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

Abstract

The application discloses a kind of based on big language model and intelligent prediction model's side slope stability evaluation method, specifically is: establish side slope training database and big language model training database;Side slope stability prediction artificial intelligence model and side slope field big language model are constructed;Through side slope field big language model, the side slope parameter extraction of user's uploaded data is carried out, and whether the side slope parameter extracted meets the input requirement of side slope stability prediction artificial intelligence model is judged;If the side slope parameter in the data uploaded by user is not complete, side slope field big language model then feedback needs to be supplemented to user side slope parameter;After user supplements, side slope field big language model reextracts and judges again, until it is monitored that user has uploaded all side slope parameters;The complete side slope parameter is sent into the side slope stability prediction artificial intelligence model trained by the side slope field big language model, and the side slope stability prediction artificial intelligence model outputs the stability coefficient of side slope.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent geotechnical engineering evaluation, and particularly relates to a slope stability evaluation method based on a large language model and an intelligent prediction model. BACKGROUND

[0002] The slope stability evaluation method based on artificial intelligence improves the evaluation method in the engineering field. This method enables people to accurately obtain the stability state of the slope at a lower cost, broadens the access to professional analysis results, and reduces the skill and time cost required for evaluation. The slope stability evaluation driven by artificial intelligence mainly predicts the stability of the slope by analyzing the geometric characteristics and material properties of the slope. The core goal is to ensure that these models can provide accurate prediction results in comparison with the mechanical analysis method.

[0003] However, the existing method for evaluating the slope stability based on artificial intelligence often ignores the challenges that users may encounter in actual operation: users may not be familiar with the operation of artificial intelligence models, the cumbersome problem decomposition and reporting process, and the high requirements for the reporting personnel, which may lead to a decline in analysis quality and affect the effect of the artificial intelligence method in the evaluation of slope stability. Users still face many obstacles in actual application. For example, the geometric size coordinates and material parameters of the slope need to be arranged in advance to be input into the slope stability prediction model. The arrangement position of these parameters usually has a certain template, and all the parameters must be placed in the correct position to accurately reflect the actual working condition of the slope to be predicted. Before these operations are realized, the user must be trained to have relevant skills.

[0004] The advent of ChatGPT has changed people's understanding of the potential of artificial intelligence in industrial applications. Large language models follow human instructions in conversations, exhibiting performances comparable to those of normal or even expert humans. However, the dialogue scenarios in the field of slope engineering are often extremely complex, beyond the processing capacity of general domain large language models. Therefore, relevant vertical domain datasets should be used for training to obtain large language models in the field of slope engineering, while also needing to be consistent with the dialogue distribution of real-world slope knowledge. The development of large language models in vertical fields such as medicine, law, finance, and education provides inspiration for the content of the invention. These studies fine-tune instructions through professional knowledge question and answer datasets, improving the model's understanding and execution capabilities for specific content. These advances indicate that it is possible to develop a specialized dialogue system in the field of slope engineering. However, large language models for slope stability assessment have not been developed, and there has been insufficient attention to technical development of scarce datasets, development of evaluation indicators, data extraction techniques, and the impact of training strategies, sample size, and iteration number on model performance. These factors are crucial for improving the domain adaptation, accuracy, and generalization capabilities of large language models, especially in specific professional fields where data is scarce. SUMMARY

[0005] The present application provides a slope stability evaluation method based on a large language model and an intelligent prediction model, which has higher evaluation accuracy and efficiency, and reduces the need for professional knowledge, making slope stability evaluation more convenient and reliable.

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

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

[0008] Step S1: Establish a slope training database according to typical slope stability cases;

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

[0010] Step S3: Based on the slope training database, construct a slope stability prediction artificial intelligence model;

[0011] Step S4: Based on the large language model training database, construct a slope domain large language model;

[0012] Step S5: The trained slope domain big data language model extracts slope parameters from the user's uploaded data and determines whether the extracted slope parameters meet the input requirements of the slope stability prediction AI model. If the slope parameters in the user's uploaded data do not meet the input requirements of the slope stability prediction AI model, the slope domain big data language model provides feedback to the user regarding the slope parameters that need to be supplemented. After the user supplements the parameters, the slope domain big data language model re-extracts and re-determines the parameters until it detects that the user has uploaded all the slope parameters that meet the input requirements of the slope stability prediction AI model. The slope domain big data language model then sends the complete slope parameters into the trained slope stability prediction AI model, which outputs the slope stability coefficient.

[0013] Preferably, in step S1, the construction step of the slope training database is as follows:

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

[0015] Step S110: Based on the collected slope parameters, using a single slope stability case as the base sample, randomly select slope parameters from multiple slope stability cases and combine them to construct multiple different layered slopes, where each slope layer has different slope angles and soil parameters.

[0016] Step S120: 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 random structural surfaces for the layered slope. The parameters of the structural surface include cohesion, internal friction angle, centroid coordinates of the structural surface, inclination angle of the structural surface, length of the structural surface, and centroid spacing between different structural surfaces.

[0017] Step S130: Combine the constructed layered slope and the generated structural surface in various ways to create multiple different slope cases;

[0018] Step S140: Using the extended finite element method combined with the strength reduction method, calculate the stability coefficient of each slope case to form a "slope case - stability coefficient" correspondence pair, 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] We constructed a slope knowledge corpus dataset and a slope prediction corpus dataset, and then performed human supervision and data optimization on the constructed slope knowledge corpus dataset and slope prediction corpus dataset.

[0021] Preferably, in step S2, the step of constructing the slope knowledge corpus dataset is: step S201: retrieve the literature related to slope stability evaluation, artificial intelligence slope analysis and geotechnical parameter inversion through academic platforms, establish an original literature library after downloading the PDF full text, use Python library to extract PDF text page by page, filter non-technical content, formulas and chart descriptions, retain text paragraphs, and separate sentences through regular expressions, remove duplicate sentences, short sentences with length less than 15 characters and sentences with formula containing garbled code;

[0022] Step S202: Call the general large language model, generate a multi-round question and answer pair containing "background knowledge-question-answer" based on the single sentence or context-related sentence cleaned in step S201; then convert the question and answer pair into Chat-ML format to clearly define the corresponding relationship between context, user input and model reply, and finally store it in a JSON file;

[0023] Step S203: Take "user input-model reply" as input, use a general reward model to score from five dimensions of usefulness, correctness, coherence, complexity and length of the answer, and according to the scoring results, sort them from high to low according to the total score.

[0024] Preferably, in step S2, the step of constructing the slope prediction corpus dataset is: step S221: For different types of slopes, respectively formulate data input formats suitable for slope stability artificial intelligence models;

[0025] Step S222: Construct a scenario of missing slope parameter information, based on the data input format corresponding to different slope types, design the dynamic response logic of the slope domain large language model to the missing information, that is, when there is missing slope parameter in the user's input slope evaluation request, the slope domain large language model can automatically identify the missing slope parameter and generate targeted follow-up questions to obtain complete slope information.

[0026] Preferably, in step S2, the step of human supervision and data optimization for the constructed slope knowledge corpus dataset and slope prediction corpus dataset is:

[0027] Step S231: The field experts manually review and label the scoring and sorting results output by the general reward model, including: correcting the scores of answers that use technical terms correctly but are misjudged as low scores, and supplementing the content of professional answers;

[0028] Step S232: Store the corrected "high-quality answer-poor-quality answer" pair in the constructed preference database;

[0029] Step S233: selecting the highest-scored question-answer pair 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 construction step of the slope stability prediction artificial intelligence model is:

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

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

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

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

[0035] Preferably, in step S4, the construction step of the slope domain large language model is:

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

[0037] Step S420: performing engineering scene performance evaluation on the trained large language model to select an optimal large language model adapted to the slope field, specifically: scoring the slope knowledge corpus by using the general reward model fine-tuned by the slope knowledge corpus preference data set, and scoring the large language models trained by the three training strategies of supervised fine-tuning, direct preference optimization and near neighbor preference optimization, and ranking them according to the scores; comparing the model responses of the large language models trained by different training strategies with the real labels specified by humans by using the ROUGE-L index to evaluate the slope prediction corpus, and ranking them according to the ROUGE-L evaluation scores; combining the comprehensive scores of the slope knowledge corpus and the slope prediction corpus, and selecting the large language model with the optimal score as the trained slope field large language model.

[0038] Preferably, in step S5, the trained slope field large language model is used to extract the user's uploaded data, including calling the weight of the trained slope field large language model through the transformers.pipeline library of Python, specifically:

[0039] The transformers.pipeline library is used to directly call the slope field large language model for the folder path of the model weight; after the calling, a generated public URL link is obtained, the public URL link is accessed to enter a visualization interface, and data is input and model responses are viewed in the visualization interface.

[0040] Preferably, in step S5, the trained slope stability prediction artificial intelligence model is used to obtain the stability coefficient of the slope, including the following steps:

[0041] The Python algorithm is used to automatically detect whether there is a newly added Txt file in the process of calling the large language model; if there is a newly added Txt file, a column vector of slope parameters contained in the Txt file is read and input into the trained slope stability prediction artificial intelligence model, and the corresponding stability coefficient is output by the slope stability prediction artificial intelligence model;

[0042] The stable coefficient is automatically read by a Python algorithm, and according to the preset standard of Technical Code of Building Slope Engineering (GB50330-2013), an evaluation result including the stable coefficient value of the slope, the stability state of the slope and the measures that can be taken for the stability state is returned through the large language model in the field of slope.

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

[0044] 1. The slope stability evaluation method based on the large language model and the intelligent prediction model in the present application is based on the large language model in the field of slope, automatically analyzes the slope data submitted by the user, and calls the slope stability prediction artificial intelligence model to feed back the evaluation result through the natural language interaction interface; the traditional manual experience interpretation is replaced by the automatic process, thereby significantly reducing the professional threshold and improving the evaluation efficiency. Meanwhile, the high-precision prediction capability of the slope stability evaluation method based on the large language model and the intelligent prediction model in the present application provides an innovative technical path for the intelligent monitoring and management of infrastructure.

[0045] 2. The present application proposes a dual-function optimization architecture for the vertical field of geotechnical engineering, and overcomes the application bottleneck of the slope stability prediction and evaluation artificial intelligence model in the professional scene; by constructing the "parameter structured extraction - field knowledge dialogue" dual-channel processing mechanism, the full-link intelligentization from unstructured data collection to stability evaluation is creatively realized.

[0046] 3. The present application can enable the user to complete the evaluation of the stability of the soil, rock and rock-soil mixed slope in the form of dialogue. In addition, the large language model proposed in the present application has deep and reliable mechanical knowledge and related artificial intelligence algorithms to understand the dialogue content, carry out the dialogue in the field of slope professional knowledge, and make appropriate responses in each step, thereby greatly improving the efficiency of slope stability evaluation. BRIEF DESCRIPTION OF DRAWINGS

[0047] Figure 1 The flowchart of the slope stability evaluation method based on the large language model and the intelligent prediction model in the present application;

[0048] Figure 2 The flowchart of constructing the slope training database;

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

[0050] Figure 4 The flowchart of constructing the training data of the large language model;

[0051] Figure 5 The training and evaluation schematic diagram of the large language model in the field of slope;

[0052] Figure 6 A schematic diagram of a dialogue interface of a large language model in the field of slope;

[0053] Figure 7 A flowchart of a collaborative mechanism of a large language model in the field of slope and an artificial intelligence model for slope stability prediction. DETAILED DESCRIPTION

[0054] The present application will be further described in conjunction with the embodiments and the accompanying drawings, but the embodiments of the present application are not limited thereto.

[0055] Referring to Figures 1-7 The slope stability evaluation method based on the large language model and the intelligent prediction model of the present application comprises the following steps:

[0056] Step S1: Establish a slope training database according to typical slope stability cases;

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

[0058] Homogeneous sedimentary slope: mainly composed of loose deposits, showing single-layer or multi-layer (≥2 layers) stratified sedimentary structure, and each sedimentary unit should have obvious stratification geological characteristics;

[0059] Structure surface controlled slope: two or more dominant structure surface combination modes are developed, and the dominant structure surface occurrence combination should meet the geometric conditions of plane sliding or spatial wedge instability;

[0060] Composite superimposed slope: it has the geological structure characteristics of the previous two types, and there is obvious rock-soil interface differentiation phenomenon, and it needs to meet the non-coordinated intersection relationship between the lithological contact surface and the original topographic line.

[0061] In this embodiment, the construction step of the slope training database is:

[0062] Step S100: According to the existing research results, collect typical slope stability cases, and extract the slope parameters of each slope stability case;

[0063] Step S110: According to the collected slope parameters, taking a single slope stability case as a basic sample, randomly extract slope parameters from multiple slope stability cases for combination to construct multiple different layered slopes, wherein each slope layer has different slope angles and soil parameters; The purpose is to retain all relevant information by avoiding simplifying the shape and rock-soil layer (i.e. slope layer) parameters of the multi-layer rock-soil slope, which can better reflect the actual working condition of the slope, and also enable the slope stability evaluation artificial intelligence model to learn / predict more accurately to obtain the stability state of the slope to be evaluated.

[0064] Step S120: According to the existing actual cases and the research of slope stability analysis, the probability distribution function of the slope structural plane parameters and the corresponding mean value and variance are determined, and the probability density function is used to generate random structural planes for the layered slope;

[0065] In this embodiment, the probability and mathematical statistics method is used to generate structural planes at each slope level; that is, a specific sampling technique is used to collect multiple sets of structural planes from different positions of the slope; according to the mean value and variance, the probability distribution function of the structural planes in the slope can be determined, and the structural planes are generated in the slope according to the probability distribution function; wherein the form of the probability distribution function can be: lognormal distribution, normal distribution, uniform distribution and negative exponential distribution;

[0066] It should be noted that the generated structural plane of the slope is different from the actual structural plane found on the actual rock slope; the structural plane obtained by sampling from the probability density function is only a random sample with a proper distribution pattern, and the physical and mechanical parameters obtained from the distribution pattern are also random samples. However, according to the principle of the Monte Carlo method, if the sampling is repeated, it will provide the average value of the obtained physical and mechanical parameters. These values represent the corresponding values of the structural planes in the actual rock slope; after multiple operations, such as Figure 3 As shown in the figure, each generated structural plane contains different parameters such as cohesive force, internal friction angle, structural plane centroid x and y coordinates, structural plane inclination θ, structural plane length l and centroid spacing between structural planes; α refers to the structural plane inclination.

[0067] Step S130: The constructed layered slope and the generated structural planes are diversified combined to construct multiple different slope cases;

[0068] In this embodiment, according to the number of layers of the slope, the structural planes can be arranged in the lower position of the rock-soil layer; and the top layer or the upper position of the soil layer can not be arranged with structural planes to simulate the rock-soil mixed slope; in addition, no structural plane can be arranged for any layer of rock-soil layer to simulate the soil layered slope.

[0069] Step S140: The extended finite element method combined with the strength reduction method is used to calculate the stability coefficients of each slope case, and a corresponding relationship pair of "slope case-stability coefficient" is formed to obtain the slope training database.

[0070] The above-mentioned slope parameters include slope height, slope angle, slope rock-soil layer unit weight, slope rock-soil layer internal friction angle and slope rock-soil layer cohesive force; these slope stability cases can be found in the related literature library, and after collecting the corresponding slope parameters, they can be stored in Excel table or Txt file; it should be noted that the slopes in these collected slope stability cases do not have structural plane data.

[0071] Step S2: Establish a large language model training database based on existing research literature on intelligent evaluation of slope stability. 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 slope stability, intelligent evaluation methods, and other content, and for knowledge popularization and understanding functions. The slope prediction corpus dataset is used for data extraction to extract material parameters and geometric shape parameters that can evaluate slope stability.

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

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

[0075] Through academic platforms such as CNKI, VIP, Wanfang, etc., with keywords such as "slope stability", "evaluation", "artificial intelligence", "slope stability coefficient", etc., relevant literature on "slope stability evaluation", "artificial intelligence slope analysis", and "geotechnical parameter inversion" is retrieved. After downloading the PDF full text, an original literature database is established. Using the Python library (PyPDF2), PDF text is extracted page by page, non-technical content (such as references, author information), formulas, and chart descriptions are filtered, text passages are retained, and sentences are divided by regular expressions, removing duplicate sentences, short sentences less than 15 characters, and sentences with formula code;

[0076] Step S202: Training sample generation and enhancement;

[0077] A general large language model (such as Llama 3Chinese Chat) is called, and based on the processed single sentence or context-related sentence, a multi-round question and answer pair containing "background knowledge-question-answer" is generated. For example:

[0078] User input: Suppose there is a description of slope stability knowledge now, please construct a scenario question and answer based on the context, question, and answer, where the context should be the background knowledge of the question and answer. The initial 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 reply: Context: With the development of computer technology and numerical analysis methods, engineers and researchers can use these new technologies to predict and evaluate the stability of slopes, making reasonable design decisions before construction.

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

[0081] Answer: The application of numerical simulation methods in slope stability prediction mainly involves the following aspects: 1. Geological condition analysis: numerical simulation can simulate the behavior of slopes 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 structures to improve the overall stability of the slope.

[0082] The question and answer pair is then converted into Chat-ML (Chat Markup Language) format, clearly defining the context (background knowledge or system prompt words for the question and answer), user input (question), and model response (answer). The corresponding relationship is finally 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] Using the general reward model, the usefulness, correctness, coherence, complexity, and verbosity of the answers are scored. Based on the scoring results, the answers are ranked from high to low according to the total score.

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

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

[0087] Step S221: User question normal scenario design;

[0088] For different types of slopes (such as soil slopes, rock slopes, and rock-soil mixed slopes), data input formats suitable for slope stability artificial intelligence models are developed, where,

[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 data of soil slopes, and each layer of the slope layer is assigned structural surface parameters.

[0091] For rock-soil mixed slopes, the data input format is consistent with that of rock slopes, but not every layer of the slope layer needs to provide structural surface parameters. The slope layer without structural surface parameters is considered as a soil layer.

[0092] Step S222: User question abnormal scenario design;

[0093] Construct the scene of missing slope parameter information, based on the data input format corresponding to different slope types, design the dynamic response logic of large language model in slope field to missing information, that is, when there is missing slope parameter in the user's input slope evaluation request, the large language model in slope field can automatically identify the missing slope parameter and generate targeted follow-up content to obtain complete slope information, for example:

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

[0095] Model reply: Please supplement the slope height, slope angle, and unit weight of the slope rock-soil layer to perform the evaluation.

[0096] In addition to the above examples, all data missing scenarios should be considered in the design of the user question exception scene; if the user provides complete information, perform keyword extraction function.

[0097] Among them, the steps of human supervision and data optimization for the constructed slope knowledge corpus dataset and slope prediction corpus dataset are:

[0098] Step S231: The field experts (geotechnical engineers) manually review and label the score ranking results output by the general reward model, including: correcting the scores of answers that use technical terms correctly but are misjudged as low scores, and supplementing "high authority" content labels for highly professional answers;

[0099] Step S232: Store the corrected "high-quality answer - low-quality answer" pairs in the constructed preference database for model fine-tuning in the reinforcement learning stage (such as nearest neighbor preference optimization, direct preference optimization algorithm).

[0100] The format example of the preference dataset is:

[0101] User input: Why is slope stability evaluation considered very important?

[0102] Model reply, high-quality answer: Slope stability analysis is very important because it can predict potential landslides or other types of stability problems in slopes. By understanding the stability of the slope, potential safety risks can be avoided during construction or design, ensuring personal safety and property safety.

[0103] Model reply, low-quality answer: Slope stability analysis is crucial for ensuring the safety and stability of engineering structures.

[0104] Step S233: Select the highest-scored question-answer pair 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, engineering case references).

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

[0106] Step S300: The data in the slope stability case constructed in the slope training database is encoded into a training form of column vectors; each column vector represents the information of a slope; according to the number of slope layers, from top to bottom, the slope heel angle, layer height, unit weight, cohesion, and internal friction angle of each slope layer are sequentially filled in the column vector; and this is repeated until the information of all slope layers is filled in; and then the layer number, cohesion, internal friction angle, structure surface centroid x and y coordinates, structure surface inclination angle, and structure surface length of each structure surface are sequentially filled in; wherein, the layer number is sequentially numbered from top to bottom as one to the total number of slope layers;

[0107] Step S310: Since the number of slope layers and the number of structure surfaces are randomly generated, the dimensions (or sizes) of the column vectors are inconsistent. In order to facilitate the training of the artificial intelligence model for predicting slope stability, a dimension alignment operation is performed on the column vectors of the slope stability cases, specifically: taking the column vector with the largest size as a reference, the column vectors smaller than the largest size are zero-padded to make the lengths of all column vectors of the slope stability cases completely consistent;

[0108] The training of the artificial intelligence model for predicting slope stability 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 with its corresponding stability coefficient (FOS) to form an input-output pair, train the support vector machine, random forest, and back feedback neural network, and evaluate the prediction results of the support vector machine, random forest, and back feedback neural network using evaluation indicators, 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-model, the best number of sub-models and related training parameters such as learning rate, iteration number, and minimum batch size are searched using the Bayesian optimization algorithm;

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

[0112] Step S4: Based on the large language model training database, a large language model in the field of slope is constructed, and the process is as shown in Figure 5 Figure 5 CPBao in the figure refers to a large language model trained using a professional domain dataset), mainly including:

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

[0114] The supervised fine-tuning training is the most commonly used training strategy for building a vertical domain large language model based on a general domain large language model. The Llama 3Chinese Chat in the general field is used as the pre-trained model, and the chat-ML style data is used as the data. The training parameters used are: number of iterations: 30 rounds; minimum batch size: 2; initial learning rate: 2e-4; learning rate reduction strategy: Cosine;

[0115] The reinforcement learning based on human feedback includes two strategies: direct preference optimization and near neighbor preference optimization. Both strategies are based on a pre-trained large language model in the general field; wherein,

[0116] The direct preference optimization uses the Llama 3Chinese Chat in the general field as the 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 reduction strategy: Cosine;

[0117] ​The neighbor preference optimization also needs to train a reward model in advance; the reward model uses Llama3Chinese Chat in the general field as the pre-trained model and is trained using the preference data set; the training parameters used are: number of iterations: 10 rounds; minimum batch size: 1; initial learning rate: 1e-4; learning rate reduction strategy: Cosine; after the reward model is trained, the neighbor preference optimization strategy training is started; the neighbor preference optimization uses Llama 3Chinese Chat in the general field as the pre-trained model, is trained using chat-ML style data, 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 reduction strategy: Cosine.

[0118] Step S420: Perform performance evaluation on the large language model after training in the engineering scene, and select the optimal large language model adapted to the slope field, specifically:

[0119] The general reward model fine-tuned by the slope knowledge corpus preference data set is used to score the slope knowledge corpus, and the large language models trained by the three training strategies of supervised fine-tuning, direct preference optimization and neighbor preference optimization are scored and sorted according to the scores; the ROUGE-L (Longest Common Subsequence Recall Oriented Summary Evaluation) index is used to compare and evaluate the model responses of the large language models trained by different training strategies with the human-specified true labels, and the large language models are sorted according to the ROUGE-L evaluation scores; 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 slope field large language model.

[0120] Step S5: The user uploads data to the slope parameter extraction through the trained slope field large language model, and judges 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 field large language model feeds back the slope parameters that need to be supplemented to the user; after the user supplements, the slope field large language model re-extracts and judges again until it is monitored that the user has uploaded all the slope parameters that meet the input requirements of the slope stability prediction artificial intelligence model; the slope field large 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] Wherein, the extraction of the user's uploaded data using the trained large language model in the slope field mainly includes calling the weight of the trained large language model in the slope field through the transformers.pipeline library of Python, specifically:

[0122] The transformers.pipeline library is used to directly call the model weight folder path of the large language model; after calling, the generated public URL link is obtained, as shown in Figure 6 , access the visualization interface by accessing the public URL link, input the relevant question in the "input your question" of the pop-up visualization interface, click "Submit", and then return the model reply.

[0123] As shown in Figure 7 , the trained large language model in the slope field is used to extract the slope parameters of the user's uploaded data; then it is judged whether the information is sufficient. The sufficient information includes the geometric shape of the slope: slope height, broken angle, structure surface position and material parameters, i.e. unit weight, cohesion, internal friction angle, structure surface material parameters, etc. In addition, the large language model in the slope professional field needs to standardize the parameter format, such as unifying the units of parameters;

[0124] When the user provides sufficient information, the large language model in the slope field generates a fixed format response, for example:

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

[0126] Then, after the Python detects the "being evaluated" keyword, the parameters extracted by the large language model in the slope field are stored in the form of column vector, which is consistent with the input format of the artificial intelligence model for predicting the slope stability, and written into a Txt file;

[0127] When the user provides insufficient information, the large language model in the slope field will automatically deduce the parameters that have not been provided according to the provided slope parameters, and make a corresponding model reply to prompt the user to supplement the missing slope parameters.

[0128] Using the trained artificial intelligence model for predicting the slope stability to obtain the stability coefficient of the slope includes the following steps:

[0129] The Python algorithm is used to automatically detect whether a new Txt file is added in the process of calling the large language model; if the new Txt file exists, a 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, and a corresponding stability coefficient is output by the slope stability prediction artificial intelligence model;

[0130] The stability coefficient is automatically read by the Python algorithm, and according to the preset “Technical Code for Building Slope Engineering” (GB50330-2013) standard, the slope field large language model returns an evaluation result containing the following contents: the stability coefficient value of the slope, the stability state of the slope and the measures that can be taken for the stability state.

[0131] The present application is based on the parameter fine-tuning technology and artificial intelligence regression prediction method, and the slope field large language model and the slope stability prediction artificial intelligence model are trained by extracting the slope data, and the collaborative mechanism of the slope field large language model and the slope stability prediction artificial intelligence model is constructed, which simplifies the slope stability evaluation process and improves the efficiency; the core advantage is that the large language model and the slope stability prediction algorithm are integrated, the large language model is fine-tuned by specific slope knowledge and stability prediction dialogue data, the slope stability interactive interface of man-machine dialogue is realized, and the status quo of difficult popularization of slope stability intelligent evaluation algorithm is broken through.

[0132] Suppose an engineer needs to evaluate a slope with complex geological conditions; traditionally, the engineer must collect slope stability parameters from geological reports and field data based on experience to determine FS using numerical or mechanical analysis methods. This mainly involves establishing a slope model, performing stability calculations, and extracting and analyzing results. Once the LLM-driven CNN is implemented, the engineer only needs to ask questions such as “What parameters are needed to evaluate slope stability?” and input relevant parameters for inquiry, such as “I want to evaluate the stability of a slope with a slope angle of 45° and an internal friction angle of 30°.” Then, the model can automatically extract key parameters and provide preliminary stability analysis results, greatly reducing data preparation time and the need for expert knowledge, providing a new perspective for slope safety decision-making.

[0133] The above is only a specific embodiment of the present application, which enables those skilled in the art to understand or implement the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined in the present application can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to these embodiments shown in the present application, but will conform to the widest scope consistent with the principles and novel features of the present application.

Claims

1. A slope stability assessment method based on a large language model and an intelligent prediction model, characterized in that, Includes the following steps: Step S1: Establish a slope training database based on typical slope stability cases; Step S2: Establish a large language model training database based on existing research literature on intelligent slope stability assessment; Step S3: Based on the slope training database, construct an artificial intelligence model for slope stability prediction; the construction steps of the artificial intelligence model for slope stability prediction are as follows: Step S300: Encode the data from the slope stability cases constructed in the slope training database into a training form of column vectors; each column vector represents information about a slope; according to the number of slope layers, fill in the column vectors sequentially from top to bottom with the slope toe angle, layer height, unit weight, cohesion, and internal friction angle of each slope layer; 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 centroid of the structural surface, the dip angle of the structural surface, and the length of the structural surface; Step S310: Perform dimension alignment on the column vectors of the slope stability cases. Specifically, using the column vector with the largest size as a reference, fill the column vectors smaller than the largest size with zeros to make the length of the column vectors of all slope stability cases 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 feedback neural network respectively, and use evaluation metrics to evaluate the prediction results of support vector machine, random forest and feedback neural network, and select the model with the best prediction performance as the ensemble learning sub-model. Step S330: For the three ensemble strategies of Stacking, Boosting, and Bagging, based on the ensemble learning sub-models 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 ensemble strategies, select the ensemble model with the best overall performance as the final artificial intelligence model for slope stability prediction. Step S4: Construct a large language model for the slope domain based on the large language model training database; the construction steps of the large language model for the slope domain are as follows: Step S410: Train the general-purpose large language model based on the aforementioned large language model training database. The training strategies include supervised fine-tuning training and reinforcement learning training based on human feedback. The supervised fine-tuning training is based on the general-domain pre-trained large language model and uses chat-ML style data for training. 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: Evaluate the engineering scenario performance of the trained large language model and select the optimal large language model suitable for the slope domain. Specifically, the following steps are taken: The slope knowledge corpus is scored using a general reward model fine-tuned with a slope knowledge corpus preference dataset. The large language models trained using three training strategies—supervised fine-tuning, direct preference optimization, and nearest neighbor preference optimization—are scored and ranked according to their scores. The slope prediction corpus is evaluated using the ROUGE-L metric, comparing the model responses of the large language models trained with different strategies with manually specified real labels, and ranked according to their ROUGE-L scores. Combining the comprehensive scores from the slope knowledge corpus and the slope prediction corpus, the large language model with the highest score is selected as the trained large language model for the slope domain. Step S5: The trained slope domain big data language model extracts slope parameters from the user's uploaded data and determines whether the extracted slope parameters meet the input requirements of the slope stability prediction AI model. If the slope parameters in the user's uploaded data do not meet the input requirements of the slope stability prediction AI model, the slope domain big data language model provides feedback to the user regarding the slope parameters that need to be supplemented. After the user supplements the parameters, the slope domain big data language model re-extracts and re-determines the parameters until it detects that the user has uploaded all the slope parameters that meet the input requirements of the slope stability prediction AI model. The slope domain big data language model then sends the complete slope parameters into the trained slope stability prediction AI model, which outputs the slope stability coefficient.

2. The slope stability assessment method based on a large language model and an intelligent prediction model according to claim 1, characterized in that, In step S1, the steps for constructing the slope training database are as follows: Step S100: Collect typical slope stability cases and extract slope parameters for each slope stability case; the slope parameters include slope height, slope angle, unit weight of slope soil and rock layer, internal friction angle of slope soil and rock layer, and cohesion of slope soil and rock layer; Step S110: Based on the collected slope parameters, using a single slope stability case as the base sample, randomly select slope parameters from multiple slope stability cases and combine them to construct multiple different layered slopes, where each slope layer has different slope angles and soil parameters. Step S120: Determine the probability distribution function of the slope structural surface parameters and its corresponding mean and variance, and use the probability distribution function to generate random structural surfaces for the layered slope. The parameters of the structural surface include cohesion, internal friction angle, centroid coordinates of the structural surface, inclination angle of the structural surface, length of the structural surface, and centroid spacing between different structural surfaces. Step S130: Combine the constructed layered slope and the generated structural surface in various ways to create multiple different slope cases; Step S140: Using the extended finite element method combined with the strength reduction method, calculate the stability coefficient of each slope case to form a "slope case - stability coefficient" correspondence pair, 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, characterized in that, In step S2, the construction of the large language model training database includes the following steps: We constructed a slope knowledge corpus dataset and a slope prediction corpus dataset, and then performed 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, characterized in that, In step S2, the steps for constructing the slope knowledge corpus dataset are as follows: Step S201: Search for literature related to slope stability assessment, artificial intelligence slope analysis, and geotechnical parameter inversion through academic platforms, download the full text PDFs, and establish an original literature database; use Python libraries to extract PDF text page by page, filter out non-technical content, formulas, and chart descriptions, retain the main text paragraphs, and use regular expressions to segment sentences, removing repeated sentences, short sentences with a length of less than 15 characters, and garbled sentences containing formulas; Step S202: Call the general large language model to generate multi-turn question-answer pairs containing "background knowledge - question - answer" based on the single sentence or context-related sentence cleaned in step S201; then convert the question-answer pairs into Chat-ML format, clarify the correspondence between context, user input and model response, and finally store them in JSON file; Step S203: Using "user input - model response" as input, a general reward model is used to score the response from five dimensions: usefulness, correctness, coherence, complexity, and redundancy. Based on the scoring results, the responses are sorted from highest to lowest according to the total score.

5. The slope stability assessment method based on a large language model and an intelligent prediction model according to claim 4, characterized in that, In step S2, the steps for constructing the slope prediction corpus dataset are as follows: Step S221: For different slope types, formulate data input formats that are compatible with the artificial intelligence model for slope stability; Step S222: Construct a scenario where slope parameter information is missing. Based on the data input format corresponding to different slope types, design the dynamic response logic of the large language model in the slope domain to the missing information. That is, when there are missing slope parameters in the slope evaluation request entered by the user, the large language model in the slope domain can automatically identify the missing slope parameters and generate targeted follow-up questions 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, 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 and ranking results output by the general reward model. This includes correcting the scores of answers that use technical terms correctly but are misjudged as low, and adding annotations to professional answers. Step S232: Store the revised "high-quality answers - low-quality answers" pairs into the constructed preference database; Step S233: Select the highest-scoring question-answer pairs from the modified slope knowledge corpus dataset and 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, characterized in that, In step S5, extracting user-uploaded data using the trained slope domain large language model includes calling the weights of the trained slope domain large language model through Python's transformers.pipeline library, specifically: The transformers.pipeline library is used to directly call the large language model for the slope domain from the folder path where the model weights are located. After the call, the generated public URL link is obtained. Accessing the public URL link leads to the visualization interface, where data can be entered and the model's response can be viewed.

8. The slope stability assessment method based on a large language model and an intelligent prediction model according to claim 7, 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 process of calling 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, and the slope stability prediction artificial intelligence model outputs the corresponding stability coefficient. The stability coefficient is automatically read using a Python algorithm. Based on the preset "Technical Specification for Building Slope Engineering" (GB50330-2013) standard, the algorithm returns an evaluation result containing the stability coefficient value of the slope, the stability status of the slope, and possible measures to be taken in response to the stability status through a large language model in the slope field.

Images