A ground penetrating radar road underground hidden disease identification method combining LLM and RAG

Abstract

The application relates to a ground penetrating radar road underground hidden disease identification method combining LLM and RAG, and the method comprises the following steps: acquiring scene characteristic information of a current detection scene; based on the scene characteristic information, adopting an RAG retrieval method to retrieve a plurality of matching algorithm processes and configuration parameters from a pre-constructed knowledge base; based on the plurality of matching algorithm processes and configuration parameters, combining the scene characteristic information and identification task requirements, and adopting a large language model to generate a plurality of complete identification algorithm schemes; acquiring a complete identification algorithm scheme selected by a user from the plurality of complete identification algorithm schemes, and performing road underground hidden disease identification under the current detection scene according to the selection result of the user to obtain an identification result. Compared with the prior art, the application has the advantages of improving the self-adaptive capability, improving the reliability and accuracy of the identification result and the like.

Description

Technical Field

[0001] This invention relates to the field of road inspection, and in particular to a ground-penetrating radar method for identifying hidden road defects by combining LLM and RAG. Background Technology

[0002] Ground penetrating radar (GPR), as a non-destructive testing technology, is widely used for the identification and detection of hidden defects in road structures. GPR transmits high-frequency electromagnetic waves into the ground and receives reflected wave signals from the underground medium interface. Through signal processing and image interpretation, it can identify defects such as voids, loosening, cracks, and delamination in road structures.

[0003] However, ground-penetrating radar (GPR) defect detection involves a complex multi-stage processing flow, including gain processing, noise reduction, background removal, filtering, multi-feature fusion, and deep learning-based defect identification. Each stage requires the selection of appropriate algorithms and parameters, which are highly dependent on prior knowledge and expert experience. For example, gain processing requires selecting time-varying gain, automatic gain, or other methods and corresponding parameters based on signal attenuation characteristics; noise reduction requires selecting wavelet thresholding, empirical mode decomposition, or Kalman filtering, or other methods and corresponding parameters based on noise type; and filtering requires selecting bandpass filtering, FK filtering, or other methods and corresponding parameters based on interference wave characteristics.

[0004] In existing technologies, ground-penetrating radar data processing mainly relies on manual experience to determine algorithm flow and parameter settings, which has the following problems: First, it is highly dependent on expert experience. Technical personnel with different engineering backgrounds may choose different algorithm processes and parameters, which can affect the stability and repeatability of the recognition results.

[0005] Second, optimizing the algorithm parameters is difficult. Ground penetrating radar signals are affected by a variety of factors, including road structure type, ground penetrating radar frequency, detection depth, and roadbed material, making it difficult to find the optimal parameter combination through trial and error.

[0006] Third, it lacks adaptive capabilities. When the detection scenario changes, the algorithm process and parameters need to be manually readjusted, making automated adaptation impossible.

[0007] Fourth, knowledge accumulation and transmission are difficult. Successful experiences and lessons learned from past projects are difficult to accumulate and reuse effectively, forcing each project to start from scratch.

[0008] Therefore, how to reduce reliance on expert experience and automatically generate the optimal algorithm flow and parameter configuration according to specific scenarios to improve the accuracy and efficiency of ground penetrating radar defect identification is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0009] The purpose of this invention is to provide a ground-penetrating radar method for identifying hidden road defects by combining LLM and RAG, which enables targeted and more accurate identification of hidden defects.

[0010] The objective of this invention can be achieved through the following technical solutions: A ground-penetrating radar method for identifying hidden road defects combining LLM and RAG includes the following steps: Obtain scene feature information of the current detection scene; Based on the scene feature information, the RAG retrieval method is used to retrieve multiple matching algorithm processes and configuration parameters from a pre-built knowledge base; Based on the aforementioned multiple matching algorithm processes and configuration parameters, and combined with the aforementioned scene feature information and recognition task requirements, a large language model is used to generate multiple complete recognition algorithm schemes. Obtain the complete recognition algorithm scheme selected by the user from the multiple complete recognition algorithm schemes, and perform road underground hidden defects recognition in the current detection scenario based on the user's selection result to obtain the recognition result.

[0011] Furthermore, the scene feature information includes one or more of the following: road structure type, ground penetrating radar frequency, detection depth range, and roadbed material type.

[0012] Furthermore, the knowledge base adopts a structured storage format, including the following data fields: Road structure type field: Records the type information of pavement structure, base course structure, and subbase course structure; Ground penetrating radar frequency field: Records the center frequency of the ground penetrating radar antenna; Algorithm flow field: Records the algorithms selected in each stage of preprocessing, feature extraction, and disease identification; Parameter configuration field: Records the key parameters and parameter ranges of each algorithm; Success Case Field: Records the scenario characteristics and recognition performance metrics of historical successful application cases.

[0013] Furthermore, for the algorithm flow fields, the algorithms in the preprocessing stage include gain, noise reduction, background removal, and filtering algorithms; the algorithms in the feature extraction stage include time-domain feature extraction, frequency-domain feature extraction, and time-frequency feature extraction algorithms; and the algorithms in the disease identification stage are deep learning algorithms.

[0014] Furthermore, regarding the parameter configuration field, In the gain algorithm, the range of the gain coefficient, the range of the time constant of the automatic gain control, and the range of the exponential coefficient of the exponential gain are recorded. In the denoising algorithm, the parameters for wavelet threshold denoising and the range of the number of intrinsic mode functions retained for empirical mode decomposition denoising are recorded. In the background removal algorithm, the range of window size and the range of background update rate for the moving average background removal are recorded. In the filtering algorithm, the cutoff frequency range of the low-pass filter and the center frequency range of the band-pass filter are recorded. In deep learning algorithms, the range of network layers, learning rate, and batch size are recorded for each deep learning algorithm.

[0015] Furthermore, the steps of retrieving multiple matches, including the algorithm flow and configuration parameters, include: The knowledge base is vectorized using a text embedding model, and a vector index structure is constructed to obtain the vectorized knowledge base; The scene feature information is vectorized to obtain vectorized scene feature information; Calculate the cosine similarity score between the vectorized scene feature information and each entry vector in the vectorized knowledge base. _sim ; The road structure type and ground-penetrating radar frequency in the vectorized scene feature information are matched with the road structure type and ground-penetrating radar frequency in the vectorized knowledge base, respectively. If the road structure type and ground-penetrating radar frequency are completely matched, a matching score is awarded. _match =1. If both the road structure type and the ground-penetrating radar frequency are partially matched, then their respective matching scores are obtained first, and then the weighted average is taken as the matching degree score. _match If the road structure type and the ground-penetrating radar frequency do not match, the matching score will be... _match =0. If either the road structure type or the ground-penetrating radar frequency is a perfect match and the other is a partial match, then the matching score of the perfectly matched side is set to 1, and the matching score of the partially matched side is obtained. The weighted average is then taken as the matching degree score. _match ;; Calculate the recognition performance index of historical successful application cases in the vectorized knowledge base, and calculate the weight of the historical successful application cases based on the recognition performance index. The weight of each historical successful application case is obtained by taking a weighted average of the recognition performance indexes. _history Alternatively, the weight of the historical successful application case can be obtained by taking a weighted average of the number of historical successful application cases and the number of applications. The cosine similarity score _sim Match score _match Weighting of historical successful application cases_history The weighted calculation yields the overall similarity score. _total , is represented as: Score _total = α×Score _sim + β×Score _match + γ×Weight _history , α+β+γ=1, In the formula, α, β, and γ are weighting coefficients; Based on the comprehensive similarity score _total The matching algorithm process and configuration parameters are obtained from the knowledge base to find the k best matches of the scene feature information.

[0016] Furthermore, the steps for generating the complete recognition algorithm scheme include: The multiple matching algorithm processes and configuration parameters, scene feature information, and recognition task requirements are converted into structured prompt words; The structured prompt words are input into a large language model to generate multiple complete initial recognition algorithm schemes. The consistency of the multiple complete initial recognition algorithm schemes is checked, and the algorithms are ranked using comprehensive similarity to obtain the final multiple complete recognition algorithm schemes.

[0017] Furthermore, it also includes: when two or more complete recognition algorithm schemes are selected, comparing and verifying all the output recognition results, specifically including the following steps: Cross-validate all identification results and count the diseased areas that are consistently identified as high-confidence results; Obtain manually labeled disease samples and calculate the recognition performance index of each complete recognition algorithm scheme; The identification result is compared with the identification result output by the traditional identification method to obtain the comparison result; Based on the high confidence results, recognition performance indicators, and comparison results, determine whether the verification is successful. If so, output the final recognition result; otherwise, adjust the parameters of each complete recognition algorithm scheme and regenerate the recognition result.

[0018] Furthermore, it also includes updating the knowledge base, specifically including the following steps: In a manner that involves periodic or real-time updates, based on the recognition results, the complete recognition algorithm scheme that successfully completes the recognition task is added to the success case field of the knowledge base.

[0019] Furthermore, it also includes: when generating multiple complete recognition algorithm schemes, user feedback is taken into consideration, specifically including the following: Obtain the complete recognition algorithm schemes and their recognition effects selected by users in each recognition task to form a user preference model; The user preference model is also used as input to the large language model to generate multiple complete recognition algorithm schemes.

[0020] Compared with the prior art, the present invention has the following beneficial effects: (1) Improved reliability and engineering feasibility of the solution: The present invention retrieves the matching algorithm process and configuration parameters that match the current detection scene from the knowledge base according to the RAG retrieval method. It can effectively suppress the illusion problem generated by the large language model, and also use the scene feature information and recognition task requirements as input to the large language model. It can generate the most suitable recognition algorithm solution for different working conditions and scenes, thereby achieving targeted and more accurate identification of hidden diseases.

[0021] (2) Improve the efficiency and adaptability of algorithm parameter optimization: This invention, through a hybrid retrieval strategy, comprehensively considers similarity scores, precise matching of scene features, and the weight of historical successful cases. It can quickly find the algorithm solution that best matches the current scene using a large language model, avoiding the inefficiency of trial and error, reducing the cost and cycle of solution design, and ensuring the adaptability of algorithms and parameters in each stage to the current ground penetrating radar data. In addition, by introducing the weight of historical successful cases, highly reliable empirical solutions can be reused first, improving the robustness and accuracy of disease identification.

[0022] (3) Reduce reliance on expert experience: By constructing a knowledge base that includes historical project experience, expert-organized content, and literature mining results, and combining a large language model and retrieval enhancement generation technology, this invention can automatically generate the optimal algorithm flow and parameter configuration according to specific scenarios, reduce manual intervention, and improve the stability and repeatability of recognition results.

[0023] (4) Enhanced adaptability: When the detection scene changes, the present invention automatically retrieves and generates an appropriate algorithm flow and parameters based on the new scene features to achieve adaptive adjustment without manual reconfiguration.

[0024] (5) Promote knowledge accumulation and inheritance: The knowledge base of this invention continuously accumulates the successful experiences and lessons learned from historical projects, forming reusable knowledge assets, realizing the effective inheritance of knowledge, adapting to different road structures and ground-penetrating radar scenarios and avoiding repeated exploration, thereby improving overall work efficiency.

[0025] (6) Further improve the reliability and accuracy of the recognition results: The present invention further improves the reliability and accuracy of the recognition results through a comparison and verification mechanism, including cross-validation of multiple schemes, comparison with manual annotation results and traditional method results. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a schematic diagram of the knowledge base structure of the present invention; Figure 3 This is a schematic diagram of the hybrid retrieval strategy calculation process of the present invention; Figure 4 This is a schematic diagram of the algorithm flow scheme generation and user selection interface provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the comparison and verification process provided in an embodiment of the present invention. Detailed Implementation

[0027] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0028] This embodiment provides a ground-penetrating radar method for identifying hidden road defects by combining LLM and RAG. This method aims to address the technical problems of existing technologies, such as strong reliance on expert experience, difficulty in optimizing algorithm parameters, and lack of adaptive capabilities. By constructing a knowledge base, employing RAG retrieval, and generating a large language model, it achieves automatic generation of the optimal algorithm flow and parameter configuration based on specific scenarios, offering advantages such as improved identification accuracy, reduced reliance on experience, and enhanced adaptive capabilities. Specifically, for example... Figure 1 As shown, the method includes the following steps: Step S1: Building a knowledge base like Figure 2 As shown, Knowledge Base 20 uses a structured storage format and includes the following data fields: The Road Structure Type field 201 records the type information of pavement structure, base course structure, and subbase course structure. For example, pavement structure type may include asphalt concrete pavement, cement concrete pavement, etc.; base course structure type may include graded crushed stone base course, cement stabilized crushed stone base course, etc.; subbase course structure type may include natural gravel subbase course, lime-soil subbase course, etc.

[0029] The ground-penetrating radar frequency field 202 is used to record the center frequency of the ground-penetrating radar antenna. Commonly used ground-penetrating radar frequencies include 100MHz, 200MHz, 400MHz, 800MHz, and 1GHz, with different frequencies suitable for detection at different depths.

[0030] The algorithm flow field 203 records the algorithms used in each stage of preprocessing, feature extraction, and disease identification. The preprocessing stage includes algorithms such as gain, noise reduction, background removal, and filtering; the feature extraction stage includes time-domain feature extraction, frequency-domain feature extraction, and time-frequency feature extraction; and the disease identification stage includes deep learning algorithms such as convolutional neural networks, recurrent neural networks, and Transformer models.

[0031] The parameter configuration field 204 is used to record the key parameters and parameter ranges of each algorithm. Specific parameter ranges include: the gain coefficient range for the gain algorithm is 10~50dB / μs, with a preferred value of 20~30dB / μs; the time constant range for automatic gain control is 0.1~1.0μs; and the exponent coefficient range for exponential gain is 0.5~2.0, with a preferred value of 1.0~1.5. The parameters for wavelet threshold denoising include wavelet basis function type, number of decomposition levels, threshold selection method, etc.; the number of intrinsic mode functions retained in empirical mode decomposition denoising ranges from 25. In the background removal algorithm, the window size for moving average background removal ranges from 5 to 50 sampling points, with an optimal value of 10 to 20 sampling points. The background update rate ranges from 0.01 to 0.5, with an optimal value of 0.05 to 0.2. In the filtering algorithm, the cutoff frequency range of the low-pass filter is 100~800MHz, with a preferred value of 200~500MHz, and the center frequency range of the band-pass filter is 200~1500MHz, with a preferred value of 400~1000MHz. In deep learning-based disease identification algorithms, the number of layers in the convolutional neural network ranges from 8 to 32, with an optimal value of 12 to 20 layers, and the learning rate ranges from 1e. -5 ~1e -3 The preferred value is 1e -4 ~5e -4 The batch size range is 16~128, with an optimal value of 32~64.

[0032] The "Successful Cases" field (205) records the scene characteristics and recognition performance metrics of historical successful application cases. Scene characteristics include road structure type, ground-penetrating radar frequency, detection depth range, and roadbed material type; recognition performance metrics include accuracy, recall, F1 score, and distribution of defect types.

[0033] The knowledge base draws from various sources, including summaries of historical project experience, expert compilation, and mining of literature and technical documents. Historical project experience summaries extract successful algorithm flows and parameter configurations from completed projects; expert compilation involves domain experts summarizing and organizing experience; and literature and technical document mining extracts relevant knowledge from academic papers, technical standards, and user manuals.

[0034] Step S2: Obtain scene feature information The scene feature information of the current detection scenario includes at least one or more of the following: road structure type, ground penetrating radar frequency, detection depth range, and roadbed material type.

[0035] The road structure type can be obtained from road design drawings or on-site surveys; the ground-penetrating radar frequency is determined by the ground-penetrating radar equipment used; the detection depth range is determined according to the detection purpose; and the roadbed material type can be obtained through sampling tests or historical data.

[0036] For example, the scene feature information of the current detection scenario is: the road structure type is "asphalt concrete pavement + graded crushed stone base course + natural sand and gravel subbase course", the ground penetrating radar frequency is 400MHz, the detection depth range is 0.5~1.5m, and the subgrade material type is "graded crushed stone".

[0037] Step S3: RAG retrieval algorithm flow and parameters like Figure 3 As shown, the RAG retrieval method employs a hybrid retrieval strategy, including similarity score retrieval 301, scene feature precise matching retrieval 302, and historical success case weight retrieval 303.

[0038] Before retrieval, the knowledge base and scene feature information need to be vectorized. The knowledge base vectorization process is as follows: A text embedding model is used to convert text descriptions in the knowledge base into vector representations. Pre-trained models such as BERT, RoBERTa, and Sentence-BERT can be used for this purpose. For each record in the knowledge base, its text description (including road structure type, ground-penetrating radar frequency, algorithm flow, parameter configuration, etc.) is input into the text embedding model to obtain the corresponding vector representation.

[0039] Construct a vector index structure for fast retrieval of similar knowledge items. The vector index structure can utilize efficient vector retrieval tools such as FAISS (Facebook AI Similarity Search), Annoy, and HNSW.

[0040] Based on vectorization, calculate the overall similarity: Score _total = α×Score _sim + β×Score _match + γ×Weight _history , Among them, Score _sim The Score is a retrieval score based on vector similarity (i.e., cosine similarity). _matchWeight is the score for accurate matching of scene features (i.e., the degree of matching score). _history The weights for historical success cases are α, β, and γ, which are weighting coefficients, and α + β + γ = 1. The weighting coefficients can be adjusted according to the actual situation; for example, α = 0.4, β = 0.3, and γ = 0.3 can be set.

[0041] Score _sim The cosine similarity method is used. The current scene feature information is converted into a vector representation, and its cosine similarity with the vectors of each entry in the knowledge base is calculated as the score. _sim The value of .

[0042] Score _match A score is awarded based on the exact match between the road structure type and ground-penetrating radar frequency in the scene features and the records in the knowledge base. Specifically, a perfect match for the road structure type is scored as 1.0, and a partial match (different road grades) is scored as 0.6–0.8; a perfect match for the ground-penetrating radar frequency is scored as 1.0, and a partial match (frequency difference ±20%) is scored as 0.7–0.9; no match is scored as 0. The overall scene matching score is calculated by weighting the matching scores of each element. For example, if both the road structure type and the ground-penetrating radar frequency are perfectly matched, the score is 1.0. _match =1.0; if the road structure type partially matches and the ground-penetrating radar frequency partially matches, then the Score is 1.0. _match Take a weighted average (e.g., 0.7); if either the road structure type or the ground-penetrating radar frequency is a perfect match and the other is a partial match, then set the matching score of the perfectly matched side to 1, obtain the matching score of the partially matched side, and then take a weighted average as the matching degree score. _match ;

[0043] Weight _history Calculated based on the recognition performance metrics of historical successful cases. The higher the recognition performance metric, the greater the weight. For example, the weight can be calculated based on the weighted average of precision, recall, and F1 score, or based on the weighted average of the number of successful cases and the number of applications.

[0044] The principles for determining the weighting coefficients α, β, and γ include: for scenarios with abundant historical cases, the γ value can be increased, for example, by setting α=0.3, β=0.2, and γ=0.5, to fully leverage the guiding role of successful historical cases; for scenarios with high requirements for precise matching, the β value can be increased, for example, by setting α=0.3, β=0.5, and γ=0.2, to enhance the weight of precise matching retrieval based on scenario features; for general application scenarios, a balanced weighting coefficient can be set, for example, α=0.4, β=0.3, and γ=0.3.

[0045] During retrieval, the overall similarity score is used. _total The entries in the knowledge base are sorted, and the records with the highest similarity are selected as the search results. These entries contain the previously verified and complete algorithm flow and parameter configuration, thus obtaining the matching algorithm flow and configuration parameters.

[0046] This embodiment uses RAG retrieval to identify experience entries that are "most matching in terms of scenario, most similar in terms of process, and most reliable in terms of effect" from historical successful cases, thus providing a strong foundation for improving the recognition accuracy.

[0047] Step S4: Generate 2-3 algorithm flow schemes Based on the search results, 2-3 algorithm flow schemes (i.e., complete recognition algorithm schemes including configuration parameters) are generated using a large language model. Each algorithm flow scheme includes a complete process and corresponding parameters for gain, noise reduction, background removal, filtering, feature fusion, and deep learning-based disease recognition.

[0048] The process of the large language model generation algorithm includes: (1) Convert the retrieved knowledge base entries into structured prompt words. The prompt words contain information such as the current scene features, the content of the retrieved knowledge base entries, and the requirements for generating the task.

[0049] (2) Understand scene features and retrieval results through a large language model, and generate algorithm flow and parameter configuration. Large language models can be GPT-4, Claude, LLaMA, etc.

[0050] (3) Perform consistency checks on the generated algorithm flow scheme to ensure compatibility between algorithms at each stage. For example, check whether the parameters of the gain algorithm and the noise reduction algorithm conflict, and check whether the parameter range of the filtering algorithm is reasonable.

[0051] (4) Sort the algorithm flow schemes according to the comprehensive similarity, and the schemes with higher similarity are ranked first.

[0052] For example, the first generated algorithm flow scheme is as follows: - Gain: Exponential gain, with an exponent coefficient of 1.2 (the exponent coefficient ranges from 0.5). 2.0，优选值为1.01.5) - Noise Reduction: Wavelet thresholding noise reduction, wavelet basis function is db4, decomposition layer is 5, threshold uses a general threshold. - Background Removal: Moving average background removal, window size is 15 sampling points (window size range is 550 sampling points, preferred value is 1020 sampling points), background update rate is 0.1 (background update rate range is 0.01~0.5, preferred value is 0.05~0.2). - Filtering: Bandpass filtering, center frequency is 500MHz (center frequency range is 200~1500MHz, preferred value is 400~1000MHz), bandwidth is 300MHz, filter order is 4. - Feature Fusion: Time domain features (amplitude, energy) + frequency domain features (power spectrum, main frequency) + time-frequency features (wavelet coefficients). - Disease Identification: Convolutional Neural Network (ResNet-18, 18 layers, layer range is 832 layers, preferred value is 1220 layers), learning rate is 2e. -4 (Learning rate range is 1e) -5 ~1e -3 The preferred value is 1e -4 ~5e -4 The batch size is 64 (the optimal batch size range is 32-64, with a maximum of 16-128). The number of training rounds is 100. The second algorithm flow is as follows: - Gain: Time-varying gain, gain coefficient of 25dB / μs (gain coefficient range of 1050dB / μs, preferred value of 2030dB / μs) - Noise reduction: Empirical mode decomposition noise reduction, retaining the first 3 intrinsic mode functions (the number of intrinsic mode functions retained ranges from 2 to 5) - Background removal: Mean background removal, sliding window length of 30 - Filtering: Low-pass filtering, cutoff frequency of 400MHz (cutoff frequency range of 100~800MHz, preferred value of 200~500MHz), filter order of 4 - Feature fusion: Time domain features (peak value, peak value ratio) + frequency domain features (spectral centroid, bandwidth) - Disease identification: Transformer model (Vision Transformer Patch 16, 12 layers, layer range of 8~32 layers, preferred value of 12~20 layers), learning rate of 5e -4 (Learning rate range is 1e) -5 ~1e -3 The preferred value is 1e -4 ~5e -4 The batch size is 32 (the batch size range is 16~128, with an optimal value of 32~64), and the number of training rounds is 200.

[0053] Based on the search results, this embodiment generates a final solution that can be directly implemented for the current task, enabling the identification of different tasks.

[0054] Furthermore, user feedback was also considered in this step when designing the large language model generation algorithm. Specifically: The system records the algorithm flow schemes selected by users in each recognition task and their recognition results, forming a user preference model. When the large language model generates a new algorithm flow scheme, the user preference model is used as input to generate a scheme that is more in line with user habits.

[0055] For example, if a user frequently chooses to use wavelet thresholding, then when generating a new algorithm flow scheme, wavelet thresholding and its related parameters should be recommended first.

[0056] By taking user feedback into account, the personalization and applicability of algorithm flow schemes can be improved, further enhancing user satisfaction.

[0057] Step S5: User selects final solution like Figure 4 As shown, the algorithm flow schemes and matching scores are displayed to the user. Each algorithm flow scheme shows its included algorithm and parameters, as well as its matching score. The user selects the final algorithm flow scheme based on the matching score and their own experience.

[0058] For example, the interface displays: Scheme 1 (matching degree 0.92): - Gain: Time-varying gain, gain coefficient 20dB / μs - Noise reduction: Wavelet thresholding, wavelet basis db4, decomposition level 5 - … (other algorithms and parameters) Option 2 (matching degree 0.85): - Gain: Automatic gain control, time constant 0.5μs - Noise reduction: Empirical mode decomposition noise reduction, retaining the first 3 intrinsic mode functions - … (other algorithms and parameters) The user selected Option 1 as the final algorithm flow scheme.

[0059] Step S6: Perform the recognition task Based on the final algorithm flow scheme selected by the user, the collected ground-penetrating radar data is processed to identify hidden underground defects in roads.

[0060] First, the raw ground-penetrating radar data undergoes gain processing to compensate for signal energy attenuation during propagation. Then, noise reduction is performed to remove random noise. Next, background removal is carried out to eliminate direct waves and system noise. Following this, filtering is applied to remove interfering waves. Then, feature fusion is performed to extract various types of features. Finally, a deep learning model is used for disease identification, outputting information such as disease type, location, and size.

[0061] Types of defects may include cavities, loosening, cracks, and voids. The location of defects can be represented by relative positions (such as distance from the starting point or depth) or absolute coordinates. The size of defects can be represented by parameters such as length, width, area, and volume.

[0062] In addition, after step S6 is executed, the knowledge base is also updated: Specifically, the successful algorithm process and parameter configuration from this recognition task are added to the successful case field of the knowledge base, recording scene feature information and recognition performance metrics. If the accuracy, recall, F1 score, and other metrics of this recognition task are higher than those of historical cases with similar scenes in the knowledge base, the weight of this successful case is increased.

[0063] The knowledge base can be updated either periodically or in real-time. Periodic updates can be performed on a daily, weekly, or monthly basis; real-time updates can be performed immediately after each recognition task is completed.

[0064] By updating the knowledge base, new successful experiences can be continuously accumulated, leading to continuous growth and optimization of knowledge, and improving the system's adaptability and accuracy.

[0065] Step S7: Comparison and Verification like Figure 5 As shown, the recognition results are compared and verified. The comparison and verification includes the following methods: (1) Cross-validation of multiple algorithm flow schemes. When multiple algorithm flow schemes are generated, each scheme is executed separately, and the disease identification results of each scheme are compared. The disease areas that are consistently identified are counted as high-confidence results. For example, if both Scheme 1 and Scheme 2 identify a cavity at a certain location, the confidence of the cavity is high; if only Scheme 1 identifies a disease at a certain location, the confidence of the disease is low.

[0066] (2) Comparison with manually labeled results. Obtain manually labeled disease samples, calculate the recognition accuracy, recall and F1 score of each algorithm scheme, select the optimal scheme or evaluate the performance of this method.

[0067] (3) Comparison with traditional methods. The identification results of this method are compared with those of traditional threshold methods and spectrum analysis methods to verify the effectiveness of this method. For example, the differences in the number, location, and type of diseases identified by the two methods can be compared.

[0068] Based on the above three results, experts judge whether the verification is successful based on experience. If it is, the final recognition result (the above cross-validation result) is output. If not, the parameters of each complete recognition algorithm scheme are adjusted and the recognition result is regenerated.

[0069] By comparing and verifying, the reliability and accuracy of the recognition results can be further improved, and feedback can be provided for updating the knowledge base.

[0070] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0071] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0072] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0073] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0074] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0075] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0076] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A ground penetrating radar road underground hidden disease identification method combining LLM and RAG, characterized in that, Includes the following steps: Obtain scene feature information of the current detection scene; Based on the scene feature information, the RAG retrieval method is used to retrieve multiple matching algorithm processes and configuration parameters from a pre-built knowledge base; Based on the aforementioned multiple matching algorithm processes and configuration parameters, and combined with the aforementioned scene feature information and recognition task requirements, a large language model is used to generate multiple complete recognition algorithm schemes. Obtain the complete recognition algorithm scheme selected by the user from the multiple complete recognition algorithm schemes, and perform road underground hidden defects recognition in the current detection scenario based on the user's selection result to obtain the recognition result.

2. The ground penetrating radar road underground hidden disease identification method combining LLM and RAG according to claim 1, characterized in that, The scene feature information includes one or more of the following: road structure type, ground penetrating radar frequency, detection depth range, and roadbed material type.

3. The ground penetrating radar road underground hidden disease identification method combining LLM and RAG according to claim 1, characterized in that, The knowledge base uses a structured storage format and includes the following data fields: Road structure type field: Records the type information of pavement structure, base course structure, and subbase course structure; Ground penetrating radar frequency field: Records the center frequency of the ground penetrating radar antenna; Algorithm flow field: Records the algorithms selected in each stage of preprocessing, feature extraction, and disease identification; Parameter configuration field: Records the key parameters and parameter ranges of each algorithm; Success Case Field: Records the scenario characteristics and recognition performance metrics of historical successful application cases.

4. The ground penetrating radar road underground hidden disease identification method combining LLM and RAG according to claim 3, characterized in that, For the algorithm flow fields, the algorithms in the preprocessing stage include gain, noise reduction, background removal, and filtering algorithms; the algorithms in the feature extraction stage include time-domain feature extraction, frequency-domain feature extraction, and time-frequency feature extraction algorithms; and the algorithms in the disease identification stage are deep learning algorithms.

5. The ground penetrating radar road underground hidden disease identification method combining LLM and RAG according to claim 4, characterized in that, For the parameter configuration field In the gain algorithm, the range of the gain coefficient, the range of the time constant of the automatic gain control, and the range of the exponential coefficient of the exponential gain are recorded. In the denoising algorithm, the parameters for wavelet threshold denoising and the range of the number of intrinsic mode functions retained for empirical mode decomposition denoising are recorded. In the background removal algorithm, the range of window size and the range of background update rate for the moving average background removal are recorded. In the filtering algorithm, the cutoff frequency range of the low-pass filter and the center frequency range of the band-pass filter are recorded. In deep learning algorithms, the range of network layers, learning rate, and batch size are recorded for each deep learning algorithm.

6. The ground penetrating radar road underground hidden disease identification method combining LLM and RAG according to claim 1, characterized in that, The algorithm process and configuration parameters for retrieving multiple matches include: The knowledge base is vectorized using a text embedding model, and a vector index structure is constructed to obtain the vectorized knowledge base; The scene feature information is vectorized to obtain vectorized scene feature information; calculating a cosine similarity Score between the vectorized scene feature information and each entry vector in the vectorized knowledge base _sim ; The road structure type and ground penetrating radar frequency in the vectorized scene feature information are matched with the road structure type and ground penetrating radar frequency in the vectorized knowledge base respectively, if the road structure type and ground penetrating radar frequency are completely matched, the matching degree score Score _match =1, if the road structure type and ground penetrating radar frequency are partially matched, the respective matching scores are obtained first, and then the weighted average value is taken as the matching degree score Score _match , if the road structure type and ground penetrating radar frequency are not matched, the matching degree score Score _match =0, if either of the road structure type and ground penetrating radar frequency is completely matched and the other is partially matched, the matching score of the completely matched one is set to 1, and the matching score of the partially matched one is obtained, and then the weighted average value is taken as the matching degree score Score _match ; calculate an identification effect index of the historical successful application cases in the vectorized knowledge base, and calculate a weight of the historical successful application cases according to the identification effect index, wherein the weight Weight of the historical successful application cases is obtained by taking a weighted average value of each identification effect index _history or the weight of the historical successful application cases is obtained by taking a weighted average value of the number and the application times of the historical successful application cases. The cosine similarity Score _sim , the matching degree score Score _match , and the weight Weight _history of the historical successful application cases are weighted and calculated to obtain a comprehensive similarity Score _total , which is represented as: Score _total = α×Score _sim + β×Score _match + γ×Weight _history , α+β+γ=1, In the formula, α, β, and γ are weighting coefficients; According to the comprehensive similarity Score _total Obtain from the knowledge base the matching algorithm processes and configuration parameters that are the top k most matching to the scene feature information.

7. The ground penetrating radar road underground hidden disease identification method combining LLM and RAG according to claim 1, characterized in that, The steps for generating the complete recognition algorithm scheme include: The multiple matching algorithm processes and configuration parameters, scene feature information, and recognition task requirements are converted into structured prompt words; The structured prompt words are input into a large language model to generate multiple complete initial recognition algorithm schemes. The consistency of the multiple complete initial recognition algorithm schemes is checked, and the algorithms are ranked using comprehensive similarity to obtain the final multiple complete recognition algorithm schemes.

8. The ground penetrating radar road underground hidden disease identification method combining LLM and RAG according to claim 1, characterized in that, Also includes: When two or more complete recognition algorithm schemes are selected, all the output recognition results are compared and verified. The specific steps include: Cross-validate all identification results and count the diseased areas that are consistently identified as high-confidence results; Obtain manually labeled disease samples and calculate the recognition performance index of each complete recognition algorithm scheme; The identification result is compared with the identification result output by the traditional identification method to obtain the comparison result; Based on the high confidence results, recognition performance indicators, and comparison results, determine whether the verification is successful. If so, output the final recognition result; otherwise, adjust the parameters of each complete recognition algorithm scheme and regenerate the recognition result.

9. The ground penetrating radar road underground hidden disease identification method combining LLM and RAG according to claim 1, characterized in that, It also includes updating the knowledge base, specifically through the following steps: In a manner that involves periodic or real-time updates, based on the recognition results, the complete recognition algorithm scheme that successfully completes the recognition task is added to the success case field of the knowledge base.

10. The ground penetrating radar road underground hidden disease identification method combining LLM and RAG according to claim 1, characterized in that, Also includes: When generating multiple complete recognition algorithm schemes, user feedback is taken into consideration, specifically including... the following: Obtain the complete recognition algorithm schemes and their recognition effects selected by users in each recognition task to form a user preference model; The user preference model is also used as input to the large language model to generate multiple complete recognition algorithm schemes.

Images