Artificial intelligence-based ice-rock mass crack monitoring and early warning method and system

Abstract

The application provides an ice rock mass crack monitoring and early warning method and system based on artificial intelligence. A multi-modal sensor array deployed on the surface of the ice rock mass collects real-time geological fluctuation data containing sound wave reflection, stress distribution and temperature change signals. The real-time geological fluctuation data is input into a pre-trained multi-dimensional feature fusion model to extract a crack feature vector containing parameters such as crack spatial coordinates. A space-time evolution prediction model is called to perform time series analysis on the feature vector to generate a probability distribution map containing the crack propagation direction in the future preset period and the critical rupture risk level. When the critical rupture risk level in the probability distribution map exceeds the preset threshold, an early warning signal is triggered and an emergency support scheme is generated. The scheme covers support point coordinates, support strength parameters and construction timing instructions, effectively monitors and warns the ice rock mass cracks, and provides targeted emergency measures.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to an ice-rock mass crack monitoring and early warning method and system based on artificial intelligence. BACKGROUND

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

[0003] In terms of data processing and analysis, related technologies lack the ability to integrate multi-physical field information, making it difficult to reveal the internal relationship between crack initiation, propagation and multi-field coupling. For the prediction of crack evolution path, existing methods mostly use deterministic models or empirical formulas, which cannot quantify the uncertainty of prediction results, making it difficult to support decision-makers in risk probability assessment and emergency resource optimization. In terms of early warning response mechanism, traditional schemes can only provide simple alarm information, lack of intelligent linkage with support design and construction scheme, leading to delayed or insufficient emergency measures. SUMMARY

[0004] In view of the above-mentioned problems, in combination with the first aspect of the present application, the present application embodiment provides an ice-rock mass crack monitoring and early warning method based on artificial intelligence, which comprises:

[0005] Collecting real-time geological wave data through a multi-modal sensor array deployed on the surface of the ice-rock mass, the geological wave data including acoustic wave reflection signals, stress distribution signals and temperature change signals;

[0006] Inputting the geological wave data into a pre-trained multi-dimensional feature fusion model to extract a crack feature vector inside the ice-rock mass, the crack feature vector containing crack spatial coordinates, crack propagation rate and dynamic perception parameters of stress concentration areas;

[0007] Calling a spatio-temporal evolution prediction model to perform time series analysis on the crack feature vector to generate a probability distribution map of the crack evolution path, the probability distribution map containing the crack propagation direction and critical rupture risk level in a future preset period;

[0008] According to a region in the probability distribution diagram in which a critical fracture risk level exceeds a preset threshold, a warning signal is triggered and a corresponding emergency support scheme is generated, the emergency support scheme including support point coordinate, support strength parameter and construction timing instruction.

[0009] In still another aspect, the embodiment of the present application also provides an ice rock mass crack monitoring and early warning system based on artificial intelligence, comprising a processor and a machine readable storage medium, the machine readable storage medium is connected with the processor, the machine readable storage medium is used for storing programs, instructions or codes, and the processor is used for executing the programs, instructions or codes in the machine readable storage medium to realize the method described above.

[0010] Based on the above aspects, the embodiment of the present application realizes the multi-physical field coupling monitoring of the micro-damage and macro-deformation in the ice rock mass by deploying a multi-modal sensor array on the surface of the ice rock mass to collect the sound wave reflection, stress distribution and temperature change signals in real time, the pre-trained multi-dimensional feature fusion model extracts the dynamic sensing parameters of the crack spatial coordinates, expansion rate and stress concentration area, not only describes the geometric shape and mechanical evolution characteristics of the crack, but also establishes a probability distribution diagram of the crack expansion path through the spatio-temporal evolution prediction model, the probability distribution diagram can quantify the uncertainty of the crack expansion direction and the spatial distribution of the critical fracture risk in a preset time period in the future, when the critical fracture risk level exceeds the preset threshold, a warning signal can be triggered synchronously and an emergency support scheme containing support points, strength parameters and construction timing can be generated, which significantly improves the safety and reliability of the ice rock mass engineering in extreme environments. BRIEF DESCRIPTION OF DRAWINGS

[0011] Figure 1 is the execution flow diagram of the ice rock mass crack monitoring and early warning method based on artificial intelligence provided by the embodiment of the present application.

[0012] Figure 2 is the schematic diagram of exemplary hardware and software components of the ice rock mass crack monitoring and early warning system based on artificial intelligence provided by the embodiment of the present application. DETAILED DESCRIPTION

[0013] The present application will be described in detail below with reference to the accompanying drawings, Figure 1 is the flow diagram of the ice rock mass crack monitoring and early warning method based on artificial intelligence provided by an embodiment of the present application, which will be described in detail below.

[0014] In step S110, real-time geological wave data is collected by deploying a multi-modal sensor array on the surface of the ice rock mass, the geological wave data including sound wave reflection signal, stress distribution signal and temperature change signal.

[0015] In this embodiment, consider a scenario of an engineering project inside an ice-rock mixed mountain, for example, a construction project of an underground facility inside an ice-rock mountain, on the surface of which a multi-modal sensor array is deployed.

[0016] Specifically, for the collection of acoustic reflection signals, the sensor can emit acoustic waves of a specific frequency and then receive acoustic waves reflected from different structures inside the ice-rock body. For example, when the acoustic wave encounters a crack or a region of different density in the ice-rock, the characteristics of the reflected wave such as frequency, intensity, and phase will change. The sensor continuously records the relevant information of these reflected waves to form acoustic reflection signals. Assuming that at a certain time, the sensor emits an acoustic wave with a frequency of 1000 Hz, after a period of time, the sensor receives an acoustic wave reflected from a suspected crack region 5 meters away, with a frequency of 950 Hz and a reduced intensity, all of which are accurately recorded.

[0017] The collection of stress distribution signals is realized by stress sensors. Due to factors such as self-gravity, external pressure, and internal structural changes in the ice-rock body, the stress distribution is uneven. The stress sensor can accurately measure the stress at each sampling point. For example, in the area close to the surface of the mountain, the stress may be relatively small due to the influence of external environmental temperature changes and atmospheric pressure. While in the deep part of the mountain, the stress may be larger due to the weight of the ice-rock above. The sensor will collect stress data at certain time intervals, for example, every 10 minutes, to obtain the stress distribution signal. Assuming that at a sampling point with a depth of 10 meters, the initial measured stress value is 1000 Pa, over time, due to the creep of ice-rock or external interference, the stress value may change, such as to 1050 Pa, and these changes are recorded by the stress sensor.

[0018] The temperature of the ice-rock body has a significant impact on its structural stability, and the temperature sensor can monitor the temperature changes of the surface and shallow internal part of the ice-rock body in real time. Due to the different thermal conductivity of ice-rock, the temperature changes at different depths also differ. For example, in winter, the temperature of the mountain surface may rapidly drop to -10℃, while at a depth of 2 meters inside the mountain, the temperature decreases relatively slowly, from an initial temperature of -2℃ to -3℃ after a day. The temperature sensor records these temperature change data at a frequency of every hour to form the temperature change signal. Thus, the data collected by the above different types of sensors collectively constitute the geological fluctuation data.

[0019] Step S120, input the geological fluctuation data into the pre-trained multi-dimensional feature fusion model to extract the crack feature vector inside the ice-rock body, which includes the crack spatial coordinates, crack propagation rate, and dynamic perception parameters of the stress concentration area.

[0020] In this embodiment, after obtaining the above geological fluctuation data, it can be input into the pre-trained multi-dimensional feature fusion model, which has been pre-trained by a large amount of sample geological fluctuation data and corresponding sample crack annotation data.

[0021] Taking the previously mentioned underground facility construction scene as an example, the sound wave reflection signal data received by the multi-dimensional feature fusion model will be processed first. Assuming that the multi-dimensional feature fusion model has determined the relationship between the sound wave reflection signal in a certain frequency range and the crack, such as the reflection wave anomaly between the frequencies of 900-1000 Hz may be related to the crack. The multi-dimensional feature fusion model can analyze the collected sound wave reflection signal, predict the reflection wave part that meets the frequency range and has specific intensity and phase characteristics, and thus determine the approximate area where the crack may exist.

[0022] For stress distribution signals, the multi-dimensional feature fusion model can analyze in combination with the mechanical properties of the ice-rock body. When there is a significant gradient change in stress distribution, for example, the stress suddenly increases from 1000 Pa to 1500 Pa in a certain area, and this change does not conform to the normal stress distribution rule, the multi-dimensional feature fusion can judge that there may be stress concentration phenomenon in this area, and stress concentration is often related to the existence and expansion of cracks. The model will calculate the range of stress concentration area and the degree of stress concentration according to the distribution of stress, and other dynamic perception parameters.

[0023] Temperature change signals are also considered by the model. In the ice-rock body, the temperature around the crack may be different from the surrounding area due to factors such as air or water flow. If the temperature change rate in a certain area is significantly higher than that in other areas, the multi-dimensional feature fusion model can analyze it as a feature. For example, the temperature in a certain area drops by 5°C in a short time, while the surrounding area only drops by 1°C, which implies that there is a crack in this area, because the crack may affect the conduction of heat.

[0024] Through the comprehensive analysis of the above signals, the multi-dimensional feature fusion model can accurately extract the crack feature vector inside the ice-rock body. Assuming that the multi-dimensional feature fusion model finally determines that there is a crack at a position 3 meters horizontally and 5 meters vertically from the sensor array, the crack propagation rate is 0.1 meters per month, and there is a stress concentration area around the crack, and the stress gradient of the stress concentration area is 100 Pa per meter, the above parameters constitute the crack feature vector.

[0025] Step S130, calling the space-time evolution prediction model to perform time series analysis on the crack feature vector, and generating a probability distribution map of the crack evolution path, which contains the crack propagation direction and critical rupture risk level in the future preset period.

[0026] After obtaining the crack feature vector, the space-time evolution prediction model can be called for analysis. Continuing with the underground facility construction scenario as an example, the space-time evolution prediction model constructs a hybrid neural network architecture containing a long short-term memory network and a three-dimensional convolution kernel.

[0027] The long short-term memory network part starts to capture the time-dependent features of the crack propagation rate from the crack feature vector. For example, the space-time evolution prediction model can trace the changes in crack propagation rate over the past few months. If the crack propagation rate in the past three months was 0.08 meters, 0.09 meters, and 0.1 meters per month, respectively, showing a gradually rising trend, the space-time evolution prediction model can record this trend as a time-dependent feature.

[0028] The three-dimensional convolution kernel extracts the topological features of the crack spatial morphology from the crack feature vector. For example, based on the spatial coordinates of the crack and the shape of the stress concentration region, the space-time evolution prediction model can construct the approximate morphology of the crack in three-dimensional space, including the length, width, strike of the crack, and the spatial relationship with the surrounding stress concentration region, and other topological features.

[0029] Then, through the gating mechanism, the time-dependent features and topological features are dynamically fused to generate dynamic fusion features containing time-space correlation. Assuming based on previous analysis, the space-time evolution prediction model predicts that in the next three months, the crack may continue to expand along the current direction, and due to the influence of the stress concentration region, the crack propagation probability will be higher in certain directions.

[0030] Next, the dynamic fusion features are input into the discriminator network. The space-time evolution prediction model can obtain the spatial coordinate point set of the real rupture event corresponding to the crack feature vector, which may come from a database of similar ice rock mass rupture events in the past or small-scale rupture events monitored on site. Then, these spatial coordinate point sets are converted into a binary rupture mask map with the same resolution as the spatial probability distribution map output by the space-time evolution prediction model. For example, if a coordinate point is determined to be a rupture point in the past rupture event, the corresponding spatial unit in the binary rupture mask map is marked as 1, otherwise marked as 0.

[0031] Next, the cross-entropy loss of 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 can be calculated to generate an initial adversarial loss component. At the same time, the set of spatial units in the spatial probability distribution map whose probability values exceed a preset probability threshold (such as 0.5) is extracted, and the spatial intersection-over-union of the set of spatial units and the fracture region in the binary fracture mask map is calculated to generate an intersection-over-union penalty term. The initial adversarial loss component and the intersection-over-union penalty term are linearly combined according to a preset weight coefficient to generate a final adversarial training loss. In this way, the model continuously optimizes the prediction results, and finally generates a probability distribution map of the crack evolution path. The probability distribution map shows that the probability of the crack propagation direction being towards a certain direction is 0.6 in the future preset period (such as the next three months), while the probability in another direction is 0.3, and also gives the critical fracture risk levels of different regions, such as the critical fracture risk level of a certain region is high, while the risk levels of other regions are medium or low.

[0032] In step S140, according to the region in the probability distribution map whose critical fracture risk level exceeds a preset threshold, a warning signal is triggered and a corresponding emergency support scheme is generated, and the emergency support scheme includes support point coordinates, support strength parameters and construction timing instructions.

[0033] Suppose the preset threshold is the critical value of the high risk level, if the critical fracture risk level of a certain region in the probability distribution map exceeds the preset threshold, a warning signal needs to be triggered. For example, in the previously mentioned underground facility construction scenario, if it is found that the critical fracture risk level of the region 5 meters away from the top of the underground facility and 8 meters horizontally is high, it means that the ice rock structure of this region is very unstable and may crack at any time, thereby threatening the safety of the underground facility.

[0034] First, all spatial units in the probability distribution map are traversed to filter out a set of spatial units whose critical fracture risk levels exceed a preset threshold, which are marked as initial dangerous regions. Spatial connectivity analysis is performed on the initial dangerous regions, and it is assumed that several adjacent spatial units are found whose risk levels are slightly lower than the threshold individually, but their spatial distance is less than a preset safety radius (such as 1 meter), then these adjacent spatial units are merged to generate a continuously distributed joint dangerous region.

[0035] Then, the geological structure features of the joint dangerous region are extracted. For the joint dangerous region, the maximum crack propagation rate of the region is 0.12 meters per month, the average stress concentration coefficient is 1200 pascals per square meter, and the temperature gradient distribution pattern is that the temperature drops rapidly from the surface to the inside, etc. geological structure features are obtained through the previously collected geological fluctuation data and analysis results.

[0036] The geological structure features are matched with support case features in a pre-stored support scheme library. The support scheme library is constructed by obtaining a historical support case dataset, which contains support schemes under various geological structures. Suppose that in the support scheme library, the geological structure features of some support cases are similar to those of the current joint danger zone, such as the crack propagation rate, stress concentration coefficient, and temperature gradient distribution pattern in the historical cases are relatively close to the current region, and a support case set with a matching degree exceeding a similarity threshold (such as 0.8) is selected.

[0037] According to the support intensity parameters and construction timing instructions of the support case set, an initial support scheme template is generated. For example, a certain case in the support case set suggests setting a support point every 2 meters around the danger zone, the support intensity parameter is able to withstand a pressure of 15000 Newtons, and the construction timing instruction is to set the support point at the front end of the crack propagation direction first, and then set it to the two sides in turn.

[0038] Then, the real-time monitored rock mass creep rate and environmental temperature change amount are obtained. Suppose that the real-time monitoring shows that the rock mass creep rate has increased and the environmental temperature has decreased rapidly. According to these changes, the dynamic attenuation coefficient of the support intensity parameter in the initial support scheme template and the time offset of the construction timing instruction are calculated. For example, due to the increase of the rock mass creep rate, the support intensity parameter needs to be increased by 10%, i.e. from 15000 Newtons to 16500 Newtons; due to the rapid decrease of the environmental temperature, the construction timing instruction needs to be advanced, which originally planned to start the support work after 10 days, needs to be advanced to start after 5 days.

[0039] The support intensity parameter of the initial support scheme template is adjusted according to the dynamic attenuation coefficient, and the priority order of the construction timing instruction is corrected based on the time offset, to generate an emergency support scheme containing updated support point coordinates, optimized support intensity parameters, and corrected construction timing instructions. For example, the final emergency support scheme determines to set a support point every 1.8 meters around the joint danger zone (updated support point coordinates), the support intensity parameter is 16500 Newtons (optimized support intensity parameter), and the construction timing instruction is to set the support point at the front end of the crack propagation direction first, and then complete the setting of all support points within 3 days (corrected construction timing instruction).

[0040] Finally, the emergency support scheme is topologically mapped with the spatial coordinates of the joint danger zone to verify the spatial coverage consistency of the support point coordinates and the crack propagation direction. If it is found that there is an uncovered area, for example, a small part of the area on one side of the crack propagation direction is not covered by the support point, the support case set is re-iterated and matched until all joint danger zones are covered by the support scheme, thereby ensuring the safety of the underground facility.

[0041] Based on the above steps, the multi-modal sensor array deployed on the surface of the ice rock mass real-time collects the sound wave reflection, stress distribution and temperature change signals, realizes the multi-physical field coupling monitoring of the micro-damage and macro-deformation of the ice rock mass, and the pre-trained multi-dimensional feature fusion model extracts the dynamic perception parameters of the crack spatial coordinates, expansion rate and stress concentration area, not only describes the geometric shape and mechanical evolution characteristics of the crack, but also establishes a probability distribution map of the crack expansion path through the time-space evolution prediction model. The probability distribution map can quantify the uncertainty of the crack expansion direction and the spatial distribution of the critical rupture risk in the future preset period. When the critical rupture risk level exceeds the preset threshold, the early warning signal can be triggered synchronously and the emergency support scheme containing the support point, strength parameter and construction time sequence can be generated, which significantly improves the safety and reliability of the ice rock mass engineering in extreme environment.

[0042] In a possible implementation, the pre-training process of the multi-dimensional feature fusion model includes:

[0043] In step S210, sample geological fluctuation data sets and corresponding sample crack annotation data are obtained. The sample geological fluctuation data includes sample sound wave reflection signals, sample stress distribution signals and sample temperature change signals. The sample crack annotation data includes sample crack spatial coordinates, sample crack expansion rates and dynamic perception parameters of sample stress concentration areas.

[0044] For example, in the early exploration stage, sensors are installed at different positions of the ice rock mass. After a period of monitoring, sample sound wave reflection signals are obtained, such as the intensity and time information of sound waves of different frequencies reflected in the ice rock mass. The sample stress distribution signal is obtained by a stress sensor, which reflects the distribution of the stress in the ice rock mass. The sample temperature change signal is provided by a temperature sensor, such as temperature values recorded at different depths and different time periods. The corresponding sample crack annotation data includes sample crack spatial coordinates, sample crack expansion rates and dynamic perception parameters of sample stress concentration areas. These annotation data can be obtained by more accurate measurement equipment or manual analysis, such as determining the spatial coordinates of the crack by using high-precision imaging equipment, obtaining the crack expansion rate through long-term tracking and monitoring, and obtaining the dynamic perception parameters through detailed analysis of the stress concentration area.

[0045] In step S220, time-frequency transformation is performed on the sample geological fluctuation data to generate a joint feature matrix including sample sound wave spectrum features, sample stress gradient features and sample temperature change rates.

[0046] In a possible implementation, step S220 includes:

[0047] Step S221, performing short-time Fourier transform on the sample acoustic wave reflection signal to generate a time-frequency spectrogram, and extracting an energy distribution peak value of a preset frequency band in the time-frequency spectrogram as the sample acoustic wave spectrum feature.

[0048] Step S222, performing spatial gradient calculation on the sample stress distribution signal to obtain stress gradient amplitudes and direction angles of each sampling point, and taking a stress gradient amplitude difference between adjacent sampling points as the sample stress gradient feature.

[0049] Step S223, performing time window sliding average processing on the sample temperature change signal to calculate a standard deviation of a temperature change amount in each time window as the sample temperature change rate.

[0050] Step S224, aligning the sample acoustic wave spectrum feature, the sample stress gradient feature, and the sample temperature change rate according to time stamps, and splicing them into a three-dimensional tensor form of the joint feature matrix after normalization processing.

[0051] In this process, the acoustic wave reflection signal is processed in a certain time window to generate a time-frequency spectrogram. For example, set the time window to 1 second, and perform Fourier transform on the acoustic wave reflection signal in this 1 second to obtain a time-frequency spectrogram. Then extract the energy distribution peak value of the preset frequency band from the time-frequency spectrogram as the sample acoustic wave spectrum feature. Assuming that the preset frequency band is 1000-1500 Hz, by analyzing the energy distribution of the time-frequency spectrogram in this frequency band, the peak value corresponding to the highest energy is found. This value is the sample acoustic wave spectrum feature. For the sample stress distribution signal, spatial gradient calculation is performed, and for each sampling point, the stress gradient amplitude and direction angle are calculated. For example, in a two-dimensional sampling plane, taking a sampling point as the center, the stress values of the surrounding adjacent sampling points are calculated to obtain the stress gradient amplitude. The direction angle is determined according to the direction of stress change, for example, if the stress gradually increases from left to right, then the direction angle is horizontally to the right. Then, the stress gradient amplitude difference of adjacent sampling points is taken as the sample stress gradient feature. If the stress gradient amplitudes of a sampling point and its right adjacent sampling point are 1000 Pa and 1200 Pa respectively, then the difference 200 Pa is part of the sample stress gradient feature. For the sample temperature change signal, time window moving average processing is performed, and the standard deviation of the temperature change in each time window is calculated as the sample temperature change rate. For example, in a 1-hour time window, the recorded temperature values are -2℃, -1.5℃, -1℃, etc. First, calculate the difference between each temperature value and the average temperature value in the time window, then calculate the sum of the squares of these differences, and finally take the square root of the average value of the sum to obtain the sample temperature change rate. Finally, align the sample acoustic wave spectrum feature, sample stress gradient feature, and sample temperature change rate by timestamp, and concatenate them into a three-dimensional tensor form joint feature matrix after normalization processing. Assuming 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, adjust their value ranges to the appropriate interval through normalization, and then concatenate them into a three-dimensional tensor form joint feature matrix in the order set.

[0052] In step S230, the initialized deep cross-attention network is used to take the sample acoustic 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. Through the multi-head attention mechanism, the sample weight distribution coefficient of each modal data in the joint feature matrix is calculated.

[0053] In the deep cross-attention network, the multi-head attention mechanism can comprehensively consider the relationship between the modal data from different representation subspaces. For example, when interacting with the query vector of the sample acoustic spectrum feature, the key vector of the sample stress gradient feature, and the value vector of the sample temperature change rate, weights can be calculated on multiple heads (assuming 3 heads). On the first head, a weight coefficient such as 0.3 can be calculated according to the correlation between certain frequency components in the sample acoustic spectrum feature and the stress change trend in the sample stress gradient feature; on the second head, another weight coefficient such as 0.2 can be calculated based on the relationship between the sample acoustic spectrum feature and the sample temperature change rate; and on the third head, a weight coefficient of 0.1 can be calculated by comprehensively considering the more complex relationship between the three features. Then, the weight coefficients calculated on different heads are integrated to obtain the final sample weight distribution coefficients of the modal data.

[0054] In step S240, the joint feature matrix is dynamically weighted and fused based on the sample weight distribution coefficients, and a fused sample crack feature vector is output. The parameters of the deep cross-attention network are optimized by a contrast loss function to minimize the spatial distribution error between the sample crack feature vector and the sample crack label data.

[0055] For example, the joint feature matrix contains multiple elements. The weight coefficient of the sample acoustic spectrum feature is 0.3, the weight coefficient of the sample stress gradient feature is 0.3, and the weight coefficient of the sample temperature change rate is 0.4. Each element in the joint feature matrix is multiplied by the corresponding weight coefficient, and then the weighted elements are added to obtain the fused sample crack feature vector, which contains crack-related features after considering the relationship between the modal data.

[0056] In one possible implementation, step S240 includes:

[0057] In step S241, the fused sample crack feature vector and the sample crack label data are divided according to the spatial grid to generate corresponding sample crack feature matrices and sample label matrices.

[0058] In step S242, according to the crack spatial coordinates in the sample label matrix, positive sample pairs and negative sample pairs are assigned in the sample crack feature matrix. The positive sample pairs are grid cells with a spatial distance less than a preset threshold in the same sample crack feature matrix, and the negative sample pairs are grid cells with a spatial distance greater than the preset threshold in different sample crack feature matrices.

[0059] In step S243, the cosine similarity of the grid cells in the positive sample pairs is calculated, and the Euclidean distance difference of the grid cells in the negative sample pairs is extracted.

[0060] Step S244, constructing a contrast loss function according to the cosine similarity and the Euclidean distance difference, wherein a similarity maximization constraint term of a positive sample pair and a difference minimization constraint term of a negative sample pair are weighted and summed.

[0061] Step S245, using a gradient descent algorithm to back-propagate partial derivatives of the contrast loss function to update attention weight coefficients and fully connected layer parameters in the deep cross-attention network until spatial distribution errors of the sample crack feature matrix and the sample label matrix converge to a preset tolerance range.

[0062] In this embodiment, it is assumed that the research area of the ice rock mass is divided into a 10x10 spatial grid, and each spatial grid corresponds to a part of the sample crack feature vector and the sample crack label data. According to the crack spatial coordinates in the sample label matrix, positive sample pairs and negative sample pairs can be assigned in the sample crack feature matrix. For example, set the preset threshold to be 3 grid units of distance. In the same sample crack feature matrix, if the spatial distance between two grid units is less than 3 grid units, they are considered as a positive sample pair, such as the two grid units with coordinates (2, 3) and (3, 3). In different sample crack feature matrices, if the spatial distance between two grid units is greater than 3 grid units, they are considered as a negative sample pair, 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. The cosine similarity of the grid units in the positive sample pair is calculated. Assuming that the vectors of the two positive sample pair grid units are [0.5, 0.3] and [0.4, 0.4], first calculate their dot product as 0.5x0.4+0.3x0.4=0.32, then calculate the length of the two vectors respectively, the length of the first vector is √(0.5²+0.3²)=0.5831, and the length of the second vector is √(0.4²+0.4²)=0.5657, then the cosine similarity is 0.32 / (0.5831x0.5657)=0.97. At the same time, the Euclidean distance difference of the grid units in the negative sample pair is extracted. For the two grid units of the above negative sample pair, assuming that 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. According to the cosine similarity and the Euclidean distance difference, a contrast loss function is constructed. Assuming that the weight of the similarity maximization constraint term of the positive sample pair is 0.5, and the weight of the difference minimization constraint term of the negative sample pair is 0.5, then the contrast loss function is 0.5x(1-0.97)+0.5x0.566=0.323. The partial derivative of the contrast loss function is propagated in the reverse direction using the gradient descent algorithm to update the attention weight coefficients and the fully connected layer parameters in the deep cross attention network. Each time the attention weight coefficients and the fully connected layer parameters are updated, the direction and size of the partial derivative are adjusted according to a certain step size. For example, if a certain attention weight coefficient has a current value of 0.4 and needs to be increased by 0.05 according to the partial derivative, the updated value is 0.45. In this way, the process is repeated until the spatial distribution error of the sample crack feature matrix and the sample label matrix converges to a preset tolerance range. Assuming that the preset tolerance range is 0.01, when the calculated spatial distribution error is less than 0.01, it is considered that the parameters of the deep cross attention network have been optimized to an appropriate level.

[0063] In a possible implementation, the method for training the spatio-temporal evolution prediction model comprises:

[0064] In step S310, a hybrid neural network architecture comprising a long short-term memory network and a three-dimensional convolution kernel is constructed, wherein the long short-term memory network is used to capture time-dependent features of crack propagation rate from sample crack feature vectors, and the three-dimensional convolution kernel is used to extract topological features of crack spatial morphology from sample crack feature vectors.

[0065] For example, in the process of long-term monitoring of ice rock mass cracks, the sample crack feature vector contains crack propagation rate information at different time points. The long short-term memory network can analyze these data, such as a series of crack propagation rate data, for example, at the past 10 time points, the crack propagation rate is gradually increasing, such as 0.05 meters per month, 0.06 meters per month, 0.07 meters per month, etc. The long short-term memory network can capture the feature that the trend gradually rises over time. The three-dimensional convolution kernel is used to extract topological features of crack spatial morphology from sample crack feature vectors. For example, according to the crack spatial coordinates and other information in the sample crack feature vector, the three-dimensional convolution kernel can construct the three-dimensional spatial morphology of the crack in the ice rock mass, determine the topological features of the crack such as length, width, strike, and spatial relationship with the surrounding ice rock structure, etc.

[0066] In step S320, the sample crack feature vector is input into the hybrid neural network architecture in time sequence, and the time-dependent features and the topological features are dynamically fused through a gating mechanism to generate dynamic fusion features containing time-space correlation.

[0067] Taking the crack propagation rate and crack spatial morphology mentioned earlier as an example, the gating mechanism can fuse the crack propagation rate feature captured by the long short-term memory network and the crack spatial morphology feature extracted by the three-dimensional convolution kernel according to the set rules and algorithms. Assuming that at a certain time, the crack propagation rate is 0.08 meters per month and shows an accelerating trend, and the crack shows a morphology of extending in a certain direction and gradually widening in space, the gating mechanism can consider these factors comprehensively, weigh the importance of each feature, and fuse them into a dynamic fusion feature containing time-space correlation. The dynamic fusion feature can comprehensively describe the state of the crack in time and space.

[0068] In step S330, the dynamic fusion feature is input into a discriminator network, the spatial matching degree between the crack evolution path predicted by the dynamic fusion feature and the real fracture event is calculated through the discriminator network, and an 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 optimized in reverse.

[0069] In a possible implementation, step S330 comprises:

[0070] Step S331, inputting the dynamic fusion features into the full connection layer of the discriminator network to generate a spatial probability distribution map of the predicted crack evolution path output by the discriminator network.

[0071] Step S332, obtaining a set of spatial coordinate points of a real fracture event corresponding to the sample crack 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.

[0072] Step S333, calculating the cross-entropy loss of 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.

[0073] Step S334, extracting a set of spatial units in the spatial probability distribution map whose probability values exceed a preset probability threshold, and calculating the spatial intersection-over-union of the set of spatial units and the fracture region in the binary fracture mask map to generate an intersection-over-union penalty term.

[0074] Step S335, linearly combining 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.

[0075] Step S336, calculating the partial derivative of the adversarial training loss with respect to the three-dimensional convolution kernel parameters of the hybrid neural network architecture through a back propagation algorithm, and updating the weight matrix of the three-dimensional convolution kernel according to the gradient direction to maximize the spatial matching degree between the spatial probability distribution map and the binary fracture mask map.

[0076] For example, the spatial probability distribution map can show the probability of cracks at different spatial locations in the ice-rock body to propagate to the location in a future time period. Meanwhile, the spatial coordinate point sets corresponding to the sample crack feature vectors are obtained, which are obtained by detailed measurement and recording of past ice-rock body fracture events. Then these spatial coordinate point sets are converted into a binary fracture mask map with the same resolution as the spatial probability distribution map, i.e. the coordinate points of the fracture are marked as 1 and the unbroken are marked as 0. The cross-entropy loss of 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 is calculated to generate the initial adversarial loss component. For example, for a spatial unit in the spatial probability distribution map, the probability value of the predicted crack propagation to the location is 0.6, and the label value corresponding to the location in the binary fracture mask map is 1. According to the calculation method of cross-entropy loss, first calculate -(1 x log(0.6) + (1-1) x log(1-0.6)), where 1 x log(0.6) represents the loss calculation part when the true label is 1, and (1-1) x log(1-0.6) represents the loss calculation part when the true label is 0 (here it is 0 because the true label is 1). The calculation result is -log(0.6), which is the contribution of the spatial unit to the initial adversarial loss component. Calculate the intersection-over-union of the spatial unit set with the probability value exceeding the preset probability threshold (assuming 0.5) in the spatial probability distribution map and the fracture region in the binary fracture mask map to generate the intersection-over-union penalty term. Assuming that there are 10 spatial units with a probability value exceeding 0.5, and the fracture region in the binary fracture mask map contains 8 spatial units, the intersection of the 10 spatial units and the 8 spatial units is 6 spatial units, and the intersection-over-union is 6 / (10+8-6)=0.5. The intersection-over-union penalty term can be calculated based on the intersection-over-union according to a specific algorithm. Linearly combine the initial adversarial loss component and the intersection-over-union penalty term according to the preset weight coefficients to generate the final adversarial training loss. Assuming that the value of the initial adversarial loss component is 0.3 and the value of the intersection-over-union penalty term is 0.2, and the preset weight coefficients are 0.6 and 0.4 respectively, then the final adversarial training loss is 0.6 x 0.3 + 0.4 x 0.2 = 0.26. The partial derivative of the adversarial training loss with respect to the three-dimensional convolution kernel parameters of the hybrid neural network architecture is calculated by the backpropagation algorithm, and the weight matrix of the three-dimensional convolution kernel is updated according to the direction of the partial derivative, so that the spatial matching degree of the spatial probability distribution map and the binary fracture mask map is maximized.For example, if the partial derivative of a certain three-dimensional convolution kernel parameter is 0.1, indicating that the increase of the parameter will reduce the adversarial training loss, then according to a certain update step (assuming 0.01), the parameter is increased by 0.01, and after multiple such updates, the spatial probability distribution map and the binaryized crack mask map are gradually improved in spatial matching degree.

[0077] Step S340, input the dynamic fusion feature into a physical constraint loss function, calculate the angle penalty term between the maximum principal stress direction in the predicted crack evolution path and the crack propagation direction, and combine the adversarial training loss to generate a joint optimization loss.

[0078] In one possible implementation, step S340 includes:

[0079] Step S341, extract 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.

[0080] Step S342, calculate the first angle between the maximum principal stress direction and the crack surface normal direction, and the second angle between the crack propagation direction and the maximum shear stress direction.

[0081] Step S343, when the first angle is less than a preset angle threshold and the second angle is greater than a critical fracture angle, reduce the fracture risk prediction value of the area.

[0082] Step S344, when the first angle is in the brittle fracture interval and the second angle satisfies the Coulomb friction condition, increase the fracture risk prediction value of the area.

[0083] Step S345, calculate the direction consistency loss according to the spatial coincidence degree of the prediction value and the actual fracture event, and weight and sum the direction consistency loss and the adversarial loss of the discriminator network to obtain the final joint optimization loss.

[0084] In this embodiment, the stress condition of the stress concentration area is calculated by the data obtained by the stress sensor deployed previously. Assuming that in a certain stress concentration area, after detailed stress analysis, 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, and 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, that is, |30-40|=10 degrees. For the second angle between the crack propagation direction and the maximum shear stress direction, it is assumed to be 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, the fracture risk prediction value of the area is originally 0.8, and it is reduced to 0.7 according to the rule. When the first angle is in the brittle fracture interval (assuming 10-20 degrees) and the second angle satisfies the Coulomb friction condition (assuming it satisfies the condition here), the fracture risk prediction value of the area is increased, such as from 0.7 to 0.8. Calculate the direction consistency loss according to the spatial coincidence degree of the prediction value and the actual fracture event. Assuming that the spatial proportion of the area with a prediction value of 0.8 coinciding with the actual fracture event is 0.6, the direction consistency loss can be calculated based on the coincidence proportion according to a specific algorithm, and the calculation result is assumed to be 0.1. The direction consistency loss and the adversarial loss of the discriminator network are weighted and summed to obtain the final joint optimization loss. Assuming that the adversarial loss is 0.26, the weight of the direction consistency loss is 0.3, and the weight of the adversarial loss is 0.7, then the final joint optimization loss is 0.3x0.1+0.7x0.26=0.212.

[0085] 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 spatio-temporal evolution prediction model is generated.

[0086] In the iteration process, the above calculation and update steps are repeatedly performed. In each iteration, the convolution kernel parameters of the hybrid neural network architecture and the weight parameters of the discriminator network are updated according to the new joint optimization loss. For example, in an iteration, the update amount calculated according to the joint optimization loss for a certain convolution kernel parameter is -0.05, then the convolution kernel parameter is reduced by 0.05. With the increase of the number of iterations, the joint optimization loss gradually decreases, and when the value of the joint optimization loss is less than the preset error range (assuming 0.01), it is considered that the spatio-temporal evolution prediction model has been trained and can be used for accurate prediction of ice rock mass crack evolution.

[0087] In one possible implementation, step S140 includes:

[0088] 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 as an initial dangerous area.

[0089] In this embodiment, the probability distribution map is obtained by the spatio-temporal evolution prediction model, which shows the risk of crack evolution of different spatial units in the ice-rock mass. Assuming that the preset threshold is set to a certain risk level value, when traversing each spatial unit in the probability distribution map, those spatial units with a risk level higher than the preset threshold are selected. For example, each spatial unit in the probability distribution map has a corresponding fracture risk level value, from 0 to 1, indicating the risk level from low to high. If the preset threshold is 0.7, when the risk level of a certain spatial unit is 0.8, the spatial unit will be selected. All such selected spatial units constitute the initial dangerous area.

[0090] Step S142, performing spatial connectivity analysis on the initial dangerous area, merging adjacent spatial units with a spatial distance less than a preset safety radius to generate a continuously distributed joint dangerous area.

[0091] In this embodiment, the distribution of each spatial unit in the initial dangerous area may be discrete. The judgment of adjacent spatial units is based on their spatial position relationship in the ice-rock mass. The preset safety radius is a distance value determined according to the physical properties of the ice-rock mass and engineering experience. For example, the preset safety radius is set to 2 meters. If there are two adjacent spatial units in the initial dangerous area, and the spatial distance between them is 1.5 meters, then these two spatial units meet the merging condition. After such analysis and merging operation on the entire initial dangerous area, a continuously distributed joint dangerous area is obtained, which represents the region in the ice-rock mass with high fracture risk and continuous in space, which is more convenient for subsequent unified planning of support scheme.

[0092] Step S143, extracting the geological structure features of the joint dangerous area, including the maximum crack propagation rate, the average stress concentration coefficient, and the temperature gradient distribution pattern.

[0093] In this embodiment, the regional maximum crack propagation rate is obtained from the long-term monitoring data of the crack. For example, the crack propagation rate is measured at different positions in the joint danger zone, such as 0.1 meters per month at a certain position and 0.12 meters per month at another position. After comparison, it is determined that the regional maximum crack propagation rate is 0.12 meters. The average stress concentration coefficient is calculated according to the stress data collected by the stress sensor in the joint danger zone. Assuming that multiple stress sampling points are set in the joint danger zone, each sampling point has a corresponding stress value. The average stress value is obtained by adding the stress values of all sampling points and dividing by the number of sampling points. The average stress concentration coefficient is obtained according to the existing calculation method of stress concentration area. The temperature gradient distribution pattern is determined by the data obtained by the temperature sensor. The temperature is measured at different depths and positions in the joint danger zone, such as -5°C at the surface and -3°C at a depth of 5 meters. The temperature gradient distribution pattern is determined by analyzing the temperature differences at different positions, such as whether the temperature gradually increases or decreases with increasing depth, etc.

[0094] Step S144, similarity matching the geological structure features with the support case features in the pre-stored support scheme library, and screening out a support case set with a matching degree exceeding a similarity threshold.

[0095] In this embodiment, the pre-stored support scheme library is constructed by a large number of historical support cases. The support case features include various related parameters under different geological structures. When performing similarity matching, multiple geological structure feature parameters are considered comprehensively. For example, for the feature of regional maximum crack propagation rate, if the maximum crack propagation rate of the joint danger zone is 0.12 meters per month, and the crack propagation rate of a support case in the support scheme library is between 0.1 meters and 0.15 meters per month, then there is a certain 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 comprehensive comparison of all geological structure features, it is found that the similarity between a certain support case and the geological structure features of the joint danger zone reaches 0.85. Therefore, this support case can be screened out. All such screened support cases constitute a support case set.

[0096] Step S145, generating an initial support scheme template according to the support intensity parameters and construction timing instructions of the support case set.

[0097] In this embodiment, each case in the support case set has a corresponding support strength parameter and construction timing instruction. The support strength parameter refers to the stress size and other related parameters that the support structure can withstand, and the construction timing instruction specifies the sequence of support construction and other information. For example, the support strength parameter suggested by a certain support case is to withstand a pressure of 15000 Newtons, and the construction timing instruction is to first set the support structure at the center position of the joint danger area, and then expand to the surrounding. By synthesizing these 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 general arrangement method, and the basic sequence of construction, etc.

[0098] In step S146, the real-time monitored rock mass creep rate and environmental temperature change amount are obtained, and the dynamic attenuation coefficient of the support strength parameter in the initial support scheme template and the time offset of the construction timing instruction are calculated.

[0099] In this embodiment, the real-time monitored rock mass creep rate and environmental temperature change amount are obtained by deploying special monitoring equipment in the ice rock mass. For example, the rock mass creep rate increases from 0.01 meters per month to 0.02 meters per month, and the environmental temperature decreases from -5°C to -8°C. For the calculation of the dynamic attenuation coefficient of the support strength parameter, an empirical model based on the relationship between rock mass creep rate, environmental temperature and support strength is assumed. According to the model, the increase of rock mass creep rate will lead to the increase of support strength requirement, and the decrease of environmental temperature will also affect the support strength. First, calculate the influence of rock mass creep rate and environmental temperature change on support strength respectively, such as the increase of rock mass creep rate leading to the increase of support strength by 10%, and the decrease of environmental temperature leading to the increase of support strength by 5%, then the total dynamic attenuation coefficient is 1+10%+5%=1.15. For the calculation of the time offset of the construction timing instruction, it is also based on the empirical model or pre-set rules. For example, the rapid decrease of environmental temperature may accelerate the risk of ice rock mass rupture, according to the rules, the construction timing needs to be advanced. If the original plan is to start construction after 10 days, due to the influence of the decrease of environmental temperature, it is calculated that it needs to be advanced by 3 days, then the time offset is -3 days.

[0100] In step S147, the support strength parameter of the initial support scheme template is adjusted according to the dynamic attenuation coefficient, and the priority order of the construction timing instruction is corrected based on the time offset, to generate an emergency support scheme containing updated support point coordinates, optimized support strength parameters and corrected construction timing instructions.

[0101] In this embodiment, according to the dynamic attenuation coefficient 1.15 calculated in the previous step, if the support strength parameter in the initial support scheme template is 15000 Newton, then the updated support strength parameter is 15000 x 1.15 = 17250 Newton. For the construction timing instruction, it is corrected based on the time offset of -3 days. If the original construction timing instruction is to first set the support structure at the center position, and then expand it to the four directions in three steps, each step interval is 2 days, then the corrected construction timing instruction may be to first set the support structure at the center position, and then advance the second step by 1 day and the third step by 2 days to speed up the progress of support construction and cope with the increased risk of ice rock mass rupture. In this way, an emergency support scheme is generated, which includes the updated support point coordinate (determined according to the support structure arrangement mode, such as arranging the coordinates of support points at a certain interval in the joint danger area), the optimized support strength parameter 17250 Newton, and the corrected construction timing instruction.

[0102] In step S148, the emergency support scheme is topologically mapped with the spatial coordinates of the joint danger area to verify the spatial coverage consistency of the support point coordinate and the crack propagation direction. If there is an uncovered area, the support case set is re-iterated to match, until all joint danger areas are covered by the support scheme.

[0103] In this embodiment, the topological mapping is to accurately check the coverage of the emergency support scheme in space on the joint danger area. By corresponding analysis of the support point coordinate in the emergency support scheme and the spatial coordinate of the joint danger area. For example, the crack propagation direction is determined to be inclined and extended in a certain direction according to the previous monitoring and analysis of the crack. It is checked whether the support point coordinate is distributed around the crack propagation direction to ensure that it can effectively prevent the further expansion of the crack. If it is found that there is an uncovered area, for example, a small part of the area on one side of the crack propagation direction is not covered by the support point, then the support case set needs to be re-iterated to match. According to the previous steps, start from similarity matching with the support case features in the pre-stored support scheme library, adjust the support strength parameter and the construction timing instruction, etc., until all joint danger areas are covered by the support scheme, so as to protect the underground facilities from the threat of ice rock mass rupture.

[0104] In one possible implementation, the construction method of the support scheme library comprises:

[0105] In step S410, a historical support case data set is obtained, which includes support area geological structure features, support point coordinate sequence, support strength parameter group, and construction timing instruction set.

[0106] For example, in the past different ice rock mass underground engineering, there are detailed records for each project. The supporting area geological structure features cover various characteristics of the ice rock mass at that time, such as rock type, ice rock ratio, structural integrity, and other information. The supporting point coordinate sequence records the specific position coordinates of each supporting point in the supporting area, which accurately determines the layout of the supporting structure in the ice rock mass. The supporting strength parameter group contains stress, pressure and other related parameters that each supporting point can withstand to ensure the effectiveness of the supporting structure. The construction timing instruction set specifies the sequence and timing of the supporting construction, such as setting up supporting structures in which positions first, then following what order to gradually complete the entire supporting project.

[0107] Step S420, extracting the historical crack propagation rate peak value, the historical stress concentration coefficient distribution map and the historical temperature gradient direction vector in the supporting area geological structure features.

[0108] In this embodiment, the extraction of the historical crack propagation rate peak value is obtained by analyzing the long-term monitoring data of the ice rock mass crack in the past project. For example, in a certain historical project, the expansion of the ice rock mass crack is continuously monitored, and the expansion rate in different time periods is recorded, such as 0.05 meters, 0.08 meters, 0.1 meters, etc. After comparison, the maximum value 0.1 meters is found, which is the historical crack propagation rate peak value of the project. The historical stress concentration coefficient distribution map is drawn according to the data collected by the stress sensors arranged in the supporting area at that time. The sensor measures the stress value at different positions, and the stress concentration coefficient at each position is obtained by the existing calculation method, and then the coefficients are plotted into a distribution map according to the position 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 positions in the supporting area, the temperature change with position is analyzed, the direction of the temperature gradient is determined, and the historical temperature gradient direction vector is composed of these direction information.

[0109] Step S430, according to the angle between the maximum gradient direction of the historical stress concentration coefficient distribution map and the historical temperature gradient direction vector, dividing the supporting type label in the historical supporting case data set, the supporting type label includes brittle fracture supporting type, shear slip supporting type and creep coordination supporting type.

[0110] For example, when the angle between the maximum gradient direction of the historical stress concentration coefficient distribution map 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 can cause the ice rock mass to be more prone to brittle fracture, so support needs to be provided for this type of fracture. If the angle is between 30 degrees and 60 degrees, it is marked as a shear slip support type, because in this case the interaction of stress and temperature can cause shear slip phenomenon in the ice rock mass. When the angle is greater than 60 degrees, it is marked as a creep coordination support type, meaning that under this stress and temperature condition, the creep phenomenon of the ice rock mass is more significant, and support needs to be provided in coordination with the creep.

[0111] In step S440, the support type label and the corresponding support point coordinate sequence are input into the graph convolution network to generate a topological connection relationship graph for each support case, which includes stress transmission paths between support points and temperature field interference areas.

[0112] In this embodiment, for the stress transmission path, the support points are taken as nodes, and the transmission mode of stress between the nodes is determined according to the mechanical properties of the ice rock mass 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, the stress transmission 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 mass. Due to the existence of the support structure, the heat conduction inside the ice rock mass can be changed, and the temperature fields around different support points will interfere with each other. The graph convolution network accurately depicts these stress transmission paths and temperature field interference areas by analyzing the support point coordinate and support type information.

[0113] In step S450, the priority score of each support point is calculated according to the node degree centrality index of the stress transmission path and the area weight of the temperature field interference area, and the construction time sequence instruction set is reordered according to the priority score.

[0114] In this embodiment, the node degree centrality index of the stress transmission path reflects the importance of each support point in the stress transmission network. For example, if a support point has a direct stress transmission relationship with multiple other support points, the node degree centrality index of that support point will be higher. The area weight of the temperature field interference region takes into account the range of the support point's influence on the temperature field. If the temperature field interference area around a certain support point is larger, it means that it has a greater impact on the temperature field, and the corresponding area weight will also be larger. When calculating the priority score of each support point, both factors are considered. For example, if a support point has a node degree centrality index of 0.6 and a temperature field interference area weight of 0.4, the priority score is 0.5 by using existing calculation methods (such as weighted summation). According to these priority scores, the construction sequence instruction set is reordered. If the original construction sequence instruction is performed 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 a role in the construction process.

[0115] In step S460, the historical crack propagation rate peak value is time-stamped aligned with the support strength parameter group, a dynamic decay curve of the support strength parameter changing with the crack propagation rate is fitted, and a critical support strength threshold value corresponding to the curve slope mutation point is extracted.

[0116] In this embodiment, in each historical support case, there is a corresponding historical crack propagation rate peak value and support strength parameter group, and these data are time-stamped. By matching the data with the same time stamp, time stamp alignment is achieved. Then, taking the historical crack propagation rate peak value as the independent variable and the support strength parameter as the dependent variable, a suitable fitting method (such as the least squares method) is used to fit a dynamic decay curve. For example, when the historical crack propagation rate peak value is 0.05 meters per month, the corresponding support strength parameter is 10,000 Newtons; when the historical crack propagation rate peak value is 0.1 meters per month, the support strength parameter is 12,000 Newtons. By fitting these data points, a curve is obtained, which reflects the trend of the change of the support strength parameter with the increase of the crack propagation rate. The critical support strength threshold value corresponding to the curve slope mutation point is extracted from the fitted dynamic decay curve. In the curve, the slope mutation point represents the turning point of the change 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, and the support strength parameter value at this point is the critical support strength threshold value. If the threshold value is lower, the support structure may not be able to effectively resist the risk of ice rock mass rupture.

[0117] Step S470, the support type label, topological connection relationship diagram, re-sequenced construction time sequence instruction set and critical support strength threshold are stored in association to generate a support scheme library containing a multi-dimensional index key, and the multi-dimensional index key includes a crack propagation rate interval, a stress-temperature gradient angle range and a priority score level.

[0118] For example, for a support case, the support type label is brittle fracture support type, the topological connection relationship diagram shows the specific support point position relationship, the re-sequenced construction time sequence instruction set determines the construction sequence, and the critical support strength threshold is 12000 Newton. These information are stored in association. When constructing the index key, according to the crack propagation rate interval, the cases with the historical crack propagation rate peak value between 0.05 meters and 0.1 meters per month are classified into an interval. The stress-temperature gradient angle range is also classified, such as 0 degrees to 30 degrees as a range. According to the priority score level, 0.4 to 0.6 is divided into a level. In this way, the support scheme library constructed can conveniently retrieve the appropriate support scheme according to the actual situation of the ice rock mass, such as the current crack propagation rate, the stress-temperature gradient angle and the priority of the support point position, and provide effective protection for the safety of underground facilities.

[0119] Figure 1 The structure schematic diagram of the ice rock mass crack monitoring and early warning system 10 based on artificial intelligence provided by the embodiment of the present application includes a processor 102, a memory 104 and a bus 106. The memory 104 is used to store execution instructions, including internal memory and external memory, the internal memory can also be understood as internal storage, used to temporarily store operation data in the processor 102 and data exchanged with the external storage such as a hard disk, the processor 102 exchanges data with the external storage through the internal memory, when the ice rock mass crack monitoring and early warning system 10 based on artificial intelligence runs, the processor 102 and the memory 104 communicate through the bus 106, so that the processor 102 executes the ice rock mass crack monitoring and early warning method based on artificial intelligence of the embodiment of the present application.

[0120] For the convenience of description, only one processor is described in the artificial intelligence-based ice rock mass crack monitoring and early warning system 10. However, it should be noted that the artificial intelligence-based ice rock mass crack monitoring and early warning system 10 in the present application can also include multiple processors, so the steps performed by one processor described in the present application can also be jointly performed or separately performed by multiple processors. For example, if the processor of the artificial intelligence-based ice rock mass crack monitoring and early warning system 10 performs steps A and B, it should be understood that steps A and B can also be jointly performed by two different processors or separately performed in one processor. For example, a first processor performs step A, a second processor performs step B, or a first processor and a second processor jointly perform steps A and B.

[0121] In addition, the embodiment of the present application also provides a readable storage medium, wherein computer executable instructions are preset, and when a processor executes the computer executable instructions, the artificial intelligence-based ice rock mass crack monitoring and early warning method is realized.

[0122] It should be noted that, in order to simplify the description of the present application and to help understand one or more embodiments of the present application, in the foregoing description of the embodiments of the present application, various features are sometimes combined into one embodiment, figure or description thereof.

Claims

1. An artificial intelligence-based ice rock mass crack monitoring and early warning method, characterized in that, The method comprises: acquiring real-time geological wave data through a multi-modal sensor array deployed on the surface of the ice rock mass, the geological wave data including acoustic wave reflection signals, stress distribution signals and temperature change signals; inputting the geological wave data into a pre-trained multi-dimensional feature fusion model to extract a crack feature vector inside the ice rock mass, the crack feature vector including a crack spatial coordinate, a crack propagation rate and dynamic perception parameters of a stress concentration area; calling a time-space evolution prediction model to perform time series analysis on the crack feature vector to generate a probability distribution map of a crack evolution path, the probability distribution map including a crack propagation direction and a critical fracture risk level in a future preset time period; triggering an early warning signal and generating a corresponding emergency support scheme according to a region in the probability distribution map where the critical fracture risk level exceeds a preset threshold, the emergency support scheme including support point coordinate, support strength parameters and construction time sequence instructions; the training method of the time-space evolution prediction model comprises: constructing 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 time-dependent features of crack propagation rate from sample crack feature vectors, and the three-dimensional convolution kernel is used to extract topological features of crack spatial morphology from sample crack feature vectors; inputting the sample crack feature vectors in time series into the hybrid neural network architecture, dynamically fusing the time-dependent features and the topological features through a gating mechanism to generate dynamic fusion features including time-space correlation; inputting the dynamic fusion features into a discriminator network, calculating a spatial matching degree between a predicted crack evolution path of the dynamic fusion features and a real fracture event through the discriminator network, and generating an adversarial training loss according to a difference between the spatial matching degree and a preset threshold, and reversely optimizing convolution kernel parameters of the hybrid neural network architecture; inputting the dynamic fusion features into a physical constraint loss function to calculate an angle penalty term between a maximum principal stress direction and a crack propagation direction in the predicted crack evolution path, and combining the adversarial training loss to generate a joint optimization loss; iteratively updating the convolution kernel parameters of the hybrid neural network architecture and the weight parameters of the discriminator network to make the joint optimization loss converge to a preset error range, and generating a trained time-space evolution prediction model; the inputting the dynamic fusion features into a physical constraint loss function to calculate an angle penalty term between a maximum principal stress direction and a crack propagation direction in the predicted crack evolution path, and combining the adversarial training loss to generate a joint optimization loss, comprises: extracting a maximum principal stress direction angle and a crack surface normal direction angle of a current stress concentration area from the sample crack feature vectors; calculating a first angle between the maximum principal stress direction and the crack surface normal direction, 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 fracture angle, reducing a fracture risk prediction value of the region; increasing a fracture risk prediction value of the region when the first included angle is in a brittle fracture interval and the second included angle satisfies a Coulomb friction condition; calculating a direction consistency loss according to a spatial coincidence degree of the prediction value and an actual fracture event, and performing weighted summation of the direction consistency loss and an adversarial loss of the discriminator network to obtain a final joint optimization loss; the dynamic fusion feature is input into a discriminator network, the spatial matching degree of a crack evolution path predicted by the dynamic fusion feature and a real fracture event is calculated through the discriminator network, and an 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 mixed neural network architecture are reversely optimized, including: the dynamic fusion feature is input into a full connection layer of the discriminator network to generate a spatial probability distribution map of a predicted crack evolution path output by the discriminator network; obtaining a spatial coordinate point set of a real fracture event corresponding to the sample crack feature vector, and converting the spatial coordinate point set into a binary fracture mask map with the same resolution as the spatial probability distribution map; calculating the cross-entropy loss of 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 spatial unit set with a probability value exceeding a preset probability threshold in the spatial probability distribution map, and calculating the spatial intersection-over-union ratio of the spatial unit set and the fracture region in the binary fracture mask map to generate an intersection-over-union ratio penalty term; linearly combining the initial adversarial loss component and the intersection-over-union ratio penalty term according to a preset weight coefficient to generate a final adversarial training loss; the adversarial training loss is differentiated with respect to the three-dimensional convolution kernel parameters of the mixed 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, so that the spatial matching degree of the spatial probability distribution map and the binary fracture mask map is maximized.

2. The ice rock mass crack monitoring and early warning method according to claim 1, characterized in that, The pre-training process of the multi-dimensional feature fusion model includes: obtaining a sample geological fluctuation data set and corresponding sample crack annotation data, the sample geological fluctuation data including sample acoustic reflection signals, sample stress distribution signals and sample temperature change signals, and the sample crack annotation data including sample crack spatial coordinates, sample crack propagation rates and dynamic perception parameters of sample stress concentration regions; performing time-frequency transformation on the sample geological fluctuation data to generate a joint feature matrix including sample acoustic spectrum features, sample stress gradient features and sample temperature change rates; using an initialized deep cross-attention network, taking the sample acoustic spectrum features as query vectors, the sample stress gradient features as key vectors and the sample temperature change rates as value vectors, and calculating sample weight distribution coefficients of each modal data in the joint feature matrix through a multi-head attention mechanism; The joint feature matrix is dynamically weighted and fused based on the sample weight distribution coefficient, and a fused sample crack feature vector is output, and parameters of the deep cross attention network are optimized through a contrast loss function, so that a spatial distribution error between the sample crack feature vector and sample crack labeled data is minimized.

3. The ice rock mass crack monitoring and early warning method according to claim 2, characterized in that, The sample geological fluctuation data is subjected to time-frequency transformation to generate a joint feature matrix containing sample acoustic spectrum features, sample stress gradient features and sample temperature change rates, including: The sample acoustic reflection signal is subjected to short-time Fourier transform to generate a time-frequency spectrum, and an energy distribution peak value of a preset frequency band in the time-frequency spectrum is extracted as the sample acoustic spectrum feature; The sample stress distribution signal is subjected to spatial gradient calculation to obtain stress gradient amplitudes and direction angles of each sampling point, and a stress gradient amplitude difference between adjacent sampling points is taken as the sample stress gradient feature; The sample temperature change signal is subjected to time window sliding average processing to calculate a standard deviation of a temperature change amount in each time window as the sample temperature change rate; The sample acoustic spectrum feature, sample stress gradient feature and sample temperature change rate are aligned according to timestamps, and are spliced into the joint feature matrix in the form of a three-dimensional tensor after normalization processing.

4. The ice rock mass crack monitoring and early warning method according to claim 2, characterized in that, The parameters of the deep cross attention network are optimized through the contrast loss function, so that the spatial distribution error between the sample crack feature vector and the sample crack labeled data is minimized, including: The fused sample crack feature vector and the sample crack labeled data are divided according to a spatial grid to generate corresponding sample crack feature matrices and sample labeled matrices; According to crack spatial coordinates in the sample labeled matrix, positive sample pairs and negative sample pairs are distributed in the sample crack feature matrix, the positive sample pairs are grid cells with a spatial distance less than a preset threshold in the same sample crack feature matrix, and the negative sample pairs are grid cells with a spatial distance greater than the preset threshold in different sample crack feature matrices; Cosine similarity of grid cells in the positive sample pairs is calculated, and Euclidean distance differences of grid cells in the negative sample pairs are extracted; A contrast loss function is constructed according to the cosine similarity and the Euclidean distance differences, wherein a similarity maximization constraint term of the positive sample pairs and a difference minimization constraint term of the negative sample pairs are weighted and summed; A gradient descent algorithm is used to back-propagate partial derivatives of the contrast loss function to update attention weight coefficients and fully connected layer parameters in the deep cross attention network until a spatial distribution error between the sample crack feature matrix and the sample labeled matrix converges to a preset tolerance range.

5. The ice rock mass crack monitoring and early warning method according to claim 1, characterized in that, According to the region in the probability distribution map where the critical rupture risk level exceeds the preset threshold, a warning signal is triggered and a corresponding emergency support scheme is generated, including: All spatial cells in the probability distribution map are traversed to screen out a spatial cell set where the critical rupture risk level exceeds the preset threshold, which is marked as an initial dangerous region; Performing spatial connectivity analysis on the initial dangerous area, merging adjacent spatial units with a spatial distance less than a preset safety radius to generate a continuously distributed joint dangerous area; Extracting geological structure features of the joint dangerous area, including the maximum crack propagation rate, the average stress concentration coefficient, and the temperature gradient distribution pattern; Performing similarity matching between the geological structure features and the support case features in the pre-stored support scheme library, and screening a support case set with a matching degree exceeding a similarity threshold; Generating an initial support scheme template according to the support intensity parameters and the construction timing instructions of the support case set; Obtaining the real-time monitored rock mass creep rate and environmental temperature change, calculating the dynamic attenuation coefficient of the support intensity parameters in the initial support scheme template and the time offset of the construction timing instructions; Adjusting the support intensity parameters of the initial support scheme template according to the dynamic attenuation coefficient, and correcting the priority order of the construction timing instructions based on the time offset, to generate an emergency support scheme containing updated support point coordinates, optimized support intensity parameters, and corrected construction timing instructions; Topologically mapping the emergency support scheme to the spatial coordinates of the joint dangerous area, verifying the spatial coverage consistency of the support point coordinates and the crack propagation direction, and if there is an uncovered area, re-iterating the matching of the support case set until all joint dangerous areas are covered by the support scheme.

6. The ice rock mass crack monitoring and early warning method according to claim 5, characterized in that, The construction method of the support scheme library includes: Obtaining a historical support case data set, which contains support area geological structure features, support point coordinate sequences, support intensity parameter groups, and construction timing instruction sets; Extracting the historical crack propagation rate peak value, the historical stress concentration coefficient distribution map, and the historical temperature gradient direction vector from the support area geological structure features; According to the angle between the maximum gradient direction of the historical stress concentration coefficient distribution map and the historical temperature gradient direction vector, dividing the support type labels in the historical support case data set, including brittle fracture support type, shear slip support type, and creep coordination support type; Inputting the support type labels and corresponding support point coordinate sequences into a graph convolution network to generate a topological connection relationship graph for each support case, which contains stress transmission paths between support points and temperature field interference areas; According to the node degree centrality index of the stress transmission path and the area weight of the temperature field interference area, calculating the priority score of each support point, and reordering the construction timing instruction set according to the priority score; Timestamp aligning the historical crack propagation rate peak value with the support intensity parameter group, fitting the dynamic attenuation curve of the support intensity parameter with the crack propagation rate, and extracting the critical support intensity threshold corresponding to the curve slope mutation point; The support type label, the topological connection relationship diagram, the reordered construction time sequence instruction set and the critical support strength threshold are stored in association to generate a support scheme library containing a multi-dimensional index key, and the multi-dimensional index key includes a crack propagation rate interval, a stress-temperature gradient angle range and a priority score level.

7. An artificial intelligence-based ice rock mass crack monitoring and early warning system, characterized in that, The ice rock mass crack monitoring and early warning system based on artificial intelligence comprises a processor and a memory, the memory is connected with the processor, the memory is used for storing programs, instructions or codes, and the processor is used for executing the programs, instructions or codes in the memory to realize the ice rock mass crack monitoring and early warning method based on artificial intelligence in any one of claims 1-6.

Images