Ice rock mass stability prediction method and system based on machine learning

Abstract

The invention provides an ice rock mass stability prediction method and system based on machine learning, and relates to the technical field of machine learning, and the method comprises the steps: firstly obtaining a geological data set containing ice rock mass distribution, geological structure and environment monitoring parameters in a target region, and carrying out the feature fusion processing of the geological data set; the method comprises the steps of obtaining an ice rock feature set covering an ice rock mass structure, geological stress and environmental dynamic features, constructing and training an ice rock stability prediction model by using a multi-layer neural network based on the ice rock feature set, and calling the trained ice rock stability prediction model to generate a real-time stability prediction result of a target area when monitoring data is input in real time. And finally, according to the real-time stability prediction result, outputting a stability alarm signal or optimizing and reinforcing suggestion parameters, so that various parameters can be integrated, accurate prediction of the stability of the ice rock mass is realized, and countermeasures are given.

Description

Technical Field

[0001] The present invention relates to the technical field of machine learning, and in particular, to a method and system for predicting the stability of ice-rock masses based on machine learning. Background Art

[0002] In the field of geotechnical engineering, the prediction of rock mass stability has always been a key link to ensure the safety and smooth progress of projects. Traditional methods for predicting rock mass stability mainly rely on empirical formulas and the subjective judgment of on-site engineers. These methods usually based on limited geological parameters and simple mechanical models, and it is difficult to comprehensively and accurately reflect the complex geological conditions and mechanical properties of rock masses.

[0003] In the early stage, the evaluation of rock mass stability mostly adopted qualitative analysis. For example, by on-site observation of basic characteristics such as the lithology of the rock mass and the degree of joint fissure development by geological surveyors, and making a rough grading judgment on the rock mass stability based on experience. Although this method is simple to operate, it is greatly affected by human factors, and there may be significant differences in the judgment results of different personnel, lacking objectivity and accuracy.

[0004] With the development of related technologies, some methods based on quantitative analysis have gradually emerged, such as the limit equilibrium method. It usually calculates the safety factor when the rock mass reaches the limit equilibrium state by establishing the mechanical equilibrium equation of the rock mass and considering factors such as the self-weight of the rock mass, external forces, and the shear strength inside the rock mass. However, the limit equilibrium method often simplifies the rock mass into an ideal rigid body or a simple structure, ignoring the complex stress-strain relationship and the diversity of geological structures inside the rock mass, resulting in a certain deviation between the prediction result and the actual situation.

[0005] In addition, numerical simulation technologies have been widely used in the prediction of rock mass stability, such as the finite element method and the discrete element method. These methods can simulate the mechanical behavior and deformation process of rock masses in more detail, considering the non-linear characteristics and complex boundary conditions of rock masses. However, the numerical simulation method is highly dependent on geological parameters, and obtaining accurate geological parameters often requires a large number of on-site tests and laboratory tests, with high costs and long time consumption. At the same time, the establishment and solution process of the numerical simulation model are relatively complex, requiring professional technical personnel and high-performance computing equipment support, which limits its wide application in actual projects. Summary of the Invention

[0006] In view of the above-mentioned problems, in combination with the first aspect of the present invention, embodiments of the present invention provide a method for predicting the stability of ice-rock masses based on machine learning, and the method includes: Obtain a set of geological data of the target area, where the set of geological data includes ice-rock mass distribution parameters, geological structure parameters, and environmental monitoring parameters; Perform feature fusion processing on the geological data set to obtain an ice-rock feature set, where the ice-rock feature set includes ice-rock body structure features, geological stress features, and environmental dynamic features; Train an ice-rock stability prediction model based on the ice-rock feature set, where the ice-rock stability prediction model is constructed by a multi-layer neural network; In response to the real-time input monitoring data, call the trained ice-rock stability prediction model to generate a real-time stability prediction result for the target area; Output a stability warning signal or optimized reinforcement suggestion parameters according to the real-time stability prediction result.

[0007] On the other hand, an embodiment of the present invention further provides an ice-rock body stability prediction system based on machine learning, including a processor and a machine-readable storage medium. The machine-readable storage medium is connected to the processor. The machine-readable storage medium is used to store programs, instructions, or codes, and the processor is used to execute the programs, instructions, or codes in the machine-readable storage medium to implement the above method.

[0008] Based on the above aspects, the embodiments of the present application achieve full-cycle dynamic monitoring and intelligent decision support for the stability evolution of ice-rock bodies. Specifically, by collecting a geological data set of the target area, covering ice-rock body distribution parameters, geological structure parameters, and environmental monitoring parameters, a basic data layer with multi-dimensional information fusion is constructed. On this basis, feature fusion processing technology is used to transform the original data into an ice-rock feature set including ice-rock body structure features, geological stress features, and environmental dynamic features, which not only effectively excavates the potential correlations between data, but also improves the data quality through feature dimensionality reduction and reconstruction, providing high-signal-to-noise ratio feature inputs for subsequent model training. Further, the ice-rock stability prediction model constructed based on the multi-layer neural network architecture captures the complex coupling relationship between the ice-rock body stability and multiple factors through deep feature extraction and non-linear mapping mechanisms, significantly improving the generalization ability and robustness of the ice-rock stability prediction model. The dynamic prediction mechanism of the ice-rock stability prediction model driven by real-time data can quickly respond to the stability state of the target area, and intelligently output a stability warning signal or optimized reinforcement suggestion parameters based on the prediction result, thus realizing the paradigm shift from passive monitoring to active prevention and control. It not only provides a scientific and reliable quantitative evaluation tool for the safety of ice-rock body engineering, but also promotes the leapfrog development of ice-rock body stability prediction from experience-driven to data-intelligence-driven through a closed-loop feedback mechanism. Description of the Drawings

[0009] Figure 1 It is a schematic execution flowchart of the ice-rock body stability prediction method provided by the embodiment of the present invention.

[0010] Figure 2It is a schematic diagram of exemplary hardware and software components of the ice-rock mass stability prediction system based on machine learning provided by an embodiment of the present invention. Detailed implementation manners

[0011] The present invention will be specifically described below with reference to the accompanying drawings of the specification. Figure 1 It is a schematic flowchart of the ice-rock mass stability prediction method based on machine learning provided by an embodiment of the present invention. The ice-rock mass stability prediction method based on machine learning will be introduced in detail below.

[0012] Step S110: Obtain a set of geological data for the target area, where the set of geological data includes ice-rock mass distribution parameters, geological structure parameters, and environmental monitoring parameters.

[0013] In this embodiment, taking the glacier area of a high-altitude mountainous area as an example of the target area, in the glacier area of this high-altitude mountainous area, the ice-rock mass distribution parameters help to understand the ice-rock structure of this glacier area. Specifically, the ice layer thickness distribution data can be obtained by using an ice layer detection radar. For example, measurement points are set at different positions of the glacier, and measurements are taken at regular intervals (such as 50 meters) from the edge to the center of the glacier. It is monitored that the ice layer thickness near the edge may be about 10 meters, while the thickness in some areas at the center of the glacier can reach 50 meters. In addition, the rock porosity data can rely on the collection and laboratory analysis of rock samples under the glacier. Assuming that multiple rock samples are collected in different sections of the glacier, and after analysis, it is found that the rock porosity in some sections is 10%, which means that there is a certain void space inside the rock. For another example, the crack density data can be obtained through on-site exploration of the glacier surface and surrounding rocks, using high-definition image shooting technology and computer image analysis algorithms to calculate the number of visible cracks per square meter. For example, 5 obvious cracks are found per square meter in a specific area.

[0014] In terms of geological structure parameters, the formation dip angle data can be obtained through long-term monitoring by inclinometers installed on the rocks around the glacier. For example, after several months of measurement, it is found that the formation dip angle in a certain direction is 30 degrees. The fault strike data relies on geological mapping techniques, including ground survey and aerial photogrammetry, to determine that the fault in this area is roughly in the northeast-southwest direction. The rock compressive strength data is obtained by conducting compressive strength tests on rock samples and simulating the bearing capacity of rocks under different pressure environments in the laboratory. The test results show that the average rock compressive strength in this area is 200 MPa.

[0015] The acquisition of environmental monitoring parameters can also be carried out in various ways. For example, the temperature change rate data is recorded by thermometers installed at different positions on the glacier. The temperature is recorded once an hour. Through observations over a period of time, it is found that the temperature change rate may be relatively fast at sunrise in the morning within a day. During the temperature increase process from -10°C to -5°C, the temperature can rise by 1°C per hour. The humidity fluctuation data can be measured with the help of humidity sensors. These humidity sensors are distributed at different heights around and inside the glacier. It is monitored that the humidity fluctuates greatly under different weather conditions. For example, the humidity may be 30% on a sunny day, while it will rise to 80% during snowfall weather. The surface vibration frequency data can be monitored using seismographs. For example, multiple seismic monitoring points can be set around the glacier. When there is a minor earthquake or ice movement inside the glacier, the surface vibration frequency can be detected. For example, during a small-scale ice movement event, the detected surface vibration frequency is 1 hertz.

[0016] Step S120: Perform feature fusion processing on the geological data set to obtain an ice-rock feature set, where the ice-rock feature set includes ice-rock mass structure features, geological stress features, and environmental dynamic features.

[0017] Specifically, for the spatial gradient feature in the ice layer thickness distribution data to be aligned with the direction of the formation dip data to generate a first fusion feature. Assume that the change trend (spatial gradient feature) of the ice layer thickness in a certain direction is gradually thinning, and the formation dip direction is inclined towards a certain valley direction. Then these two features can be directionally aligned. For example, taking the formation dip direction as the reference, analyze the comprehensive impact of the change in ice layer thickness along this direction on the ice-rock structure, so as to obtain the first fusion feature.

[0018] In addition, the weighted superposition of the rock porosity data and the fracture density data generates a second fusion feature. For example, based on experience and analysis, a weight of 0.4 is assigned to the rock porosity data, and a weight of 0.6 is assigned to the fracture density data. If the rock porosity is 10% and the fracture density is 5 fractures per square meter, then the second fusion feature after weighted superposition comprehensively reflects the degree of vulnerability of the internal structure of the rock.

[0019] Furthermore, the time-series correlation analysis of the temperature change rate data and the surface vibration frequency data generates a third fusion feature. For example, during a period of rapid temperature rise, the ice inside the glacier may melt and expand, which may affect the surface vibration frequency. Through the analysis of historical data, it is found that when the temperature change rate reaches a rising speed of 1°C per hour, the surface vibration frequency will increase from 0.5 hertz to 1 hertz within the next few hours. This correlation relationship in the time series forms the third fusion feature.

[0020] Thus, the first fusion feature, the second fusion feature, and the third fusion feature can be concatenated across dimensions to generate the ice-rock mass structure feature, that is, information from different aspects is combined according to existing rules to form a feature that comprehensively describes the ice-rock mass structure.

[0021] The dynamic attenuation feature of the rock layer compressive strength data and the non-linear mapping of the humidity fluctuation data generate the geological stress feature. For example, with the fluctuation of humidity, the rock layer compressive strength may change. When the humidity rises from 30% to 80%, the dynamic attenuation feature of the rock layer compressive strength shows that the compressive strength may drop from 200 MPa to 150 MPa. By establishing this non-linear mapping relationship, the influence of humidity on geological stress can be better reflected.

[0022] Model the topological relationship between the fault strike data and the formation dip data to generate the fourth fusion feature. Assume that the fault strike is northeast-southwest and the formation dip is 30 degrees. Through topological relationship modeling, the spatial position relationship and interaction relationship between the fault and the formation can be analyzed. Then, jointly encode the fourth fusion feature and the third fusion feature to generate the environmental dynamic feature, which comprehensively reflects the mutual relationship between environmental factors and the dynamic impact on the ice-rock structure.

[0023] Step S130, train an ice-rock stability prediction model based on the ice-rock feature set, and the ice-rock stability prediction model is constructed by a multi-layer neural network.

[0024] Specifically, a multi-task learning framework including an ice-rock mass structure feature input layer, a geological stress feature input layer, and an environmental dynamic feature input layer can be constructed. In this multi-task learning framework, the stability classification task is to judge the stability level of the ice-rock (such as stable, relatively stable, unstable), and the stress change regression task is to predict the change value of the internal stress of the ice-rock.

[0025] Input the ice-rock mass structure feature into the ice-rock mass structure feature input layer. Assume that the ice-rock mass structure feature contains various structure information after previous fusion processing. After being processed by the input layer, a first hidden feature vector is generated. This first hidden feature vector is an abstract representation of the ice-rock mass structure feature and may contain key information about the ice-rock mass structure, such as the comprehensive relationship between the ice layer thickness and the rock structure.

[0026] Input the geological stress feature into the geological stress feature input layer, and extract the local stress gradient through a convolution kernel. For example, the geological stress feature contains the stress change conditions in different regions. The convolution kernel can focus on the local area with large stress changes, such as the stress concentration area near the fault, so as to generate a second hidden feature vector, which can highlight the local feature of the geological stress.

[0027] The environmental dynamic features are input into the environmental dynamic feature input layer, and the time series network is used to extract the dynamic change trend. For example, the rate of temperature change, humidity fluctuation, etc. in the environmental dynamic features change over time. The time series network can capture the dynamic change trends of these environmental factors and generate the third hidden feature vector.

[0028] The first hidden feature vector, the second hidden feature vector, and the third hidden feature vector are subjected to cross-attention fusion to generate a joint feature vector, thereby integrating the key information in three different aspects according to the correlation relationship between them. For example, the ice-rock mass structure information in the first hidden feature vector is correlated with the geological stress information in the second hidden feature vector and the environmental dynamic information in the third hidden feature vector to generate a joint feature vector that comprehensively reflects the overall state of the ice-rock.

[0029] The joint feature vector is respectively input into the fully connected layer for the stability classification task and the fully connected layer for the stress change regression task to generate the stability probability distribution parameters and the stress change prediction value. Suppose that after the calculation of the fully connected layer, the stability probability distribution parameters show that the probability of the ice-rock being stable is 60%, the probability of being relatively stable is 30%, and the probability of being unstable is 10%; the stress change prediction value is that the stress may increase by 5 MPa in a future period of time.

[0030] Based on the cross-entropy loss between the stability probability distribution parameters and the labeled stability labels, and the mean square error loss between the stress change prediction value and the measured stress data, the parameters of the multi-task learning framework are jointly optimized. For example, if the difference between the predicted stability probability and the actual labeled stability label is large, or the error between the stress change prediction value and the measured stress data is large, the weights and other parameters in the multi-task learning framework are adjusted. After multiple iterations of optimization, the trained ice-rock stability prediction model is obtained.

[0031] Step S140, in response to the real-time input monitoring data, call the trained ice-rock stability prediction model to generate the real-time stability prediction result of the target area.

[0032] Specifically, the dynamic data of the ice layer thickness, the instantaneous temperature change data, and the surface vibration amplitude data collected in real time can be obtained. For example, at a certain moment, the dynamic data of the ice layer thickness shows that the ice layer thickness in a certain area has decreased by 1 meter in a short period of time (such as within 1 hour), the instantaneous temperature change data indicates that the temperature has suddenly dropped by 5 °C, and the surface vibration amplitude data shows that the surface vibration amplitude has increased from 0.5 Hz to 1.5 Hz.

[0033] Spatial interpolation is performed on the dynamic data of ice layer thickness to generate updated parameters for real-time ice-rock mass distribution. Since the measurement points of ice layer thickness are limited, spatial interpolation can be used to obtain an update of the ice layer thickness distribution across the entire target area. For example, the possible values of ice layer thickness can be calculated between the measurement points according to a certain algorithm, so as to obtain more comprehensive updated parameters for real-time ice-rock mass distribution.

[0034] Differential calculation is carried out on the instantaneous temperature change data and historical temperature data to generate temperature change trend parameters. If the historical temperature data shows that the temperature usually drops slowly in this season, but now it suddenly drops by 5°C, then this temperature change trend parameter indicates that the current temperature change is abnormal.

[0035] The ground vibration amplitude data is compared with a preset vibration threshold to generate a vibration anomaly area marking parameter. Suppose the preset vibration threshold is 1 Hz, and now the ground vibration amplitude reaches 1.5 Hz, then this area is marked as a vibration anomaly area.

[0036] The updated parameters for real-time ice-rock mass distribution, temperature change trend parameters, and vibration anomaly area marking parameters are input into the input layer of the ice-rock stability prediction model. Through the real-time prediction module of the multi-task learning framework, real-time stability level parameters and real-time stress change prediction parameters are output as the real-time stability prediction results. For example, the output real-time stability level parameter is relatively unstable, and the real-time stress change prediction parameter is that the stress may increase by 10 MPa within the next few hours.

[0037] Step S150, output a stability warning signal or optimized reinforcement suggestion parameters according to the real-time stability prediction result.

[0038] Specifically, a stability warning signal can be triggered according to the comparison result between the real-time stability level parameter and the preset safety threshold. If the preset safety threshold is stable and relatively stable, and now the real-time stability level parameter is relatively unstable, then a stability warning signal is triggered to notify relevant personnel that there is a risk of ice-rock stability in this area.

[0039] Based on the deviation value between the real-time stress change prediction parameter and the historical stress data, optimized reinforcement suggestion parameters including the coordinates of the reinforcement location and material parameters are generated. Suppose the historical stress data shows that the stress change is very small under normal circumstances, and now the real-time stress change prediction parameter indicates that the stress will increase by 10 MPa. Then, according to the analysis, the area where the stress increase is relatively large (coordinates of the reinforcement location) may be determined, such as near a fault or in an area where the ice layer thickness changes significantly. At the same time, the material parameters required for reinforcement are determined according to the characteristics of the ice-rock and the amplitude of stress increase, such as the need to use high-strength anchor bolts or increase the protective materials on the ice layer surface.

[0040] Based on the above steps, the embodiments of the present application achieve full-cycle dynamic monitoring and intelligent decision-making support for the stability evolution of ice-rock masses. Specifically, by collecting the geological data set of the target area, covering ice-rock mass distribution parameters, geological structure parameters, and environmental monitoring parameters, a basic data layer with multi-dimensional information fusion is constructed. On this basis, the feature fusion processing technology is used to transform the original data into an ice-rock feature set including ice-rock mass structure features, geological stress features, and environmental dynamic features, effectively mining the potential correlations between data and improving the data quality through feature dimensionality reduction and reconstruction, providing feature inputs with high signal-to-noise ratio for subsequent model training. Further, an ice-rock stability prediction model constructed based on a multi-layer neural network architecture captures the complex coupling relationship between ice-rock mass stability and multiple factors through deep feature extraction and non-linear mapping mechanisms, significantly improving the generalization ability and robustness of the ice-rock stability prediction model. The dynamic prediction mechanism of the ice-rock stability prediction model driven by real-time data can quickly respond to the stability state of the target area and intelligently output stability warning signals or optimized reinforcement suggestion parameters based on the prediction results, thus realizing the paradigm shift from passive monitoring to active prevention and control. It not only provides a scientific and reliable quantitative evaluation tool for the safety of ice-rock mass engineering, but also promotes the leapfrog development of ice-rock mass stability prediction from experience-driven to data-intelligence-driven through a closed-loop feedback mechanism.

[0041] In a possible implementation manner, the ice-rock mass distribution parameters in the geological data set include ice layer thickness distribution data, rock porosity data, and fracture density data, the geological structure parameters include formation dip angle data, fault strike data, and rock layer compressive strength data, and the environmental monitoring parameters include temperature change rate data, humidity fluctuation data, and surface vibration frequency data. Step S120 includes: Step S121, extracting the spatial gradient feature in the ice layer thickness distribution data, and aligning the spatial gradient feature with the formation dip angle data to generate a first fusion feature.

[0042] In this embodiment, the ice layer thickness distribution data is obtained by the ice layer detection radar mentioned above at different positions of the glacier. For example, on a measurement line from the edge to the center of the glacier, the ice layer thickness is 20 meters at 100 meters from the glacier edge, 18 meters at 200 meters, and 15 meters at 300 meters. Thus, the spatial gradient characteristics of the ice layer thickness in this direction can be calculated. Assuming that the distance is the independent variable and the ice layer thickness is the dependent variable, calculate the ratio of the change in ice layer thickness between adjacent two points to the change in distance to obtain the spatial gradient characteristics. For example, from 100 meters to 200 meters, the thickness change is -2 meters and the distance change is 100 meters, so the spatial gradient is -2 / 100 = -0.02 m / m. The formation dip data is 30 degrees. Align the spatial gradient characteristics with the formation dip data in terms of direction through existing algorithms. For example, taking the formation dip direction as the positive x-axis direction, analyze the comprehensive influence of the change in ice layer thickness along the x-axis direction on the ice-rock structure. If the spatial gradient is negative, it means the ice layer thins along the formation dip direction, and this combined relationship between the thinning and the formation dip forms the first fusion feature.

[0043] Step S122: Perform weighted superposition on the rock porosity data and the fracture density data to generate a second fusion feature.

[0044] In this embodiment, the rock porosity data is obtained by collecting and analyzing rock samples under the glacier. Assume that the rock porosity in a certain area is 15%. The fracture density data is obtained through on-site exploration and image analysis, and the number of fractures per square meter in this area is 8. According to the geological characteristics and empirical analysis of this area, assign a weight of 0.3 to the rock porosity data and a weight of 0.7 to the fracture density data. When calculating the weighted superposition value, first convert the rock porosity into decimal form, that is, 0.15, and then calculate the weighted value. The weighted value of the rock porosity is 0.15×0.3 = 0.045, and the weighted value of the fracture density is 8×0.7 = 5.6. Add the two to get the value of the second fusion feature as 0.045 + 5.6 = 5.645. This value comprehensively reflects the degree of vulnerability of the internal structure of the rock, including the combined influence of two factors: porosity and fracture density.

[0045] Step S123: Perform time series correlation analysis on the temperature change rate data and the surface vibration frequency data to generate a third fusion feature.

[0046] In this embodiment, the temperature change rate data is recorded by a high-precision thermometer once per hour, and the surface vibration frequency data is monitored by a seismograph. For example, within a certain time period, from 6:00 am to 12:00 noon, the temperature rises from -10°C to -5°C over a total of 6 hours, with a temperature change of 5°C. Then the temperature change rate is 5 / 6 ≈ 0.83°C per hour. At the same time, within this period, the surface vibration frequency rises from 0.2 Hz to 0.8 Hz. By analyzing historical data, it is found that for every 1°C increase in temperature, the surface vibration frequency on average rises by 0.1 Hz. Within this time period, since the temperature has risen by 5°C, according to historical patterns, the surface vibration frequency should have risen by 0.5 Hz, but actually it has risen by 0.6 Hz, exceeding the expected value by 0.1 Hz. This correlation between the temperature change rate and the surface vibration frequency in the time series, including information such as the deviation between the actual change and the expected change, forms the third fusion feature.

[0047] Step S124: Perform cross-dimensional splicing on the first fusion feature, the second fusion feature, and the third fusion feature to generate the ice-rock mass structure feature.

[0048] In this embodiment, the first fusion feature includes the comprehensive relationship between the change in ice layer thickness along the formation dip direction and the formation dip, the second fusion feature reflects the degree of vulnerability of the internal structure of the rock, and the third fusion feature embodies the temporal correlation between the temperature change rate and the surface vibration frequency. Cross-dimensional splicing is to combine these three features with different dimensions and meanings together according to certain rules to form a feature that comprehensively describes the ice-rock mass structure. This ice-rock mass structure feature synthesizes multi-faceted information including the ice layer structure, the rock structure, and the influence of temperature and surface vibration on the ice-rock structure.

[0049] Step S125: Extract the dynamic decay feature of the rock layer compressive strength data, and perform non-linear mapping on the dynamic decay feature and the humidity fluctuation data to generate the geological stress feature.

[0050] In this embodiment, the rock formation compressive strength data is obtained through the compressive strength test of rock samples, and the average compressive strength is 200 MPa. Through long-term monitoring and analysis, it is found that when the humidity fluctuates, the rock formation compressive strength will undergo dynamic attenuation. For example, when the humidity rises from 30% to 80%, after multiple measurements and analyses, it is found that the compressive strength will drop from 200 MPa to 160 MPa. Calculate the attenuation amount of the compressive strength for every 1% change in humidity. The humidity has changed by 50%, and the compressive strength has attenuated by 40 MPa. Then, the attenuation amount of the compressive strength caused by every 1% change in humidity on average is 40 / 50 = 0.8 MPa / %. Map this dynamic attenuation characteristic to the humidity fluctuation data non-linearly. Assume an empirical-based non-linear function is adopted. When the humidity is at a certain value, calculate the corresponding adjusted value of the compressive strength attenuation according to this function, and then combine it with the original compressive strength data to obtain the geological stress characteristic. This geological stress characteristic comprehensively considers the influence of humidity fluctuation on the rock formation compressive strength and reflects the dynamic change of geological stress when the humidity changes.

[0051] Step S126: Model the topological relationship between the fault strike data and the formation dip data to generate a fourth fusion feature, and jointly encode the fourth fusion feature and the third fusion feature to generate the environmental dynamic feature.

[0052] In this embodiment, the fault strike data is in the northeast-southwest direction, and the formation dip data is 30 degrees. Through topological relationship modeling, analyze the spatial position relationship and interaction relationship between the fault and the formation. For example, the included angle between the fault and the formation, the up-down position relationship of the fault relative to the formation, etc. will all affect the stability of the ice rock. Thus, a fourth fusion feature can be constructed based on the above relationships. Then, jointly encode the fourth fusion feature and the third fusion feature. During the joint encoding process, consider the topological relationship between the fault and the formation in the fourth fusion feature and the temporal correlation relationship between the temperature change rate and the surface vibration frequency in the third fusion feature. For example, when the included angle between the fault and the formation is small, the influence of temperature change on the surface vibration frequency may be transmitted faster through the fault. Therefore, the encoding process that comprehensively considers the two relationships can generate the environmental dynamic feature, which comprehensively reflects the mutual relationship between various environmental factors such as faults, formations, temperature, and surface vibration and their dynamic impact on the ice rock structure.

[0053] In a possible implementation manner, step S130 includes: Step S131: Construct a multi-task learning framework including an ice rock mass structure feature input layer, a geological stress feature input layer, and an environmental dynamic feature input layer, where the multi-task learning framework includes a stability classification task and a stress change regression task.

[0054] In this embodiment, in the glacier area of the high-altitude mountainous area described above, the stability classification task aims to determine different stability level categories of ice rock, such as stable, relatively stable, or unstable, under the current conditions, and the stress change regression task is to predict the specific change value of the internal stress of the ice rock within a certain period in the future. The ice rock mass structure feature input layer specifically receives the ice rock mass structure features generated previously, which include comprehensive information such as the relationship between ice layer thickness and formation dip angle, the degree of internal rock structure vulnerability, and the influence of temperature and ground vibration on the ice rock structure; the geological stress feature input layer receives geological stress features, which reflect the dynamic changes of geological stress under the influence of humidity fluctuation on the compressive strength of rock layers; the environmental dynamic feature input layer is responsible for receiving environmental dynamic features, which comprehensively consider the mutual relationship between various environmental factors such as faults, formations, temperature, and ground vibration and their dynamic influence on the ice rock structure.

[0055] Step S132: Input the ice rock mass structure features into the ice rock mass structure feature input layer to generate a first hidden feature vector.

[0056] In this embodiment, the ice rock mass structure feature input layer processes the input ice rock mass structure features. For example, different information components in the ice rock mass structure features undergo operations such as weighted calculation by the neurons in this layer. Assume that when the part of the relationship between ice layer thickness and formation dip angle in the ice rock mass structure features passes through some neurons in this layer, according to the pre-set weights, this part of the information is weighted and summed, and similar operations are also performed on the part of the information such as the degree of internal rock structure vulnerability and the influence of temperature and ground vibration on the ice rock structure. After a series of the above calculation processes, a first hidden feature vector is finally generated. This first hidden feature vector is the result of abstract representation of the ice rock mass structure features and contains key information in the ice rock mass structure features that has an important impact on the stability of ice rock.

[0057] Step S133: Input the geological stress features into the geological stress feature input layer, and extract the local stress gradient through a convolution kernel to generate a second hidden feature vector.

[0058] In this embodiment, the geological stress characteristics include the dynamic changes of geological stress under the influence of humidity fluctuations. The convolutional kernel scans the geological stress characteristics in the input layer of geological stress characteristics. For example, the size of the convolutional kernel is 3×3, and it slides row by row and column by column from the upper left corner on the matrix representation of the geological stress characteristics (assuming the geological stress characteristics are converted into matrix form for processing). When the convolutional kernel is at a certain position, it performs multiplication operations and summations with the geological stress characteristic data corresponding to that position, thereby extracting local stress characteristic information. Suppose in the area near the fault, due to the large change in geological stress, the local stress gradient extracted by the convolutional kernel here is significantly higher than that in other areas. By performing such convolutional operations on the entire geological stress characteristics, all the extracted local stress characteristic information is combined, and after subsequent calculation and processing, a second hidden feature vector is generated. This second hidden feature vector highlights the local stress change information in the geological stress characteristics, such as the stress change situation in the stress concentration area near the fault.

[0059] Step S134: Input the environmental dynamic characteristics into the input layer of the environmental dynamic characteristics, and extract the dynamic change trend through a time series network to generate a third hidden feature vector.

[0060] In this embodiment, the environmental dynamic characteristics synthesize the mutual relationships among various environmental factors and their dynamic impacts on the ice-rock structure. The time series network processes the data in the environmental dynamic characteristics. For example, for the data of the temperature change rate and the surface vibration frequency over time in the environmental dynamic characteristics. Suppose there is temperature change rate data over a period of time. The time series network analyzes its dynamic change trend based on the temperature change rates at multiple past time points. For example, the temperature change rates within 10 consecutive hours are 0.5℃ / hour, 0.6℃ / hour, 0.4℃ / hour, 0.7℃ / hour, 0.3℃ / hour, 0.8℃ / hour, 0.2℃ / hour, 0.9℃ / hour, 0.1℃ / hour, 0.5℃ / hour. The time series network will calculate the change amount between adjacent time points. For example, the change amount of the temperature change rate from the first time point to the second time point is 0.6 - 0.5 = 0.1℃ / hour², and then comprehensively determines the overall dynamic change trend based on the change amounts of the above adjacent time points. Similar operations are also performed on the surface vibration frequency data. The dynamic change trend information of the temperature change rate and the surface vibration frequency and other dynamic information related to environmental factors are comprehensively processed, and finally a third hidden feature vector is generated.

[0061] Step S135: Perform cross-attention fusion on the first hidden feature vector, the second hidden feature vector, and the third hidden feature vector to generate a joint feature vector.

[0062] In this embodiment, the cross-attention mechanism performs a fusion operation based on the correlation relationships among the elements in the three hidden feature vectors. For example, there may be a correlation between the part of the ice layer thickness and formation dip angle relationship in the first hidden feature vector and the stress information of the stress concentration area near the fault in the second hidden feature vector, because the change in ice layer thickness may affect the stress distribution. At the same time, the information in these two vectors is also related to the influence of the temperature change rate in the third hidden feature vector on the surface vibration frequency, because the temperature change may indirectly affect the stress through the change in ice layer thickness, and further affect the surface vibration frequency. By calculating the correlation weights among the above elements, the relevant elements in the three hidden feature vectors are combined according to the weights. For example, the weight of a certain element in the first hidden feature vector is 0.3, the weight of the relevant element in the second hidden feature vector is 0.4, and the weight of the relevant element in the third hidden feature vector is 0.3. These three elements are added according to the weights to obtain an element in the joint feature vector. After such a calculation process for all relevant elements, a joint feature vector is finally generated. This joint feature vector synthesizes the key information in three aspects: the ice-rock mass structure, geological stress, and environmental dynamics, and comprehensively reflects the overall state of the ice-rock.

[0063] Step S136: Input the joint feature vector into the fully connected layer of the stability classification task and the fully connected layer of the stress change regression task respectively, to generate stability probability distribution parameters and stress change prediction values.

[0064] In this embodiment, in the fully connected layer of the stability classification task, the information in the joint feature vector is processed by multiple neurons. Each neuron performs a multiplication operation and summation with the elements in the joint feature vector according to the preset weights, and then obtains the output result through an activation function. For example, assume there are 5 neurons. The elements in the joint feature vector are multiplied by the weights of these 5 neurons and summed to obtain 5 intermediate results, and then through an activation function (such as the sigmoid function), the probability values of each stability level (stable, relatively stable, unstable, etc.) are obtained. The above probability values constitute the stability probability distribution parameters. In the fully connected layer of the stress change regression task, the same process is used to predict the stress change value. The elements in the joint feature vector are multiplied by the weights of the neurons in the fully connected layer and summed to obtain the stress change prediction value.

[0065] Step S137: Based on the cross-entropy loss between the stability probability distribution parameters and the labeled stability labels, and the mean square error loss between the stress change prediction value and the measured stress data, jointly optimize the parameters of the multi-task learning framework to obtain the trained ice-rock stability prediction model.

[0066] In this embodiment, the labeled stability label is an accurate stability result obtained through long-term observation and analysis of the ice-rock in the glacier area of the high-altitude mountainous region. For example, if the labeled stability label at a certain moment is relatively stable, then the labeling probability of the stable category is 0, the labeling probability of the relatively stable category is 1, the labeling probability of the unstable category is 0, and the stable probability given by the corresponding stability probability distribution parameter is 0.2, the relatively stable probability is 0.4, and the unstable probability is 0.4. When calculating the cross-entropy loss, the cross-entropy loss calculations are respectively -(0×log(0.2)), -(1×log(0.4)), and -(0×log(0.4)). Thus, the cross-entropy losses of these three categories are added together to obtain the total cross-entropy loss. For the mean square error loss between the predicted value of stress change and the measured stress data, assuming the predicted value of stress change is an increase in stress of 3 MPa and the measured stress data is an increase in stress of 2 MPa, the mean square error loss is calculated as (3 - 2)² = 1. The cross-entropy loss and the mean square error loss are added together according to a certain weight to obtain the total loss value. Based on this total loss value, various parameters in the multi-task learning framework, such as the weights of neurons in each layer, are adjusted through the backpropagation algorithm. After multiple such iterative processes, the total loss value is continuously reduced until the preset convergence condition is reached, thereby obtaining a trained ice-rock stability prediction model.

[0067] In a possible implementation manner, step S140 includes: Step S141, obtaining the dynamic data of the ice layer thickness, the instantaneous temperature change data, and the surface vibration amplitude data collected in real time.

[0068] In this embodiment, in the glacier area, the dynamic data of the ice layer thickness is obtained through ice layer thickness monitoring devices installed at different positions. These ice layer thickness monitoring devices perform measurements every certain period of time (for example, every half hour). Assume that at a certain moment, the monitoring device located in the central area of the glacier shows that the ice layer thickness has decreased by 0.5 meters compared to the previous measurement. The instantaneous temperature change data is provided by a high-precision thermometer that can be accurate to 0.1 °C. At the same moment, the instantaneous temperature change data shows that the temperature has dropped by 3 °C within 10 minutes. The surface vibration amplitude data is collected by a seismograph that can accurately detect minute surface vibrations. At this time, the detected surface vibration amplitude reaches 0.8 Hz.

[0069] Step S142, performing spatial interpolation processing on the dynamic data of the ice layer thickness to generate real-time updated parameters for the ice-rock body distribution.

[0070] Since it is impossible for the ice thickness monitoring equipment to cover the entire glacier area, spatial interpolation processing is required to obtain a more comprehensive ice thickness distribution. For example, based on the existing ice thickness monitoring points, the Kriging interpolation method is used. Suppose in a rectangular glacier area, there are ice thickness monitoring points at the four corners. The ice thickness at the upper left corner is 20 meters, the upper right corner is 18 meters, the lower left corner is 16 meters, and the lower right corner is 15 meters. Through the Kriging interpolation method, based on the ice thickness of these four points and their spatial relationships, the estimated ice thickness values at other positions in the area are calculated, thereby generating real-time ice-rock mass distribution update parameters. These real-time ice-rock mass distribution update parameters can reflect the real-time distribution of the ice thickness in the entire glacier area, including areas that have not been directly monitored.

[0071] Step S143: Perform a difference calculation on the temperature instantaneous change data and the historical temperature data to generate a temperature change trend parameter.

[0072] In this embodiment, the historical temperature data is obtained through long-term temperature monitoring records. For example, the temperature data every hour in the past week. At the current moment, the temperature instantaneous change data shows that the temperature has dropped by 3°C within 10 minutes. To calculate the temperature change trend parameter, this instantaneous change needs to be compared with the historical temperature data. Suppose the historical temperature data shows that the normal temperature drop rate during this period is 0.5°C every 10 minutes. Then the current temperature change trend parameter is (3 - 0.5) = 2.5°C. This positive value indicates that the current temperature drop rate is faster than the same period in history, which is an abnormal temperature change trend.

[0073] Step S144: Compare the surface vibration amplitude data with a preset vibration threshold to generate a vibration anomaly area marking parameter.

[0074] In this embodiment, the preset vibration threshold is determined based on long-term surface vibration monitoring and analysis of this glacier area. For example, it is set to 0.5 Hz. The current surface vibration amplitude data is 0.8 Hz. Since 0.8 Hz is greater than 0.5 Hz, this indicates that the surface vibration amplitude in this area exceeds the normal range. According to this comparison result, the currently monitored area is marked as a vibration anomaly area. This vibration anomaly area marking parameter can accurately indicate which areas have abnormal surface vibrations, which is of great significance for subsequent analysis of ice-rock stability.

[0075] Step S145: Input the real-time ice-rock mass distribution update parameter, the temperature change trend parameter, and the vibration anomaly area marking parameter into the input layer of the ice-rock stability prediction model. Through the real-time prediction module of the multi-task learning framework, output the real-time stability level parameter and the real-time stress change prediction parameter as the real-time stability prediction result.

[0076] In this embodiment, after receiving the above input parameters, the real-time prediction module of the multi-task learning framework of the ice-rock stability prediction model performs calculations according to the previously trained model structure and parameters. For example, the ice layer thickness distribution in the real-time ice-rock mass distribution update parameters affects the stability assessment of the ice-rock structure. An abnormal temperature drop in the temperature change trend parameters may cause changes in the internal stress of the ice-rock. The abnormal vibration area in the abnormal vibration area marking parameters is also closely related to the ice-rock stability. Inside the model, the above parameters are processed through each layer, including the calculations of the ice-rock mass structure feature input layer, geological stress feature input layer, environmental dynamic feature input layer, and subsequent fully connected layers mentioned before. Finally, the real-time stability level parameters are output. Suppose the output result is relatively unstable, and the real-time stress change prediction parameters, such as predicting that the stress will increase by 4 MPa in the next hour.

[0077] Step S150 includes: Step S151, triggering the stability warning signal according to the comparison result between the real-time stability level parameters and the preset safety threshold.

[0078] In this embodiment, the preset safety threshold is set to the two levels of stable and relatively stable. If the real-time stability level parameters are relatively unstable, this is lower than the preset safety threshold. At this time, the stability warning signal will be triggered, and this stability warning signal can be sent out through the communication device connected to the monitoring center to notify relevant personnel that the ice-rock stability in the glacier area of the high-altitude mountainous area is in a relatively unstable state and corresponding measures need to be taken for attention or further investigation.

[0079] Step S152, generating the optimized reinforcement recommendation parameters including the reinforcement position coordinates and material parameters according to the deviation value between the real-time stress change prediction parameters and the historical stress data.

[0080] In this embodiment, the historical stress data is obtained by long-term monitoring of the ice-rock stress in the glacier area, such as the stress change data every day in the past month. The current real-time stress change prediction parameter is that the stress will increase by 4 MPa in the next hour, while the historical stress data shows that the stress change is very small during this period under normal circumstances (for example, the average increase is 0.5 MPa). Then the deviation value is (4 - 0.5) = 3.5 MPa. According to this deviation value and factors such as the specific characteristics and geographical location of the ice rock, the optimized reinforcement suggestion parameters are generated. Suppose it is found through analysis that the areas with large stress increase are mainly concentrated near a certain fault of the glacier and the areas with large changes in ice layer thickness. The coordinates of these areas are determined as the reinforcement position coordinates. For the determination of the reinforcement material parameters, considering the large stress increase amplitude and the ice-rock characteristics of this area, high-strength anchor bolts may be required for reinforcement. The parameters such as the length and diameter of the anchor bolts are determined according to factors such as the stress increase amplitude and the bearing capacity of the ice rock, thus jointly constituting the optimized reinforcement suggestion parameters.

[0081] In a possible implementation manner, the training process of the ice-rock stability prediction model further includes a dynamic optimization step: Step S210, during the model training stage, the update frequency of the environmental monitoring parameters in the geological data set is monitored in real time. When the update frequency exceeds the preset threshold, the incremental learning mode is triggered.

[0082] In this embodiment, the environmental monitoring parameters include temperature change rate data, humidity fluctuation data, surface vibration frequency data, etc. Taking the temperature change rate data as an example, multiple temperature monitoring points are arranged in the entire glacier area. The above temperature monitoring points continuously transmit temperature data back, and then the update times of the temperature change rate data within a certain time window (such as every hour) are calculated. Suppose the preset threshold is 10 updates per hour. If it is found that the update frequency of the temperature change rate data reaches 12 times per hour during a certain period, it means that the update frequency of the environmental monitoring parameters exceeds the preset threshold, and at this time, the incremental learning mode will be triggered.

[0083] Step S220, in the incremental learning mode, the newly added geological data is divided into an incremental training set, and the ice-rock body structure incremental features, geological stress incremental features, and environmental dynamic incremental features in the incremental training set are extracted.

[0084] In this embodiment, the newly added geological data may come from newly installed monitoring devices or supplementary measurements of existing data. For the extraction of the incremental characteristics of the ice-rock mass structure, it is assumed that the newly added dynamic data of ice layer thickness shows that the ice layer thickness in a certain area at the glacier edge rapidly thins in a short period of time. By comparing and analyzing with the previous ice layer thickness distribution data, the incremental characteristics of the ice-rock mass structure corresponding to this change are extracted. For example, if the ice layer thickness in this area was relatively stable before and now shows a rapid thinning trend, it reflects an obvious change in the ice-rock mass structure locally, and the relevant characteristics of this change are the incremental characteristics of the ice-rock mass structure.

[0085] For the incremental characteristics of geological stress, if the newly added humidity fluctuation data shows that the humidity suddenly increases significantly in a certain area, this will affect the compressive strength of the rock formation and thus change the geological stress situation. By analyzing the relationship between humidity fluctuation and the compressive strength of the rock formation and comparing with historical geological stress data, the incremental characteristics of geological stress are extracted. For example, a significant increase in humidity may lead to a decrease in the compressive strength of the rock formation, and the magnitude of this strength decrease and the related stress change trend are the incremental characteristics of geological stress.

[0086] For the extraction of environmental dynamic incremental characteristics, for example, the new surface vibration frequency data shows that the surface vibration frequency abnormally increases during a certain period. Combining information such as the fault strike, formation dip angle in this area, and historical surface vibration frequency data, the environmental factor changes behind this abnormal increase are analyzed. For example, it may be due to increased ice body movement or enhanced fault activity in the vicinity. The above comprehensive information constitutes the environmental dynamic incremental characteristics.

[0087] Step S230, based on the similarity between the incremental characteristics of the ice-rock mass structure and the historical ice-rock mass structure characteristics, adjust the weight allocation ratio of the ice-rock mass structure feature input layer in the multi-task learning framework.

[0088] Suppose that through a certain feature similarity calculation method (such as a calculation method based on vector distance), the similarity between the incremental characteristics of the ice-rock mass structure and the historical ice-rock mass structure characteristics is calculated to be 0.6 (here 0.6 represents a medium degree of similarity). If the similarity is high, it indicates that the new data has a large correlation with the historical data in terms of the ice-rock mass structure. Then, it may be appropriate to increase the weight related to the historical data and reduce the weight allocation ratio for the new data features to balance the impact of new and old data on the model. Conversely, if the similarity is low, the weight of the incremental characteristics of the ice-rock mass structure may be increased to make the model pay more attention to the new structural changes.

[0089] Step S240, input the incremental characteristics of geological stress into the geological stress feature input layer, and freeze some convolutional kernel parameters through the elastic weight consolidation algorithm to retain the key stress gradient extraction ability.

[0090] For example, there are 10 convolutional kernels in the input layer of the geological stress characteristics. By analyzing the characteristics of the geological stress increment, it is found that the stress characteristics corresponding to some of these convolutional kernels have been fully learned in the historical data and play an important role in the extraction of key stress gradients, such as the convolutional kernels related to the stress concentration area near the fault. According to the elastic weight consolidation algorithm, the parameters of the above convolutional kernels are frozen so that they are not updated during the incremental training process. This can retain the ability of the above convolutional kernels to extract key stress gradients, while allowing other convolutional kernels to update their parameters to adapt to the new changes in the geological stress increment characteristics.

[0091] Step S250: After performing temporal alignment processing on the environmental dynamic increment characteristics, input them into the environmental dynamic feature input layer, and update the hidden state of the time series network through a sliding window mechanism.

[0092] Suppose the environmental dynamic increment characteristics include the time series data of the new temperature change rate and the surface vibration frequency. Temporal alignment processing is to arrange these new data and the historical environmental dynamic feature data in chronological order to ensure the consistency of the data on the time axis. Then, the aligned environmental dynamic increment characteristics are input into the environmental dynamic feature input layer. In the time series network of the environmental dynamic feature input layer, the hidden state is updated through a sliding window mechanism. For example, the size of the sliding window is 5 time points. After the new data enters the window, the hidden state of the time series network is recalculated according to the data in the window. This process enables the time series network to capture the new trends in the environmental dynamic increment characteristics in a timely manner.

[0093] Step S260: Generate an incremental joint feature vector based on the ice-rock mass structure increment characteristics, geological stress increment characteristics, and environmental dynamic increment characteristics. Calculate the loss value based on the incremental joint feature vector and the labeled data in the incremental training set, and fine-tune the parameters of the cross-attention fusion module through backpropagation.

[0094] For example, the ice-rock mass structure increment characteristics, geological stress increment characteristics, and environmental dynamic increment characteristics are fused into an incremental joint feature vector through an existing fusion algorithm. Suppose the labeled data in the incremental training set shows the correct results (such as the stability level and stress change value) corresponding to a certain ice-rock state. When calculating the loss value between the incremental joint feature vector and the labeled data, if the predicted stability level of the incremental joint feature vector is inconsistent with the stability level in the labeled data, or there is a deviation between the predicted stress change value and the stress change value in the labeled data, a loss will occur. Through the backpropagation algorithm, the parameters in the cross-attention fusion module are fine-tuned according to this loss value, such as adjusting the attention weights between different features, so that the model can better adapt to the new data after incremental training.

[0095] In a possible implementation manner, the conditions for triggering the incremental learning mode further include: Step S201: When the temperature change rate data in the newly added environmental monitoring data continuously exceeds the historical maximum value, the incremental learning mode is automatically activated.

[0096] For example, the highest temperature change rate recorded in this glacier area historically is 3°C per hour. If the new temperature monitoring data shows that the temperature change rates are 3.5°C, 3.8°C, and 4°C respectively within 3 consecutive hours, all exceeding the historical maximum value, then the incremental learning mode will be automatically activated. This is because such abnormal temperature change rates may indicate a major change in the ice-rock state, and the model needs to learn new features in a timely manner to adapt to this change.

[0097] Step S202: When the difference between the newly added fault strike data in the target area and the historical geological structure parameters is greater than the set difference, start the local parameter retraining process.

[0098] Step S203: In the local parameter retraining process, only update the gradient of the fully connected layer parameters associated with the geological stress feature input layer in the multi-task learning framework, and keep the parameters of other layers fixed.

[0099] Suppose the fault strike in the historical geological structure parameters is in the northeast-southwest direction, and the set difference is 10 degrees (here, the degree refers to the angular difference of the fault strike). If the new measurement result shows that the fault strike has changed to the north-south direction, and the angular difference from the historical fault strike reaches 45 degrees, which is much greater than the set difference. At this time, start the local parameter retraining process. In this local parameter retraining process, only update the gradient of the fully connected layer parameters associated with the geological stress feature input layer in the multi-task learning framework, and keep the parameters of other layers fixed. This is because the change in the fault strike mainly affects the distribution of geological stress, so only update the fully connected layer parameters associated with the geological stress feature input layer specifically. In this way, most of the structure and parameters of the existing model can be utilized, and the impact brought by the change in the fault strike can be adapted quickly.

[0100] Step S204: During the incremental training process, adopt the momentum optimization algorithm to adaptively adjust the learning rate of the ice-rock stability prediction model to ensure the convergence stability of the model.

[0101] In this embodiment, the momentum optimization algorithm adjusts the current learning rate according to the previous gradient update direction and size. For example, at the beginning of incremental training, the initial learning rate is set to 0.01. As the training progresses, if it is found that the loss value of the model decreases rapidly in several consecutive iterations, it means that the model is converging rapidly. According to the momentum optimization algorithm, the learning rate will be appropriately reduced, such as to 0.005, to avoid overfitting or instability during the convergence process. On the contrary, if it is found that the loss value decreases slowly or stagnates, the learning rate may be appropriately increased, such as to 0.015, so that the model can learn new data features faster.

[0102] Step S205, after completing the incremental training, the optimized model parameters of the ice-rock stability prediction model are synchronized to the real-time prediction module, and the original model parameters are overwritten to achieve seamless switching.

[0103] Assume that after incremental training, the parameters of the ice-rock stability prediction model have been updated to varying degrees in each layer, including weight adjustment of the ice-rock structure feature input layer, partial update of the convolution kernel parameters of the geological stress feature input layer, and parameter fine-tuning of the cross-attention fusion module. These optimized model parameters are fully synchronized to the real-time prediction module, overwriting the original model parameters. In this way, when performing real-time prediction, the latest model parameters can be used to ensure the accuracy and reliability of the real-time stability prediction results, and this switching process is seamless and will not interfere with the continuity and accuracy of the real-time prediction.

[0104] In a possible implementation, the ice-rock stability prediction model further includes an anomaly detection mechanism in the real-time prediction stage: Step S310, after generating the real-time stability prediction result, calculate the statistical distribution difference value between the real-time stability level parameter and the historical data of the same period.

[0105] For example, the stability level parameter in the current real-time stability prediction result is relatively unstable, while the historical data for the same period (obtained by statistics of ice and rock stability data in the same time period in the past many years) shows that the area is usually stable during this period. In order to calculate the statistical distribution difference value, multiple factors need to be considered. First, the probability distribution of the stable state in the same period of history is statistically analyzed, assuming that the probability of a stable state is 80%, the probability of a relatively stable state is 15%, and the probability of an unstable state is 5%. The current prediction is relatively unstable, which is quite different from the historical stable state. The difference value is calculated using existing statistical methods (such as distance measurement methods based on probability distribution), which reflects the degree of deviation between the current prediction result and the historical situation.

[0106] Step S320: If the statistical distribution difference value exceeds the preset anomaly threshold, start the reverse verification process, re-enter the real-time input monitoring data into the ice-rock stability prediction model, and compare the consistency of the stability level parameters output twice.

[0107] Suppose the preset anomaly threshold is 0.5 (this preset anomaly threshold is determined comprehensively based on the historical data and model performance of the ice-rock stability in this glacier area). If the calculated statistical distribution difference value is 0.6, exceeding the preset anomaly threshold. At this time, re-enter the previously real-time input dynamic data of ice layer thickness, instantaneous temperature change data, surface vibration amplitude data, etc. into the ice-rock stability prediction model. Suppose the stability level parameter output for the first time is relatively unstable, and the stability level parameter output for the second time after re-entering the data is stable, which indicates that there is a conflict between the stability level parameters output twice.

[0108] Step S330: When it is detected that there is a conflict between the stability level parameters output twice, trigger an artificial review signal, freeze the current prediction result output channel, and after the artificial review confirms the data validity, select to enable the corrected prediction parameters or re-collect the monitoring data according to the artificial review result.

[0109] Specifically, once a conflict is detected, an artificial review signal will be triggered, and this artificial review signal will notify relevant professionals (such as glacier geology experts) to conduct a review. At the same time, freeze the current prediction result output channel to prevent inaccurate results from continuing to spread. Professionals will conduct a detailed inspection of the data, including checking whether the monitoring equipment is working properly and whether there are errors in the data collection process. If it is found through the review that there is an abnormality in a certain monitoring data (for example, the data deviation is caused by a malfunction of the instantaneous temperature change data collection device), then it will be enabled after correcting the prediction parameters according to the correct data; if the data validity cannot be determined, it is necessary to re-collect the monitoring data.

[0110] Step S340: If the artificial review is not completed within the preset time, automatically switch to the downgraded prediction mode, and generate a conservative stability assessment result only based on the ice-rock body structure characteristics and geological stress characteristics.

[0111] Assume that the preset time is 1 hour. If the manual review is not completed within 1 hour, it will automatically switch to the degraded prediction mode. In this degraded prediction mode, the environmental dynamic characteristics are no longer considered (because they may be uncertain), and only the ice-rock mass structure characteristics (such as the relationship between the ice layer thickness distribution and the formation dip angle, the degree of fragility of the internal rock structure, etc.) and the geological stress characteristics (such as the dynamic change of geological stress under the influence of humidity fluctuation) are used to generate a conservative stability assessment result. For example, the result obtained based on the analysis of these two characteristics may be that the stability level is evaluated as unstable, which is a relatively conservative assessment to ensure a basic judgment of the ice-rock stability in the case of uncertain data.

[0112] In a possible implementation manner, the anomaly detection mechanism is further linked with the model interpretation module: Step S410, when the statistical distribution difference value is detected to be abnormal, call the model interpretation module to perform a feature contribution degree analysis on the current stability prediction result, and according to the feature contribution degree analysis result of the current stability prediction result, extract the top-K features that have the greatest impact on the current stability prediction result among the ice-rock mass structure characteristics, geological stress characteristics, and environmental dynamic characteristics.

[0113] For example, when the statistical distribution difference value is 0.6, which exceeds the preset anomaly threshold, the model interpretation module starts to run. Specifically, the model interpretation module uses the gradient-weighted class activation mapping method to implement the feature contribution degree analysis, which specifically includes: Step S411, record the gradient change data of each hidden layer in the ice-rock stability prediction model during the forward propagation process, and perform a gradient-weighted summation on the output feature maps of the ice-rock mass structure feature input layer, geological stress feature input layer, and environmental dynamic feature input layer.

[0114] For example, when the output feature map is obtained through neuron calculation in the ice-rock mass structure feature input layer, record the gradient change data of each neuron output for the input feature. The same operation is also performed on the geological stress feature input layer and the environmental dynamic feature input layer. Then, perform a gradient-weighted summation on the output feature maps of the ice-rock mass structure feature input layer, geological stress feature input layer, and environmental dynamic feature input layer. Assume that the gradient of a certain element in the output feature map of the ice-rock mass structure feature input layer is 0.3, and the corresponding weight is 0.2, then the weighted value is 0.3×0.2 = 0.06. Perform such calculations and summations for all elements.

[0115] Step S412: Generate a heatmap mapping of each input feature to the final prediction result based on the weighted output feature map. Based on the intensity distribution of the heatmap mapping, filter out the feature dimensions with a contribution degree exceeding a preset ratio as the Top-K features, and match the Top-K features with a preset feature semantic table to convert them into interpretable geological parameter description texts.

[0116] In this embodiment, the heatmap mapping represents the contribution degree of each input feature to the final prediction result through the depth of color or the height of numerical values. For example, in the heatmap, the areas with darker colors or higher numerical values indicate that the corresponding input features have a greater contribution to the final prediction result. Based on the intensity distribution of the heatmap mapping, filter out the feature dimensions with a contribution degree exceeding a preset ratio (such as the preset ratio is 30%) as the Top-K features. Suppose there is a feature dimension of the relationship between ice layer thickness and formation dip angle in the ice-rock mass structure characteristics, a feature dimension of the influence of humidity fluctuation on the compressive strength of rock layers in the geological stress characteristics, and a feature dimension of the correlation between temperature change rate and surface vibration frequency in the environmental dynamic characteristics. Their contribution degrees in the heatmap all exceed 30%, then these feature dimensions are filtered as the Top-K features.

[0117] The preset feature semantic table contains the geological parameter descriptions corresponding to various feature dimensions. For example, the description corresponding to the relationship between ice layer thickness and formation dip angle is "the comprehensive influence of the thickness change of the ice layer along the formation dip direction on the ice-rock structure". After matching the filtered Top-K features with the feature semantic table, interpretable geological parameter description texts are obtained, such as "factors such as the comprehensive influence of the thickness change of the ice layer along the formation dip direction on the ice-rock structure, the influence of humidity fluctuation on the compressive strength of rock layers, and the correlation between temperature change rate and surface vibration frequency have a greater impact on the current stability prediction result".

[0118] Step S420: Match the Top-K features with the historical abnormal case library, and based on the matching result, attach the risk type identifier of the potential risk type and the recommended disposal measure code to the stability warning signal. Among them, when the matching degree between the potential risk type and the historical high-risk cases exceeds the set threshold, automatically raise the warning level and activate the emergency response protocol.

[0119] In this embodiment, it is assumed that there are similar cases in the historical abnormal case library where the ice layer thickness changes rapidly, the humidity fluctuates greatly, and there are abnormal correlations between temperature and surface vibration, and these cases are marked as the risk type of ice-rock structure damage. When the Top-K features have a high degree of matching with these cases (assuming the matching degree exceeds the set threshold of 80%), the risk type identifier of ice-rock structure damage is added to the stability warning signal. At the same time, according to the disposal experience in historical cases, recommended disposal measure codes are generated. For example, code 101 indicates a detailed survey of areas with large changes in ice layer thickness, and code 102 indicates strengthening the monitoring of humidity and rock compressive strength, etc.

[0120] Among them, it is assumed that the set threshold is 90%. If the matching degree between the potential risk type and historical high-risk cases (such as cases of severe ice-rock structure damage that once caused large-scale glacier collapses) reaches 95%, then the warning level is automatically upgraded from general to emergency, and the emergency response protocol is activated. The emergency response protocol may include measures such as notifying the surrounding residents to evacuate urgently and organizing professional rescue teams to conduct emergency treatment in the glacier area to cope with the possible ice-rock stability crisis.

[0121] In a possible implementation manner, the method further includes: Step S510, injecting random noise data into the geological data set during the model training stage to generate a noise-enhanced data set, where the amplitude of the random noise data is dynamically adjusted according to the historical fluctuation range of the ice-rock body distribution parameters.

[0122] In this embodiment, taking the ice layer thickness distribution data as an example, by reviewing the historical ice layer thickness measurement data, its fluctuation range is analyzed. It is assumed that in the past years of measurements, the minimum value of the ice layer thickness in a certain specific area is 10 meters, the maximum value is 50 meters, and its fluctuation range is 40 meters. Then, the amplitude of the random noise data is determined according to this fluctuation range. If the set noise amplitude is 10% of the fluctuation range, the noise amplitude is 4 meters. For each ice layer thickness measurement value, a randomly generated noise value between -4 meters and 4 meters is added to it. For example, if the original ice layer thickness at a certain measurement point is 30 meters, it becomes 33 meters after adding a noise of 3 meters. The same method is applied to ice-rock body distribution parameters such as rock porosity data and fracture density data, as well as geological structure parameters and environmental monitoring parameters, so as to generate a noise-enhanced data set.

[0123] Step S520, inputting the noise-enhanced data set into the adversarial sample generation module, and extracting the sensitive feature direction of the ice-rock stability prediction model through the gradient sign backpropagation algorithm to generate an adversarial perturbation vector.

[0124] In the adversarial sample generation module, taking the ice layer thickness distribution data as an example, when the input is ice layer thickness data containing noise, the model performs forward propagation calculation based on the current parameters to obtain the prediction result. Then, the gradient of the prediction result with respect to the input ice layer thickness data is calculated, and this gradient represents the direction of the impact of the small change in the ice layer thickness data on the prediction result. The same operation is also performed on other parameters (such as rock porosity, formation dip angle, etc.) to obtain the gradient information of each parameter on the prediction result. According to the gradient sign backpropagation algorithm, these gradient information are integrated to determine the sensitive feature direction of the model. For example, if it is found that a small increase in the ice layer thickness will cause the stability prediction result to change from stable to less stable, and the gradient of this change is large, then the ice layer thickness in this change direction is a sensitive feature direction. All the determined sensitive feature directions are combined to generate an adversarial perturbation vector.

[0125] Step S530: Superimpose the adversarial perturbation vector on the ice-rock mass distribution parameters, geological structure parameters, and environmental monitoring parameters in the geological data set to generate an adversarial sample data set.

[0126] Suppose the perturbation value for the ice layer thickness distribution data in the adversarial perturbation vector is +5 meters (indicating a 5-meter increase in the original ice layer thickness), and the original ice layer thickness at a certain measurement point is 30 meters. After superimposing the perturbation, it becomes 35 meters. For the formation dip angle data, suppose the perturbation value is +5 degrees (indicating a 5-degree increase in the formation dip angle), and the original formation dip angle is 30 degrees. After superimposing, it becomes 35 degrees. The same operation is also performed on other ice-rock mass distribution parameters, geological structure parameters, and environmental monitoring parameters to generate an adversarial sample data set. The data in this adversarial sample data set contains deliberately constructed adversarial perturbations, which are used to test and enhance the robustness of the model.

[0127] Step S540: Merge the adversarial sample data set with the original geological data set into a mixed training set, input it into the multi-task learning framework for adversarial training, and generate anti-interference model parameters by alternately optimizing the classification loss function and the adversarial robustness loss function.

[0128] In a multi-task learning framework, the classification loss function is used to measure the prediction accuracy of the stability classification task (such as stable, relatively stable, unstable), and the adversarial robust loss function is used to measure the stability of the model under adversarial samples. For example, for the classification loss function, when the predicted stability level is inconsistent with the actual labeled stability label, a loss will be generated. For the adversarial robust loss function, when the prediction results of the model on the adversarial sample dataset are significantly different from the prediction results on the original geological dataset, a loss will be generated. During the training process, these two loss functions are alternately optimized. Suppose the classification loss function is optimized first, and the classification loss is reduced by adjusting the model's parameters (such as the weights of neurons in each layer, etc.). Then the adversarial robust loss function is optimized, and the model parameters are also adjusted to reduce this loss. After multiple such alternating optimization processes, the anti-interference model parameters are finally generated, and these parameters enable the model to make more stable predictions when facing adversarial samples.

[0129] Step S550, in the real-time prediction stage, input the real-time collected monitoring data into the anomaly filtering module, extract the instantaneous change gradient of each dimension in the monitoring data, and perform a sliding window match with the historical gradient distribution of the corresponding dimension.

[0130] Taking the dynamic data of ice layer thickness as an example, assume that the ice layer thickness data collected in real time changes from 30 meters to 32 meters in a short period (such as 10 minutes). Calculate the instantaneous change gradient of this change, that is, (32 - 30) / 10 = 0.2 meters per minute. At the same time, review the historical ice layer thickness gradient data, and statistically analyze the historical ice layer thickness gradient distribution through a sliding window (for example, with a window length of the past 1 hour). Compare the current instantaneous change gradient with the historical gradient distribution. The same operations are also performed on the temperature instantaneous change data and the surface vibration amplitude data, respectively calculating their instantaneous change gradients and performing sliding window matches with their respective historical gradient distributions.

[0131] Step S560, when it is detected that the instantaneous change gradient exceeds the probability density region of the historical gradient distribution, perform Gaussian smoothing filtering on the monitoring data dimensions of the exceeded part to generate filtered monitoring data.

[0132] For example, the probability density region of the historical ice layer thickness gradient distribution is from -0.1 meters per minute to 0.1 meters per minute, and the currently calculated instantaneous change gradient is 0.2 meters per minute, which exceeds this region. After performing Gaussian smoothing filtering on the ice layer thickness data of this exceeded part, substitute the exceeded ice layer thickness change value into the Gaussian function to calculate the filtered value, obtaining the filtered ice layer thickness data. Similar processing is also performed on the temperature instantaneous change data and the surface vibration amplitude data, thereby generating filtered monitoring data.

[0133] Step S570: Input the filtered monitoring data into the ice-rock stability prediction model loaded with the anti-interference model parameters to generate an anti-interference stability prediction result.

[0134] The anti-interference model parameters enable the model to make more stable predictions in the face of possible interferences (such as data noise, abnormal fluctuations, etc.). The ice layer thickness data, temperature data, surface vibration data, etc. in the filtered monitoring data are processed through the various layers of the model, including the ice-rock mass structure feature input layer, geological stress feature input layer, environmental dynamic feature input layer, etc., and finally an anti-interference stability prediction result is output. For example, it is predicted that the stability level of the ice-rock is relatively stable, and the predicted stress change value (such as the stress will increase by 2 MPa in the next period of time).

[0135] Step S580: Regularly obtain a preset extreme geological data set, input the extreme geological data set into the ice-rock stability prediction model for stress testing, and extract the oscillation amplitude of the prediction results of the ice-rock stability prediction model under continuous abnormal inputs.

[0136] The preset extreme geological data set contains some geological data situations that rarely occur under normal circumstances but may occur in extreme cases. For example, the ice layer thickness suddenly decreases significantly (such as decreasing by 10 meters in a short time), while the temperature rises sharply (such as rising by 10 °C per hour), and the surface vibration amplitude reaches a very high value (such as 3 Hz), etc. When these extreme data are input into the model, due to the huge difference between these data and normal data, the prediction results of the model may fluctuate greatly. Record the changes in the prediction results each time extreme data is input, such as the stability level switching back and forth between stable and unstable, and the predicted stress change value changing from increasing to decreasing, etc. Calculate the fluctuation range of these prediction results, that is, the oscillation amplitude of the prediction results. Assume that the stability level changes from stable to unstable and then back to relatively stable, and quantify the range of this change into a value (such as according to the pre-set stability level quantification method, the oscillation amplitude is 0.5).

[0137] Step S590: Dynamically adjust the number of hidden layer neurons and the activation function threshold parameters of the multi-layer neural network according to the ratio of the oscillation amplitude of the prediction results to the preset stability tolerance, fuse the adjusted neural network parameters with the anti-interference model parameters, generate the final optimized model parameters and update them to the real-time prediction module.

[0138] The preset stability tolerance is an acceptable fluctuation range determined according to the requirements for ice-rock stability and the performance of the model. Assume the preset stability tolerance is 0.3. The current prediction result oscillates with an amplitude of 0.5. Calculate the ratio of the two as (0.5 / 0.3 = 1.67). If this ratio is greater than 1, it indicates that the oscillation amplitude of the model under extreme data exceeds the acceptable range, and the model needs to be adjusted. For the adjustment of the number of neurons in the hidden layer, assume the current number of neurons in the hidden layer is 100. If the ratio is greater than 1, the number of neurons can be appropriately increased, for example, increased to 120. For the adjustment of the activation function threshold parameter, for example, if the sigmoid activation function is used and the original threshold is 0.5, it is adjusted to 0.6 according to the ratio. Such adjustments enable the model to make more stable predictions when facing extreme data.

[0139] Fuse the adjusted neural network parameters with the anti-interference model parameters to generate the final optimized model parameters and update them to the real-time prediction module. The adjusted neural network parameters (such as the number of neurons in the hidden layer and the activation function threshold parameter) and the anti-interference model parameters (parameters that are robust to interference obtained through adversarial training) both contain valuable information for improving the model performance. Fuse these two types of parameters. For example, for the weight parameters of neurons, the weighted average method can be used for fusion. Assume that a certain weight in the adjusted neural network parameters is 0.6, and the corresponding weight in the anti-interference model parameters is 0.4. The fused weight is (0.6×0.6 + 0.4×0.4 = 0.52). Perform such fusion operations on all model parameters to generate the final optimized model parameters, and then update these parameters to the real-time prediction module, enabling the real-time prediction module to use the optimized model parameters to make more accurate and stable ice-rock stability predictions.

[0140] Figure 2 FIG. shows a schematic diagram of exemplary hardware and software components of a machine learning-based ice-rock mass stability prediction system 100 that can implement the ideas of the present application provided by some embodiments of the present application. For example, the processor 120 can be used on the machine learning-based ice-rock mass stability prediction system 100 and is used to execute the functions in the present application.

[0141] The machine learning-based ice-rock mass stability prediction system 100 can be a general-purpose server or a special-purpose server, both of which can be used to implement the machine learning-based ice-rock mass stability prediction method of the present application. Although only one server is shown in the present application, for convenience, the functions described in the present application can be implemented in a distributed manner on multiple similar platforms to balance the processing load.

[0142] For example, the machine learning-based ice-rock mass stability prediction system 100 may include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and different forms of storage media 140, such as magnetic disks, ROM, or RAM, or any combination thereof. Exemplarily, the machine learning-based ice-rock mass stability prediction system 100 may also include program instructions stored in ROM, RAM, or other types of non-transitory storage media, or any combination thereof. The method of the present application can be implemented according to these program instructions. The machine learning-based ice-rock mass stability prediction system 100 further includes an input / output (I / O) interface 150 between the computer and other input / output devices.

[0143] For ease of illustration, only one processor is described in the machine learning-based ice-rock mass stability prediction system 100. However, it should be noted that the machine learning-based ice-rock mass stability prediction system 100 in the present application may also include multiple processors. Therefore, the steps performed by one processor described in the present application may also be jointly performed or separately performed by multiple processors. For example, if the processor of the machine learning-based ice-rock mass stability prediction system 100 performs step A and step B, it should be understood that step A and step B may also be jointly performed by two different processors or separately performed in one processor. For example, the first processor performs step A, the second processor performs step B, or the first processor and the second processor jointly perform steps A and B.

[0144] In addition, an embodiment of the present invention also provides a readable storage medium, in which computer-executable instructions are preset. When the processor executes the computer-executable instructions, the above-mentioned machine learning-based ice-rock mass stability prediction method is implemented.

[0145] It should be noted that, in order to simplify the description of the present invention disclosure and thus help the understanding of one or more embodiments of the invention, in the previous description of the embodiments of the present invention, sometimes multiple features are merged into one embodiment, drawing, or description thereof.

Claims

1. A method for predicting ice-rock mass stability based on machine learning, characterized in that: The method comprises: Acquire a geological data set of a target area, wherein the geological data set includes ice-rock mass distribution parameters, geological structure parameters, and environmental monitoring parameters; Performing feature fusion processing on the geological data set to obtain an ice-rock feature set, wherein the ice-rock feature set includes ice-rock body structure features, geological stress features, and environmental dynamic features; Training an ice-rock stability prediction model based on the ice-rock feature set, wherein the ice-rock stability prediction model is constructed by a multi-layer neural network; In response to the real-time input monitoring data, calling the trained ice-rock stability prediction model to generate a real-time stability prediction result of the target area; Output a stability warning signal or optimize reinforcement suggestion parameters according to the real-time stability prediction result.

2. The ice-rock mass stability prediction method based on machine learning according to claim 1 is characterized in that: The ice-rock mass distribution parameters in the geological data set include ice layer thickness distribution data, rock porosity data and crack density data, the geological structure parameters include stratum dip data, fault strike data and stratum compressive strength data, and the environmental monitoring parameters include temperature change rate data, humidity fluctuation data and surface vibration frequency data; the feature fusion processing of the geological data set to obtain the ice-rock feature set includes: Extracting spatial gradient features from the ice layer thickness distribution data, and aligning the spatial gradient features with the formation dip angle data to generate a first fusion feature; Performing weighted superposition on the rock porosity data and the fracture density data to generate a second fusion feature; Performing time series correlation analysis on the temperature change rate data and the ground surface vibration frequency data to generate a third fusion feature; Cross-dimensionally splicing the first fusion feature, the second fusion feature, and the third fusion feature to generate the ice-rock body structure feature; Extracting the dynamic attenuation characteristics of the rock formation compressive strength data, and performing nonlinear mapping between the dynamic attenuation characteristics and the humidity fluctuation data to generate the geological stress characteristics; A topological relationship model is performed on the fault strike data and the formation dip data to generate a fourth fusion feature, and the fourth fusion feature is jointly encoded with the third fusion feature to generate the environmental dynamic feature.

3. The ice-rock mass stability prediction method based on machine learning according to claim 1 is characterized in that: The step of training the ice-rock stability prediction model based on the ice-rock feature set includes: Constructing a multi-task learning framework including an ice-rock mass structure feature input layer, a geological stress feature input layer, and an environmental dynamic feature input layer, wherein the multi-task learning framework includes a stability classification task and a stress change regression task; Inputting the ice-rock body structure feature into the ice-rock body structure feature input layer to generate a first hidden feature vector; Inputting the geological stress feature into the geological stress feature input layer, and extracting the local stress gradient through a convolution kernel to generate a second hidden feature vector; Inputting the environmental dynamic features into the environmental dynamic feature input layer, and extracting dynamic change trends through a time series network to generate a third hidden feature vector; Cross-attentionally fuse the first hidden feature vector, the second hidden feature vector, and the third hidden feature vector to generate a joint feature vector; Inputting the joint feature vector into the fully connected layer of the stability classification task and the fully connected layer of the stress change regression task respectively to generate stability probability distribution parameters and stress change prediction values; Based on the cross entropy loss of the stability probability distribution parameters and the annotated stability labels, and the mean square error loss of the stress change prediction value and the measured stress data, the parameters of the multi-task learning framework are jointly optimized to obtain the trained ice rock stability prediction model.

4. The ice-rock mass stability prediction method based on machine learning according to claim 3 is characterized in that: The step of calling the trained ice-rock stability prediction model in response to the real-time input monitoring data to generate a real-time stability prediction result of the target area includes: Obtain real-time dynamic data on ice thickness, instantaneous temperature change data, and surface vibration amplitude data; Performing spatial interpolation processing on the dynamic data of ice thickness to generate real-time ice and rock mass distribution update parameters; Perform differential calculation on the instantaneous temperature change data and the historical temperature data to generate temperature change trend parameters; Comparing the surface vibration amplitude data with a preset vibration threshold to generate vibration abnormality area marking parameters; The real-time ice-rock mass distribution update parameter, temperature change trend parameter and vibration abnormal area marking parameter are input into the input layer of the ice-rock stability prediction model, and the real-time stability grade parameter and the real-time stress change prediction parameter are output as the real-time stability prediction result through the real-time prediction module of the multi-task learning framework; The step of outputting a stability warning signal or optimizing reinforcement suggestion parameters according to the real-time stability prediction result comprises: triggering the stability warning signal according to a comparison result between the real-time stability level parameter and a preset safety threshold; The optimized reinforcement suggestion parameters including reinforcement position coordinates and material parameters are generated according to the deviation value between the real-time stress change prediction parameters and the historical stress data.

5. The ice-rock mass stability prediction method based on machine learning according to claim 4 is characterized in that: The training process of the ice-rock stability prediction model further includes a dynamic optimization step: During the model training phase, the update frequency of the environmental monitoring parameters in the geological data set is monitored in real time, and when the update frequency exceeds a preset threshold, an incremental learning mode is triggered; In the incremental learning mode, the newly added geological data are divided into incremental training sets, and the incremental features of ice-rock structure, geological stress and environmental dynamics in the incremental training sets are extracted; Based on the similarity between the incremental features of the ice-rock mass structure and the historical features of the ice-rock mass structure, adjusting the weight distribution ratio of the ice-rock mass structure feature input layer in the multi-task learning framework; Input the geological stress increment feature into the geological stress feature input layer, and freeze some convolution kernel parameters through the elastic weight solidification algorithm to retain the key stress gradient extraction capability; After the environment dynamic incremental features are time-series aligned, they are input into the environment dynamic feature input layer, and the hidden state of the time series network is updated through a sliding window mechanism; An incremental joint feature vector is generated according to the incremental features of the ice-rock structure, the incremental features of the geological stress and the incremental features of the environmental dynamics. A loss value is calculated based on the incremental joint feature vector and the labeled data of the incremental training set. The parameters of the cross-attention fusion module are fine-tuned through back propagation.

6. The ice-rock mass stability prediction method based on machine learning according to claim 5 is characterized in that: The conditions for triggering the incremental learning mode further include: When the temperature change rate data in the newly added environmental monitoring data continuously exceeds the historical maximum value, the incremental learning mode is automatically activated; When the difference between the newly added fault strike data and the historical geological structure parameters in the target area is detected to be greater than the set difference, the local parameter retraining process is started; In the local parameter retraining process, only the fully connected layer parameters associated with the geological stress feature input layer in the multi-task learning framework are gradient updated, and the other layer parameters are kept fixed; During the incremental training process, a momentum optimization algorithm is used to adaptively adjust the learning rate of the ice-rock stability prediction model to ensure the convergence stability of the model; After completing the incremental training, the optimized model parameters of the ice-rock stability prediction model are synchronized to the real-time prediction module, and the original model parameters are overwritten to achieve seamless switching.

7. The ice-rock mass stability prediction method based on machine learning according to claim 6 is characterized in that: The ice-rock stability prediction model further includes an anomaly detection mechanism in the real-time prediction stage: After generating the real-time stability prediction result, calculating the statistical distribution difference value between the real-time stability level parameter and the historical data of the same period; If the statistical distribution difference value exceeds the preset abnormal threshold, a reverse verification process is initiated to re-input the real-time input monitoring data into the ice-rock stability prediction model, and compare the consistency of the stability level parameters output twice; When a conflict is detected between the stability level parameters of the two outputs, a manual review signal is triggered, and the current prediction result output channel is frozen. After the data validity is confirmed by manual review, the revised prediction parameters are selected to be enabled or the monitoring data is re-collected according to the manual review results; If the manual review is not completed within the preset time, it will automatically switch to the degraded prediction mode and generate conservative stability assessment results based only on the structural characteristics of the ice-rock mass and the geological stress characteristics.

8. The ice-rock mass stability prediction method based on machine learning according to claim 7 is characterized in that: The anomaly detection mechanism is further linked with the model interpretation module: When an abnormal statistical distribution difference value is detected, the model interpretation module is called to perform feature contribution analysis on the current stability prediction result, and based on the feature contribution analysis result of the current stability prediction result, the Top-K features that have the greatest impact on the current stability prediction result among the ice-rock mass structure features, geological stress features and environmental dynamic features are extracted; Matching the Top-K features with the historical abnormal case library, and adding a risk type identifier and a recommended disposal measure code of the potential risk type to the stability warning signal according to the matching result; Among them, when the matching degree between the potential risk type and the historical high-risk cases exceeds the set threshold, the alarm level is automatically raised and the emergency response protocol is activated; The model interpretation module uses a gradient weighted class activation mapping method to implement feature contribution analysis, specifically including: In the forward propagation process, the gradient change data of each hidden layer in the ice-rock stability prediction model is recorded, and the output characteristic graphs of the ice-rock body structure characteristic input layer, the geological stress characteristic input layer and the environmental dynamic characteristic input layer are gradient-weighted summed; A heat map mapping of each input feature to the final prediction result is generated according to the weighted output feature map, and based on the intensity distribution of the heat map mapping, feature dimensions whose contribution exceeds a preset ratio are screened out as the Top-K features. The Top-K features are matched with a preset feature semantic table and converted into interpretable geological parameter description text.

9. The ice-rock mass stability prediction method based on machine learning according to claim 8, characterized in that: The method further comprises: Injecting random noise data into the geological data set to generate a noise-enhanced data set during a model training phase, wherein the amplitude of the random noise data is dynamically adjusted according to the historical fluctuation range of the ice-rock mass distribution parameter; The noise-enhanced data set is input into an adversarial sample generation module, and the sensitive feature direction of the ice-rock stability prediction model is extracted by a gradient sign back propagation algorithm to generate an adversarial disturbance vector; Superimposing the adversarial disturbance vector onto the ice-rock mass distribution parameters, geological structure parameters and environmental monitoring parameters in the geological data set to generate an adversarial sample data set; The adversarial sample data set and the original geological data set are combined into a mixed training set, which is input into the multi-task learning framework for adversarial training, and anti-interference model parameters are generated by alternately optimizing the classification loss function and the adversarial robust loss function; In the real-time prediction stage, the real-time collected monitoring data is input into the anomaly filtering module, the instantaneous change gradient of each dimension in the monitoring data is extracted, and a sliding window matching is performed with the historical gradient distribution of the corresponding dimension; When it is detected that the instantaneous change gradient exceeds the probability density area of ​​the historical gradient distribution, Gaussian smoothing filtering is performed on the monitoring data dimension of the exceeding part to generate filtered monitoring data; Inputting the filtered monitoring data into an ice-rock stability prediction model loaded with the anti-interference model parameters to generate an anti-interference stability prediction result; Regularly acquiring a preset extreme geological data set, inputting the extreme geological data set into the ice-rock stability prediction model for pressure testing, and extracting the oscillation amplitude of the prediction result of the ice-rock stability prediction model under continuous abnormal input; Dynamically adjusting the number of hidden layer neurons and activation function threshold parameters of the multi-layer neural network according to the ratio of the oscillation amplitude of the prediction result to the preset stability tolerance; The adjusted neural network parameters are integrated with the anti-interference model parameters to generate the final optimized model parameters and update them to the real-time prediction module.

10. An ice-rock stability prediction system based on machine learning, characterized in that: The ice and rock mass stability prediction system based on machine learning includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the ice and rock mass stability prediction method based on machine learning as described in any one of claims 1 to 9 above.

Images