Intelligent identification method and system for highway subgrade radar detection based on neural network

Abstract

This invention provides a neural network-based intelligent identification method and system for highway subgrade radar detection, belonging to the field of highway quality inspection technology. The method includes: standardizing and enhancing the acquired detection image data to obtain a two-dimensional radar profile; based on the two-dimensional radar profile, identifying defects using a target detection model, and aggregating the identified defect information to obtain a structured feature vector containing defect type, quantity, size, and distribution density; and evaluating the performance of the detection unit based on the structured feature vector using a decision engine of a digital standard knowledge base. The advantage of the method and system provided by this invention is that it elevates defect identification from a single step of target detection to a systematic process of automatic highway quality grading, achieving rapid identification and location of defects in radar images, and outputting quality levels and scores according to industry standards through a built-in digital standard knowledge base.

Description

Technical Field

[0001] This invention relates to the field of highway quality inspection technology, and in particular to a method and system for intelligent identification of highway subgrade radar detection based on neural networks. Background Technology

[0002] As the foundation of the road surface, the internal quality of the highway subgrade directly determines the overall load-bearing capacity, service life, and driving safety of the road. The quality grade of the highway subgrade is closely related to common subgrade defects such as compaction degree, layer thickness, hidden voids, delamination, and weak interlayers. Traditional subgrade quality inspection methods, such as the sand cone method and the ring cutter method, have drawbacks such as being point-based, discrete, destructive, and inefficient, making it difficult to comprehensively and quickly evaluate the condition of a large area of ​​subgrade. Non-contact subgrade quality inspection methods, such as falling weight deflectometers, ground-penetrating radar, macroscopic structural depth, the International Roughness Index, and surface wave spectrum analysis, generally suffer from drawbacks such as over-reliance on engineer experience, high difficulty in designing manual feature generators, and computationally intensive candidate box selection algorithms like sliding windows, resulting in numerous redundant operations and failing to effectively detect subgrade defects. Therefore, it is necessary to establish an intelligent identification method and system for highway subgrade quality radar detection based on neural network algorithms to efficiently and accurately identify highway subgrade quality.

[0003] The following are several similar or related implementation schemes in the existing technology: 1) A method and system for highway quality inspection based on unmanned aerial vehicles (UAVs) This invention relates to the field of highway quality inspection technology. The method includes: acquiring three-dimensional coordinate data of a target highway, establishing a benchmark three-dimensional model of the highway, and dividing the target highway into multiple highway sections; setting flight parameters and flight paths for a drone; equipping the drone with multiple inspection devices to collect highway section data; determining a comprehensive crack index based on the optical images; constructing a digital elevation model of the road surface based on the three-dimensional point cloud data; determining a road surface smoothness index based on the digital elevation model of the road surface; determining a material degradation index based on the visible light band reflection energy and the infrared band reflection energy; and identifying the highway sections requiring maintenance. According to this invention, highway quality can be comprehensively inspected based on cracks, smoothness, material degradation, etc., improving the comprehensiveness, accuracy, and safety of the inspection.

[0004] 2) A smart detection and early warning system for highways

[0005] The system is characterized by comprising: a fixed monitoring unit, which includes multiple fixed monitoring devices installed on fixed high poles along the highway; a mobile monitoring unit, which includes mobile monitoring devices installed on logistics vehicles operating on the highway; an information receiving unit, which receives video data monitored by the fixed and mobile monitoring devices in real time; and a central processing unit, which uses a deep learning algorithm and a high-precision image recognition algorithm to perform real-time identification and analysis of the video data to identify potential risks and violations affecting road traffic. After identifying potential risks and violations, the unit can accurately classify the urgency of the event and send it to the corresponding traffic police and joint operations personnel for processing according to the event classification.

[0006] 3) Blockchain-based highway quality inspection system and method

[0007] The system includes: a safety level classification module, a reference table of influencing factors for different parameters during highway smoothing, a module for calculating influencing factors during quality inspection, a module for determining the fluctuation range of influencing factors for highway traffic efficiency, a module for determining the core location of quality inspection, and a module for the variation curve of the maximum value of quality inspection. By dividing the data inspection locations corresponding to various types of compaction, moisture content, density, structural strength, and smoothness parameters of highways for different purposes, and using a longitudinal comparison method, the system ultimately matches the maximum values ​​of highway bearing capacity, subgrade parameters, crack width, drainage capacity, and seepage during quality inspection at the corresponding matching zoning locations. This pre-emptively repairs data inspection locations that may have abnormally collected data, preventing them from participating in subsequent data analysis and ensuring the accuracy of the final highway traffic quality inspection.

[0008] 4) Data retrieval methods for highway quality inspection

[0009] The method includes: collecting raw data sequences, driving speed data sequences, trend term data sequences, and residual term data sequences; dividing each of these into several data segments; obtaining the trend stability parameter for each data segment in the trend term data sequence; and combining the data from each data segment in the driving speed data sequence and the residual term data sequence to obtain the road quality judgment degree and error deviation coefficient corresponding to each data segment in the raw data sequence. This yields the encoding weight for each data segment in the raw data sequence, which is then used for data retrieval within the raw data sequence. This invention improves the data retrieval efficiency for highway quality inspection by segmenting the data sequence and adaptively assigning encoding weights to each data segment for data encoding and constructing an index structure. Summary of the Invention

[0010] The embodiments of the present invention provide a method and system for intelligent identification of highway roadbed radar detection based on neural networks, which is used to solve the technical problems existing in the prior art.

[0011] To achieve the above objectives, the present invention adopts the following technical solution.

[0012] A neural network-based intelligent identification method for radar detection of highway subgrade includes: S1. Standardize and enhance the collected detection image data to obtain a two-dimensional radar profile. S2. Based on the two-dimensional radar profile, the disease information is identified through the target detection model, and the identified disease information is aggregated to obtain a structured feature vector of the disease containing the disease type, quantity, size and distribution density. S3. Based on the structured feature vectors of defects, the performance of highway subgrade and detection units is graded and evaluated through the decision engine of the digital standard knowledge base; The results of the graded evaluation of the performance of the highway subgrade and the detection unit are used for risk assessment of the target highway section and performance evaluation of the detection unit.

[0013] Preferably, the standardization process in step S1 includes sequentially performing time null correction, DC drift removal, background denoising, bandpass filtering, and gain restoration on the detected image data; The specific process of image enhancement in step S1 includes contrast stretching, directional filtering, and format normalization output operations on the standardized detection image data to obtain a two-dimensional radar profile.

[0014] Preferably, the target detection model in step S2 includes a backbone network, a feature network, a detection head, and a non-maximum suppression module arranged sequentially along the data flow direction; The backbone network has stacked convolutional layers, batch normalization layers and SiLU activation function layers, which can extract multi-level features from aggregate texture to semantic information based on two-dimensional radar profiles through stepwise downsampling operations; The feature network is used to integrate the multi-level features output by the backbone network to obtain multiple sets of enhanced feature maps; The detection head is used to predict the enhanced feature map and output class probability information and bounding box coordinate information; The nonmaximum suppression module is used to perform nonmaximum suppression on the category probability information and bounding box coordinate information output by the detection head to obtain a structured feature vector of the disease that includes the disease type, quantity, size and distribution density.

[0015] Preferably, step S3 specifically includes: Based on the structured feature vectors of road defects, the highway quality grade score is obtained through calculation. Based on the structured feature vector of the disease, the index score of the probability of risk occurrence is obtained by calculation, and a two-level index system for evaluating the probability of risk occurrence is established based on the index score of the probability of risk occurrence. Based on the structured feature vector of the disease, the index score of the risk consequence is obtained by calculation, and a first-level index system for risk consequence evaluation is established based on the index score of the risk consequence. Based on a two-level indicator system for assessing the probability of risk occurrence and a one-level indicator system for assessing the consequences of risk, a risk occurrence probability and risk consequence level system is established. The performance of the detection unit is evaluated based on the probability of risk occurrence and the risk consequence level system, and a risk assessment is also conducted on the target highway section.

[0016] Secondly, the present invention provides a roadbed radar detection intelligent identification system based on neural networks, including a hardware acquisition layer, a data preprocessing layer, a core algorithm engine layer, and an integrated application interaction layer; The hardware acquisition layer is used to acquire detection image data; The data preprocessing layer is used to standardize and enhance the acquired detection image data to obtain a two-dimensional radar profile. The core algorithm engine layer is used for: Based on two-dimensional radar profile, disease information is identified through a target detection model, and the identified disease information is aggregated to obtain a structured feature vector of diseases containing disease type, quantity, size and distribution density. Based on the structured feature vectors of diseases, the performance of the detection unit is evaluated through a decision engine of a digital standardized knowledge base; The integrated approach is applied to the interaction layer to obtain a detection report based on the structured feature vectors of the disease and the performance evaluation results of the detection unit.

[0017] Preferably, the hardware acquisition layer specifically includes: Ground-penetrating radar, deployed on a testing vehicle, is used to emit radar signals to detect roadbed defects. The positioning unit, deployed on the detection vehicle, is used to provide radar vehicle location information; A synchronous control system, deployed on the inspection vehicle, is used to adjust the vehicle speed and control the radar; The process of acquiring detection image data at the hardware acquisition layer includes: The synchronous control system controls the testing vehicle to travel along the predetermined testing line at a constant speed. The ground-penetrating radar echo signals are continuously collected at preset intervals; simultaneously, based on the location information and echo signals obtained by the positioning unit, the detection image data with location markers is output, which includes images of roadbed loosening and underground cavities.

[0018] Preferably, the data preprocessing layer specifically includes performing the following processing steps on the raw radar data sequentially: Time zero-point correction: eliminates system delay; DC drift removal: Removes the offset of the signal baseline; Background denoising: Subtract the average waveform of each channel and suppress horizontal in-phase axis noise; Bandpass filtering: Using a Butterworth bandpass filter to retain the effective frequency band and suppress high and low frequency noise; Gain recovery: Apply automatic gain control or distance gain compensation to make deep signals clearly visible; To optimize the neural network input, the preprocessed radar profile image is subjected to the following processes: contrast stretching, directional filtering, and format normalization output. The dynamic range of image grayscale values ​​is linearly expanded to 0-255 to enhance feature contrast. Apply the Laplace operator or directional filter to highlight the diffraction hyperbola or phase axis discontinuity features associated with the disease; The processed data is uniformly formatted into a two-dimensional grayscale image matrix with a regular size and an image width that varies according to the length of the survey line, and is accompanied by corresponding mileage-depth coordinate information and stored.

[0019] Preferably, the detection model of the core algorithm engine layer includes a backbone network, a feature network, a detection head, and a non-maximum suppression module arranged sequentially along the data flow direction; The backbone network has stacked convolutional layers, batch normalization layers and SiLU activation function layers, which can extract multi-level features from aggregate texture to semantic information based on two-dimensional radar profiles through stepwise downsampling operations; The feature network is used to integrate the multi-level features output by the backbone network to obtain multiple sets of enhanced feature maps; The detection head is used to predict the enhanced feature map and output class probability information and bounding box coordinate information; The nonmaximum suppression module is used to perform nonmaximum suppression on the category probability information and bounding box coordinate information output by the detection head to obtain a structured feature vector of the disease containing the disease type, quantity, size and distribution density; The process of building the detection model includes: Radar profile images were collected as a dataset. Annotation tools were used to annotate the disease target areas in the images in the form of rectangular bounding boxes and to specify the disease category. The disease categories include: cavitation, voiding, loosening, water-rich, as well as non-disease objects such as various metal pipelines and various non-metal pipelines. Set up a target detection platform for the YOLOv11 object detection algorithm in the operating system; specifically, this includes downloading and configuring a virtual environment; importing the YOLOv11 source code using a programming tool and adding the corresponding interpreter; Set evaluation metrics for the training accuracy of the detection model; The core algorithm engine layer's working process specifically includes: Based on the structured feature vectors of road defects, the highway quality grade score is obtained through calculation. Based on the structured feature vector of the disease, the index score of the probability of risk occurrence is obtained by calculation, and a two-level index system for evaluating the probability of risk occurrence is established based on the index score of the probability of risk occurrence. Based on the structured feature vector of the disease, the index score of the risk consequence is obtained by calculation, and a first-level index system for risk consequence evaluation is established based on the index score of the risk consequence. Based on a two-level indicator system for assessing the probability of risk occurrence and a one-level indicator system for assessing the consequences of risk, a risk occurrence probability and risk consequence level system is established. The performance of the detection unit is evaluated based on the probability of risk occurrence and the risk consequence level system, and a risk assessment is also conducted on the target highway section.

[0020] As can be seen from the technical solutions provided by the embodiments of the present invention above, the present invention provides a method and system for intelligent identification of roadbed radar detection based on neural networks. The method includes: standardizing and enhancing the collected detection image data to obtain a two-dimensional radar profile; based on the two-dimensional radar profile, identifying disease information through a target detection model, and aggregating the identified disease information to obtain a structured feature vector containing disease type, quantity, size, and distribution density; evaluating the performance of the detection unit based on the structured feature vector through a decision engine of a digital standard knowledge base; and obtaining a detection report based on the structured feature vector and the performance evaluation results of the detection unit. The advantage of the method and system provided by the present invention is that it elevates disease identification from a single step of target detection to a systematic process of automatic road quality grading. Existing solutions cannot accurately quantify disease morphology and area, and are not directly related to engineering quality standards. The present invention uses the YOLOv11 target detection model to achieve rapid identification and location of diseases in radar images, and then automatically outputs quality levels and scores according to industry standards through a built-in digital standard knowledge base. This method overcomes the inefficiency of traditional manual evaluation based on standards, and can effectively improve the efficiency and practicality of highway quality inspection and evaluation.

[0021] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and will become apparent from the description or may be learned by practice of the invention. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 The flowchart of the intelligent identification method for highway subgrade radar detection based on neural networks provided by the present invention is shown below. Figure 2 This is a flowchart illustrating the execution process of a preferred embodiment of the intelligent identification method for highway subgrade radar detection based on neural networks provided by the present invention. Figure 3 A schematic diagram of the driving route of the radar detection vehicle in the detection area for the intelligent identification method of roadbed radar detection based on neural network provided by the present invention, wherein the arrows represent three-dimensional survey lines and driving detection directions; Figure 4 A schematic diagram of the hierarchical architecture of the intelligent identification system for highway subgrade radar detection based on neural networks provided by the present invention; Figure 5 This is a schematic diagram illustrating the architecture and working process of a preferred embodiment of the intelligent identification system for highway subgrade radar detection based on neural networks provided by the present invention. Detailed Implementation

[0024] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0025] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or couplings. The term “and / or” as used herein includes any and all combinations of one or more of the associated listed items.

[0026] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as herein.

[0027] To facilitate understanding of the embodiments of the present invention, the following will provide further explanation and description with reference to the accompanying drawings and several specific embodiments. These embodiments do not constitute a limitation on the embodiments of the present invention.

[0028] This invention provides a neural network-based intelligent identification method and system for highway subgrade radar detection, which addresses the following technical problems existing in the prior art: Ground-penetrating radar (GPR) technology, as an efficient, continuous, and non-destructive testing method, has been widely used in roadbed quality inspection. It generates radar profiles by emitting high-frequency electromagnetic waves and receiving reflected signals from interfaces of materials with different dielectric constants underground. However, current GPR technology faces the following main shortcomings in roadbed quality assessment: Reliance on human experience in interpretation: The interpretation of radar image profiles is highly dependent on the professional experience of the inspectors, which is highly subjective. Different people may have different results in identifying the same abnormal feature, and there is a lack of objective and unified quantitative standards.

[0029] Insufficient in-depth information mining: Traditional signal processing and image analysis methods, such as filtering, offsetting, and energy analysis, are mostly focused on signal enhancement and anomaly localization. From massive and complex radar waveform data, it is difficult to automatically and accurately identify the characteristics of different types of defects, and the defect feature identification results have little correlation with the quality classification of highway subgrade.

[0030] Low level of automation and intelligence: In the existing processing flow, the automation level of key links such as target identification, disease classification and parameter inversion is limited, making it difficult to achieve rapid, batch and intelligent analysis and quality classification of large-scale detection data, and failing to meet the needs of real-time, efficient and intelligent diagnosis in modern highway construction and maintenance management.

[0031] In recent years, artificial intelligence technologies such as deep learning have demonstrated strong advantages in image recognition and signal processing. A few studies have attempted to apply neural networks to ground-penetrating radar target identification, but these have mostly focused on coarse classification of single, obvious targets such as pipelines and cavities. A comprehensive intelligent grading and evaluation system, encompassing raw radar signals and various types of hidden defects and roadbed quality indicators, has not yet been established. Existing methods generally suffer from weak model generalization ability, insensitivity to complex roadbed structures and subtle anomalies, and poor interpretability.

[0032] Therefore, there is an urgent need to invent a method and system that can deeply integrate ground-penetrating radar detection data with the characteristics of highway subgrade engineering to achieve intelligent classification and identification of subgrade quality.

[0033] See Figure 1 and 2 This invention provides a method for intelligent identification of highway subgrade radar detection based on neural networks, comprising the following steps: S1. Standardize and enhance the collected detection image data to obtain a two-dimensional radar profile. S2. Based on the two-dimensional radar profile, the disease information is identified through the target detection model, and the identified disease information is aggregated to obtain a structured feature vector of the disease containing the disease type, quantity, size and distribution density. S3. Based on the structured feature vector of the disease, the performance of the detection unit (ground radar) is evaluated through the decision engine of the digital standardized knowledge base; S4. Obtain the detection report based on the structured feature vector of the disease and the performance evaluation results of the detection unit.

[0034] The results of the graded evaluation of the performance of the roadbed and the detection unit are used for risk assessment of the target road section and performance evaluation of the detection unit, and further provide data support for timely road repair and maintenance and performance upgrade of the detection unit.

[0035] In the preferred embodiment provided by the present invention, the specific process of each step is as follows.

[0036] Step 1: Synchronous Acquisition and Transmission of Multi-Source Data. This step can be performed through the data acquisition and transmission module.

[0037] Hardware deployment and integration: A light van was selected as the carrier vehicle, on which ground-penetrating radar, positioning unit and synchronous control system were integrated.

[0038] The ground-penetrating radar is used to emit radar signals to detect roadbed defects. The positioning unit is used to provide radar vehicle location information; The synchronous control system is used to adjust vehicle speed and control radar; Data acquisition process: The testing vehicle travels along the predetermined testing line at a constant speed of 20-30 km / h (e.g., Figure 3 (As shown). The host continuously acquires raw radar echo signals at a preset 2-centimeter channel spacing. Simultaneously, based on the location information and the detection information, it outputs detection image data with location markers. This detection image data includes images of roadbed loosening and underground cavities, and transmits the detection image data to the radar image processing module.

[0039] Preferably, the system records the three-dimensional spatial coordinates and attitude of each radar track, and the odometer encoder records the travel distance. All raw data streams are transmitted to the onboard computer in real time for temporary storage, and the labelimg software is used to annotate the location, size, and range features of defects in the images.

[0040] Step 2: Radar image standardization and enhancement. This step is performed by the radar image processing module.

[0041] Standardized preprocessing steps: On an industrial computer or subsequent server, use a Python preprocessing script to sequentially process the raw radar data: Time zero-point correction: eliminates system delay.

[0042] DC drift removal: Removes the offset of the signal baseline.

[0043] Background denoising: Subtract the average waveform of each channel to suppress horizontal in-phase axis noise.

[0044] Bandpass filtering: Butterworth bandpass filters are used to preserve the effective frequency band and suppress high and low frequency noise.

[0045] Gain recovery: Apply automatic gain control or distance gain compensation to make deep signals clearly visible.

[0046] Image enhancement processing: To optimize the neural network input, the preprocessed radar profile (B-Scan) is subjected to the following: contrast stretching, directional filtering, and format normalization output.

[0047] The contrast stretching: linearly expands the dynamic range of image grayscale values ​​to 0-255 to enhance feature contrast.

[0048] The specific directional filtering: applying the Laplacian operator or directional filter to highlight the diffraction hyperbola or phase axis misalignment characteristics associated with the disease.

[0049] The standardized output format is as follows: the processed data is uniformly formatted into a two-dimensional grayscale image matrix with a regular size and an image width that varies according to the length of the survey line, and the corresponding mileage-depth coordinate information is attached and stored for use by the intelligent disease identification and feature extraction module.

[0050] Step 3: Intelligent Disease Identification and Feature Vector Extraction Based on Deep Learning. This step can be performed by the configured intelligent disease identification and feature extraction module.

[0051] Neural network model construction and training: Model Architecture: The core of the model is the YOLOv11 (You Only Look Once version 11) object detection model. This model includes a backbone network for feature extraction, a feature network for fusing multi-scale information, and a detection head for class and bounding box prediction.

[0052] Backbone Network: Employs an improved CSPNet structure, specifically CSPDarknet, for extracting multi-level features from images. This network progressively downsamples and extracts feature maps ranging from geometric texture to semantic information through stacked convolutional layers, batch normalization layers, and the SiLU activation function.

[0053] Feature Network: The PANet structure is adopted to achieve multi-scale feature fusion. This network effectively integrates feature maps from different levels of the backbone network through a bidirectional fusion path from top to bottom and bottom to top, enhancing the model's ability to detect small-target defects (such as fine cracks and shallow cavities) and defects in complex backgrounds.

[0054] Detection Head: Employing a decoupled head structure, the classification and regression tasks are separated, with each task outputting class probabilities and bounding box coordinates. The detection head makes predictions on three different scale feature maps of the feature network, corresponding to large, medium, and small targets respectively, achieving accurate localization of diseases of different sizes.

[0055] Connection structure: The multi-layer feature maps output from the backbone network are input to the feature network. The feature network performs multi-scale fusion using the PANet structure to generate three sets of enhanced feature maps, which are then fed into the detection head for prediction. The results output from the detection head are subjected to non-maximum suppression (NMS) to obtain the final lesion bounding box, category, and confidence score.

[0056] Training data preparation: 6000 radar profile images were collected as the dataset. Using annotation tools, the target areas of defects in the images were labeled with rectangular bounding boxes, and their categories were specified, including: cavities, voids, loose areas, water-rich areas, and various non-defective objects such as metal and non-metal pipelines. The radar scanning characteristics of underground pipelines and various defects based on existing standards ("Technical Specification for Detection and Evaluation of Underground Defects in Urban Roads and Pipelines" DB11 / T 1399-2017) are shown in Table 1. The annotation files were saved in YOLO format.

[0057] Model Environment Setup: A PyCharm Community Edition 2024.1 environment was set up on a Windows system, and a YOLOv11 object detection algorithm object detection platform was built on this basis. Specific steps included downloading and installing an Anaconda virtual environment and configuring Python 3.8 to the appropriate version. Then, the PyTorch environment was configured as required. Next, PyCharm was installed using Anaconda, the YOLOv11 source code was imported, and the corresponding interpreter was added. After completing these steps, the corresponding terminal environment was opened, and YOLOv11 could be run on this platform.

[0058] Model Training: Mean Average Precision (mAP_0.5) is used as the evaluation metric for model training accuracy. mAP_0.5 is a variant of mean Average Precision (mAP), calculated at a specific Intersection over Union (IoU) threshold. In object detection, mAP_0.5 typically refers to the mAP calculated when the IoU threshold between the predicted bounding box and the ground truth bounding box is set to 0.5. IoU is the ratio of the area of ​​the overlapping region between the predicted and ground truth bounding boxes to the area of ​​their joint region. If the IoU value is greater than or equal to a given threshold (0.5 in this case), the detection is considered correct. mAP_0.5 is an important metric because it reflects the model's performance well in object detection tasks, especially when the IoU threshold is 0.5. This metric does not require particularly high accuracy in the localization of the detection boxes, thus focusing more on whether the model can correctly identify the target object. When analyzing the training trend graph of mAP_0.5, if the value of mAP_0.5 gradually increases as training progresses, it indicates that the detection performance of the model is improving.

[0059] In some preferred embodiments, based on a dataset containing 6000 ground-penetrating radar images of roadbed defects, the `labelimg` function is activated in PyCharm to annotate as many defect types as possible on each image. The annotated images are saved in YOLO format to a pre-defined path. After all annotations are completed, the dataset is divided into a training set and a validation set in an 8:2 ratio. Finally, image target detection experiments are conducted using YOLOv11. By continuously increasing the number of training rounds, the model is trained until the mean accuracy (mAP@0.5) for identifying specific defect types on an independent test set reaches over 90%.

[0060] Real-time disease identification and feature extraction: The standardized radar image output from step two is input into the pre-trained YOLOv11 model. The model performs forward inference and outputs a list of bounding boxes, where each bounding box contains its top-left corner coordinates, width and height, class label, and confidence score.

[0061] Feature Vector Generation: The system aggregates and calculates all identification results within the current detection segment to generate a structured disease feature vector. This vector includes the following fields: [Main disease type codes, total number of bounding boxes, average confidence level, maximum bounding box area (pixel estimate), disease distribution density along the mileage] This vector serves as a standardized data interface connecting image recognition and quality grading.

[0062] Table 1 Radar scanning characteristics of underground pipelines and various defects

[0063] Step 4: Intelligent comprehensive classification of subgrade quality based on the standard knowledge base. This step is performed by the subgrade quality comprehensive classification module.

[0064] Based on a standardized digital knowledge base: the evaluation clauses on highway subgrade quality in the "Technical Specification for Detection and Evaluation of Underground Defects in Urban Roads and Pipelines" DB11 / T 1399-2017 (Table 1) are transformed into computer-executable decision-making criteria.

[0065] In some preferred embodiments, the highway quality rating is calculated using the following formula: (1) In the formula, R is the risk value; P is the probability of the risk occurring; and C is the consequence of the risk.

[0066] Then, a disease evaluation index system is established. Risk evaluation includes the evaluation of the probability of risk occurrence and the evaluation of risk consequences. The evaluation of the probability of risk occurrence adopts a two-level index system, as shown in Table 2. The specific calculation methods of the index scores are shown in equations (2) to (4).

[0067] Table 2. Evaluation Indicators and Weights for the Probability of Underground Road Defects

[0068] Overburden-to-span ratio: The ratio of the thickness of the overburden layer to the horizontal span of the underground disease. r = h / l ,in h —Depth of the top of the disease. l —The maximum span of the disease.

[0069] (2) (3) (4) ( i =1,2,3..., m ; j =1,2,3... n ) In the formula, P i The score is the primary indicator of the probability of risk occurring; The score is a secondary indicator of the probability of risk occurring; W i Weights for primary indicators; Weights for secondary indicators; m The number of primary indicators; n This refers to the number of secondary indicators.

[0070] The derivation of the above formula is as follows: The theoretical basis comes from Table 2 of the standard (Evaluation Indicators and Weights for the Probability of Road Underground Defects). It uses a weighted summation method to convert the scores of secondary indicators such as the span ratio into primary indicator scores, and then further weights and sums them to obtain the total score. The derivation process is as follows: First, the secondary indicator score F is obtained by looking up the table based on the defect span ratio. 11 Substitute into equation (2) to calculate the coverage ratio score P1; if there are other secondary indicators (such as road grade, traffic flow, etc.), calculate the scores of each primary indicator through equation (3); finally, obtain the total probability score P by weighted summation through equation (2).

[0071] Derivation process: Table 2 of standard DB11 / T 1399-2017 (see the above-mentioned record in this manual) provides the "Evaluation Index and Weight of the Probability of Occurrence of Road Underground Defects", the structure of which is as follows: Primary Indicators: Basic Factors P 1 (weight) W 1 =1.0) Secondary indicator: Coverage ratio P 11 (weight) W 11 =1.0), its value is determined based on the span ratio. r = h / l The interval is obtained by looking up the table.

[0072] Coverage ratio r = h / l Buried from the top of the disease h and disease level span l The calculations show that, based on the correspondence between the coverage ratio and the score in Table 2, a piecewise function can be established:

[0073] Calculation of scores for primary indicators

[0074] Primary indicators P 1: Obtained by weighted summation of secondary indicators. Since there is only one secondary indicator in this specification, therefore:

[0075] Total probability score of risk occurrence P : Weighted summation of all primary indicators:

[0076] The above process is the specific content of formulas (2) to (4), where:

[0077] The risk consequences assessment adopts a first-level indicator system, as shown in Table 3, and the calculation method is shown in equations (5) to (6).

[0078] Table 3. Evaluation Indicators and Weights for Risk Consequences of Road Underground Defects (5) (6) (i=1,2,3...,m) In the formula, C i The score for the risk consequences assessment indicator; K 1 An additional coefficient is added to the risk consequence assessment. K 1 The values ​​are determined based on the type of underground disease: 1 for cavities (voids), 0.95 for severely water-rich areas, 0.9 for severely loose areas, 0.85 for moderately water-rich areas, 0.8 for moderately loose areas, and 0.7 for slightly loose areas.

[0079] The specific derivation process is as follows: Table 3 of standard DB11 / T 1399-2017 provides "Assessment Indicators and Weights for Risk Consequences of Road Underground Defects," and the score range for each indicator is determined by referring to the table based on specific conditions. In addition, additional coefficients are applied for different defect types. K 1 (See the explanation below Table 3) This is used to correct the consequence score.

[0080] Preliminary Consequence Score Calculation: Score for each indicator F i ( i =1,2,3) After taking the values ​​from Table 3, a weighted summation is used to calculate the preliminary consequence score:

[0081] Additional coefficient correction: Based on the disease type (cavity, delamination, waterlogging, looseness, etc.), refer to the additional coefficient explanation below Table 3 to obtain the correct coefficient. K 1 Final risk consequence score C for:

[0082] In summary, formulas (5) to (6) are mathematical expressions of the evaluation rules in Table 3 of the standard, transforming the original process of looking up the table and making manual judgments into quantifiable weighted summation and correction operations, thus realizing automatic computer scoring.

[0083] Preferably, an indicator system is established based on the calculated values ​​of risk probability and risk consequences, as shown in Table 4.

[0084] Table 4 Classification of Risk Probability and Risk Consequence Levels

[0085] Decision reasoning engine: The decision reasoning engine receives the disease feature vector from step three.

[0086] Based on the type of disease and quantitative indicators, the engine queries the digital standard knowledge base and automatically calculates the deduction value for the detection unit.

[0087] Preferably, the graded output unit automatically determines and outputs the final comprehensive score and quality grade of the roadbed section based on the total deduction value and the grade threshold in the table below.

[0088] Based on highway quality rating R ( P , C A highway quality evaluation system was established. The quality grades of highway subgrades were divided into five levels: excellent, good, medium, passable, and poor, as shown in Table 5.

[0089] Table 5. Highway Subgrade Quality Grades Corresponding to Risk Probability and Risk Consequence Levels

[0090] Step 5: Visualizing Results and Generating Structured Reports

[0091] This step is performed by the visualization output module.

[0092] Multi-dimensional data fusion display: Achieving three-view linkage within a self-developed graphical user interface: Main view: Displays the original radar profile and overlays a semi-transparent disease distribution result, with different diseases highlighted in different colors.

[0093] Top view: On the highway alignment map, the quality grading results of the roadbed along the route are visually displayed by different colored stripes: Excellent - White, Good - Green, Medium - Blue, Pass - Yellow, Poor - Red.

[0094] Information panel: Displays the station number, detailed information on defects, and corresponding deductions at the cursor's location in real time.

[0095] Automated report generation: The system has a built-in Word report template. After the user clicks "Generate Report", the system automatically fills in the inspection overview, the quality level distribution map along the line, the defect statistics summary table, the detailed description of the main defect sections, the radar image screenshots and maintenance suggestions, and outputs a standardized inspection report in PDF / Word format that meets the requirements of engineering management with one click.

[0096] Secondly, this invention provides a neural network-based intelligent identification system for highway subgrade radar detection. This system employs a hierarchical architecture, such as... Figure 4 As shown, the specific components are as follows: Hardware acquisition layer: Platform: Light van.

[0097] Data sensing unit: vehicle-mounted ground-penetrating radar main unit and antenna, GPS navigation system, odometer encoder.

[0098] Onboard computing and control unit: Industrial control computer, responsible for system control, data synchronization and temporary storage.

[0099] Data preprocessing layer: Deployed on industrial computers and cloud servers, it runs the algorithm software package of the radar image processing module, is responsible for completing the standardization and enhancement process of step S1, and manages raw and intermediate data.

[0100] Core algorithm engine layer: Deployed on an internet cloud platform. This layer encapsulates two core models: The trained YOLOv11 object detection model (execute step S2).

[0101] Digitalized standard knowledge base and decision reasoning engine (execution step S3).

[0102] This layer provides a model call interface in the form of a RESTful API or gRPC service, receives image data, and returns disease feature vectors and quality grading results.

[0103] Specifically, this may include a roadbed quality specification knowledge base, a decision-making reasoning component, and a graded output unit. The built-in industry specification knowledge base automatically generates roadbed quality grading results that conform to specifications based on the described defect characteristics, radar image features, and the relationship between roadbed quality grading and specifications. Specifically: The specification knowledge base: It digitally incorporates clauses related to roadbed quality grading from the national and industry standard, "Technical Specification for Detection and Evaluation of Underground Defects in Urban Roads and Pipelines" DB11 / T 1399-2017, translating the permissible limits and grading standards for different types and scales of defects into computer-executable rules. The decision-making reasoning engine: It receives defect feature vectors from the defect intelligent identification and feature extraction module and automatically matches and logically judges them against the rules in the specification knowledge base. The graded output unit: Based on the reasoning results, it automatically generates a comprehensive quality grade evaluation score for each detection unit / meter and classifies the quality into excellent, good, medium, passable, and poor.

[0104] Integrated Application and Interaction Layer: Provides users with desktop application software based on B / S (Browser / Server) or C / S (Client / Server) architecture.

[0105] This software integrates a visualization output module, providing a full suite of functions including project creation, task scheduling, data processing workflow control, multi-dimensional visualization of results, report generation, and historical data archiving management. It serves as the sole interface for users to interact with the entire intelligent identification system. In some preferred embodiments, it may include a data visualization submodule, an automated report generator, and interactive analysis tools. It integrates and displays raw images, identification results, and grading conclusions, automatically generating a structured detection report. Specifically: The data visualization unit overlays and links raw radar images, disease segmentation results, spatial mileage coordinates, and quality grading results within the same visualization interface. The automated report generator has built-in templates and can automatically generate structured reports based on grading decision results. The interactive analysis tools provide functions such as disease detail query, historical data comparison, and statistical chart generation.

[0106] The aforementioned functional layers can also be configured as functional modules corresponding to the methods described above, according to actual needs, for example... Figure 5 As shown.

[0107] In some preferred embodiments, the hardware acquisition layer specifically includes: Ground-penetrating radar, deployed on a testing vehicle, is used to emit radar signals to detect roadbed defects. The positioning unit, deployed on the detection vehicle, is used to provide radar vehicle location information; A synchronous control system, deployed on the inspection vehicle, is used to adjust the vehicle speed and control the radar; The process of acquiring detection image data at the hardware acquisition layer includes: The synchronous control system controls the testing vehicle to travel along the predetermined testing line at a constant speed. The ground-penetrating radar echo signals are continuously collected at preset intervals; simultaneously, based on the location information and detection information obtained by the positioning unit, detection image data with location markers is output, which includes images of roadbed loosening and underground cavities.

[0108] Specifically, the vehicle-mounted roadbed defect radar detection device can be installed on a carrier vehicle, a portable ground-penetrating radar system for collecting radar signals of roadbed defects, a GPS positioning and receiving system, and a central processing system. By integrating ground-penetrating radar, positioning unit, and synchronous control system, the radar detection vehicle can perform full-coverage detection of the detection area, collect raw radar echo signals with precise spatial coordinates in real time, and transmit them to the processing terminal to realize radar image detection and positioning of roadbed defects.

[0109] The data preprocessing layer's workflow can specifically include: Use a Python preprocessing script to sequentially process the raw radar data: Time zero-point correction: eliminates system delay; DC drift removal: Removes the offset of the signal baseline; Background denoising: Subtract the average waveform of each channel and suppress horizontal in-phase axis noise; Bandpass filtering: Butterworth bandpass filters are used to preserve effective frequency bands and suppress high and low frequency noise; Gain recovery: Apply automatic gain control or distance gain compensation to make deep signals clearly visible; To optimize the neural network input, the preprocessed radar profile (B-Scan) is subjected to the following: contrast stretching, directional filtering (using a specific noise suppression algorithm to highlight the reflection features related to the disease), and format standardization output. The dynamic range of image grayscale values ​​is linearly expanded to 0-255 to enhance feature contrast. Apply the Laplace operator or directional filter to highlight the diffraction hyperbola or phase axis discontinuity features associated with the disease; The processed data is uniformly formatted into a two-dimensional grayscale image matrix with a regular size and an image width that varies according to the length of the survey line, and is accompanied by corresponding mileage-depth coordinate information and stored.

[0110] The core algorithm engine layer's detection model is a target detection neural network based on the YOLOv11 architecture. This model adopts a single-stage detection paradigm, extracting multi-scale features from radar profile images through a backbone network, and using the detection head to directly regress and predict the bounding box coordinates, class probability, and confidence level of each disease target on the feature map. In some preferred embodiments, it includes a backbone network, a feature network, a detection head, and a non-maximum suppression module arranged sequentially along the data flow direction. The backbone network has stacked convolutional layers, batch normalization layers, and SiLU (Sigmoid Linear Unit) activation function layers, which can extract multi-level features from aggregate texture to semantic information based on two-dimensional radar profiles through progressive downsampling operations. The feature network is used to integrate the multi-level features output by the backbone network to obtain multiple sets of enhanced feature maps; The detection head is used to predict the enhanced feature map and output class probability information and bounding box coordinate information; The nonmaximum suppression module is used to perform nonmaximum suppression on the category probability information and bounding box coordinate information output by the detection head to obtain a structured feature vector of the disease that includes the disease type, quantity, size and distribution density.

[0111] The output of the feature vector can contain a list of bounding boxes for multiple disease instances. This list is then integrated by subsequent processing units to generate a structured disease feature vector. This vector encodes multi-dimensional information such as disease type, spatial location, approximate size of the bounding box, confidence level, and the number of each type of disease per unit length. Offline training and model library: Contains a massive dataset of labeled radar images for model training, labeled as rectangular bounding boxes of disease regions.

[0112] The construction process of the detection model can be as follows: Radar profile images were collected as a dataset. Using annotation tools, the diseased target areas in the images were labeled with rectangular bounding boxes, and their categories were specified, including: cavities, voids, looseness, water-rich areas, and non-diseased objects such as various metal pipelines and non-metal pipelines; This document describes how to set up a PyCharm Community Edition 2024.1 environment on an operating system, and then build a YOLOv11 object detection algorithm platform on top of it. The specific steps include downloading and installing an Anaconda virtual environment and configuring Python 3.8 accordingly; then configuring the environment according to PyTorch requirements; finally, installing PyCharm using Anaconda, importing the YOLOv11 source code, and adding the corresponding interpreter. After completing these steps, opening the corresponding terminal environment allows you to run YOLOv11 on the platform. The mean accuracy mAP_0.5 was used as the evaluation metric for model training accuracy.

[0113] In summary, this invention provides a method and system for intelligent identification of roadbed radar detection based on neural networks. The method includes: standardizing and enhancing the acquired detection image data to obtain a two-dimensional radar profile; identifying disease information based on the two-dimensional radar profile using a target detection model, and aggregating the identified disease information to obtain a structured feature vector containing disease type, quantity, size, and distribution density; evaluating the performance of the detection unit based on the structured feature vector using a decision engine of a digital standardized knowledge base; and obtaining a detection report based on the structured feature vector and the performance evaluation results of the detection unit. The advantages of the method and system provided by this invention are: 1. Existing technologies (such as the Faster R-CNN solution) stop at the identification and location of diseases; this invention further realizes automatic quality assessment and grading according to national / industry standards, forming a complete solution.

[0114] 2. This invention replaces the manual evaluation process of comparing standard texts with a digital standard knowledge base, overcoming the problems of low efficiency and strong subjectivity of manual evaluation, and making the results more efficient, standardized and traceable.

[0115] 3. Through rapid identification and automated grading report generation based on YOLOv11, this invention enables the technological achievements to serve engineering acceptance and maintenance decisions more directly and conveniently, resulting in a higher degree of practical application in engineering.

[0116] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of one embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing the present invention.

[0117] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that the present invention can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of the present invention.

[0118] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for apparatus or system embodiments, since they are basically similar to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. The apparatus and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0119] 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 method for intelligent identification of highway subgrade radar detection based on neural networks, characterized in that, include: S1. Standardize and enhance the collected detection image data to obtain a two-dimensional radar profile. S2. Based on the two-dimensional radar profile, the disease information is identified through the target detection model, and the identified disease information is aggregated to obtain a structured feature vector of the disease containing the disease type, quantity, size and distribution density. S3. Based on the structured feature vectors of defects, the performance of highway subgrade and detection units is graded and evaluated through the decision engine of the digital standard knowledge base; The results of the graded evaluation of the performance of the highway subgrade and the detection unit are used for risk assessment of the target highway section and performance evaluation of the detection unit.

2. The method according to claim 1, characterized in that, The standardization process in step S1 includes sequentially performing time null correction, DC drift removal, background denoising, bandpass filtering, and gain restoration on the detected image data. The specific process of image enhancement in step S1 includes contrast stretching, directional filtering, and format normalization output operations on the standardized detection image data to obtain a two-dimensional radar profile.

3. The method according to claim 1, characterized in that, The target detection model in step S2 includes a backbone network, a feature network, a detection head, and a non-maximum suppression module arranged sequentially along the data flow direction; The backbone network is equipped with stacked convolutional layers, batch normalization layers, and SiLU activation function layers, which can extract multi-level features from aggregate texture to semantic information based on two-dimensional radar profiles through stepwise downsampling operations. The feature network is used to integrate the multi-level features output by the backbone network to obtain multiple sets of enhanced feature maps. The detection head is used to predict the enhanced feature map and output class probability information and bounding box coordinate information; The nonmaximum suppression module is used to perform nonmaximum suppression on the category probability information and bounding box coordinate information output by the detection head to obtain a structured feature vector of the disease containing the disease type, quantity, size and distribution density.

4. The method according to claim 1, characterized in that, Step S3 specifically includes: Based on the structured feature vectors of road defects, the highway quality grade score is obtained through calculation. Based on the structured feature vector of the disease, the index score of the probability of risk occurrence is obtained by calculation, and a two-level index system for evaluating the probability of risk occurrence is established based on the index score of the probability of risk occurrence. Based on the structured feature vector of the disease, the index score of the risk consequence is obtained by calculation, and a first-level index system for risk consequence evaluation is established based on the index score of the risk consequence. Based on a two-level indicator system for assessing the probability of risk occurrence and a one-level indicator system for assessing the consequences of risk, a risk occurrence probability and risk consequence level system is established. The performance of the detection unit is evaluated based on the probability of risk occurrence and the risk consequence level system, and a risk assessment is also conducted on the target highway section.

5. A highway subgrade radar detection and intelligent identification system based on neural networks, characterized in that, It includes a hardware acquisition layer, a data preprocessing layer, a core algorithm engine layer, and an integrated application layer for interaction. The hardware acquisition layer is used to acquire detection image data; The data preprocessing layer is used to standardize and enhance the acquired detection image data to obtain a two-dimensional radar profile. The core algorithm engine layer is used for: Based on two-dimensional radar profile, disease information is identified through a target detection model, and the identified disease information is aggregated to obtain a structured feature vector of diseases containing disease type, quantity, size and distribution density. Based on the structured feature vectors of diseases, the performance of the detection unit is evaluated through a decision engine of a digital standardized knowledge base; The integrated application in the interaction layer is used to obtain a detection report based on the structured feature vector of the disease and the performance evaluation results of the detection unit.

6. The system according to claim 5, characterized in that, The hardware acquisition layer specifically includes: Ground-penetrating radar, deployed on a testing vehicle, is used to emit radar signals to detect roadbed defects. The positioning unit, deployed on the detection vehicle, is used to provide radar vehicle location information; A synchronous control system, deployed on the inspection vehicle, is used to adjust the vehicle speed and control the radar; The process of acquiring detection image data by the hardware acquisition layer includes: The synchronous control system controls the detection vehicle to travel along the predetermined test line at a constant speed. The ground-penetrating radar continuously acquires echo signals at preset intervals; synchronously, based on the location information obtained by the positioning unit and the echo signals, it outputs detection image data with location markers, which includes images of roadbed loosening and underground cavities.

7. The system according to claim 5, characterized in that, The data preprocessing layer specifically includes performing the following processing steps on the raw radar data in sequence: Time zero-point correction: eliminates system delay; DC drift removal: Removes the offset of the signal baseline; Background denoising: Subtract the average waveform of each channel and suppress horizontal in-phase axis noise; Bandpass filtering: Using a Butterworth bandpass filter to retain the effective frequency band and suppress high and low frequency noise; Gain recovery: Apply automatic gain control or distance gain compensation to make deep signals clearly visible; To optimize the neural network input, the preprocessed radar profile image is subjected to the following processes: contrast stretching, directional filtering, and format normalization output. The dynamic range of image grayscale values ​​is linearly expanded to 0-255 to enhance feature contrast. Apply the Laplace operator or directional filter to highlight the diffraction hyperbola or phase axis discontinuity features associated with the disease; The processed data is uniformly formatted into a two-dimensional grayscale image matrix with a regular size and an image width that varies according to the length of the survey line, and is accompanied by corresponding mileage-depth coordinate information and stored.

8. The system according to claim 5, characterized in that, The detection model of the core algorithm engine layer includes a backbone network, a feature network, a detection head, and a non-maximum suppression module arranged sequentially along the data flow direction. The backbone network is equipped with stacked convolutional layers, batch normalization layers, and SiLU activation function layers, which can extract multi-level features from aggregate texture to semantic information based on two-dimensional radar profiles through stepwise downsampling operations. The feature network is used to integrate the multi-level features output by the backbone network to obtain multiple sets of enhanced feature maps. The detection head is used to predict the enhanced feature map and output class probability information and bounding box coordinate information; The nonmaximum suppression module is used to perform nonmaximum suppression on the category probability information and bounding box coordinate information output by the detection head to obtain a structured feature vector of the disease containing the disease type, quantity, size and distribution density. The process of constructing the detection model includes: Radar profile images were collected as a dataset. Annotation tools were used to annotate the disease target areas in the images in the form of rectangular bounding boxes and to specify the disease category. The disease categories include: cavitation, voiding, loosening, water-rich, as well as non-disease objects such as various metal pipelines and various non-metal pipelines. Set up a target detection platform for the YOLOv11 object detection algorithm in the operating system; specifically, this includes downloading and configuring a virtual environment; importing the YOLOv11 source code using a programming tool and adding the corresponding interpreter; Set evaluation metrics for the training accuracy of the detection model; The working process of the core algorithm engine layer specifically includes: Based on the structured feature vectors of road defects, the highway quality grade score is obtained through calculation. Based on the structured feature vector of the disease, the index score of the probability of risk occurrence is obtained by calculation, and a two-level index system for evaluating the probability of risk occurrence is established based on the index score of the probability of risk occurrence. Based on the structured feature vector of the disease, the index score of the risk consequence is obtained by calculation, and a first-level index system for risk consequence evaluation is established based on the index score of the risk consequence. Based on a two-level indicator system for assessing the probability of risk occurrence and a one-level indicator system for assessing the consequences of risk, a risk occurrence probability and risk consequence level system is established. The performance of the detection unit is evaluated based on the probability of risk occurrence and the risk consequence level system, and a risk assessment is also conducted on the target highway section.

Images