Slope stability assessment method based on large language model and intelligent prediction model

Abstract

The invention discloses a slope stability evaluation method based on a large language model and an intelligent prediction model. The method specifically comprises the steps of establishing a slope training database and a large language model training database; constructing a slope stability prediction artificial intelligence model and a slope field large language model; slope parameter extraction is carried out on the data uploaded by the user through a slope field big language model, and whether the extracted slope parameters meet the input requirements of a slope stability prediction artificial intelligence model or not is judged; if the slope parameters in the data uploaded by the user are incomplete, the slope field big language model feeds back the slope parameters needing to be supplemented to the user; after the users are supplemented, the slope field big language model is extracted again and judged again until it is monitored that the users have uploaded all slope parameters; the slope field big language model sends the complete slope parameters into the trained slope stability prediction artificial intelligence model, and the slope stability prediction artificial intelligence model outputs the stability coefficient of the slope.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent geotechnical engineering assessment, and in particular to a slope stability assessment method based on a large language model and an intelligent prediction model. Background Art

[0002] AI-based slope stability assessment methods are improving assessment methods in the engineering field. This approach allows for accurate and cost-effective assessment of slope stability, broadens access to professional analysis results, and reduces the skills and time required for assessment. AI-driven slope stability assessments primarily predict slope stability by analyzing the geometric characteristics and material properties of the slope. A key goal is to ensure that these models provide accurate predictions when compared to mechanical analysis methods.

[0003] However, existing AI-based slope stability assessment methods often ignore the challenges that users may encounter in actual operations: users may not be familiar with the operation of AI models, the cumbersome problem decomposition and reporting process, and the high requirements for reporting personnel, all of which may lead to a decline in analysis quality, thereby affecting the effectiveness of AI methods in slope stability assessment. Users still face many obstacles in actual application. For example, the geometric coordinates and material parameters of the slope need to be arranged in advance before they can be used to input into the slope stability prediction model. The arrangement positions of these parameters usually have a certain template, and all parameters must be placed in the correct position to accurately reflect the actual working conditions of the slope to be predicted. Before implementing these operations, users must undergo training to acquire the relevant skills.

[0004] The advent of ChatGPT has transformed public perception of the potential of artificial intelligence in industrial applications. Large language models demonstrate performance comparable to humans with normal or even expert intelligence when following human instructions during conversations. However, conversational scenarios involving specialized slope engineering expertise are often extremely complex, exceeding the processing capabilities of large language models in general domains. Therefore, large language models for slope expertise should be trained using relevant vertical domain datasets, while also being consistent with the distribution of real-world slope knowledge conversations. The development of large language models in vertical domains such as healthcare, law, finance, and education has provided inspiration for the present invention. These studies have refined instructions using specialized domain knowledge question-and-answer datasets, improving the models' ability to understand and execute specific content. These advances suggest the potential for developing specialized conversational systems for slope engineering. However, large language models specifically for slope stability assessment have yet to be developed, and insufficient attention has been paid to the development of scarce dataset technologies, the development of evaluation metrics, data extraction techniques, and the impact of training strategies, sample size, and iteration count on model performance. These factors are crucial for improving the domain adaptability, accuracy, and generalization of large language models, especially in specialized domains where data is scarce. Summary of the Invention

[0005] In order to overcome the shortcomings of the existing technology, the present invention provides a slope stability assessment method based on a large language model and an intelligent prediction model. The slope stability assessment method not only has higher assessment accuracy and efficiency, but also reduces the need for professional knowledge, making slope stability assessment more convenient and reliable.

[0006] The technical solution of the present invention to solve the above technical problems is:

[0007] A slope stability assessment method based on a large language model and an intelligent prediction model includes the following steps:

[0008] Step S1: Establish a slope training database based on typical slope stability cases;

[0009] Step S2: establishing a large language model training database based on existing research literature on intelligent slope stability assessment;

[0010] Step S3: constructing an artificial intelligence model for slope stability prediction based on the slope training database;

[0011] Step S4: constructing a large language model for the slope domain based on the large language model training database;

[0012] Step S5: The slope parameters of the user's uploaded data are extracted through the trained slope domain big language model, and it is judged whether the extracted slope parameters meet the input requirements of the slope stability prediction artificial intelligence model; if the slope parameters in the user's uploaded data do not meet the input requirements of the slope stability prediction artificial intelligence model, the slope domain big language model will feedback the slope parameters that need to be supplemented to the user; after the user supplements, the slope domain big language model re-extracts and judges again until it monitors that the user has uploaded all slope parameters that meet the input requirements of the slope stability prediction artificial intelligence model; the slope domain big language model sends the complete slope parameters to the trained slope stability prediction artificial intelligence model, and the slope stability prediction artificial intelligence model outputs the stability coefficient of the slope.

[0013] Preferably, in step S1, the steps of constructing the slope training database are:

[0014] Step S100: Collect typical slope stability cases and extract slope parameters of each slope stability case; the slope parameters include slope height, slope angle, unit weight of slope rock and soil layer, internal friction angle of slope rock and soil layer, and cohesion of slope rock and soil layer;

[0015] Step S110: Based on the collected slope parameters, taking a single slope stability case as a basic sample, randomly extracting slope parameters from multiple slope stability cases and combining them to construct multiple different layered slopes, wherein each slope layer has different slope angles and soil parameters;

[0016] Step S120: determining a probability distribution function of slope structural surface parameters and its corresponding mean and variance, and using the probability density function to generate random structural surfaces for the layered slope, wherein the structural surface parameters include cohesion, internal friction angle, centroid coordinates of the structural surface, structural surface inclination, structural surface length, and centroid spacing between different structural surfaces;

[0017] Step S130: Diversify the combinations of the constructed layered slopes and the generated structural surfaces to construct multiple different slope cases;

[0018] Step S140: using the extended finite element method combined with the strength reduction method to calculate the stability coefficient of each slope case, forming a corresponding relationship pair of "slope case-stability coefficient", thereby obtaining the slope training database.

[0019] Preferably, in step S2, the construction of the large language model training database includes the following steps:

[0020] Construct a slope knowledge corpus dataset and a slope prediction corpus dataset, and perform human supervision and data optimization on the constructed slope knowledge corpus dataset and slope prediction corpus dataset.

[0021] Preferably, in step S2, the steps of constructing the slope knowledge corpus dataset are as follows: step S201: searching for literature related to slope stability assessment, artificial intelligence slope analysis, and geotechnical parameter inversion through an academic platform, downloading the full text of the PDF, and establishing an original document library; using a Python library to extract the PDF text page by page, filtering non-technical content, formulas, and chart descriptions, retaining the main text paragraphs, and segmenting sentences using regular expressions to remove repeated sentences, short sentences with a length of less than 15 characters, and garbled sentences containing formulas;

[0022] Step S202: Calling a general large language model to generate multiple rounds of question-answer pairs consisting of "background knowledge-question-answer" based on the single sentence or context-related sentences cleaned in step S201; then converting the question-answer pairs into Chat-ML format to clarify the correspondence between context, user input, and model response, and finally storing them in a JSON file;

[0023] Step S203: Using "user input - model response" as input, a general reward model is used to score the responses based on five dimensions: usefulness, correctness, coherence, complexity, and verbosity. Based on the scoring results, the responses are ranked from high to low by total score.

[0024] Preferably, in step S2, the steps of constructing the slope prediction corpus dataset are as follows: Step S221: formulating data input formats adapted to the slope stability artificial intelligence model for different slope types;

[0025] Step S222: Construct a scenario where slope parameter information is missing. Based on the data input formats corresponding to different slope types, design the dynamic response logic of the slope domain big language model to the missing information. That is, when there are missing slope parameters in the slope assessment request input by the user, the slope domain big language model can automatically identify the missing slope parameters and generate targeted follow-up content to obtain complete slope information.

[0026] Preferably, in step S2, the steps of performing human supervision and data optimization on the constructed slope knowledge corpus dataset and slope prediction corpus dataset are:

[0027] Step S231: Domain experts manually review and annotate the scoring ranking results output by the general reward model. Specifically, this includes: correcting the scores of answers that use correct technical terms but are mistakenly judged as low-scoring, and adding content annotations to professional answers.

[0028] Step S232: storing the corrected “high-quality answer-low-quality answer” pair into the constructed preference database;

[0029] Step S233: Filter out the question-answer pairs with the highest scores from the revised slope knowledge corpus dataset and the slope prediction corpus dataset as the supervised fine-tuning training dataset, and each question-answer pair in the supervised fine-tuning training dataset contains complete technical logic.

[0030] Preferably, in step S3, the steps of constructing the slope stability prediction artificial intelligence model are:

[0031] Step S300: Encode the data of the slope stability cases constructed in the slope training database into a training form of column vectors; each column vector represents the information of a slope; fill the column vectors with the toe angle, layer height, unit weight, cohesion, and internal friction angle of each slope layer from top to bottom according to the number of layers of the slope; and so on until the information of all slope layers is filled in; then fill in the layer number, cohesion, internal friction angle, x and y coordinates of the centroid of the structure surface, the inclination angle of the structure surface, and the length of the structure surface;

[0032] Step S310: performing a dimension alignment operation on the column vectors of the slope stability cases, specifically: using the column vector with the largest size as a reference, performing zero padding on the column vectors smaller than the largest size, so that the lengths of the column vectors of all the slope stability cases are consistent;

[0033] Step S320: Bind the column vector of each slope case to its corresponding stability coefficient to form an input-output pair, train the support vector machine, random forest, and post-feedback neural network respectively, and use evaluation indicators to evaluate the prediction results of the support vector machine, random forest, and post-feedback neural network. Select the model with the best prediction performance as the ensemble learning sub-model;

[0034] Step S330: For the three integration strategies of Stacking, Boosting, and Bagging, based on the integrated learning sub-model and combined with the Bayesian optimization algorithm, search for the optimal number of sub-models and related training parameters. By comparing the optimization results of different integration strategies, select the integration model with the best comprehensive performance as the final slope stability prediction artificial intelligence model.

[0035] Preferably, in step S4, the steps of constructing the large language model in the slope field are:

[0036] Step S410: Training a general large language model based on the large language model training database, wherein the training strategies include supervised fine-tuning training and reinforcement learning training based on human feedback, wherein the supervised fine-tuning training is based on the general domain pre-trained large language model and is trained using chat-ML style data; the reinforcement learning based on human feedback includes two strategies: direct preference optimization and nearest neighbor preference optimization; both strategies are based on the general domain pre-trained large language model;

[0037] Step S420: Perform engineering scenario performance evaluation on the trained large language model to select the optimal large language model adapted to the slope field, specifically: score the slope knowledge corpus using a general reward model that has been fine-tuned with the slope knowledge corpus preference dataset, and score the large language models trained by the three training strategies of supervised fine-tuning, direct preference optimization, and neighbor preference optimization, and sort them by score; use the ROUGE-L indicator to compare and evaluate the model responses of the large language models trained by different training strategies with the manually specified true labels for the slope prediction corpus, and sort them by ROUGE-L evaluation scores; combine the comprehensive scores of the slope knowledge corpus and the slope prediction corpus to select the large language model with the best score as the trained large language model for the slope field.

[0038] Preferably, in step S5, extracting the user's uploaded data using the trained slope domain language model includes calling the weights of the trained slope domain language model through the Python transformers.pipeline library, specifically:

[0039] Use the transformers.pipeline library to directly call the large language model in the slope domain to the folder path where the model weights are located; after the call, obtain the generated public URL link, access the public URL link to enter the visualization interface, enter data in the visualization interface, and view the model response.

[0040] Preferably, in step S5, obtaining the slope stability coefficient using the trained slope stability prediction artificial intelligence model includes the following steps:

[0041] The Python algorithm automatically detects whether a new TXT file exists during the call of the large language model; if a new TXT file exists, the column vector of slope parameters contained in the TXT file is read and the column vector is input into the trained slope stability prediction artificial intelligence model, which then outputs the corresponding stability coefficient.

[0042] The stability coefficient is automatically read through the Python algorithm, and according to the preset "Technical Code for Building Slope Engineering" (GB50330-2013) standard, the slope field large language model returns the evaluation results including the slope stability coefficient value, the slope stability status, and the measures that may be taken to address the stability status.

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

[0044] 1. The present invention's slope stability assessment method based on a large language model and intelligent prediction model uses a large language model in the slope domain to automatically analyze user-submitted slope data, invoke an artificial intelligence model for slope stability prediction, and provide feedback on the assessment results through a natural language interactive interface. This automated process replaces traditional manual interpretation, significantly lowering the professional threshold and improving assessment efficiency. Furthermore, the high-precision prediction capabilities of the present invention's slope stability assessment method based on a large language model and intelligent prediction model provide an innovative technical path for intelligent monitoring and management of infrastructure.

[0045] 2. This invention proposes a dual-function optimization architecture for the vertical field of geotechnical engineering to overcome the application bottleneck of artificial intelligence models for slope stability prediction and assessment in professional scenarios; by constructing a dual-channel processing mechanism of "parameter structured extraction-domain knowledge dialogue", it creatively realizes the full-link intelligence from unstructured data collection to stability assessment.

[0046] 3. This invention enables users to complete slope stability assessments for soil, rock, and mixed rock and soil types through dialogue. Furthermore, the proposed large language model incorporates in-depth and reliable mechanical knowledge and relevant artificial intelligence algorithms to understand the conversational content, conduct dialogues on slope expertise, and provide appropriate responses at each step, significantly improving the efficiency of slope stability assessments. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a flow chart of the slope stability assessment method based on the large language model and intelligent prediction model of the present invention;

[0048] Figure 2 Flowchart for constructing slope training database;

[0049] Figure 3 is a schematic diagram of the structural surface;

[0050] Figure 4 Flowchart for building large language model training data;

[0051] Figure 5 Schematic diagram of training and evaluation of large language models in the slope domain;

[0052] Figure 6 This is a schematic diagram of the dialogue interface of the large language model in the slope domain;

[0053] Figure 7 Flowchart of the collaboration mechanism between the big language model in the slope field and the artificial intelligence model for slope stability prediction. DETAILED DESCRIPTION

[0054] The present invention will be described in further detail below with reference to the embodiments and drawings, but the embodiments of the present invention are not limited thereto.

[0055] See also Figure 1-Figure 7 The slope stability assessment method based on the large language model and the intelligent prediction model of the present invention comprises the following steps:

[0056] Step S1: Establish a slope training database based on typical slope stability cases;

[0057] Based on differences in geological causes and structural morphology, typical slope stability case types mainly include:

[0058] Homogeneous sedimentary slope: The main body is loose deposits, showing a single-layer or multi-layer (≥2 layers) layered sedimentary structure. Each sedimentary unit should have obvious stratified geological characteristics;

[0059] Structural surface controlled slope: Two or more dominant structural surface combination patterns are developed, where the dominant structural surface attitude combination must meet the geometric conditions of plane sliding or spatial wedge instability;

[0060] Composite superimposed slope: It has the characteristics of the first two types of geological structures and there is obvious rock-soil interface differentiation. It is necessary to meet the requirement that the rock contact surface and the original terrain line are in a non-coordinated intersection relationship.

[0061] In this embodiment, the steps for constructing the slope training database are:

[0062] Step S100: Based on existing research results, typical slope stability cases are collected and slope parameters of each slope stability case are extracted;

[0063] Step S110: Based on the collected slope parameters, a single slope stability case is used as a basic sample, and slope parameters are randomly extracted from multiple slope stability cases for combination to construct multiple different layered slopes, wherein each slope layer has different slope angles and soil parameters; its purpose is to retain all relevant information by avoiding simplifying the shape of the multi-layer rock and soil slope and the parameters of the rock and soil layer (i.e., slope layer), so as to better reflect the actual working conditions of the slope, and also enable the subsequent slope stability assessment artificial intelligence model to learn / predict more accurately the stability status of the slope to be assessed.

[0064] Step S120: Based on existing actual cases and research on slope stability analysis, determine the probability distribution function of the slope structural surface parameters and its corresponding mean and variance, and use the probability density function to generate a random structural surface for the layered slope;

[0065] In this embodiment, a probability and mathematical statistics method is used to generate structural surfaces at each slope level. Specifically, a specific sampling technique is first used to collect multiple sets of structural surfaces from different locations on the slope. Based on the mean and variance, a probability distribution function of the structural surfaces in the slope can be determined, and these structural surfaces are generated accordingly. The probability distribution function can be in the form of a lognormal distribution, a normal distribution, a uniform distribution, and a negative exponential distribution.

[0066] It should be noted that the generated slope with structural surfaces is different from the actual structural surfaces found on the actual rock slope; the structural surfaces obtained by sampling the probability density function are just random samples with an appropriate distribution pattern, and the physical and mechanical parameters obtained from the distribution pattern are also random samples. However, according to the principle of Monte Carlo method, if this sampling is repeated, it will provide the average value of the physical and mechanical parameters obtained. These values ​​represent the corresponding values ​​of the structural surfaces in the actual rock slope; after multiple operations, such as Figure 3 As shown in the figure, each generated structural bread will contain different parameters such as cohesion, internal friction angle, x and y coordinates of the center of mass of the structural surface, inclination angle θ of the structural surface, length l of the structural surface, and the center of mass distance between the structural surfaces; α refers to the inclination of the structural surface.

[0067] Step S130: Diversify the combinations of the constructed layered slopes and the generated structural surfaces to construct multiple different slope cases;

[0068] In this embodiment, according to the number of layers of the slope, the structural surface can be set in the rock and soil layer at a lower position; while the topmost or upper soil layer may not be set with a structural surface to simulate a rock and soil mixed slope; in addition, a structural surface may not be set for any rock and soil layer to simulate a soil stratified slope.

[0069] Step S140: using the extended finite element method combined with the strength reduction method to calculate the stability coefficient of each slope case, forming a corresponding relationship pair of "slope case-stability coefficient", thereby obtaining the slope training database.

[0070] The above-mentioned slope parameters include slope height, slope angle, unit weight of slope rock and soil layer, internal friction angle of slope rock and soil layer, and cohesion of slope rock and soil layer. These slope stability cases can be found in relevant literature libraries. After collecting the corresponding slope parameters, they can be stored in Excel tables or Txt files. It should be noted that the slopes in these collected slope stability cases do not have structural surface data.

[0071] Step S2: Establish a large language model training database based on existing research literature on intelligent slope stability assessment, specifically:

[0072] A slope knowledge corpus dataset and a slope prediction corpus dataset are constructed, and human supervision and data optimization are performed on the constructed slope knowledge corpus dataset and slope prediction corpus dataset. The slope knowledge corpus dataset is used for dialogue on topics such as slope stability and intelligent assessment methods, and for knowledge popularization and understanding. The slope prediction corpus dataset is used for data extraction, extracting material parameters and geometric shape parameters that can be used to assess slope stability.

[0073] The steps of constructing the slope knowledge corpus dataset are as follows:

[0074] Step S201: document data collection and preprocessing;

[0075] Literature related to "slope stability assessment," "artificial intelligence slope analysis," and "geotechnical parameter inversion" was retrieved through academic platforms (such as HowNet, VIP, and Wanfang) using keywords such as "slope stability," "assessment," "artificial intelligence," and "slope stability coefficient." Full-text PDFs were downloaded and a library of original documents was established. The Python library (PyPDF2) was used to extract PDF text page by page, filtering out non-technical content (such as references, author information), formulas, and diagram descriptions. The main text paragraphs were retained, and sentences were segmented using regular expressions to remove duplicate sentences, short sentences less than 15 characters in length, and garbled sentences containing formulas.

[0076] Step S202: generating and enhancing training samples;

[0077] Call a general large language model (such as Llama 3Chinese Chat) to generate multiple rounds of question-answer pairs containing "background knowledge-question-answer" based on the processed single sentences or context-related sentences; for example:

[0078] User input: Assume that there is a description of slope stability knowledge. Please construct a scenario question and answer based on the context, question, and answer. The context should be the background knowledge of the question and answer. The original statement is: With the development of computer technology and numerical analysis methods, it has become possible to use numerical simulation methods for slope stability analysis.

[0079] Model Response: Context: With the development of computer technology and numerical analysis methods, engineers and researchers can use these new technologies to predict and evaluate slope stability, thereby making reasonable design decisions before construction.

[0080] Question: In modern engineering practice, what are the main aspects of the application of numerical simulation methods in slope stability prediction?

[0081] Answer: The application of numerical simulation methods in slope stability prediction is mainly reflected in the following aspects: 1. Geological condition analysis: Numerical simulation can be used to simulate the slope behavior under different geological conditions, including the shear strength, viscosity and other physical properties of different rocks; 2. Design optimization: Simulation results can guide engineering design, such as determining the optimal location and type of support structure to improve the overall stability of the slope.

[0082] The question-answer pairs are then converted into Chat-ML (Chat Markup Language) format, clarifying the correspondence between the context (background knowledge or system prompts for the question and answer), user input (question), and model response (answer). The data is then stored in a JSON file. It is important to note that multiple different answers should be generated for each sentence.

[0083] Step S203: reward model scoring and ranking;

[0084] Taking "user input - model response" as input, a general reward model is used to score the answers based on five dimensions: usefulness, correctness, coherence, complexity, and length. Based on the scoring results, the answers are ranked from high to low by total score.

[0085] In this embodiment, Armo-RM (Absolute Rating Multi-Objective Reward Model) is used to evaluate the performance of the large language model on the slope dialogue corpus.

[0086] The steps of constructing the slope prediction corpus dataset are as follows:

[0087] Step S221: User asks about normal scene design;

[0088] For different slope types (such as soil slopes, rock slopes, and rock-soil mixed slopes), data input formats adapted to the slope stability artificial intelligence model are formulated respectively, among which,

[0089] For soil slopes, the data input format includes slope height, slope angle, cohesion, internal friction angle and unit weight;

[0090] For rock slopes, the data input format adds structural surface parameters based on the soil slope data, and structural surface parameters need to be assigned to each slope layer;

[0091] For rock-soil mixed slopes, the data input format is the same as that for rock slopes, but not every slope layer needs to provide structural surface parameters. Slope layers without structural surface parameters are considered soil layers.

[0092] Step S222: User asks about abnormal scenario design;

[0093] A scenario with missing slope parameter information was constructed. Based on the data input formats corresponding to different slope types, the slope domain big language model was designed to dynamically respond to missing information. That is, when a user inputs a slope assessment request with missing slope parameters, the slope domain big language model can automatically identify the missing slope parameters and generate targeted follow-up questions to obtain complete slope information. For example:

[0094] User Input: Please evaluate the stability of a slope with a cohesion of 30 kPa and an internal friction angle of 15°.

[0095] The model should reply: Please add slope height, slope angle, and slope geotechnical layer unit weight parameters to perform the evaluation.

[0096] In addition to the above examples, all scenarios with missing data should be considered in the design and construction of user question exception scenarios; if the information provided by the user is complete, the keyword extraction function is performed.

[0097] The steps for human supervision and data optimization of the constructed slope knowledge corpus dataset and slope prediction corpus dataset are as follows:

[0098] Step S231: Domain experts (geotechnical engineers) manually review and annotate the scoring ranking results output by the general reward model. Specifically, this includes: correcting the scores of answers that use correct technical terms but are mistakenly judged as low-scoring, and adding "highly authoritative" content annotations to highly professional answers;

[0099] Step S232: The corrected “high-quality answer-low-quality answer” pair is stored in the constructed preference database for model fine-tuning in the reinforcement learning phase (such as nearest neighbor preference optimization and direct preference optimization algorithm).

[0100] An example format of a preference dataset is:

[0101] User Input: Why is slope stability assessment considered important?

[0102] Model response, high-quality answer: Slope stability analysis is important because it can predict potential landslides and other types of stability issues on a slope. By understanding slope stability, potential safety risks can be avoided during construction or design, ensuring the safety of people and property.

[0103] Model reply, poor quality answer: Slope stability analysis is crucial to ensure the safety and stability of engineering structures.

[0104] Step S233: Filter out the question-answer pairs with the highest scores from the revised slope knowledge corpus dataset and the slope prediction corpus dataset as the supervised fine-tuning training dataset, and each question-answer pair in the supervised fine-tuning training dataset contains complete technical logic (such as parameter calculation steps and engineering case references).

[0105] Step S3: Based on the slope training database, an artificial intelligence model for slope stability prediction is constructed, specifically:

[0106] Step S300: Encode the data in the slope stability case constructed in the slope training database into a training form of column vectors; each column vector represents the information of a slope; fill the column vector with the toe angle, layer height, unit weight, cohesion, and internal friction angle of each slope layer from top to bottom according to the number of layers of the slope; and so on until the information of all slope layers is filled in; subsequently fill in the layer number, cohesion, internal friction angle, x and y coordinates of the center of mass of the structure surface, the inclination angle of the structure surface, and the length of the structure surface, wherein the layer number format is numbered sequentially from one to the total number of rock and soil layers of the slope from top to bottom;

[0107] Step S310: Since the number of layers and structural surfaces of the slope is randomly generated, the dimensions (or sizes) of the column vectors will be inconsistent. To facilitate the subsequent training of the artificial intelligence model for slope stability prediction, it is necessary to perform a dimension alignment operation on the column vectors of the slope stability cases. Specifically, using the largest column vector as a reference, zero-padded the column vectors smaller than the largest size to ensure that the lengths of the column vectors of all slope stability cases are completely consistent.

[0108] The training of the slope stability prediction artificial intelligence model mainly includes selecting a suitable ensemble learning strategy and selecting a suitable ensemble learning sub-model, specifically:

[0109] Step S320: Bind the column vector of each slope case to its corresponding stability coefficient (FOS) to form an input-output pair, train the support vector machine, random forest, and feedback neural network respectively, and use evaluation indicators to evaluate the prediction results of the support vector machine, random forest, and feedback neural network, and select the model with the best prediction performance as the ensemble learning sub-model;

[0110] Step S330: For different ensemble learning strategies, namely Stacking, Boosting, and Baggin, based on the ensemble learning sub-models, combined with the Bayesian optimization algorithm, search for the optimal number of sub-models and related training parameters, such as learning rate, number of iterations, minimum batch size, etc.;

[0111] In this embodiment, the Bagging strategy focuses on optimizing the number of sub-models, the Boosting strategy focuses on optimizing the learning rate and number of iterations, and the Stacking strategy focuses on optimizing the sub-model combination and meta-model parameters. At the same time, common parameters (such as the minimum batch size) are optimized uniformly. By comparing the optimization results of different integration strategies, the integration model with the best comprehensive performance is selected as the final slope stability prediction artificial intelligence model.

[0112] Step S4: Based on the large language model training database, build a large language model for the slope field. The process is as follows: Figure 5 As shown ( Figure 5 CPBao in this article refers to a large language model trained using a professional domain dataset, which mainly includes:

[0113] Step S410: Training the general large language model based on the large language model training database, the training strategy includes supervised fine-tuning training and reinforcement learning training based on human feedback, wherein,

[0114] Supervised fine-tuning is the most common training strategy for constructing large language models in a specific domain from general-domain models. The pre-trained model uses the general-domain Llama 3 Chinese Chat dataset, using Chat-ML-style data. The training parameters used are: 30 iterations, 2 minimum batch sizes, 2e-4 initial learning rates, and Cosine learning rate descent.

[0115] The human feedback-based reinforcement learning includes two strategies: direct preference optimization and neighbor preference optimization; both strategies are based on a general domain pre-trained large language model;

[0116] Direct preference optimization uses the general domain Llama 3Chinese Chat as a pre-trained model and directly uses the preference dataset for training. The training parameters used are: number of iterations: 30 rounds; minimum batch size: 1; initial learning rate: 1e-5; learning rate descent strategy: Cosine;

[0117] Neighbor preference optimization also requires pre-training of a reward model; the reward model uses the general domain Llama3Chinese Chat as a pre-trained model and is trained using the preference dataset; the training parameters used are: number of iterations: 10 rounds; minimum batch size: 1; initial learning rate: 1e-4; learning rate decline strategy: Cosine; after the reward model training is completed, the neighbor preference optimization strategy training is started; neighbor preference optimization uses the general domain Llama 3ChineseChat as a pre-trained model, uses chat-ML style data for training, and calls the trained reward model for assistance; the training parameters used are: number of iterations: 5 rounds; minimum batch size: 1; initial learning rate: 5e-5; learning rate decline strategy: Cosine.

[0118] Step S420: Perform engineering scenario performance evaluation on the trained large language model to select the optimal large language model suitable for the slope field, specifically:

[0119] The slope knowledge corpus is scored using a general reward model that has been fine-tuned using the slope knowledge corpus preference dataset, and the large language models trained using the three training strategies of supervised fine-tuning, direct preference optimization, and nearest neighbor preference optimization are scored and sorted by score. The slope prediction corpus is evaluated using the ROUGE-L (Recall-Oriented Assisted Evaluation of the Longest Common Subsequence) indicator to compare and evaluate the model responses of the large language models trained using different training strategies with the manually specified true labels, and sorted by the ROUGE-L evaluation score. The comprehensive scores of the slope knowledge corpus and the slope prediction corpus are combined to select the large language model with the best score as the trained large language model for the slope domain.

[0120] Step S5: The slope parameters of the user's uploaded data are extracted through the trained slope domain big language model, and it is judged whether the extracted slope parameters meet the input requirements of the slope stability prediction artificial intelligence model; if the slope parameters in the user's uploaded data do not meet the input requirements of the slope stability prediction artificial intelligence model, the slope domain big language model will feedback the slope parameters that need to be supplemented to the user; after the user supplements, the slope domain big language model re-extracts and judges again until it monitors that the user has uploaded all slope parameters that meet the input requirements of the slope stability prediction artificial intelligence model; the slope domain big language model sends the complete slope parameters to the trained slope stability prediction artificial intelligence model, and the slope stability prediction artificial intelligence model outputs the stability coefficient of the slope.

[0121] Among them, using the trained slope domain language model to extract the user's uploaded data mainly includes calling the weights of the trained slope domain language model through Python's transformers.pipeline library, specifically:

[0122] Use the transformers.pipeline library to directly call the large language model in the folder path where the model weights are located; after calling, obtain the generated public URL link, such as Figure 6 As shown, access the public URL link to enter the visualization interface. In the "Enter your question" box of the pop-up visualization interface, enter the relevant question and click "Submit" to return the model response.

[0123] like Figure 7 As shown in the figure, the slope parameters are extracted from the user's uploaded data using a trained big language model for slopes. The system then determines whether the information is sufficient. Sufficient information includes the slope geometry, including slope height, break angle, structural surface location, and material parameters, such as unit weight, cohesion, internal friction angle, and structural surface material parameters. Furthermore, the big language model for slopes needs to standardize the parameter format, such as unifying the parameter units.

[0124] When the user provides sufficient information, the slope domain language model generates a fixed-format response, such as:

[0125] The slope stability is being evaluated. The input parameters are: slope height 20m, slope angle 45°, cohesion 25kPa, internal friction angle 22°, unit weight 10kN / m 3 ;

[0126] Then, after Python detects the keyword "evaluating", it stores the parameters extracted from the large language model in the slope field in the form of a column vector. The form of the column vector is consistent with the input format of the slope stability prediction artificial intelligence model and is written into the TXT file.

[0127] When the user provides insufficient information, the slope domain language model will automatically infer the parameters that have not been provided based on the slope parameters that have been provided, and make corresponding model responses, prompting the user to supplement the missing slope parameters.

[0128] Obtaining the slope stability coefficient using the trained slope stability prediction AI model involves the following steps:

[0129] The Python algorithm automatically detects whether a new TXT file exists during the call of the large language model; if a new TXT file exists, the column vector of slope parameters contained in the TXT file is read and the column vector is input into the trained slope stability prediction artificial intelligence model, which then outputs the corresponding stability coefficient.

[0130] The stability coefficient is automatically read through the Python algorithm. According to the preset "Technical Code for Building Slope Engineering" (GB50330-2013) standard, the slope field large language model returns an assessment result containing the following content: the slope stability coefficient value, the stability status of the slope, and the possible measures to be taken to address the stability status.

[0131] The present invention is based on parameter fine-tuning technology and artificial intelligence regression prediction method. By training the slope field large language model and the slope stability prediction artificial intelligence model extracted from slope data, a collaborative mechanism of the slope field large language model and the slope stability prediction artificial intelligence model is constructed, thereby simplifying the slope stability assessment process and improving efficiency. Its core advantages are reflected in: integrating the large language model and the slope stability prediction algorithm, fine-tuning the large language model through specific slope knowledge and stability prediction dialogue data, realizing the slope stability interactive interface of human-computer dialogue, and breaking through the current situation where the slope stability intelligent assessment algorithm is difficult to promote.

[0132] Suppose an engineer needs to evaluate a slope with complex geological conditions. Traditionally, engineers must empirically gather slope stability parameters from geological reports and field data to determine the FS using numerical or mechanical analysis methods. This primarily involves building a slope model, performing stability calculations, and extracting and analyzing the results. Once an LLM-driven CNN is implemented, the engineer simply asks questions such as, "What parameters are needed to evaluate slope stability?" and enters the relevant parameters for a query, such as, "I want to evaluate the stability of a slope with a slope angle of 45° and an internal friction angle of 30°." The model then automatically extracts key parameters and provides preliminary stability analysis results, significantly reducing data preparation time and the need for expert knowledge, providing a new perspective for slope safety decision-making.

[0133] The foregoing description is intended only to provide specific embodiments of the present invention, which will enable those skilled in the art to understand and implement the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein, but is intended to be construed in the widest manner consistent with the principles and novel features claimed herein.

Claims

1. A slope stability assessment method based on a large language model and an intelligent prediction model, characterized in that: The following steps are involved: Step S1: Establish a slope training database based on typical slope stability cases; Step S2: establishing a large language model training database based on existing research literature on intelligent slope stability assessment; Step S3: constructing an artificial intelligence model for slope stability prediction based on the slope training database; Step S4: constructing a large language model for the slope domain based on the large language model training database; Step S5: The slope parameters of the user's uploaded data are extracted through the trained slope domain big language model, and it is judged whether the extracted slope parameters meet the input requirements of the slope stability prediction artificial intelligence model; if the slope parameters in the user's uploaded data do not meet the input requirements of the slope stability prediction artificial intelligence model, the slope domain big language model will feedback the slope parameters that need to be supplemented to the user; after the user supplements, the slope domain big language model re-extracts and judges again until it monitors that the user has uploaded all slope parameters that meet the input requirements of the slope stability prediction artificial intelligence model; the slope domain big language model sends the complete slope parameters to the trained slope stability prediction artificial intelligence model, and the slope stability prediction artificial intelligence model outputs the stability coefficient of the slope.

2. The slope stability assessment method based on a large language model and an intelligent prediction model according to claim 1 is characterized in that: In step S1, the steps of constructing the slope training database are: Step S100: Collect typical slope stability cases and extract slope parameters of each slope stability case; the slope parameters include slope height, slope angle, unit weight of slope rock and soil layer, internal friction angle of slope rock and soil layer, and cohesion of slope rock and soil layer; Step S110: Based on the collected slope parameters, taking a single slope stability case as a basic sample, randomly extracting slope parameters from multiple slope stability cases and combining them to construct multiple different layered slopes, wherein each slope layer has different slope angles and soil parameters; Step S120: determining a probability distribution function of slope structural surface parameters and its corresponding mean and variance, and using the probability density function to generate random structural surfaces for the layered slope, wherein the structural surface parameters include cohesion, internal friction angle, centroid coordinates of the structural surface, structural surface inclination, structural surface length, and centroid spacing between different structural surfaces; Step S130: Diversify the combinations of the constructed layered slopes and the generated structural surfaces to construct multiple different slope cases; Step S140: using the extended finite element method combined with the strength reduction method to calculate the stability coefficient of each slope case, forming a corresponding relationship pair of "slope case-stability coefficient", thereby obtaining the slope training database.

3. The slope stability assessment method based on a large language model and an intelligent prediction model according to claim 2 is characterized in that: In step S2, the construction of the large language model training database includes the following steps: Construct a slope knowledge corpus dataset and a slope prediction corpus dataset, and perform human supervision and data optimization on the constructed slope knowledge corpus dataset and slope prediction corpus dataset.

4. The slope stability assessment method based on a large language model and an intelligent prediction model according to claim 3 is characterized in that: In step S2, the steps of constructing the slope knowledge corpus dataset are as follows: Step S201: Retrieve literature related to slope stability assessment, artificial intelligence slope analysis, and geotechnical parameter inversion through an academic platform, download the full-text PDF, and establish an original literature library; use a Python library to extract PDF text page by page, filter out non-technical content, formulas, and chart descriptions, retain the main text paragraphs, and segment sentences using regular expressions to remove repeated sentences, short sentences less than 15 characters in length, and garbled sentences containing formulas; Step S202: The general large language model is invoked to generate multiple rounds of question-answer pairs consisting of "background knowledge-question-answer" based on the single sentences or context-related sentences cleaned in step S201. The question-answer pairs are then converted to the Chat-ML format to clarify the correspondence between context, user input, and model response, and finally stored in a JSON file. Step S203: Using "user input - model response" as input, a general reward model is used to score the responses based on five dimensions: usefulness, correctness, coherence, complexity, and verbosity. Based on the scoring results, the responses are ranked from high to low by total score.

5. The slope stability assessment method based on a large language model and an intelligent prediction model according to claim 4 is characterized in that: In step S2, the steps of constructing the slope prediction corpus dataset are: Step S221: for different slope types, respectively formulate data input formats that are compatible with the slope stability artificial intelligence model; Step S222: Construct a scenario where slope parameter information is missing. Based on the data input formats corresponding to different slope types, design the dynamic response logic of the slope domain big language model to the missing information. That is, when there are missing slope parameters in the slope assessment request input by the user, the slope domain big language model can automatically identify the missing slope parameters and generate targeted follow-up content to obtain complete slope information.

6. The slope stability assessment method based on a large language model and an intelligent prediction model according to claim 5 is characterized in that: In step S2, the steps of human supervision and data optimization for the constructed slope knowledge corpus dataset and slope prediction corpus dataset are as follows: Step S231: Domain experts manually review and annotate the scoring ranking results output by the general reward model. Specifically, this includes: correcting the scores of answers that use correct technical terms but are mistakenly judged as low-scoring, and adding content annotations to professional answers. Step S232: storing the corrected "high-quality answer-low-quality answer" pair into the constructed preference database; Step S233: Filter out the question-answer pairs with the highest scores from the revised slope knowledge corpus dataset and the slope prediction corpus dataset as the supervised fine-tuning training dataset, and each question-answer pair in the supervised fine-tuning training dataset contains complete technical logic.

7. The slope stability assessment method based on a large language model and an intelligent prediction model according to claim 6 is characterized in that: In step S3, the steps for constructing the slope stability prediction artificial intelligence model are as follows: Step S300: Encode the data of the slope stability cases constructed in the slope training database into a training form of column vectors; each column vector represents the information of a slope; fill the column vectors with the toe angle, layer height, unit weight, cohesion, and internal friction angle of each slope layer from top to bottom according to the number of layers of the slope; and so on until the information of all slope layers is filled in; then fill in the layer number, cohesion, internal friction angle, x and y coordinates of the centroid of the structure surface, the inclination angle of the structure surface, and the length of the structure surface; Step S310: performing a dimension alignment operation on the column vectors of the slope stability cases, specifically: using the column vector with the largest size as a reference, performing zero padding on the column vectors smaller than the largest size, so that the lengths of the column vectors of all the slope stability cases are consistent; Step S320: Bind the column vector of each slope case to its corresponding stability coefficient to form an input-output pair, train the support vector machine, random forest, and post-feedback neural network respectively, and use evaluation indicators to evaluate the prediction results of the support vector machine, random forest, and post-feedback neural network. Select the model with the best prediction performance as the ensemble learning sub-model; Step S330: For the three integration strategies of Stacking, Boosting, and Bagging, based on the integrated learning sub-model and combined with the Bayesian optimization algorithm, search for the optimal number of sub-models and related training parameters. By comparing the optimization results of different integration strategies, select the integration model with the best comprehensive performance as the final slope stability prediction artificial intelligence model.

8. The slope stability assessment method based on a large language model and an intelligent prediction model according to claim 7 is characterized in that: In step S4, the steps for constructing the large language model in the slope domain are as follows: Step S410: Training a general large language model based on the large language model training database, wherein the training strategies include supervised fine-tuning training and reinforcement learning training based on human feedback, wherein the supervised fine-tuning training is based on the general domain pre-trained large language model and is trained using chat-ML style data; the reinforcement learning based on human feedback includes two strategies: direct preference optimization and nearest neighbor preference optimization; both strategies are based on the general domain pre-trained large language model; Step S420: Perform engineering scenario performance evaluation on the trained large language model to select the optimal large language model adapted to the slope field, specifically: score the slope knowledge corpus using a general reward model that has been fine-tuned with the slope knowledge corpus preference dataset, and score the large language models trained by the three training strategies of supervised fine-tuning, direct preference optimization, and neighbor preference optimization, and sort them by score; use the ROUGE-L indicator to compare and evaluate the model responses of the large language models trained by different training strategies with the manually specified true labels for the slope prediction corpus, and sort them by ROUGE-L evaluation scores; combine the comprehensive scores of the slope knowledge corpus and the slope prediction corpus to select the large language model with the best score as the trained large language model for the slope field.

9. The slope stability assessment method based on a large language model and an intelligent prediction model according to claim 8 is characterized in that: In step S5, extracting the user's uploaded data using the trained slope domain language model includes calling the weights of the trained slope domain language model through the Python transformers.pipeline library, specifically: Use the transformers.pipeline library to directly call the large language model in the slope domain to the folder path where the model weights are located; after the call, obtain the generated public URL link, access the public URL link to enter the visualization interface, enter data in the visualization interface, and view the model response.

10. The slope stability assessment method based on a large language model and an intelligent prediction model according to claim 9 is characterized in that: In step S5, obtaining the slope stability coefficient using the trained slope stability prediction artificial intelligence model includes the following steps: The Python algorithm automatically detects whether a new TXT file exists during the call of the large language model; if a new TXT file exists, the column vector of slope parameters contained in the TXT file is read and the column vector is input into the trained slope stability prediction artificial intelligence model, which then outputs the corresponding stability coefficient. The stability coefficient is automatically read through the Python algorithm, and according to the preset "Technical Code for Building Slope Engineering" (GB50330-2013) standard, the slope field large language model returns the evaluation results including the slope stability coefficient value, the slope stability status, and the measures that may be taken to address the stability status.

Images