Intelligent pavement friction performance evaluation method and system

Abstract

The application discloses a kind of intelligent pavement friction performance evaluation method and system, method includes multi-source data acquisition, pavement image depth feature extraction, vehicle-road interaction friction performance evaluation and pavement friction performance dynamic evaluation.The application relates to the field of data processing, specifically refers to a kind of intelligent pavement friction performance evaluation method and system, the scheme of the application innovatively proposes to carry out vehicle-road interaction friction performance evaluation first, then the vehicle-road interaction result is combined with environmental data to analyze, accurate and comprehensive pavement friction performance evaluation is realized;Active shape model is used to extract pavement macrostructure features, and improved local binary texture method is used to extract pavement microstructure features, to improve the accuracy of vehicle-road interaction friction performance evaluation results;The calculation efficiency is improved, and the accuracy of the model output result is improved by introducing probability sparse self-attention mechanism, distillation mechanism encoder and one-time generation decoder to improve the Transform algorithm.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to an intelligent method and system for evaluating road surface friction performance. Background Technology

[0002] The intelligent road surface friction performance evaluation method and system refers to a method that uses artificial intelligence technology to accurately evaluate the road surface friction coefficient and safety by collecting multi-dimensional data in real time and combining machine learning and data mining algorithms. This realizes the transformation of road surface friction performance evaluation from traditional static measurement to intelligent and automated real-time monitoring, effectively improving the accuracy and efficiency of road traffic safety assessment, providing decision support for road maintenance and traffic management, and laying the foundation for the development of intelligent driving and autonomous driving technologies.

[0003] However, traditional pavement friction performance assessment suffers from several technical problems. Firstly, the joint analysis of vehicle behavior data, pavement condition data, and environmental data overlooks the direct interaction between vehicles and the pavement, making it difficult to effectively separate the complex relationships between different data sources. This leads to inaccurate pavement friction performance assessment results. Secondly, existing pavement image feature extraction methods suffer from fragmented extraction of macroscopic structural features and microscopic texture features, failing to comprehensively cover all key features related to friction performance. Furthermore, these methods are weakly resistant to external interference during extraction, resulting in low feature reliability and an inability to handle complex data acquisition scenarios, further contributing to inaccurate vehicle-road interaction friction performance assessment results. Finally, existing models for vehicle-road interaction friction performance assessment suffer from high computational complexity, low efficiency in processing long-term time-series data, and an inability to effectively capture key time-series features, all of which contribute to inaccurate model output results. Summary of the Invention

[0004] To address the aforementioned issues and overcome the shortcomings of existing technologies, this invention provides an intelligent method and system for assessing road surface friction performance. Addressing the technical problem in traditional road surface friction performance assessments where the joint analysis of vehicle behavior data, road surface condition data, and environmental data overlooks the direct interaction between vehicles and the road surface, leading to inaccurate assessment results due to the difficulty in effectively separating the complex relationships between different data sources, this invention innovatively proposes first assessing the friction performance of vehicle-road interaction, and then combining the vehicle-road interaction results with environmental data for analysis. By first assessing the friction performance of the vehicle-road interaction, the core friction performance characteristics of vehicle-road interaction can be captured, and the main influencing factors of friction performance can be accurately extracted. Subsequently, the fusion of environmental data can structurally separate the direct impact of vehicle-road interaction from external environmental factors, avoiding confusion between the complex relationships between different data sources, enhancing the hierarchy and logic of the analysis, improving the adaptability and flexibility of friction performance assessment, ensuring accuracy under various road conditions and weather conditions, and significantly improving performance. This solution improves the accuracy and practicality of road surface friction performance assessment, achieving a more precise and comprehensive evaluation. Addressing the technical problems of existing road surface image feature extraction methods—namely, the fragmented extraction of macroscopic structural features and microscopic texture features, which fails to fully cover all key features related to friction performance and exhibits weak resistance to external interference during extraction, resulting in low feature reliability and inability to handle complex acquisition scenarios, ultimately leading to inaccurate vehicle-road interaction friction performance assessment results—this innovative solution employs a multi-dimensional feature extraction method. It uses an active shape model to extract macroscopic structural features of the road surface, utilizes an improved local binary texture method to extract microscopic structural features, and extracts multi-dimensional texture statistical features. This solves the problem of fragmented macroscopic and microscopic texture features in traditional methods, achieving comprehensive coverage of all key features of friction performance. It effectively improves the resistance to external interference during feature extraction, significantly enhances feature reliability, improves the accuracy of vehicle-road interaction friction performance assessment, and enhances the accuracy and real-time performance of the assessment results.To address the technical problems of high computational complexity, low efficiency in processing long-term data, and inability to effectively capture key temporal features in existing models for evaluating vehicle-road interaction friction performance, which lead to inaccurate model outputs, this solution improves the Transformer algorithm by introducing a probabilistic sparse self-attention mechanism, a distillation mechanism encoder, and a one-time generation decoder. The introduction of the probabilistic sparse self-attention mechanism significantly reduces computational complexity and improves the efficiency of processing long-term data by focusing on key temporal intervals. The distillation mechanism encoder, utilizing one-dimensional convolution and max pooling operations, compresses feature maps and enhances the processing capability of temporal data. The one-time generation decoder solves the problem of slow model result generation in long-term prediction, enhancing the model's computational efficiency and effectively improving the accuracy and stability of the model's output, thus achieving efficient and accurate vehicle-road interaction friction performance evaluation.

[0005] The technical solution adopted by this invention is as follows: This invention provides an intelligent method for evaluating road surface friction performance, which includes the following steps:

[0006] Step S1: Acquisition of multi-source data;

[0007] Step S2: Extraction of depth features from road surface image;

[0008] Step S3: Vehicle-road interaction friction performance evaluation;

[0009] Step S4: Dynamic evaluation of road surface friction performance.

[0010] Furthermore, in step S1, the multi-source data acquisition specifically involves obtaining the original road surface friction assessment data through data acquisition operations, and performing data optimization operations on the original road surface friction assessment data to obtain optimized road surface friction assessment data.

[0011] The raw data for road surface friction assessment includes historical friction performance assessment data and real-time friction performance assessment data; both the historical friction performance assessment data and the real-time friction performance assessment data include road surface data, vehicle data, and environmental data.

[0012] The historical friction performance evaluation data also includes historical vehicle-road interaction friction performance evaluation results and historical road surface friction performance dynamic evaluation results; the data optimization operations include image data optimization processing and structural data optimization processing.

[0013] Further, in step S2, the extraction of depth features from the road surface image specifically includes the following steps:

[0014] Step S21: Extraction of macroscopic structural features of the road surface, specifically including steps S211 to S214:

[0015] Step S211: Calculate the road surface shape weight parameters. Specifically, based on the iterative matching algorithm of the active shape model, key structural points are marked on the road surface optimization image. Using the marked key structural points as vertices, the road surface optimization image is divided into non-overlapping triangular structural units. The two-dimensional coordinate information of all key structural points is extracted and stitched together to form a road surface shape vector. The standard mean shape vector of the road surface is introduced and combined with the preset road surface structure feature vector. The road surface shape weight parameters are obtained through feature vector projection operation.

[0016] Step S212: Road surface shape and texture feature fusion, specifically, extracting grayscale texture information within each triangular structural unit, converting it into a one-dimensional grayscale texture vector, then concatenating the grayscale texture vector with the road surface shape weight parameters in dimensional order to form a road surface macroscopic shape and texture joint vector, and then performing dimensionality reduction processing on the road surface macroscopic shape and texture joint vector through principal component analysis algorithm to obtain the road surface shape and texture fusion features.

[0017] Step S213: Road surface fusion feature edge enhancement, specifically, performing translation-invariant wavelet transform on the road surface shape and texture fusion features to decompose and obtain low-frequency coefficients that characterize the core information of the road surface macrostructure. Then, performing pixel-by-pixel convolution operation on the low-frequency coefficients through the edge detection operator to calculate the edge intensity value of each pixel coordinate. Subsequently, high-confidence pixel coordinates with edge intensity higher than a preset threshold are selected. The above pixel coordinates are then structurally integrated based on spatial location correlation to obtain the road surface structure enhancement features.

[0018] Step S214: Output the road surface macrostructure feature set. Specifically, the road surface structure enhancement features are quantified and statistically analyzed to generate various structural sub-features. After normalizing each structural sub-feature, they are aligned and stitched together according to the feature dimensions to obtain the road surface macrostructure feature set.

[0019] Step S22: Road surface microstructure feature extraction, specifically, extracting road surface microstructure features using an improved local binary texture extraction algorithm; including steps S221 to S224:

[0020] Step S221: Road surface texture unit construction, specifically, taking the optimized road surface image as the processing object, selecting a 3×3 pixel block as the road surface texture unit, defining the center pixel coordinates and grayscale value of each road surface texture unit, and then marking the R neighboring pixels surrounding the center pixel in the road surface texture unit in a clockwise direction.

[0021] Step S222: Calculate the local binary code value. Specifically, based on the gray value of the center pixel as the threshold, the gray values ​​of the R neighboring pixels are binarized respectively. The binarization results are accumulated with the weights to obtain the local binary code value of the unit, where i is the label number of the neighboring pixel.

[0022] Step S223: Obtain the cross-scale fusion feature vector. Specifically, the local binary encoding values ​​of all road surface texture units are concatenated in order to form a high-dimensional micro-texture vector. This micro-texture vector is then concatenated with the road surface shape weight parameters in the order of feature dimensions to form a cross-scale joint vector. The cross-scale joint vector is then dimensionality-reduced using the principal component analysis algorithm to obtain the cross-scale fusion feature vector.

[0023] Step S224: Output the pavement microstructure feature set, specifically by standardizing all feature values ​​of the cross-scale fused feature vector to obtain the pavement microstructure feature set;

[0024] Step S23: Texture statistical feature extraction, specifically, constructing a gray-level co-occurrence matrix based on pixel pair distance and angle, calculating the second-order statistics of pixel pairs through the gray-level co-occurrence matrix, extracting core texture features, and obtaining the road surface texture statistical feature set;

[0025] Step S24: Road surface image depth feature set integration, specifically, after aligning the road surface macrostructure feature set, road surface microstructure feature set, and road surface texture statistical feature set by dimensions, they are stitched together according to the feature dimensions to obtain the road surface image depth feature set.

[0026] Furthermore, in step S3, the vehicle-road interaction friction performance evaluation specifically includes the following steps:

[0027] Step S31: Construct a vehicle-road interaction friction performance evaluation model, specifically by constructing a vehicle-road interaction friction performance evaluation model using an improved Transformer algorithm; including steps S311 to S314:

[0028] Step S311: Design the model input layer. Specifically, firstly, the vehicle data and the road surface image depth features are fused through feature dimension concatenation to obtain vehicle-road interaction features. Then, the vehicle-road interaction features are mapped to the model hidden layer dimension through linear transformation to generate vehicle-road interaction dimension adaptation features. Finally, the vehicle-road interaction dimension adaptation features and the position encoding are added element-wise to obtain time-aligned vehicle-road interaction input features.

[0029] Step S312: Design the encoding processing layer, specifically, input the time-aligned vehicle-road interaction input features into the N-layer encoder. Each encoder layer sequentially performs multi-head probabilistic sparse self-attention calculation, residual connection and layer normalization, self-attention distillation operation, feedforward network and secondary normalization. After processing by the N-layer encoder, the core time-series features of vehicle-road interaction are obtained.

[0030] The multi-head probabilistic sparse self-attention computation specifically involves generating a query matrix, key matrix, and value matrix based on time-aligned vehicle-road interaction input features through a learnable weight matrix mapping. Then, the sparsity evaluation score of each query vector is calculated using the KL divergence formula. Based on the scores, the top u query vectors with the highest sparsity evaluation scores are selected to form a sparse query matrix. The sparse query matrix, key matrix, and value matrix are then split according to the number of attention heads, and each head independently calculates sparse attention. Finally, the outputs of all attention heads are concatenated along the feature dimension, and the weight matrix is ​​linearly transformed to obtain the encoder's multi-head sparse attention output.

[0031] The residual connection and layer normalization are specifically achieved by residually connecting the sparse attention output of the encoder multi-head to the original input features of the encoder layer, and then performing layer normalization on the residually connected features to obtain intermediate features for vehicle-road interaction.

[0032] The self-attention distillation operation specifically involves first extracting the local temporal dependencies of intermediate features of vehicle-road interaction through one-dimensional convolution, then enhancing the nonlinear expressive power of the features through the ELU activation function, and finally downsampling the features through max pooling to obtain vehicle-road interaction distilled features.

[0033] Step S313: Design the decoding processing layer, specifically by sequentially performing decoder input construction, masked multi-head probabilistic sparse self-attention calculation, and evaluation-specific feature extraction on the core temporal features of vehicle-road interaction to obtain vehicle-road interaction evaluation features; including steps S3131 to S3133:

[0034] Step S3131: Decoder input construction, specifically, extracting tk time steps of features from the core temporal features of vehicle-road interaction as the starting features, then setting the prediction step size, constructing zero-filled placeholders, and finally concatenating the starting features and the zero placeholders in the temporal dimension to obtain the decoder input features;

[0035] Step S3132: Calculate the multi-head probabilistic sparse self-attention of the mask, specifically by calculating the multi-head sparse attention of the decoder through the input features of the decoder, and obtaining the multi-head sparse attention output of the decoder;

[0036] Step S3133: Evaluation of dedicated feature extraction, specifically, the multi-head sparse attention output of the decoder is concatenated with the residual features of the decoder input features, and then layer normalization is performed on the features after residual concatenation to obtain vehicle-road interaction decoding optimization features, and the time step features of zero placeholder positions are extracted from them to obtain vehicle-road interaction evaluation features.

[0037] Step S314: Design the evaluation output layer, specifically by inputting the vehicle-road interaction evaluation features into the fully connected layer to generate initial evaluation values, and using the Min-Max scaling method to map the initial evaluation values ​​to a continuous range of 1 to 100 to obtain the friction performance evaluation output results;

[0038] Step S32: Evaluate the model training, specifically by using historical vehicle-road interaction data as training data for the evaluation model, and then training the model to obtain the trained vehicle-road interaction friction performance evaluation model.

[0039] Step S33: Real-time evaluation of vehicle-road interaction friction performance, specifically, inputting vehicle-road interaction data into the trained vehicle-road interaction friction performance evaluation model to obtain real-time vehicle-road interaction friction performance evaluation results.

[0040] Furthermore, in step S4, the dynamic evaluation of road surface friction performance specifically includes the following steps:

[0041] Step S41: Construct and train the road surface friction performance evaluation model. Specifically, construct the road surface friction performance evaluation model based on the bidirectional long short memory neural network, use the environmental data and historical vehicle-road interaction friction performance evaluation results in the historical friction performance evaluation data as the training data of the model, train the road surface friction performance evaluation model, and obtain the trained road surface friction performance evaluation model.

[0042] Step S42: Real-time assessment of road surface friction performance. Specifically, environmental data and real-time vehicle-road interaction friction performance assessment results from the real-time friction performance assessment data are input into the trained road surface friction performance assessment model to obtain the real-time dynamic assessment results of road surface friction performance. Based on the real-time dynamic assessment results of road surface friction performance, the current road surface performance is judged in real time, potential safety risks are promptly identified, and measures are taken.

[0043] The technical solution adopted by the present invention is as follows: The present invention provides an intelligent road surface friction performance evaluation system, including a multi-source data acquisition module, a road surface image depth feature extraction module, a vehicle-road interaction friction performance evaluation module, and a road surface friction performance dynamic evaluation module;

[0044] The multi-source data acquisition module obtains road surface friction evaluation optimization data through data acquisition and data optimization operations, and sends the data to the road surface image depth feature extraction module, the vehicle-road interaction friction performance evaluation module, and the road surface friction performance dynamic evaluation module.

[0045] The road surface image depth feature extraction module receives data sent by the multi-source data acquisition module, extracts road surface macroscopic structure features, road surface microscopic structure features and texture statistical features, and finally integrates the extracted features to obtain a road surface image depth feature set, and sends the data to the vehicle-road interaction friction performance evaluation module.

[0046] The vehicle-road interaction friction performance evaluation module receives data from the multi-source data acquisition module and the road surface image depth feature extraction module. Specifically, it constructs a vehicle-road interaction friction performance evaluation model by introducing an improved Transformer algorithm with a probabilistic sparse self-attention mechanism, a distillation mechanism encoder, and a one-time generation decoder. The model is trained using historical data, and finally, real-time data is input into the trained vehicle-road interaction friction performance evaluation model to obtain real-time vehicle-road interaction friction performance evaluation results. The data is then sent to the road surface friction performance dynamic evaluation module.

[0047] The road surface friction performance dynamic evaluation module receives data sent by the multi-source data acquisition module and the vehicle-road interaction friction performance evaluation module. Specifically, it constructs a road surface friction performance evaluation model based on a bidirectional long short memory neural network, trains the model, and performs real-time evaluation of road surface friction performance to obtain real-time dynamic evaluation results of road surface friction performance. Based on the results, corresponding safety measures are taken.

[0048] The beneficial effects achieved by adopting the above solution are as follows:

[0049] (1) In response to the technical problem in traditional pavement friction performance assessment that the joint analysis of vehicle behavior data, pavement condition data and environmental data ignores the direct interaction between vehicles and pavement, making it difficult to effectively separate the complex relationships between different data sources and ultimately leading to inaccurate pavement friction performance assessment results, this solution innovatively proposes to first conduct vehicle-road interaction friction performance assessment, and then combine the vehicle-road interaction results with environmental data for analysis. By first conducting friction performance assessment on the interaction between vehicles and pavement, the core friction performance characteristics of vehicle-road interaction can be captured first, and the main influencing factors of friction performance can be accurately extracted. Then, the environmental data is fused, which can structurally distinguish the direct impact of vehicle-road interaction from external environmental factors, avoid confusing the complex relationships between different data sources, enhance the hierarchy and logic of the analysis, improve the adaptability and flexibility of friction performance assessment, ensure accuracy under various road conditions and weather conditions, significantly improve the accuracy and practicality of pavement friction performance assessment, and achieve a more accurate and comprehensive pavement friction performance assessment.

[0050] (2) In view of the technical problems in the existing road surface image feature extraction methods, such as the separation of macroscopic structural features and microscopic texture features, which cannot fully cover all key features related to friction performance, and the weak resistance to external interference factors during the extraction process, resulting in low feature reliability and inability to cope with complex acquisition scenarios, leading to inaccurate vehicle-road interaction friction performance evaluation results, this solution innovatively adopts a multi-dimensional feature extraction method. It uses an active shape model to extract the macroscopic structural features of the road surface, uses an improved local binary texture method to extract the microscopic structural features of the road surface, and extracts multi-dimensional features of texture statistics. This solves the problem of the separation of macroscopic structural features and microscopic texture features in the traditional method, achieves full coverage of all key features of friction performance, effectively improves the resistance to external interference factors during the feature extraction process, significantly improves the reliability of features, improves the accuracy of vehicle-road interaction friction performance evaluation, and improves the accuracy and real-time performance of vehicle-road interaction friction performance evaluation results.

[0051] (3) In view of the technical problems of high computational complexity, low efficiency of long time series data processing and inability to effectively capture key time series features in the existing vehicle-road interaction friction performance evaluation model, which leads to inaccurate model output results, this solution innovatively adopts an improved Transformer algorithm by introducing a probabilistic sparse self-attention mechanism, a distillation mechanism encoder and a one-time generation decoder. The introduction of the probabilistic sparse self-attention mechanism significantly reduces computational complexity and improves the efficiency of long time series data processing by focusing on key time series intervals. The distillation mechanism encoder is designed to compress feature maps and improve the processing capability of time series data by using one-dimensional convolution and max pooling operations. The one-time generation decoder solves the problem of slow model result generation speed in long time series prediction, enhances the computational efficiency of the model, effectively improves the accuracy and stability of the model output results, and realizes efficient and accurate vehicle-road interaction friction performance evaluation. Attached Figure Description

[0052] Figure 1 A flowchart illustrating an intelligent road surface friction performance evaluation method provided by the present invention;

[0053] Figure 2 This invention provides a schematic diagram of a module for an intelligent road surface friction performance evaluation system.

[0054] Figure 3 This is a flowchart illustrating step S2;

[0055] Figure 4 This is a flowchart illustrating step S3;

[0056] Figure 5 This is a flowchart illustrating step S21;

[0057] Figure 6 This is a flowchart illustrating step S22;

[0058] Figure 7 This is a flowchart illustrating step S31;

[0059] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0060] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0061] In the description of this invention, it should be understood that the terms upper, lower, front, back, left, right, top, bottom, inner, and outer, etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the system or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0062] Example 1, see Figure 1 The technical solution adopted by the present invention is as follows: The present invention provides an intelligent method for evaluating road surface friction performance, which includes the following steps:

[0063] Step S1: Multi-source data acquisition, specifically, obtaining optimized road surface friction assessment data through data acquisition and data optimization operations;

[0064] Step S2: Road surface image depth feature extraction, used to extract road surface image depth features that affect friction performance from the road surface optimization image. Specifically, it involves extracting road surface macroscopic structure features, road surface microscopic structure features, and texture statistical features, and finally integrating the extracted features to obtain a road surface image depth feature set.

[0065] Step S3: Vehicle-road interaction friction performance evaluation, which is used to evaluate the friction performance between the vehicle and the road surface. Specifically, an improved Transformer algorithm with a probabilistic sparse self-attention mechanism, a distillation mechanism encoder and a one-time generation decoder is introduced to construct a vehicle-road interaction friction performance evaluation model. The model is trained with historical data, and finally, real-time data is input into the trained vehicle-road interaction friction performance evaluation model to obtain the real-time vehicle-road interaction friction performance evaluation results.

[0066] Step S4: Dynamic assessment of road surface friction performance. This step combines external environmental data and vehicle-road interaction friction performance assessment results to analyze the changes in road surface friction performance under different environmental conditions. It provides a comprehensive and dynamic assessment of road surface friction performance. Specifically, it constructs a road surface friction performance assessment model based on a bidirectional long short-term memory neural network. By simultaneously processing forward and reverse information, it understands the influencing factors of road surface friction performance, trains the model, and conducts real-time assessment of road surface friction performance to obtain real-time dynamic assessment results. Based on these results, corresponding safety measures are taken.

[0067] By performing the above operations, this solution addresses the technical problem in traditional pavement friction performance assessments where the joint analysis of vehicle behavior data, pavement condition data, and environmental data overlooks the direct interaction between vehicles and the pavement, making it difficult to effectively separate the complex relationships between different data sources and ultimately leading to inaccurate pavement friction performance assessment results. This innovative approach proposes first assessing the friction performance of vehicle-road interaction, and then combining the vehicle-road interaction results with environmental data for analysis. By first assessing the friction performance of the vehicle-road interaction, the core friction performance characteristics of vehicle-road interaction can be captured, and the main influencing factors of friction performance can be accurately extracted. Subsequently, the fusion of environmental data can structurally distinguish the direct impact of vehicle-road interaction from external environmental factors, avoiding confusion between the complex relationships between different data sources, enhancing the hierarchy and logic of the analysis, improving the adaptability and flexibility of friction performance assessment, ensuring accuracy under various road conditions and weather conditions, significantly improving the accuracy and practicality of pavement friction performance assessment, and achieving a more precise and comprehensive pavement friction performance assessment.

[0068] Example 2, see Figure 1 and Figure 2 This embodiment is based on the above embodiment. In step S1, the multi-source data acquisition specifically involves collecting data through the vehicle platform to obtain the original road surface friction assessment data, and performing data optimization operations on the original road surface friction assessment data to obtain optimized road surface friction assessment data.

[0069] The raw data for road surface friction assessment includes historical friction performance assessment data and real-time friction performance assessment data; both the historical friction performance assessment data and the real-time friction performance assessment data include road surface data, vehicle data, and environmental data.

[0070] The historical friction performance evaluation data also includes historical vehicle-road interaction friction performance evaluation results and historical road surface friction performance dynamic evaluation results.

[0071] The historical vehicle-road interaction friction performance evaluation result is a friction performance evaluation result obtained based on the interaction analysis of past vehicle behavior and road conditions, specifically a value between 1 and 100.

[0072] The historical road surface friction performance dynamic evaluation results are specifically divided into five levels according to different environments and road conditions: excellent, good, average, poor, and very poor.

[0073] The data optimization operations include image data optimization processing and structural data optimization processing;

[0074] The road surface data is collected through the vehicle's onboard cameras and sensors, including road surface image data and road surface condition data.

[0075] The road surface condition data includes road surface slippage, road surface icing status, road surface water accumulation, and road surface rut depth.

[0076] The vehicle data includes vehicle speed, driving behavior, and tire condition;

[0077] The environmental data includes temperature, humidity, rainfall, wind speed, and solar radiation intensity;

[0078] The image data optimization processing is used to optimize road surface image data to obtain an optimized road surface image. Specifically, it includes image denoising and image enhancement. The image denoising specifically removes random noise and background interference from the image using a median filtering algorithm, reducing interference caused by low light, shadows, or other external factors, and ensuring the clarity and effectiveness of the image. The image enhancement specifically enhances the visual features of the road surface image through contrast enhancement, brightness adjustment, and edge enhancement techniques, making the road surface texture features more prominent.

[0079] The structural data optimization processing is used to optimize road condition data, vehicle data, and environmental data, specifically including data cleaning and standardization processing.

[0080] The data cleaning process involves handling missing and outlier values ​​in structured data, including missing value imputation and outlier removal.

[0081] The missing value imputation specifically involves filling in missing values ​​using the mean imputation method; the outlier removal specifically involves detecting and removing extreme values ​​and logical outliers in the original data using the IQR method.

[0082] The standardization process is based on the minimum-max normalization method to standardize the data, and the one-hot encoding method is used to encode the category fields in the original data to give them a unified dimension and range.

[0083] Example 3, see Figure 1 , Figure 2 , Figure 3 , Figure 5 and Figure 6This embodiment is based on the above embodiment. In step S2, the extraction of the road surface image depth features specifically includes the following steps:

[0084] Step S21: Extraction of macroscopic structural features of the road surface, used to extract macroscopic structural features from the optimized road surface image, specifically including steps S211 to S214:

[0085] Step S211: Calculate the road surface shape weight parameters. Specifically, based on the iterative matching algorithm of the active shape model, key structural points are marked on the road surface optimization image. Using the marked key structural points as vertices, the road surface optimization image is divided into non-overlapping triangular structural units. The two-dimensional coordinate information of all key structural points is extracted in sequence and stitched together according to the extraction order to form a road surface shape vector representing the macroscopic structural morphology of the actual road surface. The standard mean shape vector of the road surface is introduced and combined with the preset road surface structure feature vector. The road surface shape weight parameters are obtained through feature vector projection operation.

[0086] The triangular structural unit is specifically divided using a triangulation algorithm;

[0087] The formula used is as follows:

[0088] ;

[0089] In the formula, Represents the road surface shape vector. This represents the standard mean shape vector of the road surface, which is a preset baseline parameter obtained through statistical modeling of historical road surface image samples. This represents the pavement structure feature vector, which is a preset parameter. It is extracted by principal component analysis of the standard mean shape vector of the pavement and is used to quantify the difference between the actual pavement and the standard mean shape. Represents the road surface shape weight parameter. The transpose matrix representing the pavement structure feature vectors;

[0090] Step S212: Road surface shape and texture feature fusion, used to eliminate redundant information in road surface shape and texture features, and enhance the comprehensive representation of the physical properties of the road surface structure. Specifically, it extracts grayscale texture information within each triangular structural unit using a bilinear interpolation algorithm, converts it into a one-dimensional grayscale texture vector, and then concatenates this grayscale texture vector with the road surface shape weight parameters in dimensional order to form a joint vector of the road surface macroscopic shape and texture. Then, it uses a principal component analysis algorithm to reduce the dimensionality of the joint vector of the road surface macroscopic shape and texture, obtaining a road surface shape and texture fusion feature that can simultaneously represent the location and physical properties of the road surface structure; the formula used is as follows:

[0091] ;

[0092] In the formula, This indicates the fusion characteristics of road surface shape and texture. The principal component matrix represents the joint vector of the road surface's macroscopic shape and texture. Using preset parameters, it is trained using PCA on the joint vector of the road surface's macroscopic shape and texture from historical road surface image samples. The variance contribution ratio of each principal component is calculated, a preset threshold is set, and the feature vectors corresponding to the principal components whose cumulative variance contribution ratio exceeds this threshold are selected to form the principal component matrix. Represents the principal component matrix The transpose of the matrix, This represents the joint vector of the macroscopic shape and texture of the road surface. The mean vector representing the joint vector of the macroscopic shape and texture of the road surface is a preset benchmark parameter used to eliminate baseline differences between different road surface samples;

[0093] Step S213: Road surface fusion feature edge enhancement, used to filter noise interference in the road surface shape and texture fusion features, and select high-confidence features that can truly reflect the macroscopic structure of the road surface. Specifically, the road surface shape and texture fusion features are subjected to translation-invariant wavelet transform to decompose and obtain low-frequency coefficients that represent the core information of the macroscopic structure of the road surface. Then, the low-frequency coefficients are subjected to pixel-by-pixel convolution operation by the edge detection operator to calculate the edge intensity value of each pixel coordinate. Subsequently, high-confidence pixel coordinates with edge intensity higher than a preset threshold are selected. The above pixel coordinates are structurally integrated based on spatial location correlation to obtain the road surface structure enhancement features.

[0094] The edge detection operator selection is as follows: a 3×3 kernel Sobel operator is used for linear pavement structures with cracks and joints to enhance the continuity of linear edges; a 3×3 kernel Laplacian operator is used for potholes and settlement pavement areas to enhance the edge contours; and a 3×3 kernel Prewitt operator is used for pavement corner transition structures. The formulas used are as follows:

[0095] ;

[0096] In the formula, Represents the surface shape and texture blending features in pixel coordinates Edge strength value, This represents the low-frequency coefficient, used to carry core information about the macroscopic structure of the road surface. , and These represent edge detection operators adapted to different road surface macrostructures, corresponding to the Sobel, Laplacian, and Prewitt operators, respectively. This represents the convolution operator;

[0097] Step S214: Output the road surface macrostructure feature set. Specifically, the road surface structure enhancement features are quantified and statistically analyzed to generate various structural sub-features. After normalizing each structural sub-feature, they are aligned and stitched together according to the feature dimensions to obtain the road surface macrostructure feature set.

[0098] The quantitative statistics specifically involve statistically analyzing the core characterization indicators of the cracks, potholes, joints, and overall pavement morphology structure information contained in the pavement structure enhancement features, and obtaining structural sub-features such as crack density, pothole coverage, joint smoothness, and overall pavement smoothness, thereby realizing the quantitative transformation of structural features.

[0099] The normalization process specifically involves using the min-max normalization algorithm to standardize and transform each structural sub-feature, mapping all sub-feature values ​​uniformly to the [0,1] interval, and eliminating the dimensional differences between different sub-features.

[0100] Step S22: Extraction of road surface microstructure features, used to accurately represent microtexture attributes such as road aggregate distribution and surface roughness; specifically, extracting road surface microstructure features using an improved local binary texture extraction algorithm; including steps S221 to S224:

[0101] Step S221: Road surface texture unit construction, specifically, using the optimized road surface image as the processing object, selecting 3×3 pixel blocks as road surface texture units, and defining the center pixel coordinates of each road surface texture unit as... Its grayscale value Then, the R neighboring pixels surrounding the center pixel within the road surface texture unit are marked sequentially in a clockwise direction;

[0102] The definition of the center pixel coordinates is specifically that the geometric center of a 3×3 pixel block is used as the unit center;

[0103] The grayscale value is specifically calculated through grayscale processing;

[0104] Step S222: Calculate the local binary code value, specifically the gray value of the center pixel. Using the baseline threshold, the gray values ​​of R neighboring pixels are binarized, and the binarization results are accumulated with weights to obtain the local binary code value of the cell, where i is the label number of the neighboring pixel; the formula used is as follows:

[0105] ;

[0106] ;

[0107] In the formula, This represents the unit step function, used to binarize the grayscale difference between neighboring pixels and the center pixel. When the difference between the grayscale values ​​of the neighboring pixels and the center pixel is... When the output is 1, it indicates that the position is convex relative to the center, such as the surface of the aggregate particles. When the output is 0, it indicates that the location is concave relative to the center, such as aggregate gaps or tiny grooves. This represents the local binary code value of the road surface texture unit at the center pixel coordinates. This represents the grayscale value of the i-th neighboring pixel. This represents the weight of the i-th neighboring pixel;

[0108] Step S223: Obtain cross-scale fusion feature vectors to eliminate redundant information between road surface micro-texture features and macro-shape features, fuse micro-texture and macro-shape features across scales, strengthen the correlation between micro-texture and macro-structure, and achieve a comprehensive representation of the cross-scale physical properties of the road surface; specifically, the local binary encoding values ​​of all road surface texture units are concatenated in sequence to form a high-dimensional micro-texture vector, and the micro-texture vector is concatenated with the road surface shape weight parameters in the order of feature dimensions to form a cross-scale joint vector, and the cross-scale joint vector is reduced in dimensionality by principal component analysis algorithm to obtain the cross-scale fusion feature vector;

[0109] ;

[0110] In the formula, This represents a cross-scale fused feature vector. Represents a cross-scale joint vector. The mean vector representing the cross-scale joint vector is a preset parameter obtained through statistical analysis of the cross-scale joint vectors of historical road surface image samples. Represents the principal component matrix The transpose of the matrix, The principal component matrix is ​​a preset parameter, obtained by PCA training on the cross-scale joint vector of historical road surface image samples.

[0111] Step S224: Output the road surface microstructure feature set, which is used to transform the cross-scale fused feature vector into structured microstructure feature data. Specifically, the min-max normalization algorithm is used to standardize all feature values ​​of the cross-scale fused feature vector to obtain the road surface microstructure feature set.

[0112] Step S23: Texture statistical feature extraction, used to extract second-order statistical information of road surface texture, supplementing the texture complexity, depth difference, and continuity statistical attributes not covered by macroscopic structural features and microscopic fusion features, adapting to the multi-dimensional evaluation needs of road surface texture; specifically, constructing a gray-level co-occurrence matrix according to pixel pair distance and angle, calculating the second-order statistics of pixel pairs through the gray-level co-occurrence matrix, extracting core texture features, and obtaining the road surface texture statistical feature set;

[0113] The distance between the pixel pairs is 1; the angle represents the relative direction between the pixel pairs, including... , , and Four perspectives; the core texture features include entropy, contrast, correlation, energy, variance, homogeneity, inverse moment, clustering shadow, and clustering saliency; entropy represents texture complexity, reflecting the degree of disorder in the texture structure, with a larger value indicating a more complex texture; contrast represents the difference in texture depth, reflecting the degree of grayscale difference between pixels, with a larger value indicating a coarser texture and a more obvious contrast; correlation represents texture continuity, describing the linear dependence of grayscale values ​​between adjacent pixels, reflecting the continuity of the texture structure; energy represents texture uniformity, reflecting the smoothness and uniformity of the texture, with a larger value indicating a smoother texture. The more regular and uniform the texture, the better; the variance is the degree of dispersion of texture grayscale values, used to reflect the degree of dispersion of texture grayscale values, and the larger the value, the more dispersed the texture grayscale distribution; the homogeneity is the local uniformity of texture, used to measure the degree to which the distribution of matrix elements is close to the diagonal, reflecting the local uniformity of texture; the inverse difference moment is the texture smoothness, used in the opposite way to contrast, and the larger the value, the smoother the texture; the clustering shadow is the skewness characteristic of texture grayscale distribution, used to characterize the symmetry of texture grayscale distribution, reflecting the skewness characteristic; the clustering salience is the degree of concentration of texture grayscale, used to describe the steepness of texture grayscale distribution, reflecting the concentration characteristic;

[0114] Step S24: Road surface image depth feature set integration, used to fuse three types of heterogeneous features: macroscopic structure, microscopic texture, and traditional statistics, to eliminate feature redundancy and dimensional conflicts, forming a comprehensive feature set that fully covers the macroscopic structure morphology, microscopic physical properties, and texture statistical laws of the road surface; specifically, after dimensionally aligning the road surface macroscopic structure feature set, road surface microscopic structure feature set, and road surface texture statistical feature set, they are stitched together according to the feature dimensions to obtain the road surface image depth feature set.

[0115] By performing the above operations, this solution addresses the technical problems of existing road surface image feature extraction methods, which suffer from the disconnect between macroscopic structural features and microscopic texture features, failing to comprehensively cover all key features related to friction performance, and exhibiting weak resistance to external interference factors during the extraction process, resulting in low feature reliability, inability to cope with complex acquisition scenarios, and inaccurate vehicle-road interaction friction performance evaluation results. This solution innovatively adopts a multi-dimensional feature extraction method, using an active shape model to extract macroscopic structural features of the road surface, an improved local binary texture method to extract microscopic structural features, and extracting multi-dimensional texture statistical features. This solves the problem of the disconnect between macroscopic structural and microscopic texture features in traditional methods, achieving comprehensive coverage of all key features of friction performance, effectively improving the resistance to external interference factors during feature extraction, significantly enhancing feature reliability, improving the accuracy of vehicle-road interaction friction performance evaluation, and improving the accuracy and real-time performance of the evaluation results.

[0116] Example 4, see Figure 1 , Figure 2 , Figure 4 and Figure 7 This embodiment is based on the above embodiment. In step S3, the vehicle-road interaction friction performance evaluation is used to evaluate the friction performance between the vehicle and the road surface, and specifically includes the following steps:

[0117] Step S31: Construct a vehicle-road interaction friction performance evaluation model, used to build a friction performance evaluation model based on the interaction between vehicle behavior and road surface state. Specifically, this involves constructing the vehicle-road interaction friction performance evaluation model using an improved Transformer algorithm. The improved Transformer algorithm is specifically improved by introducing a probabilistic sparse self-attention mechanism, a distillation mechanism encoder, and a one-time generation decoder. This includes steps S311 to S314.

[0118] Step S311: Design the model input layer to fuse and align vehicle behavior data with road surface image depth features, generating unified, temporally aligned input features that adapt to the model's hidden dimensions. Specifically, firstly, through feature dimension concatenation, vehicle data and road surface image depth features are fused to obtain vehicle-road interaction features. Then, a linear transformation is used to map the vehicle-to-infrastructure (V2I) interaction features to the hidden layer dimension of the model, generating V2I-adapted features. Finally, the V2I-adapted features are element-wise added to the location encoding to obtain the temporally aligned V2I input features. The temporal length representing the vehicle-to-infrastructure interaction characteristics. Dimensions representing vehicle-road interaction characteristics;

[0119] The location encoding is specifically generated using sine and cosine functions and uniquely bound to the temporal location, enhancing the model's understanding of temporal information;

[0120] Step S312: Design an encoding processing layer to perform deep temporal encoding on the time-aligned vehicle-road interaction input features to extract the core temporal features in the vehicle-road interaction process, filter redundant information, and compress and abstract the temporal dependencies; specifically, the time-aligned vehicle-road interaction input features are input into an N-layer encoder. Each encoder layer sequentially performs multi-head probabilistic sparse self-attention calculation, residual connection and layer normalization, self-attention distillation operation, feedforward network and secondary normalization. After processing by the N-layer encoder, the core temporal features of vehicle-road interaction are obtained.

[0121] The multi-head probabilistic sparse self-attention computation is used to solve the problems of redundancy and key information overload in long-term computations, focusing on the core interaction time sequence. Specifically, based on time-aligned vehicle-road interaction input features, a query matrix, key matrix, and value matrix are generated through a learnable weight matrix mapping. Then, the sparsity evaluation score of each query vector is calculated using the KL divergence formula. Based on the scores, the top u query vectors with the highest sparsity evaluation scores are selected to form a sparse query matrix. The sparse query matrix, key matrix, and value matrix are then split according to the number of attention heads, and each head independently calculates sparse attention. Finally, the outputs of all attention heads are concatenated along the feature dimension, and the weight matrix is ​​linearly transformed to obtain the encoder's multi-head sparse attention output. The formula used is as follows:

[0122] ;

[0123] ;

[0124] In the formula, This represents the sparsity evaluation score of the i-th query vector. This represents the i-th query vector in the query matrix. Represents the key matrix. This represents the j-th key vector in the key matrix. Indicates the hidden layer dimension. This indicates the multi-head sparse attention output of the encoder. Represents the feature concatenation function. Represents a sparse query matrix. This represents the transpose of the key matrix of the m-th attention head. This represents the submatrix containing the values ​​of the m-th attention head. This represents the concatenated linear transformation weight matrix. Indicates the number of attention heads;

[0125] The residual connection and layer normalization are used to address information loss and instability in training heterogeneous features, preserving core information and standardizing the distribution. Specifically, the sparse attention output of the encoder multi-head is residually connected to the original input features of the encoder layer, and then layer normalization is performed on the residually connected features to obtain intermediate features for vehicle-road interaction. ;

[0126] The self-attention distillation operation is used to address the issues of weak local dependencies and numerous redundant features, strengthening local correlations and compressing temporal length. Specifically, it first extracts the local temporal dependencies of intermediate features in vehicle-road interaction through one-dimensional convolution, then enhances the nonlinear expressive power of the features using the ELU activation function to adapt to the nonlinear changes in the friction coefficient in vehicle-road interaction, and finally downsamples the features using max pooling to obtain the distilled features of vehicle-road interaction. The formula used is as follows:

[0127] ;

[0128] In the formula, Indicates the characteristics of vehicle-road interactive distillation. This indicates a max pooling operation. This represents a one-dimensional convolution operation;

[0129] Specifically, the feedforward network and the second normalization are performed by using the feedforward network to perform deep feature enhancement on the vehicle-road interaction distillation features, and then performing a second-level normalization on the output of the feedforward network to obtain the single encoder layer vehicle-road interaction output features.

[0130] Step S313: Design a decoding processing layer to extract core temporal features of vehicle-road interaction based on the encoding processing layer and construct the decoding input. Specifically, this involves sequentially performing decoder input construction, masked multi-head probabilistic sparse self-attention calculation, and evaluation-specific feature extraction on the core temporal features of vehicle-road interaction to obtain vehicle-road interaction evaluation features; including steps S3131 to S3133:

[0131] Step S3131: Decoder input construction, which addresses the problem that traditional decoder inputs rely on only a single feature and lack anchoring for the dynamic temporal correlation of vehicle-road interaction. Specifically, it extracts features from the core temporal features of vehicle-road interaction at tk time steps as the starting features. Then set the prediction step size. Construct zero-filled placeholders Finally, the initial features are concatenated with the zero placeholders in the temporal dimension to obtain the decoder input features;

[0132] Step S3132: Mask multi-head probabilistic sparse self-attention calculation, used to solve the problem of future information leakage during decoder calculation, while focusing on key temporal features of vehicle-road interaction to reduce long temporal computation redundancy, and realize the accurate association between the core features of the coding layer and the decoding requirements. Specifically, the decoder multi-head sparse attention is calculated through the decoder input features to obtain the decoder multi-head sparse attention output.

[0133] The computational decoder multi-head sparse attention specifically involves generating a decoding query matrix by linear mapping based on the decoder input features, and generating a decoding key matrix and a decoding value matrix by linear mapping based on the core temporal features of vehicle-road interaction. Then, the sparsity evaluation score of each query vector in the decoding query matrix is ​​calculated using the KL divergence formula. Next, the top u query vectors with the highest sparsity evaluation scores are selected to form the decoding sparse query matrix. Subsequently, a lower triangular mask matrix is ​​constructed. Then, the decoding sparse query matrix, decoding key matrix, and decoding value matrix are split according to the number of attention heads. Each head is superimposed with the lower triangular mask matrix to independently calculate the sparse attention output. Finally, the outputs of all attention heads are concatenated along the feature dimension, and the multi-head sparse attention output of the decoder is obtained by linearly transforming the weight matrix.

[0134] The lower triangular mask matrix is ​​used to ensure that the decoder can only focus on the current and historical time points when calculating attention, and cannot access future time points corresponding to zero placeholders; the formula used is as follows:

[0135] ;

[0136] ;

[0137] In the formula, This represents the lower triangular mask matrix, with dimensions matching the temporal length of the decoder input. This represents the timing index of the decoder input, where i is the current timing point and j is the timing point to be accessed. This indicates that access to the current and historical time series points is permitted. This indicates that future time series points are being masked. This represents the output of the m-th decoding attention head. This represents decoding the sparse query matrix. This represents the transpose of the key decoding matrix of the m-th attention head. This represents the sub-decoding matrix of the m-th attention head;

[0138] Step S3133: Evaluate the dedicated feature extraction, which generates dedicated features focused on the current friction performance evaluation, replacing the traditional serial decoding, while avoiding future information leakage and long-term computational redundancy; specifically, the multi-head sparse attention output of the decoder is concatenated with the residual features of the decoder input features, and then layer normalization is performed on the concatenated features to obtain the vehicle-road interaction decoding optimization features, from which the time step features of the zero placeholder positions are extracted to obtain the vehicle-road interaction evaluation features; the formula used is as follows:

[0139] ;

[0140] In the formula, Indicates vehicle-road interaction evaluation characteristics, This indicates the optimized features for vehicle-to-infrastructure (V2I) interaction decoding.

[0141] Step S314: Design the evaluation output layer. Specifically, the vehicle-road interaction evaluation features are input into the fully connected layer. The high-dimensional features are mapped to 1-dimensional initial evaluation values ​​through linear transformation. The initial evaluation values ​​are then mapped to a continuous range of 1 to 100 using the Min-Max scaling method to obtain the friction performance evaluation output results.

[0142] Step S32: Evaluate the model training, specifically by using historical vehicle-road interaction data as training data for the evaluation model, and then training the model to obtain the trained vehicle-road interaction friction performance evaluation model.

[0143] The historical vehicle-road interaction data includes road surface condition data, vehicle data, and historical road surface image depth feature set from historical friction performance evaluation data.

[0144] The historical road surface image depth feature set is specifically obtained by extracting road surface image depth features from historical friction performance evaluation data;

[0145] The model training specifically involves using mean squared error as the loss function to optimize the gap between the model's prediction results and the true values. The parameters of the evaluation model are jointly optimized through the backpropagation algorithm. The loss function value is continuously monitored and optimized until it converges, at which point the model training is stopped, and the trained vehicle-road interaction friction performance evaluation model is obtained.

[0146] Step S33: Real-time evaluation of vehicle-road interaction friction performance, specifically, inputting vehicle-road interaction data into the trained vehicle-road interaction friction performance evaluation model to obtain real-time vehicle-road interaction friction performance evaluation results;

[0147] The real-time vehicle-road interaction data includes road surface condition data, vehicle data, and real-time road surface image depth feature set from the real-time friction performance evaluation data.

[0148] The real-time road surface image depth feature set is specifically obtained by extracting road surface image depth features from the real-time friction performance evaluation data.

[0149] By performing the above operations, this solution addresses the technical problems of high computational complexity, low efficiency in processing long-term data, and inability to effectively capture key temporal features in existing models for evaluating vehicle-road interaction friction performance, which lead to inaccurate model output results. This solution improves the Transformer algorithm by introducing a probabilistic sparse self-attention mechanism, a distillation mechanism encoder, and a one-time generation decoder. The introduction of the probabilistic sparse self-attention mechanism significantly reduces computational complexity and improves the efficiency of processing long-term data by focusing on key temporal intervals. The distillation mechanism encoder uses one-dimensional convolution and max pooling operations to compress feature maps and enhance the processing capability of temporal data. The one-time generation decoder solves the problem of slow model result generation in long-term prediction, enhances the model's computational efficiency, and effectively improves the accuracy and stability of the model output results, achieving efficient and accurate vehicle-road interaction friction performance evaluation.

[0150] Example 5, see Figure 1 and Figure 2 This embodiment is based on the above embodiment. In step S4, the dynamic evaluation of road surface friction performance is used to comprehensively and dynamically evaluate the road surface friction performance by combining external environmental data; specifically, it includes the following steps:

[0151] Step S41: Construct and train the road surface friction performance evaluation model. Specifically, construct the road surface friction performance evaluation model based on the bidirectional long short memory neural network, use the environmental data and historical vehicle-road interaction friction performance evaluation results in the historical friction performance evaluation data as the training data of the model, train the road surface friction performance evaluation model, and obtain the trained road surface friction performance evaluation model.

[0152] The road surface friction performance evaluation model training specifically uses cross-entropy loss as the loss function to optimize the gap between the model's prediction results and the true label distribution. The evaluation model parameters are jointly optimized through backpropagation algorithm. The loss function value is continuously monitored and optimized until it converges, at which point the model training is stopped, and the trained road surface friction performance evaluation model is obtained.

[0153] Step S42: Real-time evaluation of road surface friction performance. Specifically, environmental data and real-time vehicle-road interaction friction performance evaluation results from the real-time friction performance evaluation data are input into the trained road surface friction performance evaluation model to obtain the real-time dynamic evaluation results of road surface friction performance. Based on the real-time dynamic evaluation results of road surface friction performance, the current road surface performance is judged in real time, potential safety risks are detected in a timely manner and measures are taken.

[0154] The specific measures to be taken are as follows: If the assessment result is excellent, it means that the road surface friction performance is very good, and no additional safety measures are required; the vehicle can drive normally. If the assessment result is good, it means that the road surface friction performance is relatively good, and the driver is advised to maintain a normal speed. If the assessment result is average, it means that the road surface friction performance has a certain risk, and the driver is advised to reduce speed. If the assessment result is poor, it means that the road surface friction performance is poor and there is a high safety risk; the driver should be promptly warned of a speed limit and advised to drive cautiously. If the assessment result is very poor, it means that the road surface friction performance is extremely poor and there is an extremely high risk of skidding; a strong safety warning should be issued immediately, and the driver should be advised to immediately reduce speed, avoid high-speed driving, and possibly detour or close some road sections to ensure driving safety.

[0155] Example 6, see Figure 1 and Figure 2 Based on the above embodiments, the technical solution adopted by the present invention is as follows: The present invention provides an intelligent road surface friction performance evaluation system, including a multi-source data acquisition module, a road surface image depth feature extraction module, a vehicle-road interaction friction performance evaluation module, and a road surface friction performance dynamic evaluation module;

[0156] The multi-source data acquisition module obtains road surface friction evaluation optimization data through data acquisition and data optimization operations, and sends the data to the road surface image depth feature extraction module, the vehicle-road interaction friction performance evaluation module, and the road surface friction performance dynamic evaluation module.

[0157] The road surface image depth feature extraction module receives data sent by the multi-source data acquisition module, extracts road surface macroscopic structure features, road surface microscopic structure features and texture statistical features, and finally integrates the extracted features to obtain a road surface image depth feature set, and sends the data to the vehicle-road interaction friction performance evaluation module.

[0158] The vehicle-road interaction friction performance evaluation module receives data from the multi-source data acquisition module and the road surface image depth feature extraction module. Specifically, it constructs a vehicle-road interaction friction performance evaluation model by introducing an improved Transformer algorithm with a probabilistic sparse self-attention mechanism, a distillation mechanism encoder, and a one-time generation decoder. The model is trained using historical data, and finally, real-time data is input into the trained vehicle-road interaction friction performance evaluation model to obtain real-time vehicle-road interaction friction performance evaluation results. The data is then sent to the road surface friction performance dynamic evaluation module.

[0159] The road surface friction performance dynamic evaluation module receives data sent by the multi-source data acquisition module and the vehicle-road interaction friction performance evaluation module. Specifically, it constructs a road surface friction performance evaluation model based on a bidirectional long short memory neural network, trains the model, and performs real-time evaluation of road surface friction performance to obtain real-time dynamic evaluation results of road surface friction performance. Based on the results, corresponding safety measures are taken.

[0160] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "include," "contain," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.

[0161] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0162] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. An intelligent method for evaluating road surface friction performance, characterized in that: The method includes the following steps: Step S1: Multi-source data acquisition, specifically, obtaining optimized road surface friction assessment data through data acquisition and data optimization operations; Step S2: Road surface image depth feature extraction, used to extract road surface image depth features that affect friction performance from the optimized road surface image. Specifically, this involves extracting road surface macroscopic structure features, road surface microscopic structure features, and texture statistical features. Finally, the extracted features are integrated to obtain a road surface image depth feature set. The road surface macroscopic structure feature extraction includes calculating road surface shape weight parameters, road surface shape and texture feature fusion, road surface fusion feature edge enhancement, and outputting the road surface macroscopic structure feature set. Specifically, it includes the following steps: Step S211: Calculate the road surface shape weight parameters. Specifically, based on the iterative matching algorithm of the active shape model, key structural points are marked in the road surface optimization image. Using the marked key structural points as vertices, the road surface optimization image is divided into non-overlapping triangular structural units. The two-dimensional coordinate information of all key structural points is extracted in sequence and stitched together to form a road surface shape vector. The standard mean shape vector of the road surface is introduced and combined with the preset road surface structure feature vector. The road surface shape weight parameters are obtained through feature vector projection operation. Step S212: Road surface shape and texture feature fusion, specifically, extracting grayscale texture information within each triangular structural unit, converting it into a one-dimensional grayscale texture vector, then concatenating the grayscale texture vector with the road surface shape weight parameters in dimensional order to form a road surface macroscopic shape and texture joint vector, and then performing dimensionality reduction processing on the road surface macroscopic shape and texture joint vector through principal component analysis algorithm to obtain the road surface shape and texture fusion features. Step S213: Road surface fusion feature edge enhancement, specifically, performing translation-invariant wavelet transform on the road surface shape and texture fusion features to decompose and obtain low-frequency coefficients that characterize the core information of the road surface macrostructure. Then, performing pixel-by-pixel convolution operation on the low-frequency coefficients through the edge detection operator to calculate the edge intensity value of each pixel coordinate. Subsequently, high-confidence pixel coordinates with edge intensity higher than a preset threshold are selected. The above pixel coordinates are then structurally integrated based on spatial location correlation to obtain the road surface structure enhancement features. Step S214: Output the road surface macrostructure feature set. Specifically, the road surface structure enhancement features are quantified and statistically analyzed to generate various structural sub-features. After normalizing each structural sub-feature, they are aligned and stitched together according to the feature dimensions to obtain the road surface macrostructure feature set. Step S3: Vehicle-road interaction friction performance evaluation, which is used to evaluate the friction performance between the vehicle and the road surface. Specifically, an improved Transformer algorithm with a probabilistic sparse self-attention mechanism, a distillation mechanism encoder and a one-time generation decoder is introduced to construct a vehicle-road interaction friction performance evaluation model. The model is trained with historical vehicle-road interaction data. Finally, real-time data is input into the trained vehicle-road interaction friction performance evaluation model to obtain the real-time vehicle-road interaction friction performance evaluation result. The historical vehicle-road interaction data includes road surface condition data, vehicle data, and historical road surface image depth feature set from historical friction performance evaluation data. Step S4: Dynamic evaluation of road surface friction performance. This step combines external environmental data and vehicle-road interaction friction performance evaluation results to analyze the changes in road surface friction performance under different environmental conditions. It provides a comprehensive and dynamic evaluation of road surface friction performance. Specifically, it involves constructing a road surface friction performance evaluation model based on a bidirectional long short-term memory neural network, training the model, and finally conducting a real-time evaluation of road surface friction performance to obtain the real-time dynamic evaluation results of road surface friction performance. Based on these results, corresponding safety measures are taken.

2. The intelligent road surface friction performance evaluation method according to claim 1, characterized in that: In step S2, the extraction of depth features from the road surface image specifically includes the following steps: Step S21: Extraction of macroscopic structural features of the road surface; Step S22: Extraction of road surface microstructure features, specifically including steps S221 to S224: Step S221: Road surface texture unit construction, specifically, taking the optimized road surface image as the processing object, selecting a 3×3 pixel block as the road surface texture unit, defining the center pixel coordinates and grayscale value of each road surface texture unit, and then marking the R neighboring pixels surrounding the center pixel in the road surface texture unit in a clockwise direction. Step S222: Calculate the local binary code value. Specifically, with the gray value of the center pixel as the benchmark threshold, binarize the gray values ​​of R neighboring pixels respectively, and accumulate the binarization results with the weights to obtain the local binary code value of the unit. Step S223: Obtain the cross-scale fusion feature vector. Specifically, the local binary encoding values ​​of all road surface texture units are concatenated in order to form a high-dimensional micro-texture vector. This micro-texture vector is then concatenated with the road surface shape weight parameters in the order of feature dimensions to form a cross-scale joint vector. The cross-scale joint vector is then dimensionality-reduced using the principal component analysis algorithm to obtain the cross-scale fusion feature vector. Step S224: Output the pavement microstructure feature set, specifically by standardizing all feature values ​​of the cross-scale fused feature vector to obtain the pavement microstructure feature set; Step S23: Texture statistical feature extraction, specifically, constructing a gray-level co-occurrence matrix based on pixel pair distance and angle, calculating the second-order statistics of pixel pairs through the gray-level co-occurrence matrix, extracting core texture features, and obtaining the road surface texture statistical feature set; Step S24: Road surface image depth feature set integration, specifically, after aligning the road surface macrostructure feature set, road surface microstructure feature set, and road surface texture statistical feature set by dimensions, they are stitched together according to the feature dimensions to obtain the road surface image depth feature set.

3. The intelligent road surface friction performance evaluation method according to claim 1, characterized in that: In step S3, the vehicle-road interaction friction performance evaluation specifically includes the following steps: Step S31: Construct a vehicle-road interaction friction performance evaluation model; Step S32: Evaluate the model training, specifically by using historical vehicle-road interaction data as training data for the evaluation model, and then training the model to obtain the trained vehicle-road interaction friction performance evaluation model. Step S33: Real-time evaluation of vehicle-road interaction friction performance, specifically, inputting vehicle-road interaction data into the trained vehicle-road interaction friction performance evaluation model to obtain real-time vehicle-road interaction friction performance evaluation results.

4. The intelligent road surface friction performance evaluation method according to claim 3, characterized in that: In step S31, the construction of the vehicle-road interaction friction performance evaluation model specifically includes the following steps: Step S311: Design the model input layer. Specifically, firstly, the vehicle data and the road surface image depth features are fused through feature dimension concatenation to obtain vehicle-road interaction features. Then, the vehicle-road interaction features are mapped to the model hidden layer dimension through linear transformation to generate vehicle-road interaction dimension adaptation features. Finally, the vehicle-road interaction dimension adaptation features and the position encoding are added element-wise to obtain time-aligned vehicle-road interaction input features. Step S312: Design the encoding processing layer; Step S313: Design the decoding processing layer, specifically by sequentially performing decoder input construction, masked multi-head probabilistic sparse self-attention calculation, and evaluation-specific feature extraction on the core temporal features of vehicle-road interaction to obtain vehicle-road interaction evaluation features; including steps S3131 to S3133: Step S3131: Decoder input construction, specifically, extracting tk time steps of features from the core temporal features of vehicle-road interaction as the starting features, then setting the prediction step size, constructing zero-filled placeholders, and finally concatenating the starting features and the zero placeholders in the temporal dimension to obtain the decoder input features; Step S3132: Calculate the multi-head probabilistic sparse self-attention of the mask, specifically by calculating the multi-head sparse attention of the decoder through the input features of the decoder, and obtaining the multi-head sparse attention output of the decoder; Step S3133: Evaluation of dedicated feature extraction, specifically, the multi-head sparse attention output of the decoder is concatenated with the residual features of the decoder input features, and then layer normalization is performed on the features after residual concatenation to obtain vehicle-road interaction decoding optimization features, and the time step features of zero placeholder positions are extracted from them to obtain vehicle-road interaction evaluation features. Step S314: Design the evaluation output layer. Specifically, input the vehicle-road interaction evaluation features into the fully connected layer to generate initial evaluation values. Then, use the Min-Max scaling method to map the initial evaluation values ​​to a continuous range of 1 to 100 to obtain the friction performance evaluation output results.

5. The intelligent road surface friction performance evaluation method according to claim 4, characterized in that: In step S312, the design of the coding processing layer specifically involves inputting the time-aligned vehicle-road interaction input features into an N-layer encoder. Each encoder layer sequentially performs multi-head probabilistic sparse self-attention calculation, residual connection and layer normalization, self-attention distillation operation, feedforward network and secondary normalization. After being processed sequentially by the N-layer encoder, the core time-series features of vehicle-road interaction are obtained. The multi-head probabilistic sparse self-attention computation specifically involves generating a query matrix, key matrix, and value matrix based on time-aligned vehicle-road interaction input features through a learnable weight matrix mapping. Then, the sparsity evaluation score of each query vector is calculated using the KL divergence formula. Based on the scores, the top u query vectors with the highest sparsity evaluation scores are selected to form a sparse query matrix. The sparse query matrix, key matrix, and value matrix are then split according to the number of attention heads, and each head independently calculates sparse attention. Finally, the outputs of all attention heads are concatenated along the feature dimension, and the weight matrix is ​​linearly transformed to obtain the encoder's multi-head sparse attention output. The residual connection and layer normalization are specifically achieved by residually connecting the sparse attention output of the encoder multi-head to the original input features of the encoder layer, and then performing layer normalization on the residually connected features to obtain intermediate features for vehicle-road interaction. The self-attention distillation operation specifically involves first extracting the local temporal dependencies of intermediate features in vehicle-road interaction through one-dimensional convolution, then enhancing the nonlinear expressive power of the features through the ELU activation function, and finally downsampling the features through max pooling to obtain the distilled features of vehicle-road interaction.

6. The intelligent road surface friction performance evaluation method according to claim 1, characterized in that: In step S4, the dynamic evaluation of road surface friction performance specifically includes the following steps: Step S41: Construct and train the road surface friction performance evaluation model. Specifically, construct the road surface friction performance evaluation model based on the bidirectional long short memory neural network, use the environmental data and historical vehicle-road interaction friction performance evaluation results in the historical friction performance evaluation data as the training data of the model, train the road surface friction performance evaluation model, and obtain the trained road surface friction performance evaluation model. Step S42: Real-time assessment of road surface friction performance. Specifically, environmental data and real-time vehicle-road interaction friction performance assessment results from the real-time friction performance assessment data are input into the trained road surface friction performance assessment model to obtain the real-time dynamic assessment results of road surface friction performance. Based on the real-time dynamic assessment results of road surface friction performance, the current road surface performance is judged in real time, potential safety risks are promptly identified, and measures are taken.

7. The intelligent road surface friction performance evaluation method according to claim 1, characterized in that: In step S1, the multi-source data acquisition specifically involves obtaining the original road surface friction assessment data through data acquisition operations, and performing data optimization operations on the original road surface friction assessment data to obtain optimized road surface friction assessment data. The raw data for road surface friction assessment includes historical friction performance assessment data and real-time friction performance assessment data; both the historical friction performance assessment data and the real-time friction performance assessment data include road surface data, vehicle data, and environmental data. The historical friction performance evaluation data also includes historical vehicle-road interaction friction performance evaluation results and historical road surface friction performance dynamic evaluation results; the data optimization operations include image data optimization processing and structural data optimization processing.

8. An intelligent road surface friction performance evaluation system, used to implement the intelligent road surface friction performance evaluation method as described in any one of claims 1-7, characterized in that: It includes a multi-source data acquisition module, a road surface image depth feature extraction module, a vehicle-road interaction friction performance evaluation module, and a road surface friction performance dynamic evaluation module; The multi-source data acquisition module obtains road surface friction evaluation optimization data through data acquisition and data optimization operations, and sends the data to the road surface image depth feature extraction module, the vehicle-road interaction friction performance evaluation module, and the road surface friction performance dynamic evaluation module. The road surface image depth feature extraction module receives data sent by the multi-source data acquisition module, extracts road surface macroscopic structure features, road surface microscopic structure features and texture statistical features, and finally integrates the extracted features to obtain a road surface image depth feature set, and sends the data to the vehicle-road interaction friction performance evaluation module. The vehicle-road interaction friction performance evaluation module receives data from the multi-source data acquisition module and the road surface image depth feature extraction module. Specifically, it constructs a vehicle-road interaction friction performance evaluation model by introducing an improved Transformer algorithm with a probabilistic sparse self-attention mechanism, a distillation mechanism encoder, and a one-time generation decoder. The model is trained using historical data, and finally, real-time data is input into the trained vehicle-road interaction friction performance evaluation model to obtain real-time vehicle-road interaction friction performance evaluation results. The data is then sent to the road surface friction performance dynamic evaluation module. The road surface friction performance dynamic evaluation module receives data sent by the multi-source data acquisition module and the vehicle-road interaction friction performance evaluation module. Specifically, it constructs a road surface friction performance evaluation model based on a bidirectional long short memory neural network, trains the model, and performs real-time evaluation of road surface friction performance to obtain real-time dynamic evaluation results of road surface friction performance. Based on the results, corresponding safety measures are taken.

Images