A multi-stage intelligent circulation iteration optimization method for road disease causes

Abstract

The present application relates to a kind of road disease cause multi-stage intelligent circulation iteration optimization method, belong to artificial intelligence and road engineering field.The method includes the road information data input prompt learning big model, generates embedding vector, and constructs the natural language prompt word covering disease original data background, involves specific cause analysis task and disease evolution trend;Through big language model, diagnose road disease cause and output potential cause and possibility order;Generate the summary type disease cause analysis result with disease type main cause and secondary cause order as core;Based on disease cause analysis result, the analysis result after multiple rounds of circulation iteration optimization is summarized, and the disease analysis report containing cause order and possibility evaluation is generated.The present application realizes multidimensionality, efficiency and accuracy in road disease cause analysis by combining prompt learning big model, big language model, text summary model and circulation iteration optimization mechanism, has significant advantages.

Description

Technical Field

[0001] This invention relates to a multi-stage intelligent iterative optimization method for the causes of road defects, belonging to the fields of artificial intelligence and road engineering. Background Technology

[0002] With the continuous growth of road traffic volume and the increasing complexity of climate conditions, the frequent occurrence of road defects poses a serious threat to traffic safety and road lifespan. Therefore, it is essential to analyze the causes of road defects, which will provide strong data support for subsequent scientific decision-making and effective maintenance measures.

[0003] Currently, the analysis of road damage causes mainly relies on manual inspection and experience-based judgment. This method is not only inefficient but also highly subjective, making it difficult to comprehensively analyze the complex causes of road damage. Traditional methods also face significant bottlenecks in data processing, failing to fully utilize multi-dimensional information such as pavement inspection data, traffic flow, and environmental factors. Especially when dealing with large-scale road sections and complex environmental conditions, existing methods struggle to provide real-time, accurate causal diagnosis and optimization solutions.

[0004] In recent years, the application of artificial intelligence technology in the field of road engineering has been increasing, especially the performance of deep learning-based methods in disease identification and data mining. However, most existing technologies focus on improving the accuracy of disease detection, while research on disease cause analysis and multi-factor correlation inference remains relatively limited. For example, although deep learning has made significant breakthroughs in image recognition, its ability to model multi-dimensional and multi-factor interaction relationships in cause inference is still insufficient, making it difficult to achieve efficient and accurate cause inference.

[0005] Furthermore, the current lack of a systematic approach to directly translate the results of road defect analysis into decision support hinders the scientific rigor and efficiency of road maintenance. Most existing systems focus on single-task processing (such as being limited to defect identification or data monitoring) and lack integrated, systematic analysis platforms, failing to effectively combine causal analysis and decision support. This results in a lack of comprehensive optimization guidance for road maintenance, making it unable to adequately address the increasingly complex needs for road defect management.

[0006] Therefore, proposing a method for analyzing the causes of road defects by combining multiple intelligent analysis technologies is of great significance for improving the scientific rigor and timeliness of road defect management. By introducing intelligent and automated diagnostic and decision support methods, the shortcomings of existing technologies can be overcome, providing more accurate and efficient analysis of the causes of road defects and decision support for maintenance. Summary of the Invention

[0007] This invention provides a multi-stage intelligent iterative optimization method for the causes of road defects, addressing the aforementioned problems. It integrates automated prompt generation, deep analysis of large language models, and text summarization techniques, combined with a data-driven iterative optimization mechanism, to realize a closed-loop system for the entire process of road defect diagnosis, from data input to cause analysis and decision support, providing precise data-driven support for the diagnosis and treatment of road defects.

[0008] The present invention adopts the following technical solution:

[0009] The multi-stage intelligent iterative optimization method for the causes of road defects described in this invention comprises the following steps:

[0010] S1. Input road information data into the prompt learning model, generate embedding vectors, and construct natural language prompt words that cover the background of the original disease data, the analysis tasks involving specific causes, and the disease evolution trend, as input to the large language model.

[0011] S2. Input the embedded vectors, natural language prompts, and pre-organized knowledge graph of common road disease causes generated in step S1 into the large language model to diagnose the causes of road diseases and output the potential causes and probability ranking. The output results are used as the input of the subsequent text summary model.

[0012] S3. Based on the text summary model, the detailed causal analysis content output by the large language model in step S2 is compressed and restructured to generate a summary-style disease cause analysis result with the ranking of primary and secondary causes of disease types as the core.

[0013] S4. Based on the results of the disease cause analysis, the precise optimization analysis results under specific environmental conditions are obtained through multiple rounds of iterative analysis targeting specific disease causes.

[0014] S5. Summarize the analysis results after multiple rounds of iterative optimization to generate a disease analysis report that includes cause ranking and probability assessment.

[0015] The multi-stage intelligent iterative optimization method for the causes of road defects described in this invention includes road information data in step S1, such as basic road information, defect information, traffic volume information, environmental information, and historical maintenance information.

[0016] Road information data is transformed into multidimensional feature vectors through a multi-layer embedding network. These vectors include basic road information factors, road damage information factors, traffic volume information factors, environmental information factors, and historical maintenance information factors. The multidimensional embedding vectors are generated as follows:

[0017] V i =Embedding(x1,x2,...,x s ),x i ∈Rd

[0018] Where x i Let be the i-th input feature, s be the token length, d be the embedding dimension, and R represent the set of real numbers. d Let x represent a vector space consisting of d real numbers. i = (Traffic volume, precipitation, ...), V i The generated embedding vector;

[0019] Generate natural language prompts that cover the background of the original disease data, the analysis task involving specific causes, and the disease evolution trend:

[0020] Prompt = f(x1,x2,...,x) s )

[0021] Where Prompt represents the generated natural language prompt word, f(·) represents the mapping function executed by the prompt learning large model, and the input is the embedding vector sequence (x1, x2, ..., x) of the road information data. s The output consists of natural language prompts that cover the background of the original disease data, the specific cause analysis task, and the disease evolution trend.

[0022] The multi-stage intelligent iterative optimization method for the causes of road defects described in this invention includes an input embedding vector sequence in step S2:

[0023] X = (x1, x2, ..., x T ),x t ∈R d

[0024] Natural language prompts are represented as a sequence of tokens:

[0025] Prompt=(ω1,ω2,...,ω T )

[0026] Where, ω t This represents the t-th token in the natural language prompt word sequence, where T represents the total length of the prompt word sequence;

[0027] Knowledge graphs serve as important contextual information:

[0028] K = {(c i ,r ij ,c j )}

[0029] Among them, c i c j For the concept, r ij To explain the relationship between them.

[0030] The multi-stage intelligent iterative optimization method for the causes of road defects described in this invention uses a large language model whose core computation consists of a multi-layer Transformer encoder. The multi-layer Transformer encoder includes a self-attention mechanism layer, a multi-head attention mechanism layer, and a position encoding layer. The self-attention mechanism layer is as follows:

[0031]

[0032] Where Q = XW Q K = XW K V = XW V It is a linear mapping of Query, Key, and Value, W Q W K W V These are learnable parameters;

[0033] Multi-head attention mechanism layer:

[0034] Multihead(X)=Concat(head1,...,head h W O

[0035] head i =Attention(XW i Q XW i K XW i V )

[0036] Location coding layer:

[0037] PE (pos,2i) =sin(pos / 10000) 2i / d )

[0038] PE (pos,2i+1) =cos(pos / 10000) 2i / d )

[0039] The token embedding vector is added to the positional encoding and then used as input:

[0040] H (0) =X+PE

[0041] The final output after L layers of Transformer computation is:

[0042] H (L) =Transformer L (H (0) ).

[0043] The multi-stage intelligent iterative optimization method for the causes of road defects described in this invention diagnoses the causes of road defects and outputs potential causes and probability rankings.

[0044] The large language model is based on the last layer representation H of the output layer after computation by an L-layer Transformer. (L) For the candidate set of causes C = {c1, c2, ..., c...} m}Score each cause c i The score is expressed as:

[0045] s i =MLP(H (L) ,context i )

[0046] context i For prompt words or knowledge graphs related to c i The relevant context;

[0047] Score converted to probability form:

[0048]

[0049] The sorting result is thus obtained, and the final output structure is as follows:

[0050]

[0051] The output cause report is a paragraph that covers disease types, main causes, and secondary causes of diseases. Its length and structure are somewhat complex, and it urgently needs to be summarized by a text summary model.

[0052] In the multi-stage intelligent iterative optimization method for the causes of road diseases described in this invention, in step (3), the text summary model performs information compression and recombination. The detailed causal analysis content output by the large language model is compressed by the text summary model, and the disease type, main cause, secondary cause and their ranking information are extracted. The structure is as shown in the formula:

[0053] Summary=MainCause+SubCauses+Likelihood

[0054] Among them, MainCause is the primary cause, SubCauses are other secondary causes, and Likelihood is the probability ranking.

[0055] The multi-stage intelligent iterative optimization method for the causes of road defects described in this invention uses an objective function in step S4 to measure the optimization effect of each iteration. The optimization objective is to minimize the loss function.

[0056]

[0057] in, y is the model's predicted value. i Let L(θ) be the true value, L(θ) be the loss function, and θ be the model parameters;

[0058] In each iteration, the model parameters are updated using gradient descent:

[0059]

[0060] Where, θ (t) and θ (t+1) These represent the current and updated model parameters, respectively, where η is the learning rate. This is the gradient of the loss function.

[0061] Beneficial effects

[0062] This invention combines a prompting learning model, a large language model, a text summarization model, and a cyclical iterative optimization mechanism to achieve multi-dimensionality, efficiency, and accuracy in the analysis of the causes of road defects, which has significant advantages.

[0063] Specifically, it includes:

[0064] 1. Precise Cause Diagnosis: Through multi-stage intelligent analysis technology, this invention can perform precise cause diagnosis of road defects based on a large amount of road information data. Unlike traditional methods that rely on manual judgment, the system uses an automated, data-driven approach to comprehensively analyze the type, cause, and evolution trend of defects, eliminating human error and subjective bias, and improving the accuracy of cause analysis.

[0065] 2. Highly efficient multi-factor analysis capability: This invention employs a multi-model collaborative approach (such as combining a large language model with a prompting learning model) to efficiently process multi-dimensional data in road information (such as traffic volume, environmental information, disease types, historical maintenance records, etc.). This multi-factor analysis capability enables the system to identify complex disease causes and perform multi-dimensional cross-validation, thereby providing more comprehensive and in-depth diagnostic results.

[0066] 3. Dynamic Optimization and Feedback Mechanism: Through iterative optimization, this invention can self-adjust and optimize the model based on previous analysis results. This mechanism can dynamically address the causes of road damage under different road sections, seasons, or climatic conditions, ensuring the timeliness and relevance of the analysis results. Simultaneously, the system can continuously improve prediction accuracy based on historical data and real-time feedback, enabling precise scheduling and optimization of road maintenance work according to specific circumstances.

[0067] 4. Integration of Decision Support and Practical Application: This invention not only provides accurate results for disease cause analysis but also generates structured analysis reports, directly transforming them into actionable decision support. By combining text summary models, the analysis results of this invention are clear and concise, facilitating quick understanding and appropriate measures by maintenance personnel and decision-makers, effectively improving the scientific rigor and efficiency of maintenance work.

[0068] 5. Comprehensively Improve the Efficiency of Road Defect Management: Through intelligent automatic analysis and optimization processes, this invention significantly reduces the need for manual intervention and improves the overall efficiency of road defect management. Since the analysis results directly guide maintenance decisions, this invention not only improves the accuracy of defect diagnosis but also enables precise management in actual road maintenance, reducing maintenance costs and extending road service life. Attached Figure Description

[0069] Figure 1 This is a system framework diagram of the present invention;

[0070] Figure 2 This invention provides a prompt for learning the large model structure diagram;

[0071] Figure 3 This is a schematic diagram of the large language model analysis process of the present invention;

[0072] Figure 4 This is a schematic diagram of the text summary model processing of the present invention. Detailed Implementation

[0073] To make the objectives and technical solutions of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0074] To facilitate understanding of the implementation process of this invention, a specific embodiment is provided below for detailed explanation. Assume that on a section of a highway within Anhui Province, the following road information has been collected: traffic volume (e.g., average daily passenger / freight vehicle traffic), weather conditions (e.g., maximum temperature, precipitation), typical pavement distress types (e.g., longitudinal cracks, transverse cracks, alligator cracks), road structure information, and historical maintenance records. This section faces a complex diagnostic problem involving multiple overlapping types of composite distresses, urgently requiring intelligent analysis methods for causal identification and maintenance decision support.

[0075] Step (1) prompts learning the large model

[0076] This model borrows from the T5 (Text-to-Text Transfer Transformer) structure. The encoder sequentially embeds road defect data (such as traffic volume, precipitation, and crack types) from a section of a highway in Anhui Province, generating context-aware vector representations. The decoder, with road defect type as the task objective, generates structured natural language prompts. The model structure mainly consists of three parts: an input embedding layer, a bidirectional Transformer encoder, and a task-guided natural language decoder. During model training, an instruction-tuning paradigm is introduced to ensure that the generated prompts cover defect locations, environmental conditions, and data context, thus providing inference-oriented input for subsequent larger models.

[0077] First, input the road information data table based on the detection results into the prompting learning model. This table contains the following information dimensions:

[0078] (1) Basic information factors of the road itself, such as structural layer thickness, road grade, and pavement type;

[0079] (2) Disease information factors, such as disease type, disease area, and location;

[0080] (3) Traffic volume information factors, such as average daily traffic volume;

[0081] (4) Environmental information factors, such as temperature and humidity, rainfall, seasonal distribution, etc.;

[0082] (5) Historical maintenance information factors, such as previous maintenance measures and frequency.

[0083] In this step, the data is first fed into a large-scale cue learning model. This model transforms the road defect data table into embedding vectors and generates natural language cue messages. For example, it generates the following cue words:

[0084] "Based on the following road damage data, please analyze the causes of the damage from four aspects: road materials, road structure, traffic load, and external environment, combining engineering standards and actual data, and provide reasonable evidence and probability ranking."

[0085] Data Background

[0086] • Road sections with defects: {starting and ending chainage}

[0087] • Types of damage: {Types of damage, such as ruts, cracks}

[0088] • Affected area: {Area data}

[0089] Traffic data:

[0090] Daily average traffic volume: {N, O, P, ...} vehicles (representing the daily average traffic volume of different vehicle types).

[0091] o Design load capacity vs. actual load: Please assess whether the load exceeds the limit.

[0092] • Environmental data:

[0093] Maximum temperature: {T_max}℃, Minimum temperature: {T_min}℃

[0094] o Asphalt aging temperature threshold vs. temperature of the area: Is aging likely to be accelerated?

[0095] Given the precipitation amount of {P_avg} mm and the maximum precipitation amount of {P_max} mm, is it possible that this will affect the stability of the base course?

[0096] • Historical preservation:

[0097] Recent maintenance measures, such as asphalt overlay and milling repaving.

[0098] Was crack-resistant material used? If not, could it exacerbate the cracking?

[0099] Analysis task

[0100] 1. Material factors (such as asphalt aging, aggregate gradation, etc.):

[0101] By combining temperature data, it can be determined whether the asphalt may be aging.

[0102] Was the maintenance effective? Are there any issues with material performance degradation?

[0103] 2. Structural factors (such as base course thickness, subgrade settlement, etc.):

[0104] If subsidence occurs, could it lead to longitudinal cracks?

[0105] Does the maintenance record include any mention of base layer reinforcement? If not, could this have led to reflective cracking?

[0106] 3. Traffic load (e.g., percentage of overloaded vehicles, frequent starts and stops, etc.):

[0107] Does the truck traffic volume exceed the design standard?

[0108] Could traffic loads lead to structural fatigue and crack propagation?

[0109] 4. Environmental factors (such as diurnal temperature range, rainfall infiltration, etc.):

[0110] Could temperature stress lead to thermal shrinkage cracks?

[0111] Is it possible for precipitation to seep into the base layer, causing structural damage?

[0112] Please output the disease cause analysis, including a structured analysis of 'disease type → possible causes → evidence → probability ranking'.

[0113] • Ensure that the causal analysis is based on data, rather than general discussions.

[0114] • Please analyze the causal relationships between the defects. For example, if roadbed settlement leads to longitudinal cracks, please clearly indicate this relationship.

[0115] • Please use engineering specifications (such as JTG D50-2017) for reasoning, and avoid relying solely on subjective judgment.

[0116] The key to this step is to automatically generate prompt words to transform the raw data into a natural language form that is easy for subsequent large language models to process, thus providing structured input for causal analysis.

[0117] Step (2) Large Language Model Analysis

[0118] Next, the embedded vectors generated in step (1), natural language prompts, and a pre-organized knowledge graph of common road disease causes are input into a large language model (such as GPT or DeepSeek). The large language model uses its deep learning algorithm to analyze the input data and, in conjunction with the known knowledge base of road disease causes, diagnoses the causes of the diseases.

[0119] Taking this embodiment as an example, the large language model outputs the following analysis results:

[0120] "Analysis results of the causes of road section XX: "

[0121] The main reason for the longitudinal cracks is likely that the road section has a high proportion of heavy traffic loads, coupled with poor roadbed bearing capacity, which leads to their formation under long-term heavy loads.

[0122] Secondly, a possible reason is that the semi-rigid base layer cracks under the combined action of temperature stress and load stress, and the stress concentration at the bottom of the surface layer causes the cracks to gradually extend upward.

[0123] The main cause of the transverse cracks is likely the water infiltration caused by long-term rainfall in this section of the road, which leads to the expansion of the base layer and further causes cracks.

[0124] Secondly, it could be that temperature stress causes cracks in the base layer to reflect onto the surface layer.

[0125] Through this step, the system can analyze potential causes of road defects based on road data and provide a possible ranking of causes for each type of defect.

[0126] Step (3) Text summarization model

[0127] In step (2), the analysis results obtained from the large language model may be quite lengthy and difficult to use directly for decision support. Therefore, this invention introduces a text summarization model to refine and condense the analysis results output by the large language model. The model is based on the BART (Bidirectional and Auto-Regressive Transformers) architecture and uses an encoder-decoder mechanism to semantically compress the long text causal analysis output by the large language model. The encoder performs contextual modeling on the input analysis statement, and the decoder generates a standardized causal summary (such as 'disease type - primary cause - secondary cause - probability') based on the target summary template. The model optimization introduces R3F robustness loss and RLHF fine-tuning strategy to ensure that the output language is concise, highlights the key points, and has the reliability for engineering judgment. To ensure model stability, SpanBERT embedding is used to improve the ability to abstract the relationships between causal entities.

[0128] For example, after processing by the text summary model, the following concise causal analysis report is generated:

[0129] "Analysis Report on the Causes of Road Section XX:"

[0130] The most likely cause of longitudinal cracks is long-term heavy load, followed by the reflection of cracks in the base layer to the surface layer.

[0131] The most likely cause of transverse cracks is moisture infiltration due to precipitation, with temperature stress being the second most probable contributing factor.

[0132] The report is concise and clear, allowing decision-makers to quickly obtain an overview of the causes of the disease.

[0133] Step (4) Iterative optimization:

[0134] To improve the accuracy of large language models in analyzing the causes of diseases under specific environmental factors (such as load, temperature, precipitation, etc.), this system introduces a cyclical iterative optimization mechanism. This mechanism continuously improves and adjusts the model's output by optimizing natural language prompt templates and guidance strategies.

[0135] (4.1) Dynamic Adjustment of Prompt Templates: After each causal analysis, the system dynamically adjusts the structure, instructions, and emphasis of the natural language prompts based on the quality and coverage of the large language model output. For example, if the model fails to identify the typical cause "base layer reflective cracks" in multiple consecutive samples, the system will enhance the analytical guidance related to "structural layer status" in the prompts, such as adding the instruction: "Please pay special attention to the integrity of the base layer and its possible reflective effect." This process is completed through a rule system or based on a small amount of manual feedback, without the need to train or update the model weights.

[0136] (4.2) Output Quality Assessment and Template Feedback Closed Loop: After each round of analysis, the system performs a structured evaluation of the output of the large language model. Evaluation indicators include the rationality of Top-N causes, keyword recall rate, and consistency between the analysis content and input data. If the output causes are found to lack explanations of key disease characteristics or to be poorly ranked, the problem of the current template version is recorded, and a new template is generated based on a preset strategy. For example, if the original prompts fail to guide the model to consider the precipitation infiltration mechanism, the system will add the analysis requirement of "whether the structure is loose due to long-term water accumulation" to the next round of prompts.

[0137] (4.3) Domain strategy knowledge integration: Prepare a set of typical reasoning paths for disease types and causes in advance, and integrate them into the prompt words in the form of supplementary corpus or question guidance.

[0138] For example, adding the question "Does a diurnal temperature difference exceeding 15°C constitute a temperature stress risk?" to the transverse crack prompt can enhance the model's ability to identify causes in key contexts.

[0139] (4.4) Template Iteration Number Control and Termination Mechanism: The system performs 3-5 rounds of template updates by default, regenerating and evaluating the analysis results after each iteration. If the output results of two consecutive rounds do not show significant improvement in accuracy, coverage, or other indicators, or if the indicators meet the preset standards (e.g., main cause matching rate ≥ 90%), the iteration is terminated, and the final cause report is output. The entire process does not involve model parameter training, ensuring that the system is simple to deploy and has good generalization ability and practicality. Through the above multi-dimensional optimization, the system can dynamically adapt to the distribution of disease causes in different regions, seasons, or road conditions, achieving "learning on demand and optimizing according to causes," ensuring that the final output cause analysis report is more reliable and has greater decision-making value.

[0140] Through the above optimization process based on the prompting learning paradigm, the system achieves the goal of improving analysis quality without fine-tuning, simply through prompting optimization, and significantly enhances the model's interpretability and adaptability under different environmental conditions and disease types.

[0141] Step (5) Final Output

[0142] After multiple rounds of iterative optimization, the system can finally generate an accurate causal analysis report. This report not only includes a ranking of causes but also provides probability assessments based on different scenarios analyzed by the model. For example, after final analysis and optimization, the system outputs the following report:

[0143] "Final Cause Analysis Report for Road Section XX:"

[0144] • Longitudinal cracks:

[0145] 1. Possible cause: Prolonged heavy load, likelihood: 85%

[0146] 2. Possible cause: Cracks in the base layer reflect onto the surface layer; probability: 70%.

[0147] • Horizontal cracks:

[0148] 1. Possible cause: Rainwater infiltration, probability: 90%

[0149] 2. Possible cause: Temperature stress, probability: 65%

[0150] This report can provide a scientific basis for subsequent road maintenance and repair, enabling more efficient maintenance decisions.

[0151] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A multi-stage intelligent iterative optimization method for the causes of road defects, characterized in that: The steps are as follows: S1. Input road information data into the prompt learning model to generate embedding vectors. The road information data includes basic road information, damage information, traffic volume information, environmental information, and historical maintenance information. The system constructs natural language prompts that cover the background of the original disease data, the analysis tasks involving specific causes, and the disease evolution trend as input to the large language model. The natural language prompts guide the analysis from four aspects: road materials, road structure, traffic load, and external environment, and the output paradigm is ranked by disease type, possible cause, evidence, and probability. S2. Input the embedded vectors, natural language prompts, and pre-organized common disease cause knowledge graph generated in step S1 into the large language model to diagnose the causes of road diseases and output the potential causes and probability ranking. The output results are used as the input for the subsequent text summary model. S3. Based on the text summary model, the detailed causal analysis content output by the large language model in step S2 is compressed and restructured to extract the disease type, main cause, secondary cause and their probability ranking, and generate a summary-style disease cause analysis result. S4. Based on the results of the disease cause analysis, the precise optimization analysis results under specific environmental conditions are obtained through multiple rounds of iterative analysis targeting specific disease causes. After each round of analysis, the output of the large language model will be structured and evaluated. Evaluation metrics include the rationality of Top-N causes, keyword recall rate, and consistency between the analysis content and the input data. When the output causes lack explanations of key disease characteristics or the cause ranking is unreasonable, the structure, instructions and emphasis of the natural language prompts are dynamically adjusted according to the evaluation results. The typical reasoning paths of disease types and causes are integrated into the adjusted natural language prompts in the form of supplementary prompts or question guidance, and the cause analysis is performed again. By default, 3 to 5 rounds of template updates are performed. The iteration is terminated when the output results in two consecutive rounds do not show a significant improvement in accuracy and coverage, or when the main cause matching rate reaches the preset standard. The model weights of the large language model are optimized in a loop without training or updating. S5. Summarize the analysis results after multiple rounds of iterative optimization to generate a disease analysis report that includes cause ranking and probability assessment.

2. The multi-stage intelligent iterative optimization method for the causes of road defects according to claim 1, characterized in that: In step S1, Road information data is transformed into multidimensional feature vectors through a multi-layer embedding network. These vectors include basic road information factors, road damage information factors, traffic volume information factors, environmental information factors, and historical maintenance information factors. The multidimensional embedding vectors are generated as follows: ； in Let be the i-th input feature, s be the token length, d be the embedding dimension, and R represent the set of real numbers. d Let d represent a vector space consisting of d real numbers. The generated embedding vector; Generate natural language prompts that cover the background of the original disease data, the analysis task involving specific causes, and the disease evolution trend: ； Here, Prompt represents the generated natural language prompt. This represents the mapping function executed by the prompting learning large model, with the input being a sequence of embedded vectors from road information data. The output consists of natural language prompts that cover the background of the original disease data, the specific cause analysis task, and the disease evolution trend.

3. The multi-stage intelligent iterative optimization method for the causes of road defects according to claim 1, characterized in that: In step S2, the input embedding vector sequence is: ； Natural language prompts are represented as a sequence of tokens: ； Where, ω t This represents the t-th token in the natural language prompt word sequence, where T represents the total length of the prompt word sequence; Knowledge graphs serve as important contextual information: ； in, , For the concept, To explain the relationship between them.

4. The multi-stage intelligent iterative optimization method for the causes of road defects according to claim 1, characterized in that: The core computation of a large language model consists of a multi-layer Transformer encoder; the multi-layer Transformer encoder includes a self-attention mechanism layer, a multi-head attention mechanism layer, and a positional encoding layer; among which the self-attention mechanism layer: ； in, , , It is a linear mapping of Query, Key, and Value. , , These are learnable parameters; Multi-head attention mechanism layer: ； ； Location coding layer: ； ； The token embedding vector is added to the positional encoding and then used as input: ； The final output after L layers of Transformer computation is: 。 5. The multi-stage intelligent iterative optimization method for the causes of road defects according to claim 1 or 4, characterized in that: The method describes diagnosing the causes of road defects and outputting potential causes and their probability ranking. The large language model is based on the last layer of the output layer after computation by an L-layer Transformer. For the candidate set of causes Scoring is performed for each cause. The score is expressed as: ； in For prompt words or knowledge graphs related to The relevant context; Score converted to probability form: ； This yields the sorting results, and the output cause report is a paragraph covering the disease type, the main cause of the disease, and the secondary cause.

6. The multi-stage intelligent iterative optimization method for the causes of road defects according to claim 1, characterized in that: In step S3, the text summary model performs information compression and recombination. It compresses the detailed causal analysis output by the large language model, extracting disease type, primary cause, secondary cause, and their ranking information. The structure is shown in the formula: Summary = MainCause + SubCauses + Likelihood; Among them, MainCause is the primary cause, SubCauses are other secondary causes, and Likelihood is the probability ranking.

Images