A slope stability rapid evaluation method based on artificial intelligence

Abstract

The present application belongs to the technical field of intelligent identification and evaluation, and provides a slope stability rapid evaluation method based on artificial intelligence, in order to solve the problems of slope stability evaluation depending on complex parameters, long parameter acquisition and calculation period, and difficulty in fusing multi-source information. The present application converts the slope stability evaluation into a neural network forward calculation process. After the model training is completed, the evaluation of any new slope sample only needs millisecond level time, compared with the traditional numerical simulation method, the efficiency is improved by hundreds of thousands of times, and the "rapid evaluation" is truly realized. The present application does not require high-cost accurate mechanical parameters. Even if only the qualitative description (such as "joint development, opening and filling with cohesive soil") and basic geometric information obtained by field investigation are available, the model can also be effectively evaluated through the text vectorization method.

Description

Technical Field

[0001] This invention belongs to the field of intelligent identification and assessment technology, specifically relating to a rapid slope stability assessment method based on artificial intelligence. Background Technology

[0002] Slope stability analysis is a crucial topic in geotechnical engineering, widely applied in the construction and operation of highways, railways, water conservancy and hydropower projects, and mines. Traditional slope stability assessment methods mainly fall into two categories: First, the engineering geological analogy method, relying on engineers' experience to compare the slope to be assessed with similar existing slopes to determine its stability. This method is highly subjective, heavily dependent on expert experience, and difficult to quantify and generalize. Second, the limit equilibrium method and numerical simulation methods (such as the finite element method and discrete element method), which obtain the slope's geotechnical physical and mechanical parameters (such as density, cohesion, and internal friction angle) through field investigation and laboratory tests, establish a mechanical model, calculate the slope's safety factor, and then assess its stability (stable, basically stable, understability, unstable). However, the above-mentioned existing technical solutions have the following significant drawbacks and problems in practical applications:

[0003] 1. High cost and long cycle of data acquisition: Accurate physical and mechanical parameters of soil and rock need to be obtained through complex indoor tests or in-situ tests, which is time-consuming and costly. Moreover, it is extremely difficult to obtain representative parameters for large slopes or slopes with complex geological conditions.

[0004] 2. The calculation process is complex and time-consuming: Especially for heterogeneous slopes with complex joints and weak surfaces, establishing accurate numerical models and performing calculations requires specialized software and a large amount of computing resources, which cannot meet the needs of rapid emergency response or large-scale road network slope surveys.

[0005] 3. Insufficient utilization of dynamic information: Traditional methods are unable to effectively integrate qualitative descriptive information (such as joint morphology and weathering status) obtained from field geological surveys with dynamic monitoring information (such as crack development and seepage changes) in real time, resulting in assessment results lagging behind the actual changes in the slope.

[0006] 4. Poor generalization ability: Mechanical models established based on specific projects are difficult to apply directly to other slopes with different geological conditions, lacking versatility.

[0007] Therefore, there is an urgent need for a method that can quickly, efficiently, and cost-effectively utilize multi-source information for slope stability assessment. Summary of the Invention

[0008] To address the problems of slope stability assessment relying on complex parameters, long parameter acquisition and calculation cycles, and difficulty in integrating multi-source information, this invention provides a rapid slope stability assessment method based on artificial intelligence. It utilizes the powerful nonlinear mapping and multi-source data fusion capabilities of artificial intelligence models to achieve rapid and accurate judgment of slope stability.

[0009] To solve the technical problem, the technical solution adopted by this invention is as follows:

[0010] A rapid slope stability assessment method based on artificial intelligence includes:

[0011] S1: Obtain multi-source heterogeneous data from multiple historical slope cases, including field geological survey data, exploration and test data, and monitoring data, and label each slope case with a stability level label;

[0012] S2: Preprocess the multi-source heterogeneous data, the preprocessing including at least missing value handling, outlier detection and correction, and data consistency verification;

[0013] S3: The preprocessed multi-source heterogeneous data is divided into numerical features, categorical features, and textual descriptive features. Feature engineering is then performed on the classified multi-source heterogeneous data to form a unified structured feature vector. The constituent information in the structured feature vector includes: geological environment features, slope geometric features, rock and soil physical and mechanical features, structural surface features, external load features, and response monitoring features.

[0014] S4: Construct a multi-branch fusion deep neural network model, and train, validate, and test the constructed multi-branch fusion deep neural network model;

[0015] The multi-branch fusion deep neural network model includes an input layer, a specific positive lifting subnetwork, a feature fusion layer, and a classification output layer. The classification output layer includes an output layer, the number of neurons in which is equal to the number of slope stability level categories, and outputs the stability level. The output layer also uses the Softmax activation function to output the probability distribution of the slope belonging to different stability levels.

[0016] The feature vectors obtained in step S3 are divided into training set, validation set and test set, and the constructed multi-branch fusion deep neural network is trained, validated and tested.

[0017] S5: Obtain multi-source information of the target slope and convert the multi-source information of the multi-target slope into the structured feature vector in S3;

[0018] S6: Input the structured feature vectors obtained in S5 into the trained multi-branch fusion deep neural network model for evaluation;

[0019] S7: Output an evaluation report and update the model: The evaluation report includes: basic information of the target slope, stability level and probability output by the model, and contribution of the main features affecting slope stability; after obtaining new monitoring data of the target slope or after an actual landslide event occurs, the new data of the target slope is added to S1, and incremental learning or retraining is performed on the model in S4 to achieve dynamic updating of the model.

[0020] In some embodiments, the construction of the multi-branch fusion deep neural network model in step S4 includes:

[0021] S411: Input layer construction: The number of neurons in the input layer is the same as the dimension of the structured feature vector in step S3;

[0022] S412: Feature extraction subnetwork construction: Construct multiple parallel fully connected subnetworks, each containing 2-3 hidden layers to process feature groups of different types or physical meanings. Each hidden layer uses the ReLU activation function.

[0023] S413: Feature fusion layer construction: The high-level feature vectors output by each sub-network in step S412 are concatenated and fused.

[0024] S414: Classification output layer construction: Connect 2-3 fully connected layers after the feature fusion layer, and the last layer is the output layer. The number of neurons in the output layer is equal to the number of stability level categories (for example, divide the slope stability level into 4 categories: stable, basically stable, unstable, and unstable); and use the Softmax activation function to output the probability distribution of the slope belonging to different stability levels.

[0025] In some embodiments, the raw scores (logits) output by the fully connected layer are vectors Z, Z=[z1,z2,…,z...]. k ], where K is the number of slope stability level categories; the Softmax function transforms Z into a probability distribution p=[p1,p2,…,p k The formula for calculating its ×i×th component is:

[0026] ;

[0027] Wherein, K is the number of slope stability categories (e.g., K=4 in this invention). For category index ( =1 → Stable =2→Basically stable =3→Unstable =4→unstable); For the output layer The original output values ​​of each neuron; An exponential function with the natural constant e (e≈2.71828) as its base, Mapping to positive values ​​amplifies the differences. For the summation of all K exponent values, normalizing the denominator ensures that all The sum is 1; The model predicts that the slope belongs to the first The probability of a class satisfies 0 ≤ ≤1, and ∑ =1.

[0028] In some embodiments, the step S4 of training, validating, and testing the constructed multi-branch fusion deep neural network includes:

[0029] S421: Forward propagation computation; Input the feature vectors of the training set into the model and calculate the predicted probability distribution of the output layer by layer;

[0030] S422: Loss function calculation; using classification cross-entropy as the loss function, calculate the error between the predicted probability distribution and the labeled true labels (i.e., the stability level of expert labeling);

[0031] S423: Backpropagation and Parameter Update; Using the Adam optimizer, backpropagation is performed based on the calculated loss value to update the model's weights and bias parameters layer by layer. In the specific implementation, the initial learning rate is set to 0.001, and the batch size is 32.

[0032] S424: Iterative training and validation; Repeat steps S421 to S423 to perform multiple rounds of iterative training on all samples in the training set; After each round of training, use the validation set to evaluate the accuracy of the current model and save the model parameters with the highest accuracy in the validation set.

[0033] S425: Model Testing; After training is complete, the final model is evaluated once using the step test set. The accuracy, precision, recall, and F1 score of the model on unseen data are calculated as the final performance metrics of the model.

[0034] In some embodiments: the preprocessing of multi-source heterogeneous data in S2 includes:

[0035] S21: Missing value handling. For missing data in each feature field, fill in the missing data using the mean, median, mode, or interpolation method based on the K-nearest neighbor algorithm.

[0036] S22: Outlier detection and correction; use box plots or the 3σ principle to identify outliers for each numerical feature, and correct or remove them in conjunction with business logic;

[0037] S23: Data consistency check; check for logical conflicts between different features within the same case (e.g., the slope type is "soil slope" but the cohesion value is abnormally high), and manually review or correct conflicting data.

[0038] In some embodiments, step S3, which involves classifying the preprocessed multi-source heterogeneous data into numerical features, categorical features, and textual descriptive features, and performing feature engineering on the classified multi-source heterogeneous data to form a unified structured feature vector, includes:

[0039] S31: Standardization of numerical features; For continuous numerical features such as slope height, slope, density, cohesion, internal friction angle, precipitation intensity, and displacement rate, the Z-Score standardization method is used to convert them into a distribution with a mean of 0 and a standard deviation of 1, so as to eliminate the influence of different dimensions on the training of the model (i.e., the multi-branch fusion deep neural network model).

[0040] S32: One-hot encoding of categorical features; For discrete categorical features such as slope type, landform features, and joint type, one-hot encoding is used to convert each category into a binary vector;

[0041] S33: Text-descriptive feature vectorization; For text-descriptive information such as joint filling status, weathering status, vegetation and landform response, construct a predefined professional bag-of-words model or use the TF-IDF algorithm to quantize it into a fixed-dimensional feature vector.

[0042] S34: Feature vector concatenation; All feature vectors processed in steps S31 to S33 are concatenated in a fixed order to form a unified one-dimensional feature vector.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] 1. Significantly Reduces the Data Acquisition Threshold: Traditional slope stability assessments heavily rely on a large number of detailed physical and mechanical parameters of the soil and rock mass (such as cohesion and internal friction angle). These parameters often require complex and expensive laboratory tests, resulting in long assessment cycles and high costs. This invention, by constructing a multi-source heterogeneous database, can effectively integrate field geological surveys, preliminary investigations, and monitoring data, reducing the dependence on single high-precision parameters and enabling effective assessments even in situations with scarce data or emergency rescue scenarios.

[0045] This invention does not require costly and precise mechanical parameters. Even with only qualitative descriptions (such as "joints are well-developed, open and filled with cohesive soil") and basic geometric information available from field surveys, the model can be effectively evaluated through text vectorization methods. This greatly reduces the reliance on professional tests and expensive equipment for evaluation.

[0046] 2. Achieving deep fusion and utilization of multi-source heterogeneous data: Existing technologies often struggle to simultaneously process multiple data types, including numerical, categorical, and textual descriptions. This invention utilizes feature engineering techniques, particularly employing a specialized bag-of-words model or TF-IDF algorithm, to vectorize and quantize textual descriptions of joint filling states and weathering states. This solves the problem of unstructured textual data being difficult to incorporate into models, fully leveraging the implicit information contained in the experience of geological experts, and improving the comprehensiveness and accuracy of the assessment results.

[0047] Furthermore, through feature engineering and multi-molecular network design, this invention can organically combine qualitative, quantitative, static, and dynamic information to comprehensively judge the slope condition from multiple dimensions. Its information utilization rate and decision-making logic are closer to those of experts, while avoiding the subjectivity of expert evaluation.

[0048] 3. Fast evaluation speed and strong real-time performance: Traditional numerical calculation methods are time-consuming when dealing with complex geological conditions, making it difficult to meet the needs of high-frequency monitoring and real-time early warning. The multi-branch fusion deep neural network model constructed in this invention only requires one forward propagation after training to output results, achieving a second (millisecond) response to slope stability, which is particularly suitable for real-time analysis of mass monitoring and automated monitoring systems.

[0049] Unlike commonly used fully connected black-box networks, the multi-branch sub-network structure constructed in this invention can independently extract and fuse high-level features for feature groups with different physical meanings, such as geometric physical parameters and structural surface features. This enhances the model's ability to map complex nonlinear geomechanical relationships and improves the model's generalization performance and physical interpretability.

[0050] 4. Supports dynamic model evolution and full lifecycle management: Static models built using existing methods are difficult to adapt to environmental changes and the generation of new data. This invention introduces a dynamic update and incremental learning mechanism, which can continuously update model parameters based on newly added monitoring data and actual landslide cases. This enables the assessment system to have self-optimization and self-adaptation capabilities, extends the lifecycle of the technology, and ensures the reliability of long-term application. Attached Figure Description

[0051] Figure 1 This is a flowchart of an embodiment of the present invention. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. The described embodiments are some, but not all, of the embodiments of the present invention. The following embodiments are only for specific illustration of the implementation methods of the present invention and do not limit the scope of protection of the present invention.

[0053] Referring to the accompanying drawings, the artificial intelligence-based rapid slope stability assessment method of the present invention includes:

[0054] S1: Obtain multi-source heterogeneous data of multiple historical slope cases. The multi-source heterogeneous data includes field geological survey data, exploration test data and monitoring data, and label each slope case with a stability level (for example, divide the slope stability level into 4 categories: stable, basically stable, understability and unstable).

[0055] S2: Preprocess the multi-source heterogeneous data. The preprocessing includes at least missing value handling, outlier detection and correction, and data consistency verification. This effectively eliminates the interference of low-quality data on model training, improving the robustness of the model and the reliability of the evaluation results.

[0056] S3: The preprocessed multi-source heterogeneous data is divided into numerical features, categorical features, and textual descriptive features. Feature engineering is performed on the classified multi-source heterogeneous data to form a unified structured feature vector. The constituent information in the structured feature vector includes: geological environment features, slope geometric features, rock and soil physical and mechanical features, structural surface features, external load features, and response monitoring features.

[0057] S4: Construct a multi-branch fusion deep neural network model, and train, validate, and test the constructed multi-branch fusion deep neural network model;

[0058] The multi-branch fusion deep neural network model includes an input layer, a specific positive lifting subnetwork, a feature fusion layer, and a classification output layer. The classification output layer includes an output layer, the number of neurons in which is equal to the number of slope stability level categories, and outputs the stability level. The output layer also uses the Softmax activation function to output the probability distribution of the slope belonging to different stability levels.

[0059] The feature vectors obtained in step S3 are divided into training set, validation set and test set, and the constructed multi-branch fusion deep neural network is trained, validated and tested.

[0060] S5: Obtain multi-source information of the target slope and convert the multi-source information of the multi-target slope into the structured feature vector in S3.

[0061] S6: Input the structured feature vectors obtained in S5 into the trained multi-branch fusion deep neural network model for evaluation.

[0062] S7: Output an evaluation report and update the model: The evaluation report includes: basic information of the target slope, stability level and probability output by the model, and contribution of the main features affecting slope stability; after obtaining new monitoring data of the target slope or after an actual landslide event occurs, the new data of the target slope is added to S1, and incremental learning or retraining is performed on the model in S4 to achieve dynamic updating of the model.

[0063] In practical implementation, the contribution of key features affecting stability can be output through model interpretability methods (such as SHAP value calculation), providing a basis for engineers' decision-making and enhancing the credibility of artificial intelligence evaluation results.

[0064] In some embodiments, the construction of the multi-branch fusion deep neural network model in step S4 includes:

[0065] S411: Input layer construction: The number of neurons in the input layer is the same as the dimension of the structured feature vector in step S3;

[0066] S412: Feature extraction subnetwork construction: Construct multiple parallel fully connected subnetworks, each containing 2-3 hidden layers to process feature groups of different types or physical meanings. Each hidden layer uses the ReLU activation function.

[0067] S413: Feature fusion layer construction: The high-level feature vectors output by each sub-network in step S412 are concatenated and fused.

[0068] S414: Classification output layer construction: Connect 2-3 fully connected layers after the feature fusion layer, and the last layer is the output layer. The number of neurons in the output layer is equal to the number of stability level categories (for example, divide the slope stability level into 4 categories: stable, basically stable, unstable, and unstable); and use the Softmax activation function to output the probability distribution of the slope belonging to different stability levels.

[0069] The definition of ReLU (Rectified Linear Unit) is: ReLU(x) = max(0,x). After each fully connected layer completes the linear transformation, a ReLU nonlinear transformation is immediately followed to introduce nonlinear mapping capability. This enables each feature extraction subnetwork to learn the complex nonlinear relationship between the soil mass and the slope geometry, structural surface and monitoring features, while avoiding the gradient vanishing problem in the training process of deep networks.

[0070] In the slope stability assessment scenario of this invention, the problem is to deal with highly heterogeneous, multi-source heterogeneous engineering feature data with significant differences in dimensions. Based on the grouping processing requirements of the feature extraction sub-network, the ReLU activation function is selected in a targeted manner to effectively suppress the gradient vanishing problem in the deep network training process while ensuring nonlinear expression capability, thereby achieving stable modeling of complex geomechanical relationships.

[0071] This invention employs the ReLU activation function in every hidden layer of the multi-branch fully connected sub-network, rather than simply following general deep learning conventions. It fully considers the practical engineering characteristics of slope stability assessment, such as limited sample size and significant differences in the physical meaning of features. The sparse activation properties of ReLU are highly compatible with the requirements for screening and fusing multi-source heterogeneous features in this invention, which can improve the model's generalization performance while reducing redundant computation, thereby significantly improving inference efficiency while ensuring assessment accuracy.

[0072] The ReLU activation function of this invention is not used in isolation, but rather forms a collaborative optimization whole with the multi-branch feature extraction sub-network, feature fusion layer, and Softmax output layer. Its introduced nonlinear mapping capability, combined with the sub-network's group feature learning mechanism, enables the model to capture the complex nonlinear relationships between geometric parameters, structural surface features, and monitoring responses, thereby achieving higher stability discrimination accuracy and robustness than traditional single fully connected networks.

[0073] In some embodiments, the raw scores (logits) output by the fully connected layer are vectors Z, Z=[z1,z2,…,z...]. k ], where K is the number of slope stability level categories (for example, in this invention, slope stability levels are divided into 4 categories: stable, basically stable, slightly unstable, and unstable, therefore, K=4); the Softmax function transforms Z into a probability distribution p=[p1,p2,…,p k The formula for calculating its ×i×th component is:

[0074] ;

[0075] Wherein, K is the number of slope stability categories (e.g., K=4 in this invention). For category index ( =1 → Stable =2→Basically stable =3→Unstable =4→unstable); For the output layer The original output values ​​of each neuron; An exponential function with the natural constant e (e≈2.71828) as its base, Mapping to positive values ​​amplifies the differences. For the summation of all K exponent values, normalizing the denominator ensures that all The sum is 1; The model predicts that the slope belongs to the first The probability of a class satisfies 0 ≤ ≤1, and ∑ =1.

[0076] For example: If the original output of the output layer is z=[2.0,1.0,0.1,0.5], then:

[0077] The denominator in the above formula ( )=e 2 +e 1 +e 0.1+ e 0.5 ≈7.389 + 2.718 + 1.105 + 1.649 = 12.861

[0078] p1 = 7.389 / 12.861 ≈ 0.574 (stable, probability 57.4%)

[0079] p2 = 2.718 / 12.861 ≈ 0.211 (basically stable, probability 21.1%)

[0080] p3 = 1.105 / 12.861 ≈ 0.086 (Unstable, probability 8.6%)

[0081] p4 = 1.649 / 12.861 ≈ 0.128 (unstable, probability 12.8%).

[0082] The model (i.e., a deep neural network model with multiple branches) ultimately outputs the class with the highest probability, namely "stable" (maximum p1).

[0083] In some embodiments, the step S4 of training, validating, and testing the constructed multi-branch fusion deep neural network includes:

[0084] S421: Forward propagation computation; Input the feature vectors of the training set into the model and calculate the predicted probability distribution of the output layer by layer;

[0085] S422: Loss function calculation; using classification cross-entropy as the loss function, calculate the error between the predicted probability distribution and the labeled true labels (i.e., the stability level of expert labeling);

[0086] S423: Backpropagation and Parameter Update; Using the Adam optimizer, backpropagation is performed based on the calculated loss value to update the model's weights and bias parameters layer by layer. In the specific implementation, the initial learning rate is set to 0.001, and the batch size is 32.

[0087] S424: Iterative training and validation; Repeat steps S421 to S423 to perform multiple rounds of iterative training on all samples in the training set; After each round of training, use the validation set to evaluate the accuracy of the current model and save the model parameters with the highest accuracy in the validation set.

[0088] S425: Model Testing; After training is complete, the final model is evaluated once using the step test set. The accuracy, precision, recall, and F1 score of the model on unseen data are calculated as the final performance metrics of the model.

[0089] In the specific implementation process, the unified structured feature vectors obtained in step S3 are divided into three subsets. 70% of the samples are randomly selected as the training set for learning and optimizing model parameters. 20% of the samples are randomly selected as the validation set for adjusting model hyperparameters and monitoring overfitting during training. The remaining 10% of the samples are used as the test set for the final evaluation of the model's generalization performance after training. The samples in these three subsets do not overlap.

[0090] In some embodiments: the preprocessing of multi-source heterogeneous data in S2 includes:

[0091] S21: Missing value handling. For missing data in each feature field, fill in the missing data using the mean, median, mode, or interpolation method based on the K-nearest neighbor algorithm.

[0092] S22: Outlier detection and correction; use box plots or the 3σ principle to identify outliers for each numerical feature, and correct or remove them in conjunction with business logic;

[0093] S23: Data consistency check; check for logical conflicts between different features within the same case (e.g., the slope type is "soil slope" but the cohesion value is abnormally high), and manually review or correct conflicting data.

[0094] In some embodiments, step S3, which involves classifying the preprocessed multi-source heterogeneous data into numerical features, categorical features, and textual descriptive features, and performing feature engineering on the classified multi-source heterogeneous data to form a unified structured feature vector, includes:

[0095] S31: Standardization of numerical features; For continuous numerical features such as slope height, slope, density, cohesion, internal friction angle, precipitation intensity, and displacement rate, the Z-Score standardization method is used to convert them into a distribution with a mean of 0 and a standard deviation of 1, so as to eliminate the influence of different dimensions on the training of the model (i.e., the multi-branch fusion deep neural network model).

[0096] S32: One-hot encoding of categorical features; For discrete categorical features such as slope type, landform features, and joint type, one-hot encoding is used to convert each category into a binary vector;

[0097] S33: Text-descriptive feature vectorization; For text-descriptive information such as joint filling status, weathering status, vegetation and landform response, construct a predefined professional bag-of-words model or use the TF-IDF algorithm to quantize it into a fixed-dimensional feature vector.

[0098] S34: Feature vector concatenation; All feature vectors processed in steps S31 to S33 are concatenated in a fixed order to form a unified one-dimensional feature vector.

[0099] This invention proposes a systematic preprocessing process from data preprocessing (S2), feature engineering (S3) to dataset partitioning (S4), which ensures the quality and consistency of the input data for the model (i.e., the multi-branch fusion deep neural network model).

[0100] This invention ensures the trainability and reliability of the model's generalization performance by dividing the training set, validation set, and test set, and by implementing a complete training process including forward propagation, loss calculation, backpropagation, and iterative verification.

[0101] The feature engineering process of this invention differentiates multiple types of features before merging them: instead of simply inputting all data directly into the model, it uses ZScore standardization, one-hot encoding, and professional bag-of-words / TFIDL algorithm vectorization to perform targeted quantization based on the data characteristics of numerical, categorical, and textual descriptive features, and then concatenates them into a fixed-dimensional feature vector, thus realizing the effective digitization of semi-structured and unstructured information in the geotechnical field.

[0102] Introducing a domain-adaptive text vectorization approach: For geotechnical engineering text descriptions such as "joint filling state," "weathering state," and "vegetation and geomorphological response," a predefined professional bag-of-words model or TFIDF algorithm is used for vectorization, rather than general natural language processing methods. This preserves the semantic experience of domain experts and improves the model's ability to perceive qualitative descriptive information.

[0103] The feature engineering process of this invention forms a unified structured feature vector: by splicing the feature vectors in a fixed order, a standard input vector is clearly constructed that includes geological environment features, slope geometric features, rock and soil physical and mechanical features, structural surface features, external load features, and response monitoring features, so that subsequent models can be directly and systematically trained and inferred.

[0104] In the specific implementation process, for ease of understanding, the following supplementary explanation is provided for the discrete categorical features described in step S32:

[0105] Slope types: soil slope, rock slope;

[0106] Landform features: high mountains, medium mountains, low mountains, and hills;

[0107] Joint types: genetic type (primary / structural / non-structural), mechanical type (shear joint / tension joint / conjugate shear joint), attitude relationship (strike / dip / oblique / bedding), assemblage morphology (columnar / plate-like / rhomboid / en echelon / dendritic), degree of opening (wide (>5mm) / open (3~5mm) / slightly open (1~3mm) / closed (<1mm)).

[0108] Slope stability varies with the condition of the soil and rock mass itself and the surrounding environment. Slope type, geomorphological features, and joint types describe the objective environment in which the slope exists. These characteristics form the basic material basis for the spatiotemporal evolution of slopes and play a crucial role in predicting future slope stability and potential risks. Based on these descriptions of geological environmental characteristics, the potential risk of a slope developing into a landslide can be assessed, improving the accuracy of stability assessments for slopes that currently appear stable but may evolve into landslides by at least 30%.

[0109] For ease of understanding, the joint filling state, weathering state, and vegetation and landform response in the text-descriptive feature vectorization of step S33 are explained as follows:

[0110] Joint filling condition refers to the material composition, thickness, degree of cementation, and water content of the filling material in joints and fissures, specifically including the following sub-items:

[0111]

[0112] Weathering state refers to the degree to which a rock mass has been altered by physical, chemical, and biological weathering processes in a surface or near-surface environment. It is usually classified into the following levels:

[0113]

[0114] Vegetation and geomorphological response refers to the indicative characteristics of slope stability on the surface vegetation development and geomorphological features, specifically including:

[0115]

[0116]

[0117] Qualitative descriptions are an indispensable source of information for slope stability analysis. In traditional engineering geology practice, qualitative descriptions such as joint filling status, weathering degree, and vegetation and geomorphological characteristics are the core basis for geological engineers to make preliminary judgments on slope stability.

[0118] For example:

[0119] Joints filled with moist clay have low shear strength and are prone to sliding along the joint surface.

[0120] Strongly weathered rock mass → large reduction in mechanical parameters → significantly reduced stability;

[0121] Slope cracks or ridge fissures indicate slow creep on the slope, which is in an unstable state.

[0122] These qualitative information information contain rich indicative meanings of stability, but traditional numerical methods struggle to quantify them and incorporate them into computational models. This invention, however, can achieve the automated reuse of empirical knowledge by learning the mapping relationship between these textual descriptions and stability results.

[0123] The table below shows the accuracy of the model in assessing slope stability under different data conditions. Specifically, it shows the accuracy of the model using numerical features, the accuracy of the model using both numerical and categorical features, and the accuracy of the model using numerical, categorical, and textual descriptive features.

[0124] Comparison table of model accuracy when using different data features:

[0125]

[0126] The table above clearly shows that this invention categorizes multi-source heterogeneous data into numerical features, categorical features, and textual descriptive features, which significantly improves the accuracy of slope stability assessment. Furthermore, many historical slope data only contain qualitative descriptions based on geological logging, lacking precise mechanical test data. By incorporating textual features, these "old data" can still be used for model training and evaluation.

[0127] Textual descriptions comprehensively reflect engineers' on-site intuition and experience, essentially making the experts' tacit knowledge explicit, which helps the model learn general rules under different geological conditions. Furthermore, when the model outputs evaluation results, methods such as SHAP values ​​can be used to infer which textual description features (such as "jointed infill clay") contributed most to the current judgment, making it easier for engineers to understand and accept.

[0128] In the specific implementation process, the applicant's research team's previous test on 120 slope cases of a highway in Southwest China showed that after adding text features of joint filling state and weathering state, the model's recall rate for "understood" slopes increased from 71% to 84%.

[0129] Example 1: Implementation scenario of this example: Rapid stability screening of 50 red bed slopes in a county in northeastern Sichuan, specifically including:

[0130] 1. Constructing multi-source heterogeneous slope data: Collecting data on 500 slope cases from the past 20 years in this province and surrounding areas that have experienced clear instability events or have remained stable for a long period. The data includes all eight categories of information provided by the user: regional geology, slope type, geometric information, physical information, joint characteristics, weak surface characteristics, rainfall information, and monitoring information. The final stability state (stable, basically stable, understability, unstable) of each case is labeled and determined by three senior geotechnical experts through joint review.

[0131] 2. Preprocessing of multi-source heterogeneous data: Missing value handling: For example, if there are 30 missing values ​​in the "cohesion" field out of 500 samples, the median of this field is used for filling. Outlier detection and correction: When plotting the "slope height" feature, a box plot was drawn, revealing that the slope height of one sample was 800m (significantly exceeding the normal range). After verification, it was found to be an input error; the actual value should be 80m, and this was corrected. Data consistency check: It was found that the "slope type" of one sample was "soil slope," but the "internal friction angle" was 45° (which is usually only possible for rock slopes). After manual verification, it was confirmed that the sample should be a "rock slope," and this was corrected.

[0132] 3. The preprocessed data is divided into numerical features, categorical features, and textual descriptive features. Feature engineering is performed on the classified multi-source heterogeneous data to form a unified structured feature vector: Numerical feature standardization: For numerical features such as slope height (unit: m), slope (unit: °), and cohesion (unit: kPa), their mean and standard deviation are calculated, and Z-Score standardization is performed. Categorical feature one-hot encoding: "Slope type" has 3 values: soil slope, rock slope, and soil-rock combined slope, which are encoded into 3-bit binary vectors, such as soil slope → [1,0,0]. Textual descriptive feature vectorization: For the text description of "joint filling status" (such as "no filling, closed", "filled with 5mm thick clay, open"), a professional bag of words is pre-constructed, containing 50 keywords such as {closed, open, filled, no filling, clay, rock debris...}, and the frequency of each keyword is counted to form a 50-dimensional TF-IDF feature vector. Feature vector concatenation: After all features are processed, each sample forms a one-dimensional feature vector with a fixed length of 280 dimensions.

[0133] 4. Construct a multi-branch fusion deep neural network model, and train, validate, and test the constructed multi-branch fusion deep neural network model; the multi-branch fusion deep neural network model includes an input layer, a specific positive lifting subnetwork, a feature fusion layer, and a classification output layer, wherein the classification output layer includes an output layer, the number of neurons in the output layer is equal to the number of slope stability level categories, and outputs the stability level; the output layer uses the Softmax activation function to output the probability distribution of the slope belonging to different stability levels;

[0134] The feature vectors obtained in step S3 are divided into a training set (350 samples, 70%), a validation set (100 samples, 20%), and a test set (50 samples, 10%) to train, validate, and test the constructed multi-branch fusion deep neural network.

[0135] Constructing a deep neural network model that integrates multiple branches includes:

[0136] Input layer; set the number of neurons in the input layer to 280.

[0137] Feature extraction subnetwork; three parallel subnetworks are constructed: Subnetwork 1: processes geometric and physical features (input dimension 80), containing two hidden layers with 64 and 32 neurons respectively. Subnetwork 2: processes joint and weak surface features (input dimension 120), containing two hidden layers with 128 and 64 neurons respectively. Subnetwork 3: processes monitoring and environmental features (input dimension 80), containing two hidden layers with 64 and 32 neurons respectively. Each hidden layer is followed by a ReLU activation function and Dropout (0.2).

[0138] Feature fusion layer: The output vectors of the three sub-networks (32+64+32=128 dimensions) are concatenated to obtain 128-dimensional fused features.

[0139] Classification output layer: The fusion layer is followed by two fully connected layers with 64 and 4 neurons respectively. The last layer uses the Softmax activation function to output the probabilities of four stable categories.

[0140] Training, validating, and testing the model (i.e., the constructed multi-branch fusion deep neural network model) includes:

[0141] Forward propagation: Input the 280-dimensional features of a batch (batch size 32) of the training set into the model to obtain the predicted probability.

[0142] Loss function calculation: The error between the predicted probability and the true label is calculated using the classification cross-entropy.

[0143] Backpropagation and parameter update: Update model parameters using the Adam optimizer (learning rate 0.001).

[0144] Iterative training and validation: The training proceeds for 200 epochs. After each epoch, the accuracy is calculated on the validation set, and the model parameters with the highest accuracy on the validation set are saved. Training is stopped prematurely when the validation set accuracy no longer improves for 10 consecutive epochs.

[0145] Model testing: After training, the final model was evaluated on the test set, achieving an accuracy of 92%, precision of 0.91, recall of 0.92, and an F1 score of 0.915.

[0146] 5. Obtain multi-source information of the target slope and convert the multi-source information of the multi-target slope into a structured feature vector in S3.

[0147] The multi-source information for the target slope may include the following: Engineering technicians bring tablets to the site of slope No. 1 to be evaluated. A drone is used to quickly acquire information on the slope height of 35m, slope angle of 42°, and geomorphological features classified as "erosion and denudation." On-site observation and recording show that the slope type is "rock slope," joint development is moderate, and the joint filling state is "a small amount of clay filling, slightly tensile." The density is obtained from the road section design data as 2.2g / cm³. 3 The cohesion is 20 kPa, and the internal friction angle is 28°. An automated monitoring system was connected to obtain the cumulative rainfall of 65 mm over the past 7 days and the slope toe crack displacement change rate of 2 mm / day.

[0148] 6. The obtained structured feature vectors are input into a trained multi-branch fusion deep neural network model for evaluation; the model computation time is 0.3 seconds, and the output results are: [Stable: 0.02, Basically stable: 0.05, Understable: 0.78, Unstable: 0.15]. The slope is determined to be "understable", and the main risk factors are the recent heavy rainfall and the accelerated development rate of cracks at the slope toe.

[0149] 7. Output an evaluation report and update the model: The evaluation report includes: basic information of the target slope, the stability level and probability output by the model, and the contribution of the main features affecting slope stability; after obtaining new monitoring data of the target slope or after an actual landslide event occurs, the new data of the target slope is added to S1, and the model in S4 is incrementally learned or retrained to achieve dynamic updates of the model, specifically including:

[0150] Report generation: The tablet computer automatically generates a PDF assessment report, which includes the slope location, basic information, model output probability, and the contribution of each feature calculated by SHAP value (showing rainfall, crack displacement rate, and joint filling status as the most important factors).

[0151] Dynamic Update: One month later, a localized shallow landslide occurred on the slope. Engineers added the monitoring data from this event and the period prior to the landslide to the database in step S1 as new samples. The model is incrementally trained quarterly to enhance its sensitivity when evaluating similar slopes in the future.

[0152] In practice, for cases with a small sample size, data augmentation can be performed by adding minor noise to numerical features or replacing text features with synonyms to expand the training set and enhance the data.

[0153] In addition to choosing MLP (Multilayer Perceptron / Fully Connected Network), a graph neural network (GNN)-based model can also be used, which treats different geological units of the slope as nodes and transmits information through the relationship of edges, which is more in line with the spatial structural characteristics of the slope.

[0154] For situations with abundant monitoring data, LSTM or Transformer models can be used to replace some of the fully connected subnetworks to specifically handle time series features such as displacement and rainfall, and capture dynamic change patterns.

[0155] In addition to outputting stability levels, the model can also be designed to output continuous safety coefficient values ​​(for regression tasks). This can be achieved by modifying the output layer to a single neuron (without an activation function) and changing the loss function to mean squared error (MSE).

[0156] It should be understood that the above description of the preferred embodiments is quite detailed, but it should not be considered as a limitation on the scope of protection of this invention. Those skilled in the art, under the guidance of this invention, can make substitutions or modifications without departing from the scope of protection of the claims of this invention, and all such substitutions or modifications fall within the scope of protection of this invention. The scope of protection of this invention should be determined by the appended claims.

Claims

1. A rapid slope stability assessment method based on artificial intelligence, characterized in that, include: S1: Obtain multi-source heterogeneous data from multiple historical slope cases, including field geological survey data, exploration and test data, and monitoring data, and label each slope case with a stability level label; S2: Preprocess the multi-source heterogeneous data, the preprocessing including at least missing value handling, outlier detection and correction, and data consistency verification; S3: The preprocessed multi-source heterogeneous data is divided into numerical features, categorical features, and textual descriptive features. Feature engineering is then performed on the classified multi-source heterogeneous data to form a unified structured feature vector. The constituent information in the structured feature vector includes: geological environment features, slope geometric features, rock and soil physical and mechanical features, structural surface features, external load features, and response monitoring features. S4: Construct a multi-branch fusion deep neural network model, and train, validate, and test the constructed multi-branch fusion deep neural network model; The multi-branch fusion deep neural network model includes an input layer, a specific positive lifting subnetwork, a feature fusion layer, and a classification output layer. The classification output layer includes an output layer, the number of neurons in which is equal to the number of slope stability level categories, and outputs the stability level. The output layer also uses the Softmax activation function to output the probability distribution of the slope belonging to different stability levels. The feature vectors obtained in step S3 are divided into training set, validation set and test set, and the constructed multi-branch fusion deep neural network is trained, validated and tested. S5: Obtain multi-source information of the target slope and convert the multi-source information of the multi-target slope into the structured feature vector in S3; S6: Input the structured feature vectors obtained in S5 into the trained multi-branch fusion deep neural network model for evaluation; S7: Output an evaluation report and update the model: The evaluation report includes: basic information of the target slope, stability level and probability output by the model, and contribution of the main features affecting slope stability; after obtaining new monitoring data of the target slope or after an actual landslide event occurs, the new data of the target slope is added to S1, and incremental learning or retraining is performed on the model in S4 to achieve dynamic updating of the model.

2. The rapid slope stability assessment method based on artificial intelligence according to claim 1, characterized in that, The construction of the multi-branch fusion deep neural network model in step S4 includes: S411: Input layer construction: The number of neurons in the input layer is the same as the dimension of the structured feature vector in step S3; S412: Feature extraction subnetwork construction: Construct multiple parallel fully connected subnetworks, each containing 2-3 hidden layers to process feature groups of different types or physical meanings. Each hidden layer uses the ReLU activation function. S413: Feature fusion layer construction: The high-level feature vectors output by each sub-network in step S412 are concatenated and fused. S414: Classification output layer construction: Connect 2-3 fully connected layers after the feature fusion layer, and the last layer is the output layer. The number of neurons in the output layer is equal to the number of stability level categories. The Softmax activation function is used to output the probability distribution of the slope belonging to different stability levels.

3. The rapid slope stability assessment method based on artificial intelligence according to claim 2, characterized in that, The original score output by the fully connected layer is a vector Z, Z=[z1,z2,…,z]. k The Softmax function transforms Z into a probability distribution p=[p1,p2,…,p] k The formula for calculating its ×i×th component is: ； Where K is the number of slope stability categories; For category index ( =1 → Stable =2→Basically stable =3→Unstable =4→unstable); For the output layer The original output values ​​of each neuron; An exponential function with the natural constant e (e≈2.71828) as its base, Mapping to positive values ​​amplifies the differences. For the summation of all K exponent values, normalizing the denominator ensures that all The sum is 1; The model predicts that the slope belongs to the first The probability of a class satisfies 0 ≤ ≤1, and ∑ =1.

4. The rapid slope stability assessment method based on artificial intelligence according to claim 3, characterized in that, The step S4 of training, validating, and testing the constructed multi-branch fusion deep neural network includes: S421: Forward propagation computation; Input the feature vectors of the training set into the model and calculate the predicted probability distribution of the output layer by layer; S422: Loss function calculation; using classification cross-entropy as the loss function, the error between the predicted probability distribution and the labeled true label is calculated. S423: Backpropagation and parameter update; Using the Adam optimizer, backpropagation is performed based on the calculated loss value to update the model's weights and bias parameters layer by layer; S424: Iterative training and validation; Repeat steps S421 to S423 to perform multiple rounds of iterative training on all samples in the training set; After each round of training, use the validation set to evaluate the accuracy of the current model and save the model parameters with the highest accuracy in the validation set. S425: Model Testing; After training is complete, the final model is evaluated once using the step test set. The accuracy, precision, recall, and F1 score of the model on unseen data are calculated as the final performance metrics of the model.

5. The rapid slope stability assessment method based on artificial intelligence according to claim 1, characterized in that, The preprocessing of multi-source heterogeneous data in S2 includes: S21: Missing value handling. For missing data in each feature field, fill in the missing data using the mean, median, mode, or interpolation method based on the K-nearest neighbor algorithm. S22: Outlier detection and correction; use box plots or the 3σ principle to identify outliers for each numerical feature, and correct or remove them in conjunction with business logic; S23: Data consistency check; check for logical conflicts between different features within the same case, and manually review or correct conflicting data.

6. The rapid slope stability assessment method based on artificial intelligence according to any one of claims 1-5, characterized in that, Step S3 involves classifying the preprocessed multi-source heterogeneous data into numerical features, categorical features, and textual descriptive features. Feature engineering is then performed on each of these classified multi-source heterogeneous data to form a unified structured feature vector. This includes: S31: Standardization of numerical features; For continuous numerical features such as slope height, slope, density, cohesion, internal friction angle, precipitation intensity, and displacement rate, the Z-Score standardization method is used to convert them into a distribution with a mean of 0 and a standard deviation of 1, so as to eliminate the influence of different units on model training. S32: One-hot encoding of categorical features; For discrete categorical features such as slope type, landform features, and joint type, one-hot encoding is used to convert each category into a binary vector; S33: Text-descriptive feature vectorization; For text-descriptive information such as joint filling status, weathering status, vegetation and landform response, construct a predefined professional bag-of-words model or use the TF-IDF algorithm to quantize it into a fixed-dimensional feature vector. S34: Feature vector concatenation; All feature vectors processed in steps S31 to S33 are concatenated in a fixed order to form a unified one-dimensional feature vector.

Images