Mine slope dynamic monitoring method, system, equipment and medium

Abstract

The invention relates to a mine slope dynamic monitoring method, system and device and a medium, and the method comprises the steps: carrying out the space-time registration and feature derivation of multi-source data from a GNSS, an inclinometer, a strain gauge and a microseismic monitor, and generating a fusion data set; based on the data set, solving a slope full-field displacement and stress field through mechanical inversion to form displacement-stress coupling data; constructing a differentiable physical constraint loss function of the quantitative prediction result and the rock mechanics principle deviation; performing joint training on the deep learning model in combination with a gradient back propagation algorithm to generate a displacement prediction model constrained by a physical mechanism; and a graded early warning signal fusing the displacement predicted value and the physical confidence coefficient is generated through physical consistency verification. According to the method, a physical mechanism is used as a priori knowledge depth embedded model training front end, so that the problem of insufficient early warning reliability caused by neglecting a mechanical rule of a pure data driving model is effectively solved, and the early warning accuracy and interpretability of a monitoring system under a complex working condition are improved.

Description

Technical Field

[0001] This invention relates to the field of mine safety monitoring and early warning technology, specifically to a method, system, equipment, and medium for dynamic monitoring of mine slopes. Background Technology

[0002] Mine slope stability monitoring is a crucial aspect of ensuring safe mine production. With advancements in sensing technology, modern monitoring systems commonly employ multiple sensors for collaborative observation. For example, they acquire surface displacement data via global navigation satellite systems, capture internal rock fracture signals through microseismic monitoring networks, and combine this data with sensor data such as dip angle and strain to form a multi-source, heterogeneous monitoring data stream. To extract effective early warning information from this massive amount of data, existing solutions typically employ data-driven analysis, particularly deep learning-based time-series prediction models. These models fuse and learn from historical monitoring data from various sensors, attempting to construct a mapping relationship between data and slope condition to achieve advanced deformation prediction and landslide warning.

[0003] However, this "end-to-end" prediction method, which relies on statistical correlation of data, has inherent technical flaws. Because deep learning models are essentially "black boxes," their training objective is solely to minimize prediction errors based on historical data. The learning process within the model is completely disconnected from the physical and mechanical mechanisms of slope instability. This leads to the model learning surface statistical characteristics from the data but failing to understand and adhere to fundamental rock mechanics principles such as stress balance, energy conservation, and rock fracture criteria. In engineering practice, this deficiency manifests as predictions that sometimes contradict common sense, such as predicting rapid displacement acceleration without significant stress adjustment or energy release, or failing to provide warnings when the mechanical state has significantly deteriorated. This disconnect between "data" and "physics" results in poor generalization ability and severely insufficient reliability and interpretability of warning results when dealing with complex conditions or rare scenarios not covered by training data, posing a risk of false alarms or missed alarms. Summary of the Invention

[0004] Based on this, the purpose of this invention is to provide a method, system, equipment, and medium for dynamic monitoring of mine slopes that can deeply embed rock mechanics mechanisms into the front end of a prediction model, thereby achieving an essential fusion of data-driven and physical-driven approaches.

[0005] The objective of this invention is achieved through the following solution:

[0006] In a first aspect, the present invention provides a method for dynamic monitoring of mine slopes, comprising the following steps:

[0007] S1: Spatiotemporal registration and physical feature derivation processing are performed on multi-source sensor monitoring data from GNSS, inclinometers, strain gauges, and microseismic monitoring instruments. The original data from different spatiotemporal references are aligned to a unified grid and strain and energy release rate derivation features are calculated to generate a multi-source spatiotemporal fusion dataset.

[0008] S2: Based on a multi-source spatiotemporal fusion dataset, point displacement and linear strain observations are used as boundary conditions and substituted into the slope mechanics model to solve the full-field state, generating displacement-stress coupling data containing the full-field displacement field and full-field stress field of the slope.

[0009] S3: Based on the microseismic energy data in displacement-stress coupling data and multi-source spatiotemporal fusion dataset, physical law consistency quantification is performed to generate a differentiable physical constraint loss function;

[0010] S4: Based on the gradient backpropagation algorithm, a slope displacement prediction model constrained by physical mechanism is generated by jointly training the multi-source spatiotemporal fusion dataset and displacement-stress coupling data.

[0011] S5: Based on the slope displacement prediction model, forward prediction and physical consistency verification are performed on the real-time multi-source sensor monitoring data to generate a graded early warning signal that integrates the displacement prediction value and physical confidence level. The graded early warning signal is used to indicate different slope risk levels and treatment priorities.

[0012] Secondly, the present invention provides a dynamic monitoring system for mine slopes, which is configured with the following modules:

[0013] The multi-source data fusion processing module is used to perform spatiotemporal registration and physical feature derivation processing on monitoring data from multiple sources such as GNSS, inclinometers, strain gauges, and microseismic monitors. It aligns the original data from different spatiotemporal references to a unified grid and calculates strain and energy release rate derived features to generate a multi-source spatiotemporal fusion dataset.

[0014] The slope coupled field solution module is used to solve the whole field state of the slope mechanical model by substituting point displacement and linear strain observations as boundary conditions into the multi-source spatiotemporal fusion dataset, and generating displacement-stress coupled data containing the whole field displacement field and the whole field stress field of the slope.

[0015] The physical constraint loss generation module is used to quantify the consistency of physical laws based on microseismic energy data in displacement-stress coupling data and multi-source spatiotemporal fusion datasets, and generate a differentiable physical constraint loss function.

[0016] The prediction model joint training module is used to jointly train multi-source spatiotemporal fusion datasets and displacement-stress coupling data based on the gradient backpropagation algorithm to generate slope displacement prediction models constrained by physical mechanisms.

[0017] The graded early warning signal generation module is used to perform forward prediction and physical consistency verification on real-time multi-source sensor monitoring data based on the slope displacement prediction model, and generate graded early warning signals that fuse displacement prediction values ​​and physical confidence levels. The graded early warning signals are used to indicate different slope risk levels and handling priorities.

[0018] Thirdly, this application provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement any of the above-mentioned dynamic monitoring methods for mine slopes.

[0019] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the above-mentioned methods for dynamic monitoring of mine slopes.

[0020] In summary, the dynamic monitoring method for mine slopes provided in this application constructs a displacement-stress coupling field and loss function that can be directly verified for physical consistency by fusing multi-source heterogeneous monitoring data and performing mechanical inversion, thereby constraining the optimization direction of the model. This ensures that the output of the final trained prediction model not only fits historical data but also inherently follows fundamental mechanical principles such as static equilibrium, energy conservation, and fracture criteria, significantly enhancing the model's extrapolation and generalization capabilities and the interpretability of prediction results under data-scarce or unknown extreme conditions. Furthermore, this method introduces a dual verification mechanism based on physical confidence in the early warning decision-making stage, which can proactively identify and warn of abnormal predictions that, while conforming to statistical trends, contradict physical laws. This effectively overcomes the risks of false alarms and missed alarms caused by "black box" decision-making in complex conditions under existing technologies, achieving a systematic improvement in the accuracy, reliability, and engineering practicality of early warnings.

[0021] To better understand and implement this invention, the following detailed description is provided in conjunction with the accompanying drawings. Attached Figure Description

[0022] Figure 1 A flowchart illustrating a dynamic monitoring method for mine slopes provided in this application embodiment;

[0023] Figure 2 This is a schematic diagram of the structure of a dynamic monitoring system for mine slopes, provided as another embodiment of this application. Detailed Implementation

[0024] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Preferred embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a thorough and complete understanding of the disclosure of the invention.

[0025] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0026] In one embodiment, such as Figure 1 As shown, a dynamic monitoring method for mine slopes is provided. This embodiment illustrates the method applied to a terminal, but it is understood that the method can also be applied to a server, or to a system including both a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0027] S1: Spatiotemporal registration and physical feature derivation processing are performed on monitoring data from multiple sources such as GNSS, inclinometers, strain gauges, and microseismic monitors. The original data from different spatiotemporal references are aligned to a unified grid and strain and energy release rate derivation features are calculated to generate a multi-source spatiotemporal fusion dataset.

[0028] Specifically, the system acquires multi-source raw monitoring data from the Global Navigation Satellite System (GNSS), inclinometers, strain gauges, and microseismic monitoring instruments. GNSS data records the three-dimensional coordinates and displacement increments of discrete points on the slope surface; inclinometer data reflects changes in the tilt angle of the monitored section; strain gauge data records changes in linear strain in a specific monitoring segment; and microseismic monitoring instrument data includes the occurrence time, spatial coordinates, magnitude, and amplitude of microseismic events. The spatiotemporal references of different sensor types differ. GNSS data uses a geodetic coordinate system, while microseismic monitoring instrument data uses event-triggered time sampling. The spatial installation locations of inclinometers and strain gauges are discretely distributed. These differences prevent the raw data from being directly used for subsequent fusion analysis; therefore, the system needs to perform spatiotemporal registration processing to eliminate data heterogeneity.

[0029] In the time registration stage, the system first identifies the sampling frequency of each sensor's data, determines the highest sampling frequency as the time reference, and generates a preset time interval sequence. For low sampling frequency data and event-based microseismic data, the system uses an adaptive interpolation algorithm to complete and synchronize the time dimension, ensuring that the timestamps of all data are uniformly aligned to the preset time interval sequence, thus ensuring consistency of data from different sensors in the time dimension. In the spatial registration stage, the system constructs a unified three-dimensional grid coordinate system based on the three-dimensional terrain model of the slope. This coordinate system is constructed based on the geographic coordinates of the area where the slope is located, combined with terrain elevation data to divide the grid cells. The system uses a coordinate transformation algorithm to map the discrete point data of the Global Navigation Satellite System and the installation position coordinates of the inclinometers and strain gauges to this three-dimensional grid coordinate system, achieving alignment of different spatial reference data to a unified grid, ensuring data consistency in the spatial dimension.

[0030] After spatiotemporal registration, the system calculates strain characteristics based on strain tensor calculation methods from elasticity, combining spatial gradient information from global navigation satellite system displacement data with the positional relationship of grid nodes. Through vector operations, it solves for strain characteristic parameters such as normal strain and shear strain at each node within the grid. For energy release rate calculation, the system uses energy calculation models from seismology, combining the magnitude, amplitude, and propagation distance of microseismic events, to mathematically derive the corresponding energy release rate parameter. This parameter characterizes the intensity of energy release during the internal fracturing process of the rock mass. The system associates and stores the registered raw data and derived strain and energy release rate characteristic parameters according to time series and grid spatial location, forming a multi-source spatiotemporal fusion dataset containing spatiotemporal dimensions, original monitoring dimensions, and physically derived characteristic dimensions.

[0031] S2: Based on a multi-source spatiotemporal fusion dataset, point displacement and linear strain observations are used as boundary conditions and substituted into the slope mechanics model to solve the full-field state, generating displacement-stress coupling data containing the full-field displacement field and full-field stress field of the slope.

[0032] Specifically, the system extracts spatiotemporally registered point displacement observations from the Global Navigation Satellite System and linear strain observations from strain gauges from a multi-source spatiotemporal fusion dataset, and substitutes these as boundary conditions into a pre-defined slope mechanics model. The slope mechanics model is constructed based on continuum mechanics theory. The system selects the corresponding constitutive equation based on the geological structural characteristics of the slope rock mass. These geological structural characteristics include rock mass type, joint distribution, and water content. The selection of the constitutive equation must match the mechanical response characteristics of the rock mass; common constitutive equations include elastoplastic and viscoelastic constitutive equations.

[0033] Preferably, to achieve accurate solutions for the overall mechanical state of the slope, the system employs the finite element method to mesh the entire slope. During mesh generation, the system adjusts the mesh density based on the spatial differences in rock mass mechanical parameters. The mesh element size in areas with drastic changes in mechanical parameters is smaller than that in areas with stable mechanical parameters. This approach controls computational load while ensuring solution accuracy. When substituting boundary conditions, the system applies point displacement observations from the Global Navigation Satellite System to the corresponding mesh nodes, fixing the displacement degrees of freedom of these nodes through constraint equations to ensure consistency between the mesh node displacements and the observed values. Strain gauge linear strain observations are also applied to the corresponding mesh elements, and element strain constraint equations ensure that the element strain conforms to the observed data.

[0034] Furthermore, the system constructs a mechanical equilibrium equation based on the principle of virtual work. This equation uses the displacement of the grid nodes as unknowns and comprehensively considers the constitutive relationship, boundary conditions, and external forces of the rock mass. The system employs the Newton-Raphson iterative algorithm to solve the mechanical equilibrium equation. During the iteration process, the displacement values ​​of the grid nodes are continuously adjusted, and the corresponding stresses and unbalanced forces are calculated until the unbalanced forces meet the preset convergence conditions. In the solution process, the system first solves the displacement field distribution of the entire slope through displacement boundary conditions and strain boundary conditions, obtaining the displacement components of each grid node in three-dimensional space. Then, based on the constitutive equation and the spatial derivative of the displacement field, the stress components such as normal stress and shear stress of each grid element are obtained through the stress calculation equation, forming the full-field stress field distribution. Since the displacement field and stress field are interconnected through the constitutive equation, changes in displacement will trigger adjustments in stress, and the distribution of stress will constrain the evolution of displacement, thus forming a coupling relationship between the two. The system correlates the full-field displacement field data with the full-field stress field data according to grid nodes and time series, generating displacement-stress coupled data that includes the spatial location, displacement components, and stress components of the slope across the entire domain, thus realizing the deduction from local discrete observation data to the continuous mechanical state of the entire domain.

[0035] S3: Based on the displacement-stress coupling data and the microseismic energy data in the multi-source spatiotemporal fusion dataset, the physical laws are quantified to generate a differentiable physical constraint loss function.

[0036] Specifically, the system uses displacement-stress coupled data and microseismic energy data from multi-source spatiotemporal fusion datasets as its foundation. It conducts quantitative analysis of the consistency of physical laws around the core principles of rock mechanics, including the stress balance law, the law of conservation of energy, and the rock mass fracture criterion. Quantitative analysis ensures the consistency between the data-driven process and the physical mechanisms. Preferably, the system performs divergence calculations on the full-field stress field in the displacement-stress coupled data based on the stress balance equations in continuum mechanics. The stress balance equations describe the stress equilibrium relationship within the rock mass, and the result of the divergence calculation is the stress balance residual for each grid cell. The stress balance residual reflects the degree of deviation of the stress field from the equilibrium state.

[0037] In the process of quantifying the consistency of energy conservation laws, the system calculates the change in strain energy across the entire slope using displacement-stress coupling data. This change in strain energy is obtained through the integration of stress and strain components. Simultaneously, the system extracts the microseismic energy release rate from the multi-source spatiotemporal fusion dataset and constructs an energy balance equation. This equation characterizes the conservation relationship between the change in strain energy and the microseismic energy release rate. The system calculates the matching error between the change in strain energy and the microseismic energy release rate, reflecting the degree of agreement between the energy evolution process and the conservation laws. In the process of quantifying the consistency of rock mass fracture laws, the system extracts the maximum principal stress distribution from the displacement-stress coupling data and determines the stress-exceeding region by combining it with the rock mass compressive strength parameters. Simultaneously, the system extracts the spatial distribution region of microseismic events and calculates the overlap between the stress-exceeding region and the spatial distribution region of microseismic events. This overlap reflects the correspondence between stress concentration and rock mass fracture.

[0038] Based on three consistent quantitative indicators—stress balance residual, energy matching error, and fracture zone overlap error—the system constructs a differentiable physical constraint loss function. This loss function uses a weighted summation form, assigning a corresponding weight coefficient to each quantitative indicator. The weight coefficients are set based on the importance of each physical law in slope stability analysis and engineering practice experience, and the sum of all weight coefficients is 1. The system multiplies the stress balance residual, energy matching error, and fracture zone overlap error by their respective weight coefficients and then sums the results to obtain the total physical constraint loss value. Since all components of the loss function are constructed through differentiable mathematical operations, this loss function possesses differentiability and is compatible with the gradient backpropagation process of deep learning models.

[0039] S4: Based on the gradient backpropagation algorithm, a slope displacement prediction model constrained by physical mechanism is generated by jointly training a multi-source spatiotemporal fusion dataset and displacement-stress coupling data.

[0040] Specifically, the system constructs a deep learning-based temporal prediction network architecture, which includes a feature extraction part, a temporal modeling part, and a prediction output part. The feature extraction part is implemented using a convolutional neural network, which extracts features from multi-source spatiotemporal fusion data through sliding operations of convolutional kernels, capturing the spatial and physical features in the data; the temporal modeling part uses a long short-term memory network or a Transformer model, which captures the temporal dependencies in the data through gating or self-attention mechanisms; the prediction output part uses a fully connected layer, which maps the features after temporal modeling to the predicted slope displacement values ​​at future times through linear transformation.

[0041] During model training, the system employs a joint training strategy, using a multi-source spatiotemporal fusion dataset as the model's input data and the full-field displacement field from the displacement-stress coupling data as the physical supervision signal. A differentiable physical constraint loss function is also introduced as a constraint term. In the initial training phase, the system randomly initializes the parameters of the prediction network, including convolutional kernel weights and network layer biases. The system inputs the multi-source spatiotemporal fusion data into the prediction network through a forward propagation process. The feature extraction part performs convolution and pooling operations on the input data to extract high-dimensional features; the temporal modeling part performs temporal encoding on the high-dimensional features to capture the dependencies between data at different time steps; and the prediction output part outputs the initial displacement prediction values.

[0042] The system calculates the prediction loss between the initial displacement prediction value and the actual displacement field in the displacement-stress coupling data. The prediction loss is calculated using the mean square error loss function, obtained by calculating the mean square difference between the predicted and actual values. Simultaneously, the system calculates the constraint loss using a differentiable physical constraint loss function, obtained through the aforementioned physical law consistency quantification index. The system weights and sums the prediction loss and constraint loss according to a preset ratio to obtain the total training loss. Based on the gradient backpropagation algorithm, the system calculates the gradient value of the total training loss with respect to the parameters of each layer of the prediction network using the chain rule. The gradient value reflects the degree of influence of parameter changes on the total training loss. Based on the gradient value, the system adjusts the network parameters using stochastic gradient descent or adaptive moment estimation optimization algorithms to reduce the total training loss. The above processes of forward propagation, loss calculation, gradient backpropagation, and parameter update are iteratively performed until the total training loss converges to a preset threshold or the number of iterations reaches the maximum, at which point training terminates. Through this joint training process, the prediction model not only learns the statistical characteristics of multi-source monitoring data but also embeds rock mechanics mechanisms into the model learning process through the physical constraint loss function, ultimately generating a slope displacement prediction model with physical mechanism constraints.

[0043] S5: Based on the slope displacement prediction model, forward prediction and physical consistency verification are performed on the real-time multi-source sensor monitoring data to generate a graded early warning signal that integrates the displacement prediction value and physical confidence level. The graded early warning signal is used to indicate different slope risk levels and treatment priorities.

[0044] Specifically, the system acquires multi-source sensor data in real time from the Global Navigation Satellite System, inclinometers, strain gauges, and microseismic monitoring instruments. The real-time acquired data undergoes the same spatiotemporal registration and physical feature derivation processing as S1. The spatiotemporal registration process for real-time data follows the same time reference and spatial grid coordinate system, while the physical feature derivation processing employs the same strain calculation method and energy release rate derivation logic. This ensures that the spatiotemporal reference and feature dimensions of the real-time data remain consistent with the model training data, generating a real-time multi-source spatiotemporal fusion dataset.

[0045] Preferably, the system inputs a real-time multi-source spatiotemporal fusion dataset into the trained slope displacement prediction model. Through the model's forward propagation process, it outputs the predicted displacement values ​​for the entire slope over a preset time period. The predicted displacement values ​​include the displacement increment, displacement rate, and cumulative displacement value for each grid node. To verify the reliability of the prediction results, the system performs a physical consistency check. During the check, the system substitutes the predicted displacement values ​​into the slope mechanics model to solve for the corresponding predicted stress field. Subsequently, the physical consistency quantification process in step S3 is repeated to calculate the stress balance residual of the predicted stress field, the matching error between the predicted strain energy and the real-time microseismic energy, and the overlap error between the predicted stress exceedance area and the distribution of real-time microseismic events. Based on these error indicators, the system generates a physical confidence score. The physical confidence score is negatively correlated with the error indicators. Through normalization, the physical confidence score is mapped to a fixed interval, the range of which covers the confidence level corresponding to the minimum to maximum error.

[0046] Furthermore, the system constructs a tiered early warning mechanism based on displacement prediction values ​​and physical confidence levels. Multiple sets of displacement thresholds and physical confidence level thresholds are preset. The displacement thresholds are set based on the mechanical analysis results of the slope stability state, while the physical confidence level thresholds are set based on the requirements for prediction reliability in engineering practice. When the predicted displacement value is lower than the first set of displacement thresholds and the physical confidence level is higher than the first set of physical confidence level thresholds, the system generates a first-level early warning signal. When the predicted displacement value is between the first and second sets of displacement thresholds, or the physical confidence level is lower than the first set of physical confidence level thresholds, the system generates a second-level early warning signal. When the predicted displacement value is between the second and third sets of displacement thresholds, the system generates a third-level early warning signal. When the predicted displacement value is higher than the third set of displacement thresholds, the system generates a fourth-level early warning signal. Each level of early warning signal includes information such as the predicted displacement value, physical confidence level, risk level, handling priority, and specific handling suggestions. The system transmits the tiered early warning signals to the terminal equipment of the mine safety monitoring center in a standardized data format. After receiving the early warning signals, the terminal equipment displays and stores them, ensuring that relevant management personnel can obtain early warning information in a timely manner and take corresponding handling measures, thereby achieving dynamic monitoring and accurate early warning of mine slopes.

[0047] In summary, the dynamic monitoring method for mine slopes provided in this application constructs a displacement-stress coupling field and loss function that can be directly verified for physical consistency by fusing multi-source heterogeneous monitoring data and performing mechanical inversion, thereby constraining the optimization direction of the model. This ensures that the output of the final trained prediction model not only fits historical data but also inherently follows fundamental mechanical principles such as static equilibrium, energy conservation, and fracture criteria, significantly enhancing the model's extrapolation and generalization capabilities and the interpretability of prediction results under data-scarce or unknown extreme conditions. Furthermore, this method introduces a dual verification mechanism based on physical confidence in the early warning decision-making stage, which can proactively identify and warn of abnormal predictions that, while conforming to statistical trends, contradict physical laws. This effectively overcomes the risks of false alarms and missed alarms caused by "black box" decision-making in complex conditions under existing technologies, achieving a systematic improvement in the accuracy, reliability, and engineering practicality of early warnings.

[0048] In one embodiment, S1 of the dynamic monitoring method for mine slopes provided by the present invention specifically includes the following steps:

[0049] S11: Perform spatial coordinate mapping processing on monitoring data from multiple sources of GNSS, inclinometers, strain gauges, and microseismic monitors. Interpolate and align the data of each sensor to equal-interval timestamps in the time dimension, and map them to corresponding nodes of a unified grid in the spatial dimension to generate basic aligned data with a unified spatiotemporal reference.

[0050] Specifically, the system receives raw monitoring data transmitted from the Global Navigation Satellite System (GNSS), inclinometers, strain gauges, and microseismic monitoring instruments. The observation principles of different sensors determine the output format of the raw data: the GNSS outputs the three-dimensional spatial coordinates of the monitoring points and their corresponding timestamps; the inclinometers output the time-series data of the tilt angle of the monitoring section; the strain gauges output continuous records of local strain in the rock mass; and the microseismic monitoring instruments output the occurrence time, spatial location, and waveform-related parameters of microseismic events. Due to differences in sensor deployment locations and sampling mechanisms, these raw data exhibit variations in spatiotemporal references, requiring unified processing to eliminate heterogeneity.

[0051] Furthermore, the system performs spatial coordinate mapping processing, calling a preset coordinate transformation protocol to convert the original spatial coordinates of all sensors to a preset geodetic coordinate system. The selection of this coordinate system is based on the geographical location of the mine slope and engineering monitoring specifications. The system constructs a three-dimensional mesh model based on slope topographic survey data. The mesh area covers the monitored slope and surrounding affected areas, and the mesh division method is determined by combining slope topographic features and the distribution of monitoring points. The system employs a spatial interpolation algorithm to map discrete GPS monitoring point data, linearly distributed strain gauge data, and event-triggered microseismic monitoring data one by one to the corresponding nodes of the three-dimensional mesh model, achieving spatial data unification.

[0052] In the time dimension processing, the system extracts the timestamp information from the raw data of each sensor and analyzes the differences in sampling frequencies among different sensors. Based on monitoring requirements, the system determines a unified, equally spaced time step and uses a time interpolation algorithm to adjust the data from all sensors. For data with a sampling frequency higher than the unified time step, the system extracts the corresponding timestamp data through downsampling; for data with a sampling frequency lower than the unified time step or missing data, the system supplements the monitoring data corresponding to the missing timestamps through interpolation. After completing the spatiotemporal processing, the system integrates all data to generate basic aligned data with a unified spatiotemporal reference.

[0053] S12: Based on the central difference method, spatial gradient calculation is performed on the GNSS displacement sequence in the basic alignment data, the displacement change rate of adjacent grid nodes in three dimensions is calculated, and a strain tensor field characterizing the deformation distribution of the slope surface is generated.

[0054] Specifically, the system calls upon the Global Navigation Satellite System displacement sequence from the basic alignment data. This sequence contains three-dimensional displacement data of each grid node at different equally spaced timestamps. To obtain strain information characterizing the deformation distribution of the slope surface, the system uses the central difference method to calculate the spatial gradient of the displacement sequence. The central difference method solves for the displacement change rate by the displacement difference between adjacent grid nodes. Based on the topology of a unified three-dimensional spatial grid, the system determines the set of adjacent grid nodes for each grid node. The selection of adjacent nodes follows the grid spatial partitioning rules to ensure accurate reflection of the displacement changes of the target node in the three-dimensional direction. For each grid node, the system extracts its displacement data in the x, y, and z spatial directions, and simultaneously extracts the displacement data of adjacent nodes in the corresponding directions. The first-order partial derivatives of the target node's displacement in each direction are calculated using the central difference formula. These partial derivatives represent the displacement change rate of adjacent grid nodes in the corresponding three-dimensional direction.

[0055] Preferably, based on the displacement rate in three dimensions, the system solves for the normal strain and shear strain components of each grid node according to the composition relationship of the strain tensor in elasticity. Each strain component is combined according to the tensor structure to form a complete strain tensor field. Each component in the strain tensor field corresponds to the deformation characteristics of the grid node in a specific direction, which can comprehensively reflect the tensile, compressive, and shear deformation distribution of the slope surface.

[0056] S13: Based on the list of microseismic events in the basic aligned data, calculate the occurrence density and total energy released by microseismic events on the spatial grid per unit time, and generate time series data of spatial microseismic density distribution and microseismic energy release rate.

[0057] Specifically, the system extracts a list of microseismic events from the base alignment data. This list includes the spatiotemporal coordinates and energy parameters of each microseismic event. To quantify the spatiotemporal distribution characteristics and energy release intensity of microseismic events, the system uses a unified three-dimensional spatial grid as a basis, dividing the statistical regions of microseismic events by grid cells. The division of statistical regions is consistent with the previous spatial mapping grid. The system traverses the list of microseismic events, determines the grid cell to which each microseismic event belongs based on its spatial coordinates, and counts the number of microseismic events occurring in each grid cell per unit time. This number is the spatial microseismic density of the corresponding grid cell.

[0058] In the calculation of microseismic energy release rate time series data, the system groups microseismic events according to equally spaced time stamps, grouping all microseismic events corresponding to the same time stamp into one group, and accumulating the energy parameters of each group of microseismic events to obtain the total released energy of the microseismic events at each time stamp. The system correlates the total released energy with unit time intervals to obtain the microseismic energy release rate per unit time, and arranges it in time stamp order to form microseismic energy release rate time series data. The spatial microseismic density distribution reflects the spatial aggregation characteristics of fracture events within the rock mass, and the microseismic energy release rate time series data reflects the energy evolution law during the rock mass fracture process.

[0059] S14: The basic alignment data, strain tensor field, spatial microseismic density distribution and microseismic energy release rate time series data are fused and encapsulated, and multi-dimensional spatiotemporal information is integrated into a unified data structure to generate a multi-source spatiotemporal fusion dataset.

[0060] Specifically, the system collects time-series data on the spatial microseismic density distribution and microseismic energy release rate of the strain tensor field from the basic alignment data. These data reflect the slope's state information from four dimensions: original monitored deformation characteristics, microseismic distribution, and energy release. To achieve the collaborative application of multi-dimensional information, the system needs to perform data fusion and encapsulation processing to construct a unified data structure. The system designs a unified data structure adapted to multi-dimensional spatiotemporal information. This structure uses timestamps and grid nodes as core indexes, enabling rapid association and access to data from various dimensions through these core indexes. During the data fusion process, the system stores the original monitoring values ​​of each sensor in the basic alignment data according to timestamps and grid node indices. After associating each strain component in the strain tensor field with its corresponding grid node, it integrates them into the unified data structure according to timestamps. The system stores the spatial microseismic density distribution indexed by grid cells and timestamps, and stores the time-series data of microseismic energy release rate sequentially according to timestamps.

[0061] Furthermore, the system standardizes the format of each data component, clearly defining the data storage type, field definitions, and indexing rules to ensure data consistency and accessibility. Through a fusion and encapsulation process, the system integrates the original monitoring information, deformation characteristics of the strain tensor field, spatial clustering information of the microseismic density distribution, and energy evolution information of the microseismic energy release rate time series data from the basic aligned data into a unified data structure, forming a multi-source spatiotemporal fusion dataset. This dataset provides comprehensive and interconnected multi-dimensional input data for subsequent slope mechanics model solving, physical constraint loss function construction, and prediction model training.

[0062] In one embodiment, step S2 of the dynamic monitoring method for mine slopes provided by the present invention specifically includes the following steps:

[0063] S21: Based on the GNSS point displacement data and fiber optic strain gauge linear strain data in the multi-source spatiotemporal fusion dataset, the boundary conditions and observation constraints are formatted, the discrete point displacements are converted into displacement constraints of the corresponding nodes of the unified grid, and the distributed linear strains are converted into strain field constraints of the corresponding elements of the unified grid, generating a set of boundary conditions for mechanical solutions.

[0064] Specifically, the system extracts point displacement data from the global navigation satellite system and linear strain data from fiber optic strain gauges from the multi-source spatiotemporal fusion dataset. The two types of data have discreteness and format heterogeneity. Directly substituting them into the slope mechanics model will lead to insufficient solution convergence. Therefore, the system needs to perform formatting processing of boundary conditions and observation constraints.

[0065] Preferably, the system first performs spatial matching processing on the point displacement data of the Global Navigation Satellite System. Based on the node coordinate system of a unified three-dimensional spatial grid, the spatial coordinates of the discrete point displacement data are accurately matched with the grid nodes. Through constraint mapping rules, the matched point displacements are converted into displacement constraints of the corresponding grid nodes, clarifying the displacement degree-of-freedom restriction mode of the constraint nodes. For the fiber optic strain gauge linear strain data, the system first analyzes the geometric parameters of the fiber optic strain gauge deployment path. Based on the spatial overlap area between the path and the unified grid cell, the linear strain data is expanded into global strain constraints of the corresponding grid cell through strain field interpolation rules, so that the strain constraints cover all integration points within the cell. The system performs format standardization processing on the converted displacement constraints and strain constraints, unifies the data storage fields and indexing rules, and forms a boundary condition set containing a list of grid node displacement constraints and a list of grid cell strain constraints. This boundary condition set can be directly adapted to the input specifications of the slope mechanics model, providing compliant constraint input for subsequent full-field state solution.

[0066] S22: The displacement-strain joint constraint set and the preset slope rock mass mechanical parameters are processed by the full-field state solution based on the finite element method. The joint constraint set is substituted as the known boundary conditions into the control equation of the slope mechanical model to solve the displacement response of all unknown nodes in the model and generate the full-field displacement field of the slope.

[0067] Specifically, the system uses the displacement-strain joint constraint set as known boundary conditions and substitutes it into the pre-defined governing equations of the slope mechanics model. These governing equations are constructed based on the principle of virtual work and describe the intrinsic relationship between displacement, stress, and strain of the slope rock mass under static equilibrium. The system uses the finite element method to discretize the governing equations. First, the entire slope domain is divided into polyhedral elements. Then, the interpolation relationship between the displacement of the element nodes and the displacement within the elements is established through element shape functions, transforming the continuous governing equations into a discretized linear system of equations.

[0068] Preferably, the system can use the Newton-Raphson iterative method to solve the discretized linear equations, gradually correcting the nodal displacement values ​​through iterative calculations until the residuals of the equations meet the preset convergence conditions, thus obtaining the three-dimensional displacement components of all unknown grid nodes. Based on the topology of a unified three-dimensional spatial grid, the system integrates and sorts the displacement components of all nodes according to their node numbers, generating a full-field displacement field covering the entire slope region. This displacement field completely characterizes the spatial displacement distribution of each area of ​​the slope.

[0069] S23: The stress field is reconstructed based on the constitutive relationship of the rock mass for the displacement field of the entire slope. The strain of each element is calculated according to the displacement field, and the stress state of the element is calculated according to the stress-strain relationship of the material to generate the stress field of the entire slope.

[0070] Specifically, the system first calculates the strain components of each grid element using geometric equations. These equations describe the geometric relationship between displacement and strain. The system utilizes the displacement differences between adjacent nodes in the full-field displacement field, combined with the derivative matrix of the element shape function, to solve for the normal strain, shear strain, and other total strain components of each grid element. Subsequently, the system calls upon a preset rock mass constitutive relation. This constitutive relation is selected based on the mechanical properties of the slope rock mass and can be either elastoplastic or viscoelastic. The system substitutes the calculated element strain components into the constitutive equations and calculates the normal stress, shear stress, and other stress components of each grid element using the stress solution matrix.

[0071] Preferably, the system integrates the stress components of all cells according to their grid cell numbers, establishes a spatial correspondence between stress components and grid cells, and generates a full-field stress field covering the entire slope. This full-field stress field accurately reflects the stress distribution characteristics within the slope's rock mass, providing core mechanical parameters for subsequent slope stability analysis.

[0072] S24: Couple and encapsulate the whole-field displacement field and the whole-field stress field of the slope to establish a one-to-one correspondence between the displacement field and the stress field on the spatial grid, and generate displacement-stress coupled data.

[0073] Specifically, the system uses the node and element numbers of a unified three-dimensional spatial mesh as the core association key to establish a one-to-one correspondence between the displacement field and the stress field. For mesh nodes, the system binds the three-dimensional displacement components of the nodes in the displacement field with the stress components of the corresponding nodes in the stress field through the node numbers; for mesh elements, the system binds the strain distribution of the elements in the displacement field with the stress distribution of the corresponding elements in the stress field through the element numbers. During the binding process, spatial coordinate consistency verification is performed to ensure the spatial matching accuracy of the associated data.

[0074] Furthermore, the system constructs a coupled data structure with a three-level storage architecture, using timestamps as the first-level index, spatial grid numbers as the second-level index, and displacement-stress components as the third-level index. The bound displacement and stress field data are integrated and stored according to this architecture. Through the above coupling and encapsulation process, the system generates displacement-stress coupled data, which fully preserves the intrinsic mechanical relationship between displacement and stress, providing collaborative mechanical data support for subsequent consistency quantification of physical laws and training of prediction models.

[0075] In one embodiment, step S3 of the dynamic monitoring method for mine slopes provided by the present invention specifically includes the following steps:

[0076] S31: Perform equilibrium equation residual calculation on the whole field stress field in the displacement-stress coupling data, calculate the magnitude of the divergence of the stress tensor and the sum of the volume forces for each element, and generate the equilibrium constraint residual field.

[0077] Specifically, the system extracts the stress tensor of each grid cell in the full-field stress field. Simultaneously, the preset volume force parameters are invoked. The volume force parameters are determined based on the gravity field distribution characteristics of the slope, and directly reflect the influence of the slope rock mass's own gravity on mechanical equilibrium.

[0078] Based on the equilibrium equations of continuum mechanics, the system performs residual calculations element by element. The mathematical expression of the equilibrium equations is:

[0079]

[0080] in, It represents the divergence of the stress tensor, describing the rate of change of stress in space. For stress tensor, This represents the volume force vector. During the calculation, the system first spatially discretizes the stress tensor of each element using finite element shape functions, and then solves for the divergence of the stress tensor. Then, this divergence is compared with the volume force vector of the corresponding element. Perform vector summation, and take the modulus of the summation result to obtain the equilibrium equation residuals for each element. The calculation expression is as follows:

[0081]

[0082] in, This represents the residual of a single-unit equilibrium equation.

[0083] Furthermore, the system spatially integrates the equilibrium equation residuals of all grid cells. Based on the cell numbering order of the unified grid, the residual value of each cell is precisely mapped to its corresponding spatial location, constructing an equilibrium-constrained residual field. During the integration process, the system simultaneously verifies the spatial continuity of the residual data and smooths out abrupt changes in local residuals caused by numerical calculation errors, ensuring that the residual field accurately reflects the degree to which the stress field satisfies the equilibrium equations across the entire field. Each grid cell in the equilibrium-constrained residual field corresponds to a unique residual value, and the magnitude of the residual value directly characterizes the rationality of the mechanical equilibrium of the stress distribution in that region.

[0084] S32: Perform energy temporal synergy analysis on the full-field displacement field and microseismic energy data in the displacement-stress coupling data. Calculate the time change rate of the overall strain energy of the slope from the displacement field and compare it with the microseismic energy release rate within the same time window to generate an energy-constrained residual sequence.

[0085] Specifically, the system first calculates the overall strain energy of the slope based on the full-field displacement data. The strain energy calculation follows the strain energy solution principle of the finite element method, and the strain energy of each grid element is calculated. Solve using the formula:

[0086]

[0087] in, For the element strain tensor, For element stress tensor, The element volume is the sum of the strain energy of all mesh elements in the system. The overall strain energy of the slope is obtained. .

[0088] Furthermore, the system calculates the rate of change of the overall strain energy over time by using the ratio of the difference in overall strain energy between two consecutive time points to the time interval. The calculation expression is as follows:

[0089]

[0090] in, The strain energy is the rate of change over time. For the overall strain energy at the next time stamp, The overall strain energy at the current timestamp. The time interval between the two timestamps is defined. The system then sets time windows based on the frequency of microseismic events, dividing the continuous time dimension into several equally spaced time segments to ensure that each time window contains sufficient microseismic event data to guarantee the reliability of the analysis.

[0091] For each time window, the system extracts the microseismic energy release rate data within that window. Compare it with the strain energy time change rate calculated in the same period. A quantitative comparison was performed, and the energy constraint residual for a single time window was obtained by calculating the difference between the two. The expression is as follows:

[0092]

[0093] in, The system generates an energy-constrained residual sequence by sequentially sorting the energy-constrained residuals of each time window. This sequence directly reflects the temporal coordination between changes in slope strain energy and microseismic energy release, and the magnitude of the residual values ​​characterizes the degree of deviation of the energy conversion process from the conservation law.

[0094] S33: Based on the Mohr-Coulomb criterion, the full-field stress field in the displacement-stress coupling data is processed to evaluate the rupture proximity. The algebraic distance from the current stress state to the rupture envelope is calculated for each element to generate the rupture proximity index field.

[0095] Specifically, the system acquires full-field stress field data from the displacement-stress coupling data, and simultaneously calls preset Mohr-Coulomb criterion parameters. These parameters are determined based on the lithological investigation results of the slope rock mass, and the core parameters include rock mass cohesion. and internal friction angle This directly determines the accuracy of the criterion in determining the fracture state of the rock mass. The system first extracts the stress components of the entire stress field element by element, focusing on extracting the normal stress of the element. and shear stress The actual stress state of each element is clearly defined. The fracture proximity is calculated based on the Mohr-Coulomb criterion, and the limit equilibrium condition expression for the Mohr-Coulomb criterion is as follows:

[0096]

[0097] in, The shear strength of the rock mass. For cohesion, Normal stress, Let be the internal friction angle. The system uses this formula to calculate the shear strength of each element. Then calculate the shear stress of the current element. With shear strength The algebraic distance is calculated using the following expression:

[0098]

[0099] in, This is an indicator of rupture proximity. When... When, it indicates that the current stress state of the element has not reached the rupture limit; when When this occurs, it indicates that the current stress state of the element has met the rupture condition. The system will then display the rupture proximity index for each element. Following the cell distribution pattern of a unified grid, the system maps the data to the spatial locations of the corresponding cells to generate a fracture proximity index field. During the generation process, the system verifies the spatial continuity of the index data and corrects abrupt changes in the index caused by discontinuous stress transmission at cell boundaries, ensuring that the index field accurately reflects the proximity of the rock mass in each area of ​​the slope to the fracture state. This index field provides a quantitative basis for fracture risk in the subsequent construction of physical constraints, enabling a comprehensive and accurate assessment of the fracture state of the slope rock mass.

[0100] S34: The equilibrium constraint residual field, energy constraint residual sequence and rupture proximity index field are weighted and integrated in space-time. The residuals of the three physical dimensions are superimposed according to the preset residual weights and integrated in the computational domain to generate a differentiable physical constraint loss function.

[0101] Specifically, the system presets weighting coefficients for the residuals in three physical dimensions, which are respectively the weights of the balance constraints. Energy constraint weights and breakage constraint weight The weighting coefficients are determined based on the importance of each physical constraint in the slope stability analysis, and satisfy the following conditions: This ensures the mechanical rationality and scientific nature of the fusion process.

[0102] Preferably, the system performs weighted summation of the three types of residual data according to preset weight coefficients to obtain comprehensive residual data across the entire domain. The weighted summation expression is as follows:

[0103]

[0104] in, To integrate residuals, For spatial coordinate variables, For time variables, To balance the spatial distribution function of the constrained residual field, Let be the time distribution function of the energy-constrained residual sequence. This represents the spatial distribution function of the rupture proximity index field. Subsequently, spatial-temporal integration is performed, integrating the comprehensive residual data along the entire slope computational domain in the spatial dimension. To perform integration, the spatial integration result is integrated along a set time series T in the time dimension. The double integral expression is as follows: .

[0105] Furthermore, the system constructs a differentiable physical constraint loss function based on the double integration results, and the final loss function expression is as follows: This expression employs a differentiable mathematical form, ensuring that the loss function can participate in the gradient backpropagation process of subsequent deep learning models. It can accurately quantify the degree of deviation between the prediction results and the core laws of rock mechanics. Once constructed, this loss function provides strict physical constraints for subsequent model training, effectively avoiding the shortcomings of traditional data-driven models that deviate from the laws of mechanics.

[0106] In one embodiment, step S4 of the dynamic monitoring method for mine slopes provided by the present invention specifically includes the following steps:

[0107] S41: Construct training sample pairs for key physical features in multi-source spatiotemporal fusion datasets and displacement-stress coupling data. Use historical time-series fusion data as input features, future time-series GNSS displacements as prediction labels, and the statistics of the full-field displacement field and stress field in the same period as physical state labels to generate a spatiotemporal sample set for joint training.

[0108] Specifically, the system extracts key physical features from the multi-source spatiotemporal fusion dataset and displacement-stress coupled data. These key physical features encompass original monitoring features, derived strain features, microseismic energy features, and related features of the full-field displacement and stress fields, serving as the core data source for sample construction. The system initiates the construction process for training sample pairs, first setting time series division rules and using a sliding time window method to divide historical time series segments into future prediction segments. The historical time series segments are used to extract input features, while the future prediction segments are used to determine prediction labels.

[0109] Preferably, the system concatenates multi-source spatiotemporal fusion data within historical timeframes into an input feature vector in chronological order. ,in The fused data features at time t, This represents the length of the historical time series window. Simultaneously, GNSS displacement data within future time series segments will be used as prediction labels. ,in The GNSS displacement value at time p in the future. To predict the future step size, the system further extracts full-field displacement and stress field data from the same historical time series, calculates their statistics as physical state labels s, and includes the mean, extreme values, and distribution characteristic parameters of the field data. The system then matches the input feature vector X, the predicted label y, and the physical state label s one-to-one to form complete training sample pairs. By traversing all time segments that meet the criteria, batch training sample pairs are generated and integrated into a spatiotemporal sample set for joint training after format standardization. The system performs validity verification on the spatiotemporal sample set, removing sample pairs with missing features or abnormal labels.

[0110] S42: Construct a hybrid neural network that includes a data-driven branch and a physical perception branch. Perform dual-path encoding and fusion processing on the input features in the spatiotemporal sample set. Learn the temporal evolution law of the monitoring data through the data-driven branch, and encode the mechanical state characteristics of the slope in the same period through the physical perception branch. After weighted fusion by the attention mechanism, output the future multi-step displacement prediction sequence.

[0111] Specifically, the system constructs a hybrid neural network comprising a data-driven branch and a physical sensing branch, clearly defining the structural design and functional positioning of the two branches. The data-driven branch adopts a temporal neural network structure to learn the temporal evolution patterns of the monitoring data. The input of this branch is the input feature vector X of the spatiotemporal sample set. It extracts the temporal correlation features of the data through multiple temporal coding layers. The temporal coding layers adopt a gated recurrent unit or Transformer encoder structure to capture long-term temporal dependencies.

[0112] The physical perception branch employs a spatial feature encoding structure to encode the mechanical state features of the slope during the same period. The input to this branch is the physical state label 's'. Through convolutional or fully connected layers, feature mapping of the mechanical state statistics is performed to generate fixed-dimensional physical feature vectors, thus achieving quantified encoding of the mechanical state information. The system introduces an attention fusion mechanism at the dual-branch output end to calculate the attention weights of the dual-branch feature vectors. and ,in Data-driven branch feature weights For the physical perception branch feature weights, satisfying .

[0113] Furthermore, the system performs weighted fusion of the two-branch feature vectors using attention weights, and the fusion expression is as follows:

[0114]

[0115] in, To fuse feature vectors, Data-driven branch output features The physical perception branch outputs features. After the fused feature vector is decoded by the output layer, it outputs a multi-step displacement prediction sequence. This completes the dual-path encoding and fusion prediction of the input features.

[0116] S43: Substitute the predicted sequence of future multi-step displacements as virtual observations into the slope mechanics model to deduce the corresponding virtual displacement field and virtual stress field, and evaluate the physical law conformity of the predicted sequence of future multi-step displacements based on the differentiable physical constraint loss function to generate the predicted physical loss value.

[0117] Specifically, the system will output the future multi-step displacement prediction sequence from the hybrid neural network. As virtual observations, these observations are formatted and converted into a boundary condition format recognizable by the slope mechanics model. Following boundary condition processing rules, future multi-step displacement predictions are converted into virtual displacement constraints for the corresponding grid nodes. The system then substitutes these virtual displacement constraints into the preset slope mechanics model and solves the model's governing equations to obtain the corresponding virtual displacement field. With virtual stress field ,in For spatial coordinate variables, This is for predicting future moments. Furthermore, the system calls the constructed differentiable physical constraint loss function to apply the virtual displacement field... With virtual stress field Substitute the loss function into the calculation to evaluate whether the mechanical state corresponding to the future multi-step displacement prediction sequence conforms to the core laws of rock mechanics.

[0118] The result of the loss function calculation is the predicted physical loss value. This value quantitatively characterizes the degree of conformity between the predicted multi-step displacement sequence and the physical laws. The smaller the predicted physical loss value, the smaller the deviation of the prediction result from the mechanical laws, and the higher the physical rationality. The system records the predicted physical loss value.

[0119] S44: Based on the gradient backpropagation algorithm, the data loss and predicted physical loss values ​​between the future multi-step displacement prediction sequence and the real displacement labels obtained from the spatiotemporal sample set are jointly optimized and iterated. The parameters of each layer of the network are dynamically adjusted with the goal of minimizing the total loss. Through multiple iterations, the network prediction is made to approach the historical data patterns and physical mechanism constraints at the same time, and a slope displacement prediction model with physical mechanism constraints is generated.

[0120] Specifically, the system constructs a total loss function, which is composed of data loss. Compared with the predicted physical loss value Weighted composition, the expression is:

[0121]

[0122] in, This is the physical loss weighting coefficient, used to adjust the proportion of physical constraints in the total loss. Its value is determined based on the physical consistency requirements of model training. (Data loss) Predicting sequences based on future multi-step displacements With the real displacement labels in the spatiotemporal sample set The calculation yielded the result using the mean square error loss formula. ,in Predict the step size for the future.

[0123] Preferably, the system can use a gradient backpropagation algorithm to calculate the gradient of the total loss function with respect to the parameters of each layer of the hybrid neural network. The gradient calculation starts from the output layer and propagates back along the network to the input layer, obtaining the gradient vectors of all trainable parameters. Based on the gradient vectors, the system calls the gradient descent optimization algorithm to dynamically adjust the parameters of each layer of the network according to the gradient direction, reducing the value of the total loss function. After each parameter adjustment, the system uses a validation set to verify the model performance and calculates the total loss value on the validation set. The system repeats the iterative process of parameter adjustment and performance verification. When the total loss value on the validation set remains stable for multiple consecutive times or reaches a preset convergence threshold, the iteration stops. At this point, the hybrid neural network has simultaneously learned the patterns of historical data and the constraints of rock mechanics physical mechanisms, and the prediction results combine data fitting accuracy and physical rationality. The system saves the network parameters at this point and generates a slope displacement prediction model constrained by physical mechanisms. This model can be used for subsequent real-time prediction and risk warning of slope displacement.

[0124] In one embodiment, step S5 of the dynamic monitoring method for mine slopes provided by the present invention specifically includes the following steps:

[0125] S51: Perform spatiotemporal registration and feature derivation processing on the real-time multi-source sensor monitoring data to generate a real-time multi-source spatiotemporal fusion dataset. Based on the slope displacement prediction model, perform forward prediction calculation processing on the multi-source spatiotemporal fusion dataset to output the displacement prediction sequence and change trend of key points of the slope in the future period and generate displacement prediction values.

[0126] Specifically, the system continuously receives multi-source real-time monitoring data from GNSS, inclinometers, strain gauges, and microseismic monitoring instruments. The system first performs preprocessing on the real-time data, removing abnormal data caused by sensor malfunctions or transmission interference. The determination of abnormal data is based on a preset data stream stability standard. Further, the system follows the spatiotemporal registration and feature derivation process established in step S1 to process the preprocessed real-time data: in the time dimension, real-time data with different sampling frequencies are interpolated and aligned to a unified, equally spaced timestamp; in the spatial dimension, discrete real-time monitoring data are mapped to preset unified grid nodes and cells; and derived features such as the real-time strain tensor field and microseismic energy release rate are calculated based on rock mechanics principles, ultimately generating a real-time multi-source spatiotemporal fusion dataset.

[0127] Preferably, the system calls a pre-trained physical mechanism-constrained slope displacement prediction model, inputting a real-time multi-source spatiotemporal fusion dataset into the model for forward prediction calculations. During the prediction process, the model mines the temporal evolution features of real-time data through a data-driven branch, encodes the synchronous mechanical state correlation information through a physical perception branch, and after processing by an attention fusion mechanism, decodes and outputs the displacement prediction sequence of key points on the slope within a preset future time period. Based on the displacement prediction sequence, the system calculates the displacement change trend, obtains the displacement change rate and acceleration features through linear fitting or difference operations, integrates the displacement prediction sequence and change trend features, and generates a complete displacement prediction value.

[0128] S52: The predicted displacement value is substituted into the slope mechanics model as a virtual boundary condition for physical state deduction. The corresponding virtual displacement field and virtual stress field are calculated. Based on the differentiable physical constraint loss function, the deviation between the virtual stress field and the static equilibrium principle, the deviation between the energy change rate derived from the virtual displacement field and the real-time microseismic energy release rate, and the proximity deviation between the virtual stress state and the Mohr-Coulomb fracture criterion are quantified to generate a physical confidence index.

[0129] Specifically, the system substitutes virtual displacement constraints into a preset slope mechanics model, solves the model's governing equations, and derives the virtual displacement field corresponding to future moments. With virtual stress field Preferably, the system calls the constructed differentiable physical constraint loss function to quantify three types of core deviations: First, by substituting the virtual stress field into the equilibrium equation, the modulus of the sum of the stress tensor divergence and the volume force is calculated to obtain the deviation between the virtual stress field and the static equilibrium principle. Secondly, the time-varying rate of strain energy is calculated based on the virtual displacement field and compared with the energy release rate of real-time microseismic events during the same period to obtain the deviation of the energy change rate. Third, based on the virtual stress field, the algebraic distance of each grid element from the envelope surface of the Mohr-Coulomb fracture criterion is calculated to obtain the fracture proximity deviation. The system uses a weighted summation method to comprehensively quantify the three types of deviations, generating a physical confidence index. The calculation expression is as follows:

[0130]

[0131] in, This is the physical confidence index. For preset weighting coefficients, These are the preset maximum thresholds for the three types of bias. The physical confidence index ranges from [0,1], with a larger value indicating stronger physical rationality of the prediction result.

[0132] S53: Based on the magnitude and acceleration of displacement prediction and the physical confidence index, perform multi-threshold fuzzy logic comprehensive judgment, map different data combinations to specific risk levels and handling priorities according to preset rules, and generate graded early warning signals.

[0133] Specifically, the system pre-defines a multi-dimensional threshold system, covering the magnitude threshold of the displacement prediction value, the displacement change acceleration threshold, and the physical confidence index threshold. The threshold system is determined based on mine slope safety management regulations and engineering practice experience, with different threshold ranges corresponding to different levels of risk correlation. The system initiates a multi-threshold fuzzy logic comprehensive judgment process, using the displacement prediction value magnitude, displacement change acceleration, and physical confidence index as input variables to construct a fuzzy logic inference rule base. The inference rule base clarifies the mapping relationship between different combinations of input variables and risk levels: when the displacement prediction value magnitude is below the low threshold, the change acceleration is gradual, and the physical confidence index is above the high threshold, it is judged as a low-risk level; when the displacement prediction value magnitude is in the middle threshold range, the change acceleration increases, and the physical confidence index is in the middle threshold range, it is judged as a medium-risk level; when the displacement prediction value magnitude is above the high threshold, the change acceleration increases sharply, or the physical confidence index is below the low threshold, it is judged as a high-risk level.

[0134] Furthermore, based on fuzzy logic reasoning, the system maps different risk levels to corresponding response priorities: low risk levels correspond to increased frequency of routine monitoring, medium risk levels correspond to special inspections and reinforcement preparations, and high risk levels correspond to emergency evacuation and area lockdown. Finally, the system integrates risk level and response priority information to generate tiered early warning signals, which are then pushed to the mine safety monitoring center and on-site terminal equipment through preset transmission channels.

[0135] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0136] Based on the same inventive concept, this application also provides a dynamic monitoring system for mine slopes to implement the aforementioned dynamic monitoring method. The solution provided by this system is similar to the implementation described in the above method; therefore, the specific limitations of one or more embodiments of the dynamic monitoring system for mine slopes provided below can be found in the limitations of the dynamic monitoring method for mine slopes described above, and will not be repeated here.

[0137] Preferably, such as Figure 2 As shown, the present invention provides a dynamic monitoring system 600 for mine slopes, which is configured with the following modules:

[0138] The multi-source data fusion processing module 610 is used to perform spatiotemporal registration and physical feature derivation processing on multi-source sensor monitoring data from GNSS, inclinometers, strain gauges, and microseismic monitors. It aligns the original data from different spatiotemporal references to a unified grid and calculates strain and energy release rate derived features to generate a multi-source spatiotemporal fusion dataset.

[0139] The slope coupled field solution module 620 is used to solve the whole field state of the slope mechanical model based on the multi-source spatiotemporal fusion dataset, by substituting the point displacement and linear strain observations as boundary conditions into the slope mechanical model, and generating displacement-stress coupled data containing the whole field displacement field and the whole field stress field of the slope.

[0140] The physical constraint loss generation module 630 is used to perform physical law consistency quantification based on microseismic energy data in displacement-stress coupling data and multi-source spatiotemporal fusion dataset, and generate a differentiable physical constraint loss function.

[0141] The prediction model joint training module 640 is used to jointly train multi-source spatiotemporal fusion datasets and displacement-stress coupling data based on the gradient backpropagation algorithm to generate a slope displacement prediction model constrained by physical mechanism.

[0142] The graded early warning signal generation module 650 is used to perform forward prediction and physical consistency verification on real-time multi-source sensor monitoring data based on the slope displacement prediction model, and generate graded early warning signals that fuse displacement prediction values ​​and physical confidence levels. The graded early warning signals are used to indicate different slope risk levels and handling priorities.

[0143] Preferably, the multi-source data fusion processing module 610 provided in this application is configured with the following units:

[0144] The spatiotemporal reference unification processing unit is used to perform spatial coordinate mapping processing on multi-source sensor monitoring data from GNSS, inclinometers, strain gauges, and microseismic monitoring instruments. It interpolates and aligns the data of each sensor to equally spaced timestamps in the time dimension and maps them to the corresponding nodes of a unified grid in the spatial dimension, generating basic alignment data for spatiotemporal reference unification.

[0145] The strain tensor field generation unit is used to perform spatial gradient calculation on the GNSS displacement sequence in the basic alignment data based on the central difference method, calculate the displacement change rate of adjacent grid nodes in three dimensions, and generate a strain tensor field characterizing the deformation distribution of the slope surface.

[0146] The microseismic feature quantization unit is used to calculate the occurrence density and total energy released of microseismic events on the spatial grid per unit time based on the list of microseismic events in the basic aligned data, and to generate time series data of spatial microseismic density distribution and microseismic energy release rate.

[0147] The multi-source data fusion encapsulation unit is used to fuse and encapsulate basic aligned data, strain tensor field, spatial microseismic density distribution and microseismic energy release rate time series data, integrate multi-dimensional spatiotemporal information into a unified data structure, and generate a multi-source spatiotemporal fusion dataset.

[0148] Preferably, the slope coupled field solving module 620 provided in this application is configured with the following units:

[0149] The boundary condition generation unit for mechanical solutions is used to format boundary conditions and observation constraints based on GNSS point displacement data and fiber optic strain gauge linear strain data in a multi-source spatiotemporal fusion dataset. It converts discrete point displacements into displacement constraints of nodes corresponding to a unified grid and distributed linear strains into strain field constraints of elements corresponding to a unified grid, thereby generating a set of boundary conditions for mechanical solutions.

[0150] The slope full-field displacement field solution unit is used to perform full-field state solution processing based on the finite element method on the displacement-strain joint constraint set and the preset slope rock mass mechanical parameters. The joint constraint set is substituted as known boundary conditions into the control equation of the slope mechanical model to solve the displacement response of all unknown nodes in the model and generate the slope full-field displacement field.

[0151] The slope full-field stress field reconstruction unit is used to reconstruct the stress field based on the rock mass constitutive relationship of the slope full-field displacement field. It calculates the strain of each unit according to the displacement field, and then calculates the stress state of the unit according to the stress-strain relationship of the material to generate the slope full-field stress field.

[0152] The displacement-stress coupling data generation unit is used to couple and encapsulate the full-field displacement field and the full-field stress field of the slope, establish a one-to-one correspondence between the displacement field and the stress field on the spatial grid, and generate displacement-stress coupling data.

[0153] Preferably, the physical constraint loss generation module 630 provided in this application is configured with the following units:

[0154] The equilibrium constraint residual field generation unit is used to calculate the equilibrium equation residuals of the full-field stress field in the displacement-stress coupling data, and calculates the magnitude of the divergence of the stress tensor and the sum of the volume forces element by element to generate the equilibrium constraint residual field.

[0155] The energy-constrained residual sequence generation unit is used to perform energy temporal synergy analysis on the full-field displacement field and microseismic energy data in displacement-stress coupling data. It calculates the time change rate of the overall strain energy of the slope from the displacement field and compares it with the microseismic energy release rate in the same time window to generate an energy-constrained residual sequence.

[0156] The fracture proximity index field generation unit is used to evaluate the fracture proximity of the full-field stress field in the displacement-stress coupling data based on the Mohr-Coulomb criterion. It calculates the algebraic distance from the current stress state to the fracture envelope surface for each element and generates the fracture proximity index field.

[0157] The physical constraint loss function generation unit is used to perform weighted fusion and space-time integration processing on the equilibrium constraint residual field, energy constraint residual sequence and rupture proximity index field. It superimposes the residuals of the three physical dimensions according to the preset residual weights and integrates them on the computational domain to generate a differentiable physical constraint loss function.

[0158] Preferably, the prediction model joint training module 640 provided in this application is configured with the following units:

[0159] The training sample set construction unit is used to construct training sample pairs for key physical features in multi-source spatiotemporal fusion datasets and displacement-stress coupling data. It takes historical time series fusion data as input features, future time series GNSS displacement as prediction labels, and the full field displacement field and stress field statistics of the same period as physical state labels to generate a spatiotemporal sample set for joint training.

[0160] The hybrid neural network dual-path encoding unit is used to construct a hybrid neural network containing a data-driven branch and a physical sensing branch. It performs dual-path encoding and fusion processing on the input features in the spatiotemporal sample set. The data-driven branch learns the temporal evolution law of the monitoring data, and the physical sensing branch encodes the mechanical state characteristics of the slope at the same time. After weighted fusion by the attention mechanism, it outputs the future multi-step displacement prediction sequence.

[0161] The predictive physical loss value calculation unit is used to substitute the predicted sequence of future multi-step displacements as virtual observations into the slope mechanics model to deduce the corresponding virtual displacement field and virtual stress field, and evaluate the physical law conformity of the predicted sequence of future multi-step displacements based on the differentiable physical constraint loss function, and generate the predicted physical loss value.

[0162] The model joint optimization training unit is used to perform joint optimization and iterative processing on the data loss and predicted physical loss values ​​between the future multi-step displacement prediction sequence and the real displacement labels obtained from the spatiotemporal sample set based on the gradient backpropagation algorithm. With the goal of minimizing the total loss, the parameters of each layer of the network are dynamically adjusted. Through multiple iterations, the network prediction is made to simultaneously approach the historical data patterns and physical mechanism constraints, generating a slope displacement prediction model with physical mechanism constraints.

[0163] Preferably, the graded early warning signal generation module 650 provided in this application is configured with the following units:

[0164] The real-time displacement prediction value generation unit is used to perform spatiotemporal registration and feature derivation processing on the real-time multi-source sensor monitoring data, generate a real-time multi-source spatiotemporal fusion dataset, perform forward prediction calculation on the real-time multi-source spatiotemporal fusion dataset based on the slope displacement prediction model, output the displacement prediction sequence and change trend of key points of the slope in the future period, and generate displacement prediction values.

[0165] The physical confidence index calculation unit is used to substitute the displacement prediction value as a virtual boundary condition into the slope mechanics model to perform physical state deduction, calculate the corresponding virtual displacement field and virtual stress field, and quantify the deviation between the virtual stress field and the static equilibrium principle, the deviation between the energy change rate derived from the virtual displacement field and the real-time micro-seismic energy release rate, and the proximity deviation between the virtual stress state and the Mohr-Coulomb fracture criterion based on the differentiable physical constraint loss function, and generate the physical confidence index.

[0166] The graded early warning signal generation unit is used to make a comprehensive judgment based on the magnitude of displacement prediction, change acceleration and physical confidence index, and to map different data combinations to specific risk levels and handling priorities according to preset rules, thereby generating graded early warning signals.

[0167] In one embodiment, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described dynamic monitoring method for mine slopes.

[0168] In one embodiment, this application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described dynamic monitoring method for mine slopes.

[0169] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0170] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0171] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope disclosed in this application, and these should all be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for dynamic monitoring of mine slopes, characterized in that, Includes the following steps: S1: Spatiotemporal registration and physical feature derivation processing are performed on multi-source sensor monitoring data from GNSS, inclinometers, strain gauges, and microseismic monitoring instruments. The original data from different spatiotemporal references are aligned to a unified grid and strain and energy release rate derivation features are calculated to generate a multi-source spatiotemporal fusion dataset. S2: Based on the multi-source spatiotemporal fusion dataset, the point displacement and linear strain observations are substituted into the slope mechanics model as boundary conditions to solve the full field state, generating displacement-stress coupling data containing the full field displacement field and full field stress field of the slope. S3: Based on the displacement-stress coupling data and the microseismic energy data in the multi-source spatiotemporal fusion dataset, perform physical law consistency quantification to generate a differentiable physical constraint loss function; S4: Based on the gradient backpropagation algorithm, the multi-source spatiotemporal fusion dataset and the displacement-stress coupling data are jointly trained to generate a slope displacement prediction model constrained by physical mechanism. S5: Based on the slope displacement prediction model, forward prediction and physical consistency verification are performed on the real-time collected multi-source sensor monitoring data to generate a graded early warning signal that integrates the displacement prediction value and physical confidence level. The graded early warning signal is used to indicate different slope risk levels and treatment priorities.

2. The method according to claim 1, characterized in that, S1 includes: S11: Perform spatial coordinate mapping processing on multi-source sensor monitoring data from GNSS, inclinometer, strain gauge, and microseismic monitor. Interpolate and align the data of each sensor to equal-interval timestamps in the time dimension, and map them to corresponding nodes of a unified grid in the spatial dimension to generate basic aligned data with unified spatiotemporal reference. S12: Based on the central difference method, the spatial gradient of the GNSS displacement sequence in the basic alignment data is calculated, the displacement change rate of adjacent grid nodes in three dimensions is calculated, and a strain tensor field characterizing the deformation distribution of the slope surface is generated. S13: Based on the list of microseismic events in the basic alignment data, calculate the occurrence density and total energy released by microseismic events on the spatial grid per unit time, and generate time series data of spatial microseismic density distribution and microseismic energy release rate. S14: The basic alignment data, the strain tensor field, the spatial microseismic density distribution, and the time series data of the microseismic energy release rate are fused and encapsulated to integrate multi-dimensional spatiotemporal information into a unified data structure, generating a multi-source spatiotemporal fusion dataset.

3. The method according to claim 1, characterized in that, S2 includes: S21: Based on the GNSS point displacement data and fiber optic strain gauge linear strain data in the multi-source spatiotemporal fusion dataset, perform formatting processing of boundary conditions and observation constraints, convert discrete point displacements into displacement constraints of nodes corresponding to a unified grid, convert distributed linear strains into strain field constraints of elements corresponding to a unified grid, and generate a set of boundary conditions for mechanical solutions. S22: The displacement-strain joint constraint set and the preset slope rock mass mechanical parameters are subjected to full-field state solution processing based on the finite element method. The joint constraint set is substituted as a known boundary condition into the control equation of the slope mechanical model, and the displacement response of all unknown nodes in the model is solved to generate the full-field displacement field of the slope. S23: The stress field reconstruction process based on the rock mass constitutive relationship is performed on the displacement field of the slope. The strain of each element is calculated according to the displacement field, and the stress state of the element is calculated according to the stress-strain relationship of the material to generate the stress field of the slope. S24: Couple and encapsulate the full-field displacement field and the full-field stress field of the slope to establish a one-to-one correspondence between the displacement field and the stress field on the spatial grid, and generate displacement-stress coupling data.

4. The method according to claim 1, characterized in that, S3 includes: S31: Perform equilibrium equation residual calculation on the whole field stress field in the displacement-stress coupling data, calculate the magnitude of the divergence of the stress tensor and the sum of the volume forces for each element, and generate an equilibrium constraint residual field. S32: Perform energy temporal synergy analysis on the full-field displacement field and the microseismic energy data in the displacement-stress coupling data, calculate the time change rate of the overall strain energy of the slope from the displacement field, and compare it with the microseismic energy release rate in the same time window to generate an energy-constrained residual sequence. S33: Based on the Mohr-Coulomb criterion, the full-field stress field in the displacement-stress coupling data is subjected to fracture proximity assessment processing. The algebraic distance from the current stress state to the fracture envelope surface is calculated for each element to generate a fracture proximity index field. S34: Perform weighted fusion and space-time integration on the equilibrium constraint residual field, the energy constraint residual sequence, and the rupture proximity index field. Superimpose the residuals of the three physical dimensions according to the preset residual weights and integrate them on the computational domain to generate a differentiable physical constraint loss function.

5. The method according to claim 1, characterized in that, S4 includes: S41: Construct training sample pairs for key physical features in the multi-source spatiotemporal fusion dataset and the displacement-stress coupling data, using historical time-series fusion data as input features, future time-series GNSS displacement as prediction labels, and the full-field displacement field and stress field statistics of the same period as physical state labels, to generate a spatiotemporal sample set for joint training. S42: Construct a hybrid neural network that includes a data-driven branch and a physical perception branch, perform dual-path encoding and fusion processing on the input features in the spatiotemporal sample set, learn the temporal evolution law of the monitoring data through the data-driven branch, encode the mechanical state characteristics of the slope at the same time through the physical perception branch, and output the future multi-step displacement prediction sequence after weighted fusion by the attention mechanism. S43: Substitute the predicted future multi-step displacement sequence as virtual observations into the slope mechanics model to deduce the corresponding virtual displacement field and virtual stress field, and evaluate the physical law conformity of the predicted future multi-step displacement sequence based on the differentiable physical constraint loss function to generate the predicted physical loss value. S44: Based on the gradient backpropagation algorithm, the data loss between the predicted future multi-step displacement sequence and the real displacement labels obtained from the spatiotemporal sample set and the predicted physical loss value are jointly optimized and iterated. The parameters of each layer of the network are dynamically adjusted with the goal of minimizing the total loss. Through multiple iterations, the network prediction is made to approach the historical data pattern and physical mechanism constraints at the same time, and a slope displacement prediction model with physical mechanism constraints is generated.

6. The method according to any one of claims 1-5, characterized in that, S5 includes: S51: Perform spatiotemporal registration and feature derivation processing on the real-time collected multi-source sensor monitoring data to generate a real-time multi-source spatiotemporal fusion dataset. Based on the slope displacement prediction model, perform forward prediction calculation processing on the multi-source spatiotemporal fusion dataset to output the displacement prediction sequence and change trend of key points of the slope in the future period and generate displacement prediction values. S52: Substitute the predicted displacement value as a virtual boundary condition into the slope mechanics model for physical state deduction, calculate the corresponding virtual displacement field and virtual stress field, and quantify the deviation between the virtual stress field and the static equilibrium principle, the deviation between the energy change rate derived from the virtual displacement field and the real-time micro-seismic energy release rate, and the proximity deviation between the virtual stress state and the Mohr-Coulomb fracture criterion based on the differentiable physical constraint loss function, and generate a physical confidence index. S53: Based on the magnitude and acceleration of the displacement prediction value and the physical confidence index, perform multi-threshold fuzzy logic comprehensive judgment, map different data combinations to specific risk levels and handling priorities according to preset rules, and generate graded early warning signals.

7. A dynamic monitoring system for mine slopes, characterized in that, The system includes: The multi-source data fusion processing module is used to perform spatiotemporal registration and physical feature derivation processing on monitoring data from multiple sources such as GNSS, inclinometers, strain gauges, and microseismic monitors. It aligns the original data from different spatiotemporal references to a unified grid and calculates strain and energy release rate derived features to generate a multi-source spatiotemporal fusion dataset. The slope coupled field solution module is used to solve the whole field state by substituting point displacement and linear strain observations as boundary conditions into the slope mechanics model based on the multi-source spatiotemporal fusion dataset, and generating displacement-stress coupled data containing the whole field displacement field and the whole field stress field of the slope. The physical constraint loss generation module is used to perform physical law consistency quantification based on the displacement-stress coupling data and the microseismic energy data in the multi-source spatiotemporal fusion dataset, and generate a differentiable physical constraint loss function. The prediction model joint training module is used to jointly train the multi-source spatiotemporal fusion dataset and the displacement-stress coupling data based on the gradient backpropagation algorithm to generate a slope displacement prediction model constrained by physical mechanism. The graded early warning signal generation module is used to perform forward prediction and physical consistency verification on the real-time multi-source sensor monitoring data based on the slope displacement prediction model, and generate a graded early warning signal that integrates the displacement prediction value and physical confidence level. The graded early warning signal is used to indicate different slope risk levels and treatment priorities.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 6.

Images