A dynamic assessment method for slope stability based on excavation time-series multimodal fusion characteristics

By constructing an excavation sequence and utilizing a CNN-LSTM neural network model, the problems of low computational efficiency and lack of information in slope stability analysis were solved, enabling dynamic assessment of slope stability, improving assessment efficiency, and providing spatial information on potential damage areas.

CN122863465BActive Publication Date: 2026-11-10CHINA 19TH METALLURGICAL CORP
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202611365407.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-09-04
Publication Date
2026-11-10
Estimated Expiration
2046-09-04

AI Technical Summary

Technical Problem

Existing technologies have low computational efficiency in slope stability analysis, are difficult to describe the evolutionary relationship between continuous excavation states of slopes, and lack spatial information on potential failure areas in the prediction results.

Method used

A dynamic slope stability assessment method based on excavation time-series multimodal fusion features is adopted. By constructing an excavation sequence, static and dynamic parameters are extracted, and a CNN-LSTM neural network model is used to predict the starting point, ending point ratio, and safety factor of the plastic penetration zone, thereby achieving end-to-end dynamic slope stability assessment.

Benefits of technology

It significantly reduces the amount of redundant calculations required by traditional numerical methods, can reflect the overall stability of the slope and provide spatial information on potential damage areas, meeting the needs of rapid comparison of construction schemes and dynamic risk assessment throughout the entire process.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122863465B_ABST
    Figure CN122863465B_ABST
Patent Text Reader

Abstract

This invention primarily relates to the field of geotechnical engineering construction safety technology. To address the problems of low calculation efficiency in slope stability prediction and the lack of spatial information on the evolutionary relationships between continuous excavation states and potential damage areas in the prediction results, this invention provides a dynamic slope stability assessment method based on the multimodal fusion characteristics of the excavation time sequence. The core of this invention is: extracting static parameters, dynamic parameters, and a dual-channel geometric grid superimposed on the unexcavated slope body and the current remaining slope body from the slope excavation sequence; inputting these into a trained CNN-LSTM neural network model; the CNN-LSTM neural network model outputs predicted values ​​for the starting and ending proportions of the plastic penetration zone and the safety factor according to time steps; and calculating the starting and ending elevations of the plastic penetration zone based on the predicted starting and ending proportions of the plastic penetration zone and the original total slope height of the parent slope, thereby achieving dynamic risk assessment of the entire slope excavation process.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention mainly relates to the field of geotechnical engineering construction safety technology, and in particular to a dynamic assessment method for slope stability based on the multimodal fusion characteristics of excavation sequence. Background Technology

[0002] Slope excavation is widely used in open-pit mines, roads, water conservancy and hydropower projects, foundation pits, and rock slope engineering. Multi-level slopes are typically excavated in stages and layers according to a predetermined construction sequence. As the excavation process progresses, the slope's geometric boundaries continuously change, accompanied by unloading, stress redistribution, and the expansion of the plastic penetration zone. The slope's safety factor and potential failure range may change significantly at any intermediate construction stage. Therefore, stability analysis based solely on the final excavation state is insufficient to fully identify the stage-specific instability risks during construction.

[0003] Currently, slope stability analysis typically employs limit equilibrium, finite element, finite difference, or discrete element methods. The strength reduction method obtains the slope safety factor and failure zone by progressively reducing the shear strength parameters of the soil and rock mass until the model becomes unstable. This method does not require a pre-defined sliding surface and can automatically reveal the failure mechanism based on the evolution of stress, deformation, and plastic penetration zones; however, its computational cost is generally higher than that of the limit equilibrium method.

[0004] In the analysis of the entire process of graded excavation, the same parent slope typically includes an undisturbed state and multiple continuously excavated states. For each combination of material parameters, slope geometry scheme, and construction scheme, a numerical model must be established for each excavation state, iterative calculations must be performed, and safety factors and plastic penetration zone results must be extracted. As the number of parent slopes, excavation levels, and sub-steps within each level increases, the workload of repeated numerical calculations and post-processing of results increases significantly, making it difficult to directly meet the needs of rapid comparison and selection of construction schemes and dynamic evaluation of the entire process.

[0005] Existing technologies can use artificial neural networks, convolutional neural networks, or long short-term memory neural networks to predict slope stability results and, to some extent, reduce redundant calculations in traditional numerical analysis. However, existing technologies still primarily treat different slopes or different working conditions as independent static samples.

[0006] Furthermore, existing technologies primarily use safety factors or stability categories as single outputs. While safety factors reflect the overall stability of a slope, they cannot directly indicate the location and extent of potential failure zones along the slope height. During staged excavation, even if safety factors are similar across different states, the distribution and penetration trend of the plastic penetration zone may still differ. Therefore, predicting only the safety factor cannot fully reflect the stability evolution during slope excavation. Summary of the Invention

[0007] This invention provides a dynamic assessment method for slope stability based on the multimodal fusion characteristics of excavation time series, aiming to solve the problems of low computational efficiency in slope stability prediction and the lack of evolutionary relationships between continuous excavation states and spatial information on potential damage areas in the prediction results.

[0008] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0009] A dynamic assessment method for slope stability based on excavation time-series multimodal fusion features, the method comprising:

[0010] Step 1: Construct an excavation sequence by combining the unexcavated and continuously excavated states of the same parent slope according to the construction order;

[0011] Step 2: Extract the static and dynamic parameters of each time step in the excavation sequence, and construct a dual-channel geometric grid formed by superimposing the binary image of the unexcavated slope body with the current binary image of the remaining slope body.

[0012] Step 3: Construct the static parameters, dynamic parameters, and dual-channel geometric grid of the same parent slope at each time step into a parent slope input sequence in chronological order; input the parent slope input sequence into the CNN-LSTM neural network model, where the CNN is used to extract the spatial morphological features of the dual-channel geometric grid, and the LSTM network receives the spatial morphological features, static parameters, and dynamic parameters of each time step in chronological order, and outputs the temporal hidden state that integrates the current slope state and all historical excavation information at each effective time step. Based on the temporal hidden state, the predicted values ​​of the starting point ratio, ending point ratio, and safety factor of the plastic penetration zone at each time step are obtained.

[0013] Step 4: Construct a joint loss by using the error between the predicted safety factor, the predicted starting proportion of the plastic penetration zone, the predicted ending proportion of the plastic penetration zone, and the corresponding true value at each time step. Use the input sequence of the parent slope as input to train the CNN-LSTM neural network model.

[0014] Step 5: Input the static parameters, dynamic parameters, and dual-channel geometric grid of the slope to be evaluated at each time step into the trained CNN-LSTM neural network model. Output the predicted values ​​of the starting point ratio, ending point ratio, and safety factor of the plastic penetration zone according to the time step. Calculate the starting elevation and ending elevation of the plastic penetration zone based on the predicted values ​​of the starting point ratio, ending point ratio, and the original total slope height of the parent slope.

[0015] Furthermore, step 2, which involves constructing a dual-channel geometric grid by superimposing the binary image of the unexcavated slope with the current binary image of the remaining slope, includes: converting the slope region in the unexcavated state of the parent slope into a first binary channel, converting the remaining slope region after excavation at the current time step into a second binary channel, setting the pixel value inside the slope to 1, and the pixel value outside the slope to 0.

[0016] Furthermore, the static parameters include elastic modulus, Poisson's ratio, density, cohesion, internal friction angle, and original total slope height; the dynamic parameters include the current newly added excavation height, the current cumulative excavation height, the width of adjacent graded platforms, and the single-step excavation area increment.

[0017] Furthermore, the CNN described in step 3 includes five convolutional layers. Max pooling layers are set after the first four convolutional layers, and an adaptive average pooling layer is set after the fifth convolutional layer. The adaptive pooling results are flattened and fully connected to obtain the spatial morphological features corresponding to each time step.

[0018] Furthermore, the spatial morphological features include the original slope outline, the current remaining slope outline, the location of the excavation platform, and geometric transitions.

[0019] Furthermore, step 3, which involves obtaining the predicted starting point ratio and ending point ratio of the plastic penetration region at each time step based on the temporal hidden state, includes: [The text abruptly ends here, so the translation stops.] The temporal hidden states at each time step are nonlinearly mapped to predict the starting point ratio and length ratio of the plastic penetration region. The ending point ratio is then calculated based on the starting point ratio and length ratio of the plastic penetration region.

[0020] Furthermore, step 3, obtaining the predicted safety factor based on the temporal hidden state, includes:

[0021] The direct safety factor is obtained by directly predicting the branch using the temporal hidden state input;

[0022] The plastic penetration region embedding vector obtained by nonlinear mapping of the temporal hidden state and the starting point ratio, length ratio and ending point ratio of the plastic penetration region are combined into the plastic penetration region characterization vector. The temporal hidden state and the plastic penetration region characterization vector are concatenated and input into the coupling prediction branch to obtain the coupling safety factor.

[0023] The predicted safety factor is obtained by weighting the direct safety factor and the coupled safety factor through an adaptive fusion gate.

[0024] Furthermore, the predicted safety factor value is:

[0025] ;

[0026] In the formula, This is the predicted value for the safety factor. For direct safety factor, For the coupling safety factor, The fusion weights are: , In the formula, For adaptive fusion gate weights, It is a sigmoid activation function. It is a linear rectified activation function. For the first The temporal hidden state at each time step The vector representing the plastic through region. and This is the weight matrix. and This is a bias term.

[0027] Furthermore, the plastic penetration region representation vector of the input coupled prediction branch is processed by stopping gradient, so that the safety factor loss does not directly update the plastic penetration region prediction branch in reverse.

[0028] Furthermore, the joint loss mentioned in step 4 is:

[0029] ;

[0030] In the formula, For joint losses, Mean square error, This is the predicted value for the safety factor. For the true value of the safety factor, This is the predicted proportion of the starting point of the plastic penetration zone. This represents the true value of the starting proportion of the plastic penetration zone. This is the predicted proportion of the endpoint of the plastic penetration zone. This represents the true value of the endpoint proportion of the plastic penetration zone. , and These are the weighting coefficients for the corresponding loss terms.

[0031] The beneficial effects of this invention are as follows: This invention organizes the continuous excavation state of the same parent slope into a variable-length excavation sequence, which can adapt to slope engineering with different excavation levels and different numbers of excavation steps; it extracts and fuses features from static parameters, dynamic parameters, and dual-channel geometric grids respectively, while preserving the spatial features of slope material properties, excavation process information, and geometric morphology; by modeling the temporal evolution relationship of the excavation process through an LSTM network, it can realize end-to-end dynamic assessment of slope stability, significantly reducing the amount of repetitive calculations in traditional numerical methods and improving assessment efficiency; at the same time, it outputs the safety factor of each excavation step and the location range of the plastic penetration zone, which can reflect the overall stability of the slope and provide spatial information of potential damage areas, providing a more complete basis for the staged risk identification in the graded excavation process, and effectively meeting the needs of rapid comparison of construction schemes and dynamic risk assessment throughout the entire process. Attached Figure Description

[0032] Figure 1 This is a flowchart of the slope stability dynamic assessment method based on excavation time-series multimodal fusion features as described in this invention;

[0033] Figure 2 A scatter plot showing the prediction of the safety factor of excavated slopes using this invention;

[0034] Figure 3 To use this invention to generate a test set of inter-regional intersection ratio distribution diagram of the plastic penetration zone under the corresponding stability state of the excavated slope;

[0035] Figure 4 This is a comparison chart of the safety factor prediction performance of this invention with other models;

[0036] Figure 5 This is a comparison chart of the prediction performance of the intersection ratio of the plastic penetration region of this invention with other models. Detailed Implementation

[0037] Existing slope stability prediction schemes suffer from low computational efficiency, struggle to describe the evolutionary relationships between continuous excavation states of slopes, and lack spatial information on potential damage areas. Therefore, this invention provides a dynamic slope stability assessment method based on excavation time-series multimodal fusion features. The core of this invention is to organize the unexcavated state and various excavation states of the same parent slope into an excavation sequence according to construction order. Static and dynamic parameters are extracted for each time step in the sequence, and a dual-channel geometric grid is constructed by superimposing a binary image of the unexcavated slope body with a binary image of the current remaining slope body. The static parameters, dynamic parameters, and the dual-channel geometric grid are then input into a CNN-LSTM neural network model in a temporal sequence. The CNN extracts the spatial morphological features of the dual-channel geometric grid, and the LSTM network outputs the temporal hidden state fused with historical information at each effective time step through a gating mechanism. Finally, the plastic penetration prediction branch is used to output the plastic penetration. The system calculates the predicted starting and length proportions of the slope excavation zone, and the predicted ending proportion of the plastic penetration zone. It outputs the direct and coupled safety factors through the safety factor prediction branch, and obtains the predicted safety factor by weighting the direct and coupled safety factors through an adaptive fusion gate. The CNN-LSTM neural network model is trained with the joint error between the predicted safety factor, the predicted starting and ending proportions of the plastic penetration zone, and the corresponding true values. Finally, the system outputs the safety factor, the starting elevation, and the ending elevation of the plastic penetration zone for each time step of the slope to be evaluated, thus realizing the dynamic stability evaluation of the entire slope excavation process.

[0038] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0039] like Figure 1 As shown, the slope stability dynamic assessment method based on excavation time-series multimodal fusion features of the present invention includes the following steps.

[0040] Step 1: Construct an excavation sequence by combining the unexcavated state and the continuously excavated state of the same parent slope according to the construction order.

[0041] Each parent slope is assigned a unique number. The same parent slope number is used for both the unexcavated and excavated states of the same parent slope, and time indices are set according to the excavation sequence to form a continuous excavation sequence. The length of the excavation sequence is variable, allowing for different effective time steps to accommodate construction schemes with different excavation levels, number of substeps within a level, or different termination states. During CNN-LSTM neural network model training, parent slope samples are grouped according to the excavation sequence length. Each training batch only includes parent slopes with the same sequence length, without zero-padding or invalid time step masks, thus avoiding the participation of padded data in LSTM network computation.

[0042] For multiple parent slopes with different material parameters and geometric schemes, the true values ​​of the safety factor, the starting proportion of the plastic penetration zone, and the ending proportion of the plastic penetration zone are calculated for each state using the strength-reduced finite element method or the strength-reduced finite difference method. The safety factor is standardized using the mean and standard deviation of the training set. The validation set and test set both use the standardized parameters determined by the training set, without refitting.

[0043] This step organizes the unexcavated state and the excavation states at each level of the same parent slope into an excavation sequence according to the actual construction order. This allows the LSTM network to receive sequence information step by step along the time axis and use the information of the previous excavation state to assist in the stability prediction of the current state. This breaks through the limitation of existing static machine learning methods that treat each excavation state as an independent sample for prediction and realizes the temporal evolution modeling of the stability of the continuous construction process.

[0044] Step 2: Extract the static and dynamic parameters of each time step in the excavation sequence, and construct a dual-channel geometric grid by superimposing the binary image of the unexcavated slope with the current binary image of the remaining slope.

[0045] In the multimodal input construction stage, static and dynamic parameters are extracted for each time step in the excavation sequence. Static parameters are those that remain constant across time steps for the same parent slope, specifically including elastic modulus, Poisson's ratio, density, cohesion, internal friction angle, and original total slope height. Dynamic parameters are those that change with excavation time, specifically including the current newly added excavation height, the current cumulative excavation height, the width of adjacent graded platforms, and the single-step excavation area increment.

[0046] Specifically, the current cumulative excavation height is obtained by summing the excavation heights of each excavated stage in the current time step. The current newly added excavation height is the difference between the cumulative excavation heights of two adjacent time steps. In the unexcavated state, both the current newly added excavation height and the cumulative excavation height are set to zero. The width of adjacent stage platforms is calculated based on the geometric relationship between the horizontal position, stage height, excavation slope angle, and original slope angle of adjacent excavation stages. When a corresponding platform has not yet formed, its width is set to zero; after the platform is formed, its actual width is maintained in subsequent time steps. The single-step excavation area increment is obtained by comparing the slope area removed in the current time step with the slope area removed in the previous time step.

[0047] The construction of the dual-channel geometric grid includes: under a unified computational frame and a unified spatial resolution, converting the original slope body M0 in its unexcavated state as the parent slope into the first binary channel, and converting the remaining slope body M after excavation at the current time step into the second binary channel. tThe data is converted to a second binary channel. Pixels inside the slope are set to 1, and pixels outside the slope are set to 0. The first and second binary channels are superimposed to form a dual-channel geometric raster. The first binary channel remains unchanged within the same parent slope sequence, while the second binary channel changes during the excavation process. The geometric raster files are matched with parameter records according to the time step working condition number, and the file number, number of channels, image size, and duplicate files are verified. To reduce the memory and computational burden during training, while maintaining the aspect ratio and binary boundaries, the nearest neighbor method is used to reduce the number of raster cells, while keeping the raster values ​​at zero or one, avoiding intermediate values ​​without clear physical meaning at the slope boundaries.

[0048] In this step, three types of inputs are constructed for each time step: static parameters, dynamic parameters, and dual-channel geometric grids. These inputs describe the slope characteristics from three dimensions: inherent material properties, construction progress status, and slope geometry, respectively, thus achieving a more complete representation of the slope status.

[0049] Step 3: Construct the static parameters, dynamic parameters, and dual-channel geometric grid of the same parent slope at each time step into the parent slope input sequence in chronological order. The input sequence of the parent slope is fed into the CNN-LSTM neural network model. The CNN extracts the spatial morphological features of the dual-channel geometric grid, and the LSTM network receives the spatial morphological features, static parameters, and dynamic parameters of each time step in chronological order. At each effective time step, it outputs the temporal hidden state that integrates the current slope status and all historical excavation information. Based on the temporal hidden state, the predicted values ​​of the safety factor, the starting proportion of the plastic penetration zone, and the ending proportion of the plastic penetration zone are obtained for each time step.

[0050] The CNN-LSTM neural network model includes a static parameter encoder, a dynamic parameter encoder, and a convolutional geometric encoder. The static parameter encoder comprises a fully connected layer, a nonlinear activation layer, and a static feature output layer. After mapping by the static parameter encoder, the static parameters are extended along the time dimension to all effective time steps of the corresponding parent slope, ensuring that each time step obtains the corresponding material properties and overall size information. The dynamic parameter encoder performs feature mapping for different time steps, preserving the changes in the incremental excavation height, cumulative excavation height, width of adjacent graded platforms, and sub-excavation area increments as the construction process progresses. The convolutional geometric encoder internally uses a CNN as the feature extraction backbone network, comprising five convolutional layers with output channels of 16, 32, 64, 96, and 128 respectively. Max pooling layers are applied after the first four convolutional layers, and an adaptive average pooling layer is applied after the fifth convolutional layer. The convolutional layers progressively extract features such as the original slope contour, the current remaining slope contour, the excavation platform position, and geometric transitions, while the pooling layers progressively compress the spatial dimensions. After flattening and fully connected mapping, the adaptive pooling result yields a low-dimensional spatial morphological feature vector for each time step.

[0051] In the feature learning and fusion stage, static parameters, dynamic parameters, and a dual-channel geometric grid are input into the static parameter encoder, dynamic parameter encoder, and convolutional geometric encoder, respectively. Static and dynamic feature vectors are obtained through static and dynamic MLPs, and spatial morphological feature vectors are extracted using a CNN. At each time step, the static, dynamic, and spatial morphological feature vectors are concatenated along the feature dimension, and the concatenated result is normalized before being input into the LSTM network.

[0052] During the LSTM timing modeling and output phase, the LSTM cell structure receives the memory cells from the previous time step sequentially according to the actual excavation sequence of the slope. The hidden state of the previous time step and the input of the current time step Forgotten Gate go through After activation, the memory unit from the previous moment is discarded. Historical information in the input field; input gate go through After activation, it controls the writing of new information and the input of the current time step. and the hidden state of the previous time step Candidate memories are generated using the tanh function. The forget gate is accessed via the × and + operations. Retained historical information and input gate The written information is added together to update the memory unit. Output gate go through Activated control memory unit The output information after tanh activation is then processed by the × operation to generate the current temporal hidden state. Proceed to the next step. Temporal hidden states at each time step This system comprehensively characterizes the current slope material properties, overall dimensions, current excavation status, current slope geometry, and excavation history at previous time steps. The LSTM network outputs the temporal hidden state at each effective time step, rather than only using the last time step of the sequence for prediction. This enables dynamic evaluation of the entire slope excavation process step-by-step, ensuring that stability changes at each excavation stage are captured promptly.

[0053] The temporal hidden state at time step t is input into the plastic penetration region prediction branch of the CNN-LSTM neural network model. The intermediate features of the plastic penetration region are obtained through nonlinear feature mapping, and the starting proportion of the plastic penetration region is output. and length ratio Based on the starting proportion of the plastic penetration zone and length ratio Calculate the endpoint ratio of the plastic penetration zone. Simultaneously, the time-series hidden states output by the LSTM network are... After inputting the plastic penetration region prediction branch, the intermediate features of the plastic penetration region obtained by the first fully connected network are further mapped into the plastic penetration region embedding vector, which, together with the starting point ratio, length ratio, and ending point ratio of the plastic penetration region, are used for subsequent safety factor prediction (the starting point ratio and ending point ratio of the plastic penetration region are directly supervised during model training, and the length ratio is only used as an interval constraint variable).

[0054] The safety factor prediction branch of the CNN-LSTM neural network model includes a direct prediction branch and a coupled prediction branch. The direct prediction branch takes the temporal hidden state as input to obtain the direct safety factor. It combines the plastic region embedding vector with the start-point, length, and end-point proportions of the plastic region to form the plastic region representation vector. The plastic region coupled prediction branch concatenates the temporal hidden state with the plastic region representation vector to predict the coupled safety factor. An adaptive fusion gate weights the direct and coupled safety factors to obtain the predicted safety factor value.

[0055] The adaptive fusion gate calculates weights based on the fusion features, and first defines the gating function:

[0056] In the formula, For adaptive fusion gate weights, It is a sigmoid activation function. It is a linear rectified activation function. For the first The temporal hidden state at each time step The vector representing the plastic through region. and This is the weight matrix. and This is a bias term.

[0057] The predicted safety factor is:

[0058] In the formula, This is the predicted value for the safety factor. For direct safety factor, For the coupling safety factor, For weight fusion.

[0059] The plastic region representation vectors and proportional features of the input coupled prediction branch and adaptive fusion gate are processed using stopping gradients, preventing the safety factor loss from directly updating the plastic region prediction branch. The safety factor loss can still be used to update the temporal feature extraction network through shared temporal hidden states, thereby reducing the direct interference of the safety factor prediction task on the prediction results of the plastic region prediction branch.

[0060] This invention sets up a plastic penetration zone prediction branch and a safety factor prediction branch after the LSTM network to output the starting point ratio, ending point ratio and safety factor of the plastic penetration zone, respectively. This enables the synchronous acquisition of the overall stability of the slope and the location information of the local failure area, overcoming the defect of incomplete information from a single output index.

[0061] Step 4: Construct a joint loss by using the error between the predicted safety factor, the predicted starting proportion of the plastic penetration zone, the predicted ending proportion of the plastic penetration zone, and the corresponding true value at each time step. Use the input sequence of the parent slope as input to train the CNN-LSTM neural network model.

[0062] The model is trained using a multi-task joint loss. The joint loss consists of the errors between the predicted safety factor, the predicted proportions of the starting point and the predicted proportions of the plastic penetration zone, and the actual values. Each term is calculated using mean squared error. The joint loss is expressed as:

[0063] ;

[0064] In the formula, This is the predicted value for the safety factor. For the true value of the safety factor, This is the predicted proportion of the starting point of the plastic penetration zone. This represents the true value of the starting proportion of the plastic penetration zone. This is the predicted proportion of the endpoint of the plastic penetration zone. This represents the true value of the endpoint proportion of the plastic penetration zone. , and These are the weighting coefficients for the corresponding loss terms.

[0065] The error between the predicted and actual values ​​of the safety factor is calculated in the standardized space, while the errors between the predicted and actual values ​​of the starting and ending proportions of the plastic penetration zone are calculated in the proportional space relative to the original total slope height. By simultaneously supervising the safety factor and the starting and ending proportions of the plastic penetration zone, the CNN-LSTM neural network model can balance the overall slope stability with the prediction accuracy of potential failure areas.

[0066] Using the parent slope number as the dividing unit, all samples were divided into training, validation, and test sets in an 8:1:1 ratio. Unexcavated and continuously excavated states of the same parent slope could only be included in the same dataset to avoid samples from the same source appearing in both the training and test sets. The training set was used for model parameter training and standardized / normalized parameter fitting. The validation set was used for hyperparameter selection, learning rate adjustment, and early stopping detection. The test set was used only for evaluating the generalization performance after model training was completed.

[0067] Excavation sequences of different lengths are grouped according to sequence length, with each batch containing only sequences of the same length. The AdamW optimizer, gradient pruning, learning rate decay, and early stopping mechanisms are used during model training. The model is trained using a gradient descent-based optimizer with gradient pruning, adaptive learning rate decay, and early stopping mechanisms. Bayesian optimization is used to search for network hyperparameters, including static feature dimension, dynamic feature dimension, convolution output dimension, adaptive pooling size, LSTM hidden dimension, number of LSTM layers, plastic penetration region embedding dimension, learning rate, weight decay, and random deactivation ratio. Hyperparameters are searched using the tree-structured Parzen estimation method. When selecting the model, the joint loss on the validation set, the root mean square error of the safety factor, the intersection-union ratio of the plastic penetration region, and the coefficient of determination of the safety factor are comprehensively considered. The optimal parameter combination based on the comprehensive evaluation results is selected, and the model weights, standardized parameters, input feature definitions, and data partitioning information are saved.

[0068] This step uses a joint loss function composed of three parts: the predicted safety factor, the predicted starting point ratio of the plastic penetration zone, and the error between the predicted ending point ratio and their corresponding true values, to train the model. This allows the model to simultaneously consider the prediction accuracy of the safety factor and the location of the plastic penetration zone during the optimization process. The safety factor prediction uses two branches: direct prediction and plastic penetration zone coupled prediction, which are then weighted and merged through an adaptive fusion gate. This allows the model to automatically adjust the weight distribution of the two branches based on the input features. At the same time, by stopping gradient processing, the backpropagation channel of the safety factor loss to the plastic penetration zone prediction branch is cut off, so that the plastic penetration zone prediction branch is only supervised by the plastic penetration zone prediction error, avoiding interference between the safety factor prediction task and the plastic penetration zone prediction task.

[0069] Step 5: Input the static parameters, dynamic parameters, and dual-channel geometric grid of the slope to be evaluated at each time step into the trained CNN-LSTM neural network model. Output the predicted values ​​of the starting point ratio, ending point ratio, and safety factor of the plastic penetration zone according to the time step. Calculate the starting elevation and ending elevation of the plastic penetration zone based on the predicted values ​​of the starting point ratio, ending point ratio, and the original total slope height of the parent slope.

[0070] The trained model is applied to assess the dynamic stability of the slope to be evaluated. For the slope to be evaluated, a dual-channel geometric grid of static parameters, dynamic parameters, and each time step is constructed using the same method as in the training phase. This grid is then used to form an excavation sequence according to the actual construction order and input into the trained CNN-LSTM neural network model. The CNN-LSTM neural network model outputs the predicted safety factor of the current excavation state, the starting proportion and ending proportion of the plastic penetration zone at each time step, and obtains the distribution of the plastic penetration zone based on the starting proportion, ending proportion, and the original total slope height of the parent slope. Let the original total slope height of the parent slope be H, then the starting elevation and ending elevation of the plastic penetration zone are respectively... and The conversion of the starting and ending elevations of the plastic penetration zone based on the starting and ending proportions of the plastic penetration zone ensures that the ending elevation is not less than the starting elevation, and that both the starting and ending elevations are within the original slope height range. This eliminates physically impossible prediction results from a computational perspective, eliminating the need for additional constraints to correct the model output.

[0071] Simultaneously, based on the safety factor at each excavation time step, and the initial and final elevations of the plastic penetration zone, safety factor variation curves and plastic penetration zone evolution results are generated according to the continuous excavation sequence. This allows for the analysis of slope stability changes with cumulative excavation height. Based on the prediction results at each time step, excavation stages with significantly decreased safety factors or rapid expansion of the plastic penetration zone are identified, providing a reference for adjusting the construction sequence, optimizing graded platforms, and determining key monitoring areas.

[0072] Example

[0073] In this embodiment, a total of 147 parent slope sequences were created, containing 2036 effective excavation stages. The safety factor prediction results were predicted on the test set according to the method described in this invention, and the scatter plot of the prediction results is shown below. Figure 2 As shown, Figure 2 The coefficient of determination (R²) reached 0.9892, the root mean square error (RMSE) was 0.0316, and the mean absolute error (MAE) was 0.0243, indicating that the predicted safety factor obtained based on this invention is highly consistent with the actual safety factor value obtained from numerical simulation, and the prediction accuracy is good. The intersection-union ratio (IUU) distribution of the predicted and actual intervals of the plastic penetration zone in the test set is shown below. Figure 3 As shown, the average interval intersection-to-union ratio reaches 0.948, indicating that the location of the plastic penetration zone predicted by the present invention is highly consistent with the actual failure area, proving that the present invention can accurately locate the potential failure area space. Figure 4To compare the determination coefficients R² of the present invention with those of the LSTM and MLP models used directly on the training, validation and test sets, the present invention achieves determination coefficients R² of 0.998, 0.991 and 0.989 on the training, validation and test sets, respectively, which are better than the prediction results of the LSTM and MLP models. Figure 5 The invention is compared with the LSTM model and the MLP model on the training set, validation set and test set for the intersection-union ratio of plastic through intervals. The invention achieves intersection-union ratios of 0.960, 0.937 and 0.948 on the training set, validation set and test set, respectively, which are significantly better than the prediction results of the LSTM model and the MLP model.

Claims

1. A dynamic assessment method for slope stability based on excavation time-series multimodal fusion features, characterized in that, The method includes: Step 1: Construct an excavation sequence by combining the unexcavated and continuously excavated states of the same parent slope according to the construction order; Step 2: Extract the static and dynamic parameters of each time step in the excavation sequence, and construct a dual-channel geometric grid formed by superimposing the binary image of the unexcavated slope body with the current binary image of the remaining slope body. Step 3: Construct the static parameters, dynamic parameters, and dual-channel geometric grid of the same parent slope at each time step into a parent slope input sequence in chronological order; input the parent slope input sequence into the CNN-LSTM neural network model, where the CNN is used to extract the spatial morphological features of the dual-channel geometric grid, and the LSTM network receives the spatial morphological features, static parameters, and dynamic parameters of each time step in chronological order, and outputs the temporal hidden state that integrates the current slope state and all historical excavation information at each effective time step. Based on the temporal hidden state, the predicted values ​​of the starting point ratio, ending point ratio, and safety factor of the plastic penetration zone at each time step are obtained. Step 4: Construct a joint loss by using the error between the predicted safety factor, the predicted starting proportion of the plastic penetration zone, the predicted ending proportion of the plastic penetration zone, and the corresponding true value at each time step. Use the input sequence of the parent slope as input to train the CNN-LSTM neural network model. Step 5: Input the static parameters, dynamic parameters, and dual-channel geometric grid of the slope to be evaluated at each time step into the trained CNN-LSTM neural network model. Output the predicted values ​​of the starting point ratio, ending point ratio, and safety factor of the plastic penetration zone according to the time step. Calculate the starting elevation and ending elevation of the plastic penetration zone based on the predicted values ​​of the starting point ratio, ending point ratio, and the original total slope height of the parent slope.

2. The slope stability dynamic assessment method based on excavation time-series multimodal fusion features according to claim 1, characterized in that, Step 2 describes constructing a dual-channel geometric grid formed by superimposing the binary image of the unexcavated slope and the current binary image of the remaining slope. This includes: converting the unexcavated slope area of ​​the parent slope into a first binary channel, converting the remaining slope area after excavation at the current time step into a second binary channel, setting the pixel value inside the slope to 1, and the pixel value outside the slope to 0.

3. The slope stability dynamic assessment method based on excavation time-series multimodal fusion features according to claim 1, characterized in that, The static parameters include elastic modulus, Poisson's ratio, density, cohesion, internal friction angle, and original total slope height; the dynamic parameters include the current newly added excavation height, the current cumulative excavation height, the width of adjacent graded platforms, and the single-step excavation area increment.

4. The slope stability dynamic assessment method based on excavation time-series multimodal fusion features according to claim 1, characterized in that, The CNN described in step 3 includes five convolutional layers. Max pooling layers are set after the first four convolutional layers, and an adaptive average pooling layer is set after the fifth convolutional layer. The adaptive pooling results are flattened and fully connected to obtain the spatial morphological features corresponding to each time step.

5. The method for dynamic assessment of slope stability based on excavation time-series multimodal fusion features according to claim 1, characterized in that, The spatial morphological features include the original slope outline, the current remaining slope outline, the location of the excavation platform, and geometric transitions.

6. The slope stability dynamic assessment method based on excavation time-series multimodal fusion features according to claim 1, characterized in that, Step 3, which involves obtaining the predicted starting point and ending point proportions of the plastic penetration region at each time step based on the temporal hidden state, includes: [The text abruptly ends here, likely due to an incomplete sentence or a formatting error.] The temporal hidden states at each time step are nonlinearly mapped to predict the starting point ratio and length ratio of the plastic penetration region. The ending point ratio is then calculated based on the starting point ratio and length ratio of the plastic penetration region.

7. The method for dynamic assessment of slope stability based on excavation time-series multimodal fusion features according to claim 1, characterized in that, Step 3, obtaining the predicted safety factor based on the temporal hidden state, includes: The direct safety factor is obtained by directly predicting the branch using the temporal hidden state input; The plastic penetration region embedding vector obtained by nonlinear mapping of the temporal hidden state and the starting point ratio, length ratio and ending point ratio of the plastic penetration region are combined into the plastic penetration region characterization vector. The temporal hidden state and the plastic penetration region characterization vector are concatenated and input into the coupling prediction branch to obtain the coupling safety factor. The predicted safety factor is obtained by weighting the direct safety factor and the coupled safety factor through an adaptive fusion gate.

8. The slope stability dynamic assessment method based on excavation time-series multimodal fusion features according to claim 7, characterized in that, The predicted safety factor value is: ; In the formula, This is the predicted value for the safety factor. For direct safety factor, For the coupling safety factor, The fusion weights are: , In the formula, For adaptive fusion gate weights, It is a sigmoid activation function. It is a linear rectified activation function. For the first The temporal hidden state at each time step The vector representing the plastic through region. and This is the weight matrix. and This is a bias term.

9. The method for dynamic assessment of slope stability based on excavation time-series multimodal fusion features according to claim 7, characterized in that, The plastic penetration region representation vector of the input coupled prediction branch is processed by stopping gradient, so that the safety factor loss does not directly update the plastic penetration region prediction branch in reverse.

10. The method for dynamic assessment of slope stability based on excavation time-series multimodal fusion features according to claim 1, characterized in that, The joint loss mentioned in step 4 is: ; In the formula, For joint losses, Mean square error, This is the predicted value for the safety factor. For the true value of the safety factor, This is the predicted proportion of the starting point of the plastic penetration zone. This represents the true value of the starting proportion of the plastic penetration zone. This is the predicted proportion of the endpoint of the plastic penetration zone. This represents the true value of the endpoint proportion of the plastic penetration zone. , and These are the weighting coefficients for the corresponding loss terms.

Citation Information

Patent Citations

  • Tunnel monitoring method and device for data fusion, computer equipment and storage medium

    CN117272232A

  • Intelligent monitoring and early warning method and system for dangerous rock falling of high and steep slope

    CN120612801A