get the cosine similarity as 0.32 / (0.5831×0.5657)=0.97. At the same time, the Euclidean distance difference between the grid cells in the negative sample pair is extracted. For the two grid cells in the negative sample pair, assuming their corresponding vectors are [0.2, 0.3] and [0.6, 0.7], the Euclidean distance is calculated as √((0.6-0.2)²+(0.7-0.3)²)=0.566. A contrastive loss function is constructed based on the cosine similarity and the Euclidean distance difference. Assuming a weight of 0.5 for the constraint that maximizes the similarity of the positive sample pair and a weight of 0.5 for the constraint that minimizes the difference of the negative sample pair, the contrastive loss function is 0.5×(1-0.97)+0.5×0.566=0.323. Gradient descent is used to backpropagate the partial derivatives of the contrastive loss function to update the attention weights and fully connected layer parameters in the deep crisscross attention network. During each update, the attention weights and fully connected layer parameters are adjusted with a fixed step size based on the direction and magnitude of the partial derivatives. For example, if the current value of a certain attention weight coefficient is 0.4 and the partial derivative calculation shows that it needs to be increased by 0.05, the updated value will be 0.45. This process is repeated until the spatial distribution error between the sample crack feature matrix and the sample annotation matrix converges to a preset tolerance range. Assuming the preset tolerance range is 0.01, when the calculated spatial distribution error is less than 0.01, the parameters of the deep cross-attention network are considered to have been optimized to an appropriate level.

[0056] In one possible implementation, the training method of the spatiotemporal evolution prediction model includes: Step S310: construct a hybrid neural network architecture including a long short-term memory network and a three-dimensional convolution kernel, wherein the long short-term memory network is used to capture the time-dependent characteristics of the crack propagation rate from the sample crack feature vector, and the three-dimensional convolution kernel is used to extract the topological characteristics of the crack spatial morphology from the sample crack feature vector.

[0057] For example, during long-term monitoring of cracks in ice and rock masses, the sample crack feature vectors contain information about the crack growth rate at different time points. Long-short-term memory networks can analyze this data. For example, if a series of crack growth rate data shows a gradually increasing value of 0.05, 0.06, and 0.07 meters per month over the past 10 time points, the LSTM network can capture this characteristic trend of gradual increase over time. A three-dimensional convolution kernel is used to extract topological features of the crack's spatial morphology from the sample crack feature vectors. For example, based on information such as the crack's spatial coordinates in the sample crack feature vectors, the 3D convolution kernel can construct the three-dimensional spatial morphology of the crack within the ice and rock mass, determining topological features such as the crack's length, width, direction, and its spatial relationship to the surrounding ice and rock structures.

[0058] Step S320 , inputting the sample crack feature vectors into the hybrid neural network architecture in time series, dynamically fusing the time-dependent features with the topological features through a gating mechanism, and generating dynamic fusion features containing time-space correlation.

[0059] Taking the previously mentioned crack growth rate and spatial morphology as an example, the gating mechanism can fuse the time-varying crack growth rate features captured by the long-short-term memory network with the spatial morphology features extracted by the three-dimensional convolution kernel, based on predefined rules and algorithms. For example, suppose that at a certain moment, the crack growth rate is 0.08 meters per month and is accelerating. At the same time, the crack exhibits a spatial pattern of extending in a certain direction and gradually widening. The gating mechanism can comprehensively consider these factors, weighing the importance of each feature and fusing them into a dynamic fusion feature that incorporates temporal and spatial correlations. This dynamic fusion feature can comprehensively describe the crack's state in time and space.

[0060] In step S330, the dynamic fusion features are input into the discriminator network, and the spatial matching degree between the crack evolution path predicted by the dynamic fusion features and the actual fracture event is calculated by the discriminator network. The adversarial training loss is generated according to the difference between the spatial matching degree and the preset threshold, and the convolution kernel parameters of the hybrid neural network architecture are reversely optimized.

[0061] In a possible implementation, step S330 includes: Step S331: input the dynamic fusion features into the fully connected layer of the discriminator network to generate a spatial probability distribution map of the predicted crack evolution path output by the discriminator network.

[0062] Step S332 : obtaining a set of spatial coordinate points of a real fracture event corresponding to the sample fracture feature vector, and converting the set of spatial coordinate points into a binary fracture mask map with the same resolution as the spatial probability distribution map.

[0063] Step S333 , calculating the cross entropy loss between the probability value of each spatial unit in the spatial probability distribution map and the label value of the corresponding spatial unit in the binary fracture mask map to generate an initial adversarial loss component.

[0064] Step S334 , extracting a set of spatial units whose probability values exceed a preset probability threshold in the spatial probability distribution map, and calculating a spatial intersection-over-union (IoU) of the spatial unit set and the rupture area in the binary rupture mask map to generate an IoU penalty term.

[0065] Step S335: linearly combine the initial adversarial loss component and the intersection-over-union penalty term according to a preset weight coefficient to generate a final adversarial training loss.

[0066] Step S336, using the back-propagation algorithm to calculate the partial derivative of the adversarial training loss with respect to the three-dimensional convolution kernel parameters of the hybrid neural network architecture, and updating the weight matrix of the three-dimensional convolution kernel according to the gradient direction, so as to maximize the spatial matching degree between the spatial probability distribution map and the binary rupture mask map.

[0067] For example, the spatial probability distribution map can show the probability that cracks at different spatial locations within the ice rock body will extend to those locations within a certain time period in the future. At the same time, a set of spatial coordinate points corresponding to the sample crack feature vectors is obtained. These spatial coordinate points are obtained by detailed measurement and recording of past ice rock body fracture events. These spatial coordinate points are then converted into a binary fracture mask map with the same resolution as the spatial probability distribution map, marking fractured coordinate points as 1 and unfractured coordinate points as 0. The cross entropy loss is calculated between the probability value of each spatial unit in the spatial probability distribution map and the label value of the corresponding spatial unit in the binary fracture mask map to generate the initial adversarial loss component. For example, for a spatial unit in the spatial probability distribution map, the predicted probability of a crack extending to that location is 0.6, while the corresponding label value in the binary crack mask map is 1. According to the cross-entropy loss calculation method, we first calculate -(1×log(0.6)+(1-1)×log(1-0.6)), where 1×log(0.6) represents the loss calculation portion when the true label is 1 and (1-1)×log(1-0.6) represents the loss calculation portion when the true label is 0 (this portion is 0 because the true label is 1). The result is -log(0.6), which is the contribution of this spatial unit to the initial adversarial loss component. This calculation is repeated for all spatial units and summed to obtain the initial adversarial loss component. Next, we extract the set of spatial units in the spatial probability distribution map whose probability values exceed a preset probability threshold (assuming it is 0.5). We then calculate the spatial intersection-over-union (IoU) of this set of spatial units with the crack region in the binary crack mask map to generate the IoU penalty term. Assume there are 10 spatial cells with a probability value exceeding 0.5, and the fracture region in the binary fracture mask image encompasses 8 spatial cells. The intersection of these 10 spatial cells and the 8 spatial cells is 6 spatial cells. Therefore, the intersection-over-union (IoU) is 6 / (10+8-6)=0.5. The IoU penalty term can be calculated based on this IoU according to a specific algorithm. The initial adversarial loss component and the IoU penalty term are linearly combined with preset weight coefficients to generate the final adversarial training loss. Assuming the initial adversarial loss component has a value of 0.3, the IoU penalty term has a value of 0.2, and the preset weight coefficients are 0.6 and 0.4, respectively, the final adversarial training loss is 0.6×0.3+0.4×0.2=0.26. The partial derivative of the adversarial training loss with respect to the 3D convolution kernel parameters of the hybrid neural network architecture is taken via the backpropagation algorithm. The weight matrix of the 3D convolution kernel is updated based on the direction of the partial derivative to maximize the spatial match between the spatial probability distribution map and the binary fracture mask image.For example, if the partial derivative corresponding to a certain three-dimensional convolution kernel parameter is 0.1, it means that increasing this parameter will reduce the adversarial training loss. Then, according to a certain update step size (assuming it is 0.01), the parameter is increased by 0.01. After multiple such updates, the spatial matching degree between the spatial probability distribution map and the binary rupture mask map is gradually improved.

[0068] Step S340: Input the dynamic fusion feature into the physical constraint loss function, calculate the angle penalty term between the maximum principal stress direction and the crack propagation direction in the predicted crack evolution path, and generate a joint optimization loss in combination with the adversarial training loss.

[0069] In a possible implementation, step S340 includes: Step S341 : extracting the maximum principal stress direction angle and the crack surface normal direction angle of the current stress concentration area from the sample crack feature vector.

[0070] Step S342 , calculating a first angle between the maximum principal stress direction and the normal direction of the crack surface, and a second angle between the crack propagation direction and the maximum shear stress direction.

[0071] Step S343: when the first angle is smaller than a preset angle threshold and the second angle is larger than a critical rupture angle, reducing the rupture risk prediction value of the region.

[0072] Step S344: when the first angle is in the brittle fracture range and the second angle satisfies the Coulomb friction condition, increasing the fracture risk prediction value of the region.

[0073] Step S345 , calculating the directional consistency loss based on the spatial overlap between the predicted value and the actual rupture event, and performing a weighted summation of the directional consistency loss and the adversarial loss of the discriminator network to obtain a final joint optimization loss.

[0074] In this embodiment, the stress conditions in the stress concentration area are calculated using data obtained from previously deployed stress sensors. Assume that in a certain stress concentration area, after detailed stress analysis, it is determined that the maximum principal stress direction angle is 30 degrees and the crack surface normal direction angle is 40 degrees. Calculate the first angle between the maximum principal stress direction and the crack surface normal direction, as well as the second angle between the crack propagation direction and the maximum shear stress direction. The first angle is the absolute value of the difference between the two direction angles, i.e., |30-40| = 10 degrees. For the second angle between the crack propagation direction and the maximum shear stress direction, assume that it is 60 degrees after analysis and calculation. When the first angle is less than the preset angle threshold (assuming 15 degrees) and the second angle is greater than the critical fracture angle (assuming 50 degrees), the fracture risk prediction value of the area is reduced. For example, if the original fracture risk prediction value of the area was 0.8, it would be reduced to 0.7 according to this rule. When the first angle is within the brittle fracture range (assuming it's 10-20 degrees) and the second angle satisfies the Coulomb friction condition (assuming it does), increase the fracture risk prediction value for that region, for example, from 0.7 to 0.8. Calculate the directional consistency loss based on the spatial overlap between the predicted value and the actual fracture event. Assume that the spatial overlap between the region with a predicted value of 0.8 and the actual fracture event is 0.6. The directional consistency loss can be calculated based on this overlap using a specific algorithm, assuming it's 0.1. The directional consistency loss is weighted and summed with the adversarial loss of the discriminator network to obtain the final joint optimization loss. Assuming the adversarial loss is 0.26, the weight of the directional consistency loss is 0.3, and the weight of the adversarial loss is 0.7. The final joint optimization loss is 0.3 × 0.1 + 0.7 × 0.26 = 0.212.

[0075] Step S350, by iteratively updating the convolution kernel parameters of the hybrid neural network architecture and the weight parameters of the discriminator network, the joint optimization loss is converged to a preset error range, and a trained spatiotemporal evolution prediction model is generated.

[0076] During the iterative process, the above calculation and update steps are repeated. Each iteration updates the convolution kernel parameters of the hybrid neural network architecture and the weight parameters of the discriminator network based on the new joint optimization loss. For example, in one iteration, if the update calculated for a convolution kernel parameter based on the joint optimization loss is -0.05, then the convolution kernel parameter is subtracted by 0.05. As the number of iterations increases, the joint optimization loss gradually decreases. When the value of the joint optimization loss is less than the preset error range (assuming 0.01), the spatiotemporal evolution prediction model is considered fully trained and can be used to accurately predict the evolution of cracks in ice and rock masses.

[0077] In a possible implementation, step S140 includes: Step S141 , traversing all spatial units in the probability distribution map, screening out a set of spatial units whose critical fracture risk level exceeds a preset threshold, and marking them as initial dangerous areas.

[0078] In this embodiment, the probability distribution map is previously derived through a spatiotemporal evolution prediction model, which shows the risk situation related to the evolution of cracks corresponding to different spatial units in the ice rock body. Assuming that the preset threshold is set to a certain risk level value, when traversing each spatial unit of the probability distribution map, those spatial units with risk levels higher than this preset threshold are selected. For example, each spatial unit in the probability distribution map has a corresponding fracture risk level value, ranging from 0 to 1, representing the risk level from low to high. If the preset threshold is 0.7, then when the risk level of a certain spatial unit traversed is 0.8, the spatial unit will be selected, and all spatial units selected in this way constitute the initial danger zone.

[0079] Step S142 : performing spatial connectivity analysis on the initial danger zone, merging adjacent spatial units whose spatial distance is less than a preset safety radius, and generating a continuously distributed joint danger zone.

[0080] In this embodiment, in the initial danger zone, the distribution of each spatial unit may be discrete. The judgment of adjacent spatial units is based on their spatial position relationship in the ice rock body. The preset safety radius is a distance value determined based on the physical properties of the ice rock body and engineering experience. For example, the preset safety radius is set to 2 meters. If there are two adjacent initial danger zone spatial units with a spatial distance of 1.5 meters between them, then these two spatial units meet the merging conditions. After performing such analysis and merging operations on the entire initial danger zone, a continuously distributed joint danger zone is obtained. The joint danger zone represents an area in the ice rock body that has a higher risk of fracture and is continuous in space, which is more convenient for subsequent unified support plan planning.

[0081] Step S143 : extracting geological structural features of the joint hazardous area, wherein the geological structural features include the regional maximum crack growth rate, the average stress concentration factor, and the temperature gradient distribution pattern.

[0082] In this embodiment, the regional maximum crack growth rate is obtained from long-term crack monitoring data. For example, crack growth rates are measured at different locations within the joint hazard zone. For example, at one location, cracks grow at 0.1 meters per month, while at another location, they grow at 0.12 meters per month. After comparison, 0.12 meters is determined to be the regional maximum crack growth rate. The average stress concentration factor is calculated based on stress data collected by stress sensors within the joint hazard zone. Assuming multiple stress sampling points are set within the joint hazard zone, each with a corresponding stress value, the stress values of all sampling points are added together and divided by the number of sampling points to obtain the average stress value. The average stress concentration factor is then calculated using existing calculation methods for stress concentration areas. The temperature gradient distribution pattern is determined based on data acquired by the temperature sensor. Temperatures are measured at different depths and locations within the joint hazard zone. For example, if the surface temperature is -5°C and the temperature at a depth of 5 meters is -3°C, the temperature gradient distribution pattern is determined by analyzing the temperature differences at different locations. For example, whether the temperature gradually increases or decreases with depth, etc.

[0083] Step S144 , performing similarity matching between the geological structure characteristics and the support case characteristics in the pre-stored support solution library, and screening out a set of support cases whose matching degree exceeds a similarity threshold.

[0084] In this embodiment, the pre-stored support scheme library was previously constructed through a large number of historical support cases. The support case characteristics include various relevant parameters under different geological structures. When performing similarity matching, multiple geological structure characteristic parameters should be comprehensively considered. For example, for the feature of the maximum regional crack expansion rate, if the maximum crack expansion rate in the joint hazardous area is 0.12 meters per month, and the crack expansion rate corresponding to a certain support case in the support scheme library is between 0.1 meters and 0.15 meters per month, then there is a certain degree of similarity in this feature. Similar comparisons are also made for the average stress concentration coefficient and the temperature gradient distribution pattern. Assuming that the similarity threshold is set to 0.8, after comprehensively comparing all geological structure features, it is found that the similarity between a certain support case and the geological structure feature of the joint hazardous area reaches 0.85, and the support case can be screened out, and all the support cases screened out in this way constitute a support case set.

[0085] Step S145: Generate an initial support plan template based on the support strength parameters and construction sequence instructions of the support case set.

[0086] In this embodiment, each case in the support case set has corresponding support strength parameters and construction sequence instructions. The support strength parameters refer to relevant parameters such as the stress that the support structure can withstand, and the construction sequence instructions specify information such as the order of support construction. For example, the support strength parameter recommended by a certain support case is to be able to withstand a pressure of 15,000 Newtons, and the construction sequence instructions are to first set up a support structure at the center of the joint danger area and then expand it to the surrounding areas. By integrating this information in the support case set, an initial support scheme template is generated to determine the basic framework of the support, such as preliminarily determining the type of support structure, the approximate layout method, and the basic order of construction.

[0087] Step S146, obtaining the real-time monitored rock creep rate and ambient temperature change, and calculating the dynamic attenuation coefficient of the support strength parameter in the initial support scheme template and the time offset of the construction timing instruction.

[0088] In this embodiment, real-time monitoring of rock creep rate and ambient temperature changes is performed using specialized monitoring equipment deployed within the ice rock mass. For example, the rock creep rate increases from 0.01 meters per month to 0.02 meters per month, while the ambient temperature drops from -5°C to -8°C. To calculate the dynamic attenuation coefficient of the support strength parameter, an empirical model based on the relationship between rock creep rate, ambient temperature, and support strength is assumed. According to this model, an increase in rock creep rate leads to an increase in support strength requirement, while a decrease in ambient temperature also affects support strength. The effects of changes in rock creep rate and ambient temperature on support strength are first calculated separately. For example, if an increase in rock creep rate leads to a 10% increase in support strength, and a decrease in ambient temperature leads to a 5% increase in support strength, the total dynamic attenuation coefficient is 1 + 10% + 5% = 1.15. The calculation of the time offset for construction sequence instructions is also based on empirical models or pre-set rules. For example, a rapid decrease in ambient temperature may accelerate the risk of cracking in the ice rock mass, requiring the construction sequence to be advanced according to the rules. If the original plan was to start construction in 10 days, but due to the drop in ambient temperature, it is calculated that it needs to be started 3 days earlier, then the time offset is -3 days.

[0089] Step S147: Adjust the support strength parameters of the initial support scheme template according to the dynamic attenuation coefficient, and correct the priority order of the construction timing instructions based on the time offset to generate an emergency support scheme including updated support point coordinates, optimized support strength parameters and corrected construction timing instructions.

[0090] In this embodiment, based on the previously calculated dynamic attenuation coefficient of 1.15, if the support strength parameter in the initial support plan template is 15,000 Newtons, then the updated support strength parameter is 15,000 × 1.15 = 17,250 Newtons. The construction sequence instructions are corrected based on a time offset of -3 days. If the original construction sequence instructions first set up the support structure at the center, then expanded outward in three steps, with each step separated by two days, then the revised construction sequence instructions might be to first set up the support structure at the center, then advance the second step by one day, and the third step by two days, to accelerate the support construction progress and address the increased risk of ice and rock mass fracture. This generates an emergency support plan that includes the updated support point coordinates (determined by the support structure layout, such as the coordinates of support points arranged at a certain distance within a joint hazard zone), the optimized support strength parameter of 17,250 Newtons, and the revised construction sequence instructions.

[0091] In step S148, the emergency support scheme is topologically mapped with the spatial coordinates of the joint danger zone to verify the spatial coverage consistency between the support point coordinates and the crack propagation direction. If there is an uncovered area, the support case set is re-iteratively matched until all joint danger zones are covered by the support scheme.

[0092] In this embodiment, the topological mapping is to accurately check the spatial coverage of the joint danger zone by the emergency support scheme. The support point coordinates in the emergency support scheme are analyzed in correspondence with the spatial coordinates of the joint danger zone. For example, the crack expansion direction is determined to be inclined in a certain direction based on the previous monitoring and analysis of the cracks. Check whether the support point coordinates are distributed around the crack expansion direction to ensure that the cracks can be effectively prevented from further expansion. If it is found that there are uncovered areas, for example, a small area on one side of the crack expansion direction is not covered by the support points, then it is necessary to iterate the matching of the support case set again. Follow the previous steps again, starting from the similarity matching of the support case features in the pre-stored support scheme library, adjust the support strength parameters and construction timing instructions, etc., until all joint danger zones are covered by the support scheme, thereby protecting underground facilities from the threat of ice and rock rupture.

[0093] In a possible implementation, the method for constructing the support scheme library includes: Step S410: Acquire a historical support case data set, wherein the historical support case data set includes geological structure characteristics of the support area, a support point coordinate sequence, a support strength parameter group, and a construction timing instruction set.

[0094] For example, in different underground projects in ice and rock masses in the past, detailed records were kept for each project. The geological structural characteristics of the support area cover various characteristics of the ice and rock mass at that time, such as rock type, ice-rock ratio, structural integrity and other information. The support point coordinate sequence records the specific location coordinates of each support point in the support area. These coordinates accurately determine the layout of the support structure in the ice and rock mass. The support strength parameter group contains relevant parameters such as stress and pressure that each support point can withstand to ensure the effectiveness of the support structure. The construction timing instruction set specifies in detail the sequence and time arrangement of support construction, such as where to set up the support structure first, and then in what order to gradually complete the entire support project.

[0095] Step S420 , extracting the historical crack growth rate peak value, the historical stress concentration factor distribution map, and the historical temperature gradient direction vector from the geological structure characteristics of the support area.

[0096] In this embodiment, the extraction of the historical crack expansion rate peak is obtained by analyzing the long-term monitoring data of ice-rock cracks in past projects. For example, in a certain historical project, the expansion of ice-rock cracks is continuously monitored, and the expansion rate is recorded in different time periods, such as 0.05 meters, 0.08 meters, 0.1 meters, etc. per month. After comparison, the maximum value of 0.1 meter is found. This 0.1 meter is the historical crack expansion rate peak of the project. The historical stress concentration coefficient distribution map is drawn based on the data collected by the stress sensors arranged in the support area at that time. The sensors measure stress values at different locations, and the stress concentration coefficient of each location is obtained through existing calculation methods. These coefficients are then plotted into a distribution map according to the positional relationship. The historical temperature gradient direction vector is determined based on the data of the temperature sensor. The temperature is measured at different depths and locations in the support area, and the temperature change with position is analyzed to determine the direction of the temperature gradient. This direction information is combined into the historical temperature gradient direction vector.

[0097] Step S430: According to the angle between the maximum gradient direction of the historical stress concentration factor distribution diagram and the historical temperature gradient direction vector, the support type labels in the historical support case data set are divided. The support type labels include brittle fracture support type, shear slip support type and creep coordination support type.

[0098] For example, when the angle between the maximum gradient direction of the historical stress concentration factor distribution diagram and the historical temperature gradient direction vector is less than 30 degrees, the support case is marked as a brittle fracture support type. This is because in this case, the distribution relationship between stress and temperature may make the ice rock mass more susceptible to brittle fracture, so support is required for this type of fracture. If the angle is between 30 and 60 degrees, it is marked as a shear slip support type, because in this case, the interaction between stress and temperature may cause the ice rock mass to produce shear slip. When the angle is greater than 60 degrees, it is marked as a creep coordinated support type, which means that under such stress and temperature conditions, the creep phenomenon of the ice rock mass is more significant, and a coordinated creep support method is required.

[0099] In step S440, the support type label and the corresponding support point coordinate sequence are input into a graph convolutional network to generate a topological connection relationship diagram for each support case. The topological connection relationship diagram includes the stress transfer path and temperature field interference area between the support points.

[0100] In this embodiment, for the stress transfer path, the support points are used as nodes, and the stress transfer method between each support point is determined based on the mechanical properties of the ice rock body and the mechanical relationship of the support structure. For example, if the distance between two adjacent support points is close and the support structure has a certain rigidity, then the stress transfer between the two support points will be relatively direct. The temperature field interference area takes into account the influence of the support structure on the temperature field of the ice rock body. Since the existence of the support structure may change the heat conduction inside the ice rock body, the temperature fields around different support points will interfere with each other. The graph convolutional network accurately depicts these stress transfer paths and temperature field interference areas by analyzing information such as the support point coordinates and support type.

[0101] Step S450: Calculate the priority score of each support point according to the node degree centrality index of the stress transfer path and the area weight of the temperature field interference area, and reorder the construction timing instruction set according to the priority score.

[0102] In this embodiment, the node degree centrality index of the stress transfer path reflects the importance of each support point in the stress transfer network. For example, if a support point has a direct stress transfer relationship with multiple other support points, then the node degree centrality index of this support point is relatively high. The area weight of the temperature field interference area takes into account the range of influence of the support point on the temperature field. Assuming that the area of the temperature field interference area around a certain support point is large, it means that it has a greater impact on the temperature field, and the corresponding area weight will also be large. When calculating the priority score of each support point, these two factors are taken into consideration. For example, if the node degree centrality index of a support point is 0.6 and the area weight of the temperature field interference area is 0.4, the priority score of 0.5 is obtained through existing calculation methods (such as weighted summation). The construction timing instruction set is reordered according to these priority scores. If the original construction timing instructions are carried out in a fixed order, after reordering, the construction steps corresponding to the support points with high priority scores will be advanced to ensure that the support structure can better play its role during the construction process.

[0103] Step S460: align the timestamps of the historical crack growth rate peak value with the support strength parameter group, fit the dynamic attenuation curve of the support strength parameter changing with the crack growth rate, and extract the critical support strength threshold corresponding to the curve slope mutation point.

[0104] In this embodiment, each historical support case has a corresponding historical peak crack growth rate and support strength parameter set, and these data are time-stamped. Timestamp alignment is achieved by matching data with the same time stamp. Then, a dynamic attenuation curve is fitted using an appropriate fitting method (such as the least squares method) with the historical peak crack growth rate as the independent variable and the support strength parameter as the dependent variable. For example, when the historical peak crack growth rate is 0.05 meters per month, the corresponding support strength parameter is 10,000 Newtons; when the historical peak crack growth rate is 0.1 meters per month, the support strength parameter is 12,000 Newtons. A curve is fitted from these data points, reflecting the changing trend of the support strength parameter as the crack growth rate increases. From the fitted dynamic attenuation curve, the critical support strength threshold corresponding to the slope mutation point of the curve is extracted. A slope mutation point in the curve represents a turning point in the trend of the support strength parameter. For example, at a certain point, the slope of the curve suddenly changes from a smaller value to a larger value. The support strength parameter value corresponding to this point is the critical support strength threshold. If it is below this threshold, the support structure may not be able to effectively resist the risk of fracture of the ice and rock mass.

[0105] In step S470, the support type label, topological connection relationship diagram, reordered construction timing instruction set and critical support strength threshold are associated and stored to generate a support scheme library containing multidimensional index keys, wherein the multidimensional index keys include crack propagation rate range, stress-temperature gradient angle range and priority score level.

[0106] For example, for a support case, the support type label is brittle fracture support. The topological connection diagram shows the specific support point relationships. The reordered construction timing instruction set determines the construction sequence, and the critical support strength threshold is 12,000 Newtons. This information is associated and stored. When constructing the index key, categorization is performed based on the crack growth rate range. For example, cases with a historical crack growth rate peak between 0.05 meters and 0.1 meters per month are grouped into one range. The stress-temperature gradient angle range is also categorized, such as 0 to 30 degrees. It is also divided according to priority rating levels, such as 0.4 to 0.6. This support solution library, constructed in this way, can easily and quickly retrieve appropriate support solutions based on the actual conditions of the ice-rock mass, such as the current crack growth rate, stress-temperature gradient angle, and support point priority, effectively ensuring the safety of underground facilities.

[0107] Figure 1 The schematic diagram of the structure of the artificial intelligence-based ice and rock crack monitoring and early warning system 10 provided in an embodiment of the present invention includes a processor 102, a memory 104, and a bus 106. The memory 104 is used to store execution instructions and includes internal memory and external memory. The internal memory can also be understood as internal memory, which is used to temporarily store calculation data in the processor 102 and data exchanged with external memory such as a hard disk. The processor 102 exchanges data with the external memory through the internal memory. When the artificial intelligence-based ice and rock crack monitoring and early warning system 10 is running, the processor 102 communicates with the memory 104 via the bus 106, so that the processor 102 executes the artificial intelligence-based ice and rock crack monitoring and early warning method of the embodiment of the present invention.

[0108] For ease of explanation, only one processor is described in the artificial intelligence-based ice and rock crack monitoring and early warning system 10. However, it should be noted that the artificial intelligence-based ice and rock crack monitoring and early warning system 10 in this application can also include multiple processors, so the steps performed by one processor described in this application can also be performed jointly or individually by multiple processors. For example, if the processor of the artificial intelligence-based ice and rock crack monitoring and early warning system 10 performs steps A and B, it should be understood that steps A and B can also be performed jointly by two different processors or individually in one processor. For example, the first processor performs step A and the second processor performs step B, or the first processor and the second processor perform steps A and B together.

[0109] In addition, an embodiment of the present invention further provides a readable storage medium, in which computer-executable instructions are preset. When a processor executes the computer-executable instructions, the above-mentioned artificial intelligence-based ice and rock crack monitoring and early warning method is implemented.

[0110] It should be noted that in order to simplify the description of the present invention and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, multiple features are sometimes combined into one embodiment, figure or description thereof.

Claims

1. An artificial intelligence-based ice and rock crack monitoring and early warning method, characterized in that: The method comprises: A multimodal sensor array deployed on the surface of the ice and rock mass collects real-time geological fluctuation data, including acoustic wave reflection signals, stress distribution signals, and temperature change signals; Inputting the geological fluctuation data into a pre-trained multi-dimensional feature fusion model to extract the crack feature vector inside the ice rock body, wherein the crack feature vector includes the crack spatial coordinates, crack propagation rate and dynamic perception parameters of the stress concentration area; Calling a spatiotemporal evolution prediction model to perform a time series analysis on the crack characteristic vector to generate a probability distribution map of the crack evolution path, wherein the probability distribution map includes the crack propagation direction and critical fracture risk level within a preset future time period; According to the area where the critical fracture risk level in the probability distribution map exceeds a preset threshold, an early warning signal is triggered and a corresponding emergency support plan is generated. The emergency support plan includes support point coordinates, support strength parameters and construction timing instructions.

2. The ice-rock crack monitoring and early warning method according to claim 1 is characterized in that: The pre-training process of the multi-dimensional feature fusion model includes: Acquire a sample geological fluctuation data set and corresponding sample fracture annotation data, wherein the sample geological fluctuation data includes a sample acoustic wave reflection signal, a sample stress distribution signal, and a sample temperature change signal; and the sample fracture annotation data includes the spatial coordinates of the sample fracture, the sample fracture expansion rate, and dynamic perception parameters of the sample stress concentration area; Performing time-frequency transformation on the sample geological fluctuation data to generate a joint feature matrix including sample acoustic wave spectrum characteristics, sample stress gradient characteristics, and sample temperature change rate; Using an initialized deep cross-attention network, the sample sound wave spectrum feature is used as a query vector, the sample stress gradient feature is used as a key vector, and the sample temperature change rate is used as a value vector. The sample weight distribution coefficient of each modal data in the joint feature matrix is calculated through a multi-head attention mechanism; The joint feature matrix is dynamically weighted fused based on the sample weight distribution coefficient, and the fused sample crack feature vector is output. The parameters of the deep cross attention network are optimized through the contrast loss function to minimize the spatial distribution error between the sample crack feature vector and the sample crack annotation data.

3. The ice-rock crack monitoring and early warning method according to claim 2 is characterized in that: The time-frequency transformation of the sample geological fluctuation data is performed to generate a joint feature matrix including the sample acoustic wave spectrum characteristics, the sample stress gradient characteristics and the sample temperature change rate, including: Performing a short-time Fourier transform on the sample sound wave reflection signal to generate a time-frequency spectrum, and extracting an energy distribution peak value of a preset frequency band in the time-frequency spectrum as a spectrum feature of the sample sound wave; Performing spatial gradient calculation on the sample stress distribution signal to obtain the stress gradient amplitude and direction angle of each sampling point, and using the stress gradient amplitude difference between adjacent sampling points as the sample stress gradient feature; Performing time window sliding average processing on the sample temperature change signal, and calculating the standard deviation of the temperature change in each time window as the sample temperature change rate; The sample acoustic wave spectrum features, sample stress gradient features and sample temperature change rate are aligned according to timestamps, and are normalized and then spliced into the joint feature matrix in the form of a three-dimensional tensor.

4. The ice-rock crack monitoring and early warning method according to claim 2 is characterized in that: Optimizing the parameters of the deep cross attention network by using a contrastive loss function to minimize the spatial distribution error between the sample crack feature vector and the sample crack annotation data includes: Dividing the fused sample crack feature vector and the sample crack annotation data into a spatial grid to generate a corresponding sample crack feature matrix and a sample annotation matrix; According to the spatial coordinates of the cracks in the sample annotation matrix, positive sample pairs and negative sample pairs are allocated in the sample crack feature matrix, wherein the positive sample pairs are grid cells in the same sample crack feature matrix whose spatial distance is less than a preset threshold, and the negative sample pairs are grid cells in different sample crack feature matrices whose spatial distance is greater than the preset threshold; Calculating the cosine similarity of the grid cells in the positive sample pair, and extracting the Euclidean distance difference of the grid cells in the negative sample pair; Constructing a contrast loss function based on the cosine similarity and the Euclidean distance difference, wherein a weighted sum of a constraint term maximizing the similarity of positive sample pairs and a constraint term minimizing the difference of negative sample pairs is performed; The gradient descent algorithm is used to backpropagate the partial derivatives of the contrast loss function to update the attention weight coefficients and fully connected layer parameters in the deep cross-attention network until the spatial distribution error between the sample crack feature matrix and the sample annotation matrix converges to a preset tolerance range.

5. The ice-rock crack monitoring and early warning method according to claim 1 is characterized in that: The training method of the spatiotemporal evolution prediction model includes: Constructing a hybrid neural network architecture comprising a long short-term memory network and a three-dimensional convolution kernel, wherein the long short-term memory network is used to capture the time-dependent characteristics of the crack propagation rate from the sample crack feature vector, and the three-dimensional convolution kernel is used to extract the topological characteristics of the crack spatial morphology from the sample crack feature vector; Inputting the sample crack feature vectors into the hybrid neural network architecture in time series, dynamically fusing the time-dependent features with the topological features through a gating mechanism to generate dynamic fusion features containing time-space correlation; Inputting the dynamic fusion features into a discriminator network, calculating the spatial matching degree between the crack evolution path predicted by the dynamic fusion features and the actual fracture event through the discriminator network, generating an adversarial training loss based on the difference between the spatial matching degree and a preset threshold, and reversely optimizing the convolution kernel parameters of the hybrid neural network architecture; Inputting the dynamic fusion feature into a physical constraint loss function, calculating a penalty term for the angle between the maximum principal stress direction and the crack propagation direction in the predicted crack evolution path, and combining the adversarial training loss to generate a joint optimization loss; By iteratively updating the convolution kernel parameters of the hybrid neural network architecture and the weight parameters of the discriminator network, the joint optimization loss is converged to a preset error range, and a trained spatiotemporal evolution prediction model is generated.

6. The ice-rock crack monitoring and early warning method according to claim 5 is characterized in that: The step of inputting the dynamic fusion feature into a physical constraint loss function, calculating a penalty term for the angle between the maximum principal stress direction and the crack propagation direction in the predicted crack evolution path, and combining the adversarial training loss to generate a joint optimization loss includes: Extracting the maximum principal stress direction angle and the crack surface normal direction angle of the current stress concentration area from the sample crack characteristic vector; Calculating a first angle between the maximum principal stress direction and the normal direction of the crack surface, and a second angle between the crack propagation direction and the maximum shear stress direction; When the first angle is less than a preset angle threshold and the second angle is greater than a critical rupture angle, reducing a rupture risk prediction value of the region; When the first angle is in the brittle fracture range and the second angle satisfies the Coulomb friction condition, increasing the fracture risk prediction value of the region; The directional consistency loss is calculated according to the spatial overlap between the predicted value and the actual rupture event, and the directional consistency loss is weightedly summed with the adversarial loss of the discriminator network to obtain the final joint optimization loss.

7. The ice-rock crack monitoring and early warning method according to claim 5 is characterized in that: The dynamic fusion features are input into the discriminator network, the spatial matching degree between the crack evolution path predicted by the dynamic fusion features and the actual fracture event is calculated by the discriminator network, and the adversarial training loss is generated according to the difference between the spatial matching degree and a preset threshold, and the convolution kernel parameters of the hybrid neural network architecture are reversely optimized, including: Inputting the dynamic fusion features into the fully connected layer of the discriminator network to generate a spatial probability distribution map of the predicted crack evolution path output by the discriminator network; Acquire a set of spatial coordinate points of a real fracture event corresponding to the sample fracture feature vector, and convert the set of spatial coordinate points into a binary fracture mask map with the same resolution as the spatial probability distribution map; Calculating the cross entropy loss between the probability value of each spatial unit in the spatial probability distribution map and the label value of the corresponding spatial unit in the binary fracture mask map to generate an initial adversarial loss component; Extracting a set of spatial units whose probability values exceed a preset probability threshold in the spatial probability distribution map, and calculating a spatial intersection-and-union ratio (IoU) between the set of spatial units and the rupture area in the binary rupture mask map to generate an IoU penalty term; Linearly combine the initial adversarial loss component and the intersection-over-union penalty term according to a preset weight coefficient to generate a final adversarial training loss; The adversarial training loss is partially derivatived with respect to the three-dimensional convolution kernel parameters of the hybrid neural network architecture through a back-propagation algorithm, and the weight matrix of the three-dimensional convolution kernel is updated according to the gradient direction to maximize the spatial matching degree between the spatial probability distribution map and the binary rupture mask map.

8. The ice-rock crack monitoring and early warning method according to claim 1 is characterized in that: The triggering of an early warning signal and generating a corresponding emergency support plan according to an area where the critical fracture risk level in the probability distribution diagram exceeds a preset threshold value includes: Traversing all spatial units in the probability distribution map, screening out a set of spatial units whose critical fracture risk level exceeds a preset threshold, and marking them as initial danger zones; Performing a spatial connectivity analysis on the initial danger zone, merging adjacent spatial units whose spatial distance is less than a preset safety radius, and generating a continuously distributed joint danger zone; Extracting geological structural characteristics of the joint hazardous area, wherein the geological structural characteristics include regional maximum crack growth rate, average stress concentration factor, and temperature gradient distribution pattern; Performing similarity matching between the geological structure characteristics and the support case characteristics in the pre-stored support solution library, and screening out a set of support cases whose matching degree exceeds a similarity threshold; generating an initial support scheme template based on the support strength parameters and construction timing instructions of the support case set; Obtaining the real-time monitored rock creep rate and ambient temperature change, and calculating the dynamic attenuation coefficient of the support strength parameter in the initial support scheme template and the time offset of the construction sequence instruction; Adjusting the support strength parameters of the initial support scheme template according to the dynamic attenuation coefficient, and correcting the priority order of the construction sequence instructions based on the time offset, to generate an emergency support scheme including updated support point coordinates, optimized support strength parameters, and corrected construction sequence instructions; The emergency support scheme is topologically mapped with the spatial coordinates of the joint danger zone to verify the spatial coverage consistency between the support point coordinates and the crack propagation direction. If there is an uncovered area, the support case set is re-iteratively matched until all joint danger zones are covered by the support scheme.

9. The ice-rock crack monitoring and early warning method according to claim 8, characterized in that: The method for constructing the support scheme library includes: Acquire a historical support case data set, wherein the historical support case data set includes geological structure characteristics of the support area, a support point coordinate sequence, a support strength parameter group, and a construction timing instruction set; Extracting historical crack growth rate peaks, historical stress concentration factor distribution maps, and historical temperature gradient direction vectors from the geological structural characteristics of the support area; Classify the support type labels in the historical support case dataset according to the angle between the maximum gradient direction of the historical stress concentration factor distribution diagram and the historical temperature gradient direction vector, wherein the support type labels include brittle fracture support type, shear slip support type, and creep coordination support type; Input the support type label and the corresponding support point coordinate sequence into a graph convolutional network to generate a topological connection relationship graph for each support case, wherein the topological connection relationship graph includes the stress transfer path and temperature field interference area between the support points; Calculating a priority score for each support point based on the node degree centrality index of the stress transfer path and the area weight of the temperature field interference region, and reordering the construction timing instruction set according to the priority score; Aligning the timestamps of the historical crack growth rate peaks with the support strength parameter group, fitting a dynamic attenuation curve of the support strength parameter changing with the crack growth rate, and extracting the critical support strength threshold corresponding to the sudden change point of the curve slope; The support type label, topological connection relationship diagram, reordered construction timing instruction set and critical support strength threshold are associated and stored to generate a support scheme library containing multidimensional index keys, wherein the multidimensional index keys include crack propagation rate range, stress-temperature gradient angle range and priority score level.

10. An artificial intelligence-based ice and rock crack monitoring and early warning system, characterized in that: The artificial intelligence-based ice and rock crack monitoring and early warning system includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the artificial intelligence-based ice and rock crack monitoring and early warning method described in any one of claims 1 to 9 above.

Images