Road damage identification method and system based on artificial intelligence
By adopting multi-scale feature fusion and attention mechanism models in road disease identification, combining knowledge distillation technology and online update mechanism, the problems not considered in the existing technology that multi-view and multi-scale data are solved, and efficient and accurate roads are achieved Disease identification reduces costs and improves the ability to identify nuances.
Patent Information
- Application Number
- CN202411166121.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-23
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2044-08-23
AI Technical Summary
The existing technology does not consider multi-perspective and multi-scale data in road disease identification, lacks continuous learning and update mechanisms, has low generalization capabilities, and is high equipment cost, making it difficult to popularize daily maintenance inspections.
Using artificial intelligence-based road disease identification method, multi-view and multi-scale data are collected through mobile patrol vehicles or drones, and automatically dynamically annotated using the RDS system, select multi-scale feature fusion and attention mechanism models, and use knowledge distillation technology Combined with ViT large model for training, an adaptive weight allocation mechanism and an online update mechanism are introduced to improve model recognition performance.
It realizes efficient, accurate and real-time road disease identification, reduces costs, improves the nuanced ability of identification, reduces manual participation, and ensures the stable operation and update of the model.
Smart Images

Figure CN118675074B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of artificial intelligence, and in particular to a road hazard identification method and system based on artificial intelligence. Background Art
[0002] With the rapid construction of traffic roads, a large number of roads will experience overloading and the impact of the natural environment, and cracks, potholes and other defects will appear. In order to ensure the safe driving of vehicles and the service life of roads, the later road maintenance is crucial, and a large amount of manpower and material resources need to be invested in road inspections. At present, the main method of manual inspection is adopted, that is, the inspectors drive the road maintenance inspection vehicle to observe the road surface of the roads under their jurisdiction with the naked eye. When encountering a disease, the inspectors need to get off the car and take photos with a handheld terminal for reporting. However, the inspectors are prone to fatigue and omissions when observing in the car for a long time, and there are traffic hazards when getting off the car to take photos, which affects the normal operation of traffic. In addition, the industry has also invested in a small number of professional inspection vehicles, equipped with high-precision differential positioning modules, three-dimensional laser scanning radars, and industrial computers for road data collection and analysis, but the cost of this equipment is relatively high. The special vehicles on the market range from tens to millions, and are mainly used for road inspections before major, medium and minor road repairs, and it is difficult to popularize them to daily maintenance inspections.
[0003] For example, a Chinese patent application with publication number CN117422669A discloses a road disease detection method based on artificial intelligence visual recognition, including: step one: using an image module to take a photo of the road, obtain the original image, and integrate a multi-target tracking algorithm with positioning information; step two: performing image preprocessing on the collected original image through a computing module; step three: continuing to perform semantic segmentation on the image, using a semantic segmentation algorithm to divide the road surface and non-road surface areas, distinguishing the dynamic elements of vehicles, pedestrians, and non-motor vehicles, and performing pixel-level segmentation on the objects in the image to provide clean input image data for the next step of target detection; the method of this technical solution is easy to deploy in a portable inspection instrument, greatly improving the popularity of automated detection; it is easy to deploy using the existing low-cost and mature hardware environment, and at the same time uses GPS information to remove duplicate targets, which greatly simplifies the target tracking algorithm and effectively improves the computing efficiency.
[0004] The above existing technologies all have the following problems: 1) multi-view and multi-scale data are not considered; 2) there is a lack of continuous learning and updating mechanism; 3) the generalization ability is low. Summary of the invention
[0005] In view of the shortcomings of the existing technology, the present invention proposes a road disease recognition method and system based on artificial intelligence. Road images and video data are collected by mobile inspection vehicles or drones, and the RDS system is used to autonomously and dynamically annotate road disease targets. Multi-scale feature fusion and attention mechanism models are selected as the basic models, and the ViT large model is combined with knowledge distillation technology for training. The recognition performance of the model is improved by subdividing the disease data set and introducing an adaptive weight allocation mechanism. The trained model is deployed to the data processing system of the inspection vehicle or fixed camera, and the stable operation of the system is ensured through online update and performance monitoring mechanisms, thereby realizing efficient, accurate and real-time road disease recognition.
[0006] To achieve the above object, the present invention provides the following technical solutions:
[0007] The road disease identification method based on artificial intelligence includes:
[0008] Step S1: Use a mobile inspection vehicle or drone to shoot the road, collect multi-view and multi-scale road images and video data, and perform preprocessing. Based on the RDS system, use bounding boxes or key frames to autonomously and dynamically annotate road disease targets in the captured images and videos to obtain road disease data;
[0009] Step S2: Select the multi-scale feature fusion and attention mechanism model as the basic model, select the ViT large model as the teacher model, obtain soft labels by training the ViT large model, and use the soft labels to perform knowledge distillation training on the basic model. Introduce an adaptive weight allocation mechanism, and automatically adjust the knowledge transfer strategy from the ViT large model to the basic model according to the category of road disease data;
[0010] Step S3: Segment the road disease data set according to the disease type and severity to form multiple subclasses, train the knowledge transfer strategy for each subclass in a hierarchical refinement manner, input the segmented road disease data subsets into the basic model for training, optimize the basic model parameters according to the training results, and evaluate the optimized basic model;
[0011] Step S4: Deploy the trained basic model to the data processing system of the inspection vehicle or fixed camera to perform real-time road disease and traffic accident identification, integrate and operate the identification results with the road management system and traffic management system, deploy an online model update mechanism and performance degradation monitoring mechanism, and perform regular maintenance and updates;
[0012] The specific steps of step S2 include:
[0013] S2.1: Clarify the task requirements, select a model that combines multi-scale feature fusion and attention mechanism as the base model, and select the ViT large model as the teacher model;
[0014] S2.2: Use the road disease dataset to train the ViT large model, calculate the value of the loss function through forward propagation, and use the backpropagation algorithm to calculate the gradient of the loss function to the parameters of the ViT large model. At the same time, use the optimizer to update the parameters of the ViT large model according to the gradient until the preset number of iterations is reached, obtain the trained ViT large model parameters and prediction results, and use the prediction results as soft labels. The formula of the loss function is:
[0015]
[0016] Among them, L represents the total loss, represents the dynamic equilibrium parameter, Represents the dynamic temperature parameter, which is used to control the smoothness of the soft label. represents the weight vector of the i-th road disease data category, represents the soft label of the ViT large model for the i-th road disease data category, Represents a positive number, used to prevent numerical instability problems in logarithmic functions. represents the predicted probability of the basic model for the i-th road disease data category, Represents the one-hot encoding of the true label. If the i-th category is the true category, then ,otherwise , represents the knowledge distillation loss with the category importance weight vector, and t represents the current iteration number;
[0017] Said and The formula is:
[0018] ;
[0019] in, , Respectively represent the minimum and maximum values of the dynamic balance parameters. As the training progresses, Will be from Gradually decrease to , thereby gradually reducing the weight of knowledge distillation loss and increasing the weight of classification loss, , Respectively represent the minimum and maximum values of the dynamic temperature parameters, Indicates the total number of iterations;
[0020] S2.3: According to different categories of road damage data, an adaptive weight allocation mechanism is designed to dynamically adjust the knowledge transfer weight from the ViT large model to the basic model;
[0021] S2.4: Train the basic model using road damage data, soft labels, true labels, and the adaptive weight assignment mechanism, and record the prediction results.
[0022] Specifically, the specific steps of step S3 include:
[0023] S3.1: Conduct a visual inspection of the road damage data, determine the type of damage in combination with the damage classification standards, develop corresponding evaluation indicators according to different types of damage, and classify the severity of the damage according to the evaluation indicators. Subdivide the road damage data into multiple subcategories according to the type of damage, and further subdivide each subcategory according to the severity;
[0024] S3.2: For each sub-category, formulate a knowledge transfer strategy and use the ViT large model as the teacher model to predict the sub-category data and generate soft labels;
[0025] S3.3: Use a hierarchical strategy to sort and group the sub-class data, input the segmented road disease data subsets into the basic model and train it layer by layer using soft labels;
[0026] S3.4: Based on the training results, use the gradient descent optimization algorithm to adjust the parameters of the base model, and use an independent test set to evaluate the optimized base model.
[0027] Specifically, the criteria for grading the severity of the disease in S3.1 include:
[0028]
[0029] Among them, S represents the severity score of the disease, represents the weight of the jth road damage assessment index, represents the score of the jth road damage assessment index, and S j ∈ [ 1 , 1 0 ] , represents the location parameter of the jth road damage assessment index, represents the time when the jth road disease occurs or is detected, represents the jth road damage area, and takes As the unit of measurement, represents the jth road damage depth, As the unit of measurement, represents the location parameter coefficient, represents the time parameter coefficient, represents the area parameter coefficient, represents the depth parameter coefficient, n represents the number of road damage assessment indicators;
[0030] like , then the severity is mild;
[0031] like , the severity is moderate;
[0032] like , the severity is severe.
[0033] Specifically, the road damage assessment index score determination condition is:
[0034] { P = P cen , dow , S = 10 ; P = P center , S ∈ [ 5 , 9 ]; P = P subur , S ∈ [ 2 , 4 ]; P = P bian , sub , S = 1 T = T w in, rain , S = 10 ; T = T w in, sun , S ∈ [ 5 , 9 ]; T = T sum , rain , S ∈ [ 2 , 4 ]; T = T sum , sun , S = 1 D > 25 , S = 10 ; 10 < D ≤ 25 , S ∈ [ 2 , 9 ]; D ≤ 10 , S = 1 M ≥ 0 . 3 , S = 10 ; 0 . 05 ≤ M < 0 . 3 , S ∈ [ 2 , 9 ]; M < 0 . 05 , S = 1
[0035] in, Indicates that the disease is located in the center of the urban road. It means the disease is located at the edge of the urban road. Indicates that the disease is located in the center of the suburban road. Indicates that the disease is located at the edge of a suburban road. It means that the disease occurred during the winter rainy season. It means that the disease occurred on sunny days in winter. It means that the disease occurred during the summer rainy season. It indicates that the disease time is a sunny day in summer, P indicates the location of the road disease assessment index, T indicates the time when the road disease occurs or is detected, D indicates the depth of the road disease, and M indicates the area of the road disease.
[0036] The road disease identification system based on artificial intelligence includes: data processing module, model training module, knowledge transfer module and alarm module;
[0037] The data processing module is used to receive and pre-process the raw road data collected from road inspection vehicles and drone equipment, including images and videos;
[0038] The model training module is used to train the recognition model based on the deep learning method using the labeled road damage data to identify different types of road damage;
[0039] The knowledge transfer module is used to optimize the basic model by using the pre-trained large model through the knowledge distillation method;
[0040] The alarm module is used to deploy the trained basic model to the actual operating environment, input real-time road image and video data, and immediately trigger the alarm mechanism if road damage is identified.
[0041] Specifically, the knowledge transfer module includes: a data segmentation unit, a hierarchical training unit, and an anomaly detection unit;
[0042] The data segmentation unit is used to segment the road disease data set into data subsets of specific categories or features for targeted learning;
[0043] The hierarchical training unit is used to adopt a hierarchical training strategy to first train basic features and then gradually transition to complex disease recognition;
[0044] The anomaly detection unit is used to identify and remove abnormal samples or noise by analyzing data and model behavior during training.
[0045] Specifically, an electronic device includes a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of a road hazard identification method based on artificial intelligence when executing the computer program.
[0046] Specifically, a computer-readable storage medium stores computer instructions, which, when executed, execute the steps of a road hazard identification method based on artificial intelligence.
[0047] Compared with the prior art, the present invention has the following beneficial effects:
[0048] 1. The present invention proposes a road disease identification system based on artificial intelligence, and optimizes and improves the architecture, operation steps and processes. The system has the advantages of simple process, low investment and operation costs, and low production work costs.
[0049] 2. The present invention proposes an artificial intelligence-based road damage identification method, which collects multi-view, multi-scale images and video data of roads through mobile inspection vehicles or drones, and can more comprehensively identify and understand road damage and identify subtle differences in road damage; the autonomous dynamic labeling function based on the RDS system can automatically label road damage targets in images and videos, thereby improving the efficiency and accuracy of data processing, reducing manual participation, and reducing costs; multi-scale feature fusion and attention mechanism models are selected as the basic models, and through knowledge distillation training combined with the ViT large model, an efficient basic model can be quickly and accurately trained, while fully utilizing the powerful capabilities of the teacher model while maintaining the lightweight and high efficiency of the basic model.
[0050] 3. The present invention proposes a road damage identification method based on artificial intelligence, which subdivides the road damage data set according to the type and severity of the damage to form multiple subclasses, and adopts a hierarchical and refined method to train the knowledge transfer strategy. This subdivision training method can ensure that the model has a high recognition accuracy for road damages of different types and severity; an online model update mechanism and a performance degradation monitoring mechanism are deployed to regularly maintain and update the model to ensure that the model always remains in the best state. BRIEF DESCRIPTION OF THE DRAWINGS
[0051] Figure 1 A schematic diagram of a road hazard identification method based on artificial intelligence according to the present invention;
[0052] Figure 2 This is a workflow diagram of the road damage identification method based on artificial intelligence of the present invention;
[0053] Figure 3 This is a flow chart for realizing the knowledge transfer strategy of the road damage identification method based on artificial intelligence of the present invention;
[0054] Figure 4 This is an architecture diagram of the road hazard identification system based on artificial intelligence of the present invention. DETAILED DESCRIPTION
[0055] Example 1
[0056] See also Figure 1-Figure 3 , an embodiment of the present invention provides: a road disease identification method based on artificial intelligence, comprising the following steps:
[0057] Step S1: Use a mobile inspection vehicle or drone to shoot the road, collect multi-view and multi-scale road images and video data, and perform preprocessing. Based on the RDS system, use bounding boxes or key frames to autonomously and dynamically annotate road disease targets in the captured images and videos to obtain road disease data;
[0058] Furthermore, the image and video preprocessing process includes:
[0059] (1) Loading raw image and video data;
[0060] (2) De-noising images and videos to reduce the interference of background noise;
[0061] (3) Enhance images and videos to improve their clarity and contrast;
[0062] (4) Crop and scale images and videos.
[0063] The RDS system access and annotation process specifically includes:
[0064] (1) Import the preprocessed image and video data into the RDS system;
[0065] (2) Automatically identify images and videos using pre-trained models or algorithms;
[0066] (3) Automatically draw bounding boxes or extract key frames for annotation based on the recognition results;
[0067] (4) Optimize and adjust the annotation results to improve the accuracy and efficiency of annotation;
[0068] (5) Output the annotation results into a standard annotation file format.
[0069] Step S2: Select the multi-scale feature fusion and attention mechanism model as the basic model, select the ViT large model as the teacher model, obtain soft labels by training the ViT large model, and use the soft labels to perform knowledge distillation training on the basic model. Introduce an adaptive weight allocation mechanism, and automatically adjust the knowledge transfer strategy from the ViT large model to the basic model according to the category of road disease data;
[0070] Step S3: Segment the road disease data set according to the disease type and severity to form multiple subclasses, train the knowledge transfer strategy for each subclass in a hierarchical refinement manner, input the segmented road disease data subsets into the basic model for training, optimize the basic model parameters according to the training results, and evaluate the optimized basic model;
[0071] Step S4: Deploy the trained basic model to the data processing system of the inspection vehicle or fixed camera to perform real-time road disease and traffic accident identification, integrate and operate the identification results with the road management system and traffic management system, deploy an online model update mechanism and performance degradation monitoring mechanism, and perform regular maintenance and updates.
[0072] The specific steps of step S2 include:
[0073] S2.1: Clarify the task requirements, select a model that combines multi-scale feature fusion and attention mechanism as the base model, and select the ViT large model as the teacher model;
[0074] S2.2: Use the road disease dataset to train the ViT large model, calculate the value of the loss function through forward propagation, and use the backpropagation algorithm to calculate the gradient of the loss function to the parameters of the ViT large model. At the same time, use the optimizer to update the parameters of the ViT large model according to the gradient until the preset number of iterations is reached, obtain the trained ViT large model parameters and prediction results, and use the prediction results as soft labels. The formula of the loss function is:
[0075]
[0076] Among them, L represents the total loss, represents the dynamic equilibrium parameter, Represents the dynamic temperature parameter, which is used to control the smoothness of the soft label. represents the weight vector of the i-th road disease data category, represents the soft label of the ViT large model for the i-th road disease data category, Represents a positive number, used to prevent numerical instability problems in logarithmic functions. represents the predicted probability of the basic model for the i-th road disease data category, Represents the one-hot encoding of the true label. If the i-th category is the true category, then ,otherwise , represents the knowledge distillation loss with the category importance weight vector, and t represents the current iteration number;
[0077] Further, ;
[0078] in, , Respectively represent the minimum and maximum values of the dynamic balance parameters. As the training progresses, Will be from Gradually decrease to , thereby gradually reducing the weight of knowledge distillation loss and increasing the weight of classification loss, , Respectively represent the minimum and maximum values of the dynamic temperature parameters, Indicates the total number of iterations;
[0079] S2.3: According to different categories of road damage data, an adaptive weight allocation mechanism is designed to dynamically adjust the knowledge transfer weight from the ViT large model to the basic model;
[0080] S2.4: Train the basic model using road damage data, soft labels, true labels, and the adaptive weight assignment mechanism, and record the prediction results.
[0081] The specific steps of step S3 include:
[0082] S3.1: Visually inspect the road damage data, determine the type of damage in combination with the damage classification standards, develop corresponding evaluation indicators according to different types of damage, and classify the severity of the damage according to the evaluation indicators. Subdivide the road damage data into multiple subcategories according to the type of damage, and further subdivide each subcategory according to the severity. The standards for classifying the severity of the damage include:
[0083]
[0084] in, represents the severity score of the disease, and S ∈ [ 1 , 1 0 ] , represents the weight of the jth road damage assessment index, and , represents the score of the jth road damage assessment index, and S j ∈ [ 1 , 1 0 ] , representing a range from mild to severe, represents the location parameter of the jth road damage assessment index, represents the time when the jth road disease occurs or is detected, represents the jth road damage area, and takes As the unit of measurement, represents the jth road damage depth, As the unit of measurement, represents the location parameter coefficient, represents the time parameter coefficient, represents the area parameter coefficient, represents the depth parameter coefficient, n represents the number of road damage assessment indicators;
[0085] like , then the severity is mild;
[0086] like , the severity is moderate;
[0087] like , then the severity is severe;
[0088] Furthermore, the road damage assessment index score judgment condition is:
[0089] { P = P cen , dow , S = 10 ; P = P center , S ∈ [ 5 , 9 ]; P = P subur , S ∈ [ 2 , 4 ]; P = P bian , sub , S = 1 T = T w in, rain , S = 10 ; T = T w in, sun , S ∈ [ 5 , 9 ]; T = T sum , rain , S ∈ [ 2 , 4 ]; T = T sum , sun , S = 1 D > 25 , S = 10 ; 10 < D ≤ 25 , S ∈ [ 2 , 9 ]; D ≤ 10 , S = 1 M ≥ 0 . 3 , S = 10 ; 0 . 05 ≤ M < 0 . 3 , S ∈ [ 2 , 9 ]; M < 0 . 05 , S = 1
[0090] in, Indicates that the disease is located in the center of the urban road. It means the disease is located at the edge of the urban road. Indicates that the disease is located in the center of the suburban road. Indicates that the disease is located at the edge of a suburban road. It means that the disease occurred during the winter rainy season. It means that the disease occurred on sunny days in winter. It means that the disease occurred during the summer rainy season. It indicates that the disease time is in sunny summer, P indicates the location of the road disease assessment index, T indicates the time when the road disease occurs or is detected, D indicates the depth of the road disease, and M indicates the area of the road disease;
[0091] S3.2: For each sub-category, formulate a knowledge transfer strategy and use the ViT large model as the teacher model to predict the sub-category data and generate soft labels;
[0092] S3.3: Use a hierarchical strategy to sort and group the sub-class data, input the segmented road disease data subsets into the basic model and train it layer by layer using soft labels;
[0093] S3.4: Based on the training results, use the gradient descent optimization algorithm to adjust the parameters of the base model, and use an independent test set to evaluate the optimized base model.
[0094] Example 2
[0095] See also Figure 4 , another embodiment provided by the present invention: a road disease identification system based on artificial intelligence, comprising: a data processing module, a model training module, a knowledge transfer module, and an alarm module;
[0096] Data processing module, used to receive and pre-process raw road data collected from road inspection vehicles and drone equipment, including images and videos;
[0097] Model training module, which is used to train the recognition model based on deep learning method using the labeled road damage data, so that it can identify different types of road damage;
[0098] The knowledge transfer module is used to improve the recognition ability of the basic model through the knowledge distillation method using the pre-trained large model;
[0099] The alarm module is used to deploy the trained basic model to the actual operating environment, input real-time road image and video data, and immediately trigger the alarm mechanism if road damage is identified.
[0100] The data processing module includes: data acquisition unit, preprocessing unit, target detection unit, and data labeling unit;
[0101] The data collection unit is responsible for collecting images and video data of road conditions from road inspection vehicles and drone equipment to ensure the continuity and integrity of the data;
[0102] The pre-processing unit is used to perform preliminary processing on the collected raw road data, including format conversion, denoising, enhancement, and correction, so as to improve the data quality and make it suitable for model training and analysis;
[0103] The target detection unit is used to automatically detect the existing defective areas, such as cracks, pits, and fissures in the pre-processed image;
[0104] The data annotation unit is used to accurately manually or automatically annotate the damaged areas identified by the target detection unit, including the damage type, location, and size, and to construct a road damage dataset.
[0105] The model training module includes: model selection unit and optimization unit;
[0106] The model selection unit is used to select the basic model and large model architecture as the basis for disease classification and location according to the requirements of the disease identification task;
[0107] The optimization unit is used to optimize the performance of the basic model and the large model through hyperparameter adjustment, regularization, and learning rate scheduling strategies, reduce overfitting, and improve the accuracy and generalization ability of the model on the validation set.
[0108] The knowledge transfer module includes: data segmentation unit, hierarchical training unit, and anomaly detection unit;
[0109] The data segmentation unit is used to segment the road damage dataset into data subsets of specific categories or features, so as to conduct targeted learning and improve the model's ability to identify the details of various types of damage;
[0110] Hierarchical training unit, which is used to adopt hierarchical training strategy, first train basic features, and then gradually transition to complex disease recognition, so that the model can gradually learn and master deeper knowledge;
[0111] The anomaly detection unit is used to identify and remove abnormal samples or noise by analyzing the data and model behavior during the training process to ensure the stability and accuracy of model training.
[0112] The alarm module includes: model deployment unit, identification unit, online update unit, alarm and notification unit;
[0113] A model deployment unit is used to deploy the trained model to the actual operating environment, such as a cloud server or edge computing device, in preparation for real-time road disease identification;
[0114] The recognition unit is used to process the input image or video stream in real time in the deployment environment, apply the model to identify the disease, and output the recognition result;
[0115] Online update unit, used to support online learning and updating of the model, continuously optimizing the model based on newly collected data and adapting to changes in disease characteristics;
[0116] The alarm and notification unit is used to automatically trigger an alarm according to preset rules when a disease is identified, and notify staff through multiple channels such as SMS, email, and APP, while providing a detailed disease report and recommended response measures.
[0117] Example 3
[0118] An electronic device includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of a road hazard identification method based on artificial intelligence are implemented. For details, please refer to the above method embodiment, which will not be repeated here.
[0119] A computer-readable storage medium stores computer instructions, which, when executed, execute the steps of a road damage identification method based on artificial intelligence, wherein the storage medium may be a volatile or non-volatile computer-readable storage medium.
[0120] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation modes, which are merely illustrative rather than restrictive. Under the guidance of the present invention, ordinary technicians in the field may also change, modify, replace and modify the above-mentioned embodiments without departing from the purpose and scope of protection of the present invention, and all of these are within the protection of the present invention.
Claims
1. A road disease identification method based on artificial intelligence, characterized in that: include: Step S1: Use a mobile inspection vehicle or drone to shoot the road, collect multi-view and multi-scale road images and video data, and perform preprocessing. Based on the RDS system, use bounding boxes or key frames to autonomously and dynamically annotate the road disease targets in the captured images and videos to obtain a road disease dataset; Step S2: Select the multi-scale feature fusion and attention mechanism model as the basic model, select the ViT large model as the teacher model, obtain soft labels by training the ViT large model, and use the soft labels to perform knowledge distillation training on the basic model. Introduce an adaptive weight allocation mechanism, and automatically adjust the knowledge transfer strategy from the ViT large model to the basic model according to the category of the road disease dataset; Step S3: Segment the road disease dataset according to the disease type and severity to form multiple subclasses, train the knowledge transfer strategy for each subclass in a hierarchical refinement manner, input the segmented subclass data into the basic model for training, optimize the basic model parameters according to the training results, and evaluate the optimized basic model; Step S4: Deploy the trained basic model to the data processing system of the inspection vehicle or fixed camera to perform real-time road disease and traffic accident identification, integrate and operate the identification results with the road management system and traffic management system, deploy an online model update mechanism and performance degradation monitoring mechanism, and perform regular maintenance and updates; The specific steps of step S2 include: S2.1: Clarify the task requirements, select a model that combines multi-scale feature fusion and attention mechanism as the base model, and select the ViT large model as the teacher model; S2.2: Use the road disease dataset to train the ViT large model, calculate the value of the loss function through forward propagation, and use the backpropagation algorithm to calculate the gradient of the loss function to the parameters of the ViT large model. At the same time, use the optimizer to update the parameters of the ViT large model according to the gradient until the preset number of iterations is reached, obtain the trained ViT large model parameters and prediction results, and use the prediction results as soft labels. The formula of the loss function is: ; Among them, L represents the total loss, represents the dynamic equilibrium parameter, Represents the dynamic temperature parameter, which is used to control the smoothness of the soft label. represents the category weight vector of the i-th road disease dataset, represents the soft label of the ViT large model for the i-th road disease dataset category, Represents a positive number, used to prevent numerical instability problems in logarithmic functions. represents the predicted probability of the basic model for the i-th road disease dataset category, Represents the one-hot encoding of the true label. If the i-th category is the true category, then ,otherwise , represents the knowledge distillation loss with the category importance weight vector, and t represents the current iteration number; Said and The formula is: ; in, , Respectively represent the minimum and maximum values of the dynamic balance parameters. As the training progresses, Will be from Gradually decrease to , thereby gradually reducing the weight of knowledge distillation loss and increasing the weight of classification loss, , Respectively represent the minimum and maximum values of the dynamic temperature parameters, Indicates the total number of iterations; S2.3: According to the different categories of road damage datasets, an adaptive weight allocation mechanism is designed to dynamically adjust the knowledge transfer strategy from the ViT large model to the basic model; S2.4: Train the basic model using the road damage dataset, soft labels, true labels, and adaptive weight assignment mechanism, and record the prediction results.
2. The method for identifying road damage based on artificial intelligence according to claim 1, characterized in that: The specific steps of step S3 include: S3.1: Visually inspect the road damage dataset, determine the type of damage based on the damage classification standard, develop corresponding evaluation indicators based on different types of damage, and classify the severity of the damage based on the evaluation indicators. Subdivide the road damage dataset into multiple subcategories based on the type of damage, and further subdivide each subcategory by severity. S3.2: For each sub-category, formulate a knowledge transfer strategy and use the ViT large model as the teacher model to predict the sub-category data and generate soft labels; S3.3: Use a hierarchical strategy to sort and group the sub-category data, input the sorted and grouped sub-category data into the basic model and train it layer by layer using soft labels; S3.4: Based on the training results, use the gradient descent optimization algorithm to adjust the parameters of the base model, and use an independent test set to evaluate the optimized base model.
3. The method for identifying road damage based on artificial intelligence as claimed in claim 2, characterized in that: The criteria for grading the severity of the disease in S3.1 include: ; Among them, S represents the severity score of the disease, represents the weight of the jth road damage assessment index, represents the score of the jth road damage assessment index, and , represents the location parameter of the jth road damage assessment index, represents the time when the jth road disease occurs or is detected, represents the jth road disease area, with square meters as the unit of measurement. represents the jth road damage depth, measured in millimeters. represents the location parameter coefficient, represents the time parameter coefficient, represents the area parameter coefficient, represents the depth parameter coefficient, n represents the number of road damage assessment indicators; like , then the severity is mild; like , the severity is moderate; like , the severity is severe.
4. The method for identifying road damage based on artificial intelligence as claimed in claim 3, characterized in that: The road damage assessment index score determination conditions are: ; in, It means the disease is located in the center of the urban road. It means the disease is located at the edge of the urban road. Indicates that the disease is located in the center of the suburban road. Indicates that the disease is located at the edge of a suburban road. It means that the disease occurred during the winter rainy season. It means that the disease occurred on sunny days in winter. It means that the disease occurred during the summer rainy season. It indicates that the disease time is in sunny summer, P indicates the location of the road disease assessment index, T indicates the time when the road disease occurs or is detected, D indicates the depth of the road disease, and M indicates the area of the road disease.
5. A road damage identification system based on artificial intelligence, which is used to implement the road damage identification method based on artificial intelligence according to any one of claims 1 to 4, characterized in that: include: Data processing module, model training module, knowledge transfer module, alarm module; The data processing module is used to receive and pre-process the raw road data collected from road inspection vehicles and drone equipment, including images and videos; The model training module is used to train the recognition model based on the deep learning method using the labeled road damage data set to identify different types of road damage; The knowledge transfer module is used to optimize the basic model by using the pre-trained large model through the knowledge distillation method; The alarm module is used to deploy the trained basic model to the actual operating environment, input real-time road image and video data, and immediately trigger the alarm mechanism if road damage is identified.
6. The road hazard identification system based on artificial intelligence as claimed in claim 5, characterized in that: The knowledge transfer module includes: a data segmentation unit, a hierarchical training unit, and an anomaly detection unit; The data segmentation unit is used to segment the road disease data set into data subsets of specific categories or features for targeted learning; The hierarchical training unit is used to adopt a hierarchical training strategy to first train basic features and then gradually transition to complex disease recognition; The anomaly detection unit is used to identify and remove abnormal samples or noise by analyzing data and model behavior during training.
7. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the road damage identification method based on artificial intelligence described in any one of claims 1 to 4 are implemented.
8. A computer-readable storage medium, characterized in that: Computer instructions are stored thereon, and when the computer instructions are executed, the steps of the road damage identification method based on artificial intelligence described in any one of claims 1 to 4 are executed.
