Road surface wave rapid identification model construction method, electronic device and storage medium

Abstract

The invention relates to a method for constructing a road wave rapid identification model, an electronic device and a storage medium, and belongs to the technical field of road engineering. In order to solve the problem of rapid intelligent identification of road wave diseases, the invention includes collecting road images for compression processing to obtain compressed road images; labeling to obtain a data set of a road wave rapid identification model; constructing a basic model, a first auxiliary model, a second auxiliary model and an enhanced model for rapid identification of road waves; constructing a loss function between the basic model, the first auxiliary model, the second auxiliary model and the output layer of the enhanced model for rapid identification of road waves, constructing an overall loss function, and then constructing a model optimization method based on the overall loss function, and enhancing and improving the optimized basic model obtained through training to obtain a final road wave rapid identification model; constructing a digital twin model based on the final road wave rapid identification model for updating road wave diseases. The invention improves road safety.

Description

Technical Field

[0001] The invention belongs to the technical field of road engineering, and in particular relates to a method for constructing a road surface wave rapid identification model, an electronic device and a storage medium. Background Art

[0002] Pavement wave disease usually occurs when the asphalt pavement undergoes plastic deformation under the action of long-term vehicle loads, or when the base layer is not compacted enough and drainage is poor. Construction quality problems such as improper material mix design, improper paving and compaction operations during construction can also lead to pavement wave disease. The formation of this disease is the result of the interaction between the pavement structure and material mechanical properties and external loads.

[0003] Road surface wave disease will lead to uneven driving, affect ride comfort, reduce vehicle handling stability, and increase safety risks. When vehicles pass through wave sections, they will generate additional dynamic loads, which will accelerate the damage of the road surface. In addition, wave disease will affect the drainage effect, easily forming water accumulation on the road surface, further deteriorating the road conditions.

[0004] Rapid identification of pavement wave damage can be used to quickly discover and assess the severity of wave damage, so as to formulate timely and effective maintenance and repair plans, thereby extending the service life of the road, ensuring the safety of road users, while reducing maintenance costs and improving road management efficiency.

[0005] The invention patent with application number 201210127649.0 and invention name “A design and laying method for a speed-suppressing wave asphalt pavement” proposes a design and laying method for a wave asphalt pavement for traffic speed control needs, which aims to naturally encourage drivers to reduce their driving speed through the wave shape of the pavement, and is suitable for various types of roads with dense traffic and speed limits. However, the invention does not consider the rapid and intelligent identification of pavement wave diseases, resulting in the inability to detect wave diseases in time, affecting road safety and driving comfort, aggravating pavement damage, and leading to excessive maintenance costs and inefficiency in road maintenance management. Summary of the invention

[0006] The problem to be solved by the present invention is the rapid and intelligent identification of road wave diseases, and a method for constructing a rapid road wave identification model, an electronic device and a storage medium are proposed.

[0007] To achieve the above object, the present invention is implemented through the following technical solutions:

[0008] A method for constructing a road surface wave rapid identification model comprises the following steps:

[0009] S1. Using a vehicle-mounted three-dimensional structured light camera to collect road images, and then compressing the collected road images to obtain a compressed road image;

[0010] S2. The compressed road surface image obtained in step S1 is annotated to obtain a data set for a road surface wave rapid recognition model;

[0011] S3. Construct the basic model, the first auxiliary model, the second auxiliary model and the enhanced model for rapid identification of road surface waves;

[0012] S4. Construct the loss function among the basic model, the first auxiliary model, the second auxiliary model and the output layer of the enhanced model for rapid identification of road surface waves;

[0013] S5. construct an overall loss function based on the loss function between the basic model, the first auxiliary model, the second auxiliary model and the output layer of the enhanced model obtained in step S4, and then construct a model optimization method based on the overall loss function, and obtain an optimized basic model after training;

[0014] S6. The optimized basic model obtained in step S5 is enhanced and improved to obtain a final road wave rapid recognition model;

[0015] S7. Construct a digital twin model based on the final pavement wave rapid identification model for updating pavement wave damage.

[0016] Furthermore, the specific implementation method of step S1 includes the following steps:

[0017] S1.1. Use a vehicle-mounted 3D structured light camera to collect road images, and control the vehicle speed within 70 km / h during the collection process;

[0018] S1.2. compressing the collected road surface image by using Huffman coding to obtain a compressed road surface image;

[0019] S1.3. Perform a plane projection on the compressed road surface image, and obtain a projected road surface image which is a compressed road surface image.

[0020] Furthermore, the specific implementation method of step S2 includes the following steps:

[0021] S2.1. By manually labeling, the compressed pavement image obtained in step S1 is framed to select waves and three types of defects, namely, peeling, rutting, and bulging. The waves are labeled as Wave, and the other three types of defects are labeled as other1, other2, and other3, respectively, to obtain a data set for the pavement wave rapid recognition model.

[0022] S2.2. Divide the dataset of the road wave rapid recognition model into a training set and a test set in a ratio of 8:2.

[0023] Furthermore, the specific implementation method of step S3 includes the following steps:

[0024] S3.1. Constructing a basic model for rapid identification of road surface waves:

[0025] S3.1.1. The basic model A consists of 1 input layer, 9 convolutional layers, 9 activation layers, 3 pooling layers, 1 merging layer, 1 upsampling layer, 1 path aggregation network layer PANet, 1 pooling splicing layer, 1 spatial pyramid pooling layer SPP, and 1 output layer. The specific network structure is the following structures connected in sequence;

[0026] S3.1.2. Input the training set obtained in step S2 into the basic model A, calculate the weights and biases of the neural network units through the error back propagation algorithm, complete the training of the basic model A, and obtain the trained basic model A;

[0027] The calculation formula of the loss function of the basic model A is:

[0028] ;

[0029] Among them, La is the loss function of the basic model A; Na is the total number of data samples in the basic model A; For the it The true labels of samples, Based on the prediction of model A it The probability that a sample is class 1;

[0030] S3.1.3. The node output values ​​of the four output units corresponding to the basic model A are a 1. a 2. a 3. a 4. Perform initial conversion on the node output values ​​of the four output units to obtain the output values The corresponding probability value The calculation formula is:

[0031] ;

[0032] Then the node output values ​​of the four output units are further optimized and converted to obtain the output value The corresponding enhancement probability value The calculation formula is:

[0033] ;

[0034] in, G is the enhancement coefficient, and then the actual disease corresponding to each node is marked as ;

[0035] S3.1.4. Adoption The training set obtained in step S2 is combined with the error chain derivation rule to train the basic model A. After the training is completed, the output values ​​of the four types of diseases corresponding to the basic model A are obtained, which are recorded as , , , ;

[0036] S3.2. Constructing the first auxiliary model for rapid identification of road surface waves:

[0037] S3.2.1. The first auxiliary model B consists of 1 input layer, 9 convolutional layers, 9 activation layers, 5 pooling layers, 1 merging layer, 1 upsampling layer, 1 PANet layer, 1 pooling splicing layer, 1 SPP layer, and 1 output layer. The specific network structure is the following structures connected in sequence;

[0038] S3.2.2. Input the training set obtained in step S2 into the first auxiliary model B, calculate the weights and biases of the neural network units through the error back propagation algorithm, complete the training of the first auxiliary model B, and obtain the trained first auxiliary model B;

[0039] The calculation formula of the loss function of the first auxiliary model B is:

[0040] ;

[0041] Wherein, Lb is the loss function of the first auxiliary model B; Nb is the total number of data samples in the first auxiliary model B; For the ib The true labels of samples, The first auxiliary model B predicts ib The probability that a sample is class 1;

[0042] S3.2.3. The node output values ​​of the four output units corresponding to the first auxiliary model B are b 1. b 2. b 3. b 4. Perform initial conversion on the node output values ​​of the four output units to obtain the output values b i The corresponding probability value The calculation formula is:

[0043] ;

[0044] Then the node output values ​​of the four output units are further optimized and converted to obtain the output value b i The corresponding enhancement probability value The calculation formula is:

[0045] ;

[0046] S3.2.4. Adoption The training set obtained in step S2 is combined with the error chain derivation rule to train the first auxiliary model B. After the training is completed, the disease output value corresponding to the first auxiliary model B is obtained as , among the 4 nodes in the output layer, the actual disease corresponding to each node is marked as , i =1,2,3,4;

[0047] S3.3. Constructing the second auxiliary model for rapid identification of road surface waves:

[0048] S3.3.1. The second auxiliary model C consists of 1 input layer, 12 convolutional layers, 12 activation layers, 5 pooling layers, 1 merging layer, 1 upsampling layer, 1 PANet layer, 1 pooling splicing layer, 1 SPP layer, and 1 output layer. The specific network structure is the following structures connected in sequence;

[0049] S3.3.2. Input the training set obtained in step S2 into the second auxiliary model C, calculate the weights and biases of the neural network units through the error back propagation algorithm, complete the training of the second auxiliary model C, and obtain the trained second auxiliary model C;

[0050] The calculation formula of the loss function of the second auxiliary model C is:

[0051] ;

[0052] Wherein, Lc is the loss function of the second auxiliary model C; Nc is the total number of data samples in the second auxiliary model C; For the ic The true labels of samples, The second auxiliary model C predicts ic The probability that a sample is class 1;

[0053] S3.3.3. The node output values ​​of the four output units corresponding to the second auxiliary model C are , the node output values ​​of the four output units are initially converted to obtain The corresponding probability value The calculation formula is:

[0054] ;

[0055] Then the node output values ​​of the four output units are further optimized and converted to obtain the output value The corresponding enhancement probability value The calculation formula is:

[0056] ;

[0057] in, G is the enhancement coefficient, and then the actual disease corresponding to each node is marked as ;

[0058] S3.3.4. Adoption The training set obtained in step S2 is combined with the error chain derivation rule to train the second auxiliary model C. After the training is completed, the output values ​​of the four types of diseases corresponding to the second auxiliary model C are obtained as follows: ;

[0059] S3.4. Constructing an enhanced model for rapid identification of road waves:

[0060] S3.4.1. The enhanced model D consists of 1 input layer, 14 convolutional layers, 14 activation layers, 5 pooling layers, 1 merging layer, 1 upsampling layer, 1 PANet layer, 1 pooling splicing layer, 1 SPP layer, and 1 output layer. The specific network structure is the following structures connected in sequence;

[0061] S3.4.2. Input the training set obtained in step S2 into the enhanced model D, calculate the weights and biases of the neural network units through the error back propagation algorithm, complete the training of the enhanced model D, and obtain the trained enhanced model D;

[0062] The loss function of the enhanced model D is calculated as:

[0063] ;

[0064] Where Ld is the loss function of the enhanced model D; Nd is the total number of data samples in the enhanced model D; For the id The true labels of samples, To strengthen the model D to predict the id The probability that a sample is class 1;

[0065] S3.4.3. The node output values ​​of the four output units corresponding to the enhanced model D are , the node output values ​​of the four output units are initially converted to obtain The corresponding probability value The calculation formula is:

[0066] ;

[0067] Then the node output values ​​of the four output units are further optimized and converted to obtain the output value The corresponding enhancement probability value The calculation formula is:

[0068] ;

[0069] in, G is the enhancement coefficient, and then the actual disease corresponding to each node is marked as ;

[0070] S3.4.4. Adoption The training set obtained in step S2 is combined with the error chain derivation rule to train the enhanced model D. After the training is completed, the output values ​​of the four types of diseases corresponding to the enhanced model D are obtained as follows: .

[0071] Furthermore, the specific implementation method of step S4 includes the following steps:

[0072] S4.1. Based on and , establish the loss function between the first auxiliary model B and the output layer of the basic model A, and the calculation formula is:

[0073] ;

[0074] based on and , construct the output value The corresponding probability value The loss function of the base model A , the calculation formula is:

[0075] ;

[0076] based on and , establish the relationship loss function between the output layer of the first auxiliary model B and the basic model A , the calculation formula is:

[0077]

[0078] in, for The weight coefficient of for The weight coefficient of

[0079] S4.2. Establish the relationship loss function between the output layer of the second auxiliary model C and the first auxiliary model B , the expression is:

[0080] ;

[0081] in, for The weight coefficient of for The weight coefficient of Based on and The loss function of Based on and The loss function of

[0082] The calculation formula is:

[0083] ;

[0084] The calculation formula is:

[0085] ;

[0086] S4.3. Establish the relationship loss function between the output layer of the second auxiliary model C and the basic model A , the expression is:

[0087] ;

[0088] in, for The weight coefficient of for The weight coefficient of Based on and The loss function is Based on and The loss function of

[0089] The calculation formula is:

[0090] ;

[0091] The calculation formula is:

[0092] ;

[0093] S4.4. Establishing the relationship loss function between the output layer of the enhanced model D and the base model A , the expression is:

[0094] ;

[0095] in, for The weight coefficient of for The weight coefficient of Based on and The loss function of

[0096] The calculation formula is:

[0097] ;

[0098] S4.5. Establish the relationship loss function between the output layer of the reinforcement model D and the second auxiliary model C , the expression is:

[0099] ;

[0100] in, for The weight coefficient of for The weight coefficient of Based on and The loss function is Based on and The loss function of

[0101] The calculation formula is:

[0102] ;

[0103] The calculation formula is:

[0104] .

[0105] Furthermore, the specific implementation method of step S5 includes the following steps:

[0106] S5.1. Define the corresponding overall loss function between the basic model A, the first auxiliary model B, the second auxiliary model C, and the enhanced model D The calculation formula is:

[0107] ;

[0108] in, , , , , They are , , , , The weight coefficient of , , , , The value of is 0 or 1;

[0109] By changing , , , , The value of The overall loss function In the form of, thus obtaining the basic model A with 32 parameters;

[0110] S5.2. Using the data set of the road wave rapid recognition model obtained in step S2, the model is trained based on the chain rule and by error back propagation to obtain 32 basic models A;

[0111] Set the accuracy of 32 basic models A to be , i=1,2,3,…,32, the calculation formula is:

[0112] ;

[0113] in, Represents the first i The number of samples that are predicted to be cracks and are actually cracks when the model is used to identify the data in the test set; Represents the first i The number of samples that a model predicts as cracks but is not actually cracks when identifying the data in the test set;

[0114] Comparison The maximum value in ;

[0115] The corresponding basic model A is the basic model after training and optimization.

[0116] Furthermore, the specific implementation method of step S6 includes the following steps:

[0117] S6.1. For the optimized basic model obtained in step S5, the enhancement coefficient G is set to be: G=1, 3, 5, 7, 9, 11, 15, 20, 30, 40, 50;

[0118] S6.2. Based on different enhancement coefficients G, 11 working conditions are obtained, and the optimized basic model is trained to obtain 11 models. The accuracy rates corresponding to the 11 optimized basic models are set as follows: , i=1,2,3,…,11;

[0119] Comparison The maximum value in ;

[0120] The corresponding optimized basic model is the final road wave rapid identification model.

[0121] Furthermore, the specific implementation method of step S7 includes the following steps:

[0122] S7.1. Export the final road wave rapid identification model obtained in step S6 as a deployment file for independent services, including one of a pt file, a pth file, a json file, and a txt file;

[0123] S7.2. Construct a road information model;

[0124] S7.2.1. Collect information on the location, length, width, and layout of roads;

[0125] S7.2.2. The data obtained in step S7.2.1 is stored in a database, and a basic three-dimensional model of the road is constructed, wherein the basic three-dimensional model of the road includes geometric information of the road;

[0126] S7.2.3. Based on step S7.2.2, material properties, design specifications, and historical maintenance data are added to the basic three-dimensional model of the road to obtain a road information model;

[0127] S7.2.4. Export the road information model to a format file supported by the web application, including a json file and a js file;

[0128] S7.3. In the digital twin platform, import the deployment file of the final road surface wave rapid identification model obtained in step S7.1, and import the road information model format file obtained in step S7.2;

[0129] S7.4. Use Python to construct an API interface between the image data and the final road surface wave rapid recognition model for intelligent rutting identification. Use Python to construct an API interface between the final road surface wave rapid recognition model and the road information model to complete the construction of a digital twin model for updating the road surface rutting status based on the final road surface wave rapid recognition model.

[0130] An electronic device comprises a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of a method for constructing a road surface wave rapid identification model when executing the computer program.

[0131] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements a method for constructing a road surface wave rapid identification model.

[0132] Beneficial effects of the present invention:

[0133] The invention discloses a method for constructing a road wave rapid recognition model. Road wave is a continuous undulation caused by uneven road surface, which is usually manifested as relatively regular concave-convex morphology and shadow changes. In combination with the characteristics of waves, the model structure is designed to include an input layer, a convolution layer, an activation layer, a pooling layer, a merging layer, an upsampling layer, a PANet layer, a pooling splicing layer, an SPP layer, and an output layer. The functions of each structural layer are as follows: the input layer receives the road surface image and provides the model with raw data, including the concave-convex texture of the wave and its surrounding environment information. The convolution layer extracts visual features of different levels through multiple filters, including the edge, shape and texture of the wave, which is particularly important for periodic wave morphology. The activation layer introduces nonlinearity, so that the model can learn more complex feature combinations, which helps to distinguish waves from ordinary road surface textures. The pooling layer reduces the dimension of the feature map, reduces the amount of calculation, and improves the robustness of the model to small displacements and deformations (such as small undulations of waves). The merging layer is used to fuse feature maps of different scales, which is very important for identifying multi-scale wave undulation features. The upsampling layer is used to restore the spatial resolution of the feature map to ensure that the position information of the waves is accurately represented in the final output. The PANet layer improves the ability to recognize the size and shape of road waves by strengthening the contextual information transmission of the feature pyramid. The pooling splicing layer combines the pooling outputs of different sizes, allowing the model to capture wave features from small to large, which is crucial for processing road waves under different conditions. The SPP layer (spatial pyramid pooling) enhances the model's global understanding of wave texture by extracting multi-level spatial features, which is beneficial for identifying road waves of different wavelengths and wave heights. The output layer maps complex features into final judgment results, such as the classification probability of whether each pixel belongs to the wave area, ensuring accurate recognition of road waves.

[0134] The present invention discloses a method for constructing a road wave rapid identification model, and proposes a multi-model loss function combination improvement method to enable the basic model to obtain performance improvement in areas that are difficult to identify or prone to errors.

[0135] The method for constructing a road wave rapid identification model described in the present invention proposes an optimal basic model optimization method to ensure that the performance improvement effect of the basic model is optimal.

[0136] The invention discloses a method for constructing a road wave rapid identification model, and proposes a basic model improvement method based on an enhancement coefficient to enhance the generalization ability of the wave disease identification model.

[0137] The present invention discloses a method for constructing a road wave rapid identification model, and proposes a method for constructing a road digital twin model and updating road wave damage, so as to realize the construction of a digital twin model and the accurate update of the wave damage area. The integration of digital twin technology and the method for rapid identification of road wave damage can realize instant and accurate diagnosis of road conditions by combining high-precision digital models with real-time data. By integrating sensor data, image recognition technology and machine learning models, it is not only possible to detect and locate road wave damage in a timely manner, but also to predict future damage development, thereby optimizing maintenance plans and resource allocation. This helps to improve road safety, reduce maintenance costs, and extend the service life of roads, and is of great value in improving public safety and economic efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0138] Figure 1 The present invention provides a flow chart of a method for constructing a road surface wave rapid identification model. DETAILED DESCRIPTION

[0139] In order to make the purpose, technical solution and advantages of the present invention more clear, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention, that is, the specific embodiments described are only part of the embodiments of the present invention, rather than all of the specific embodiments. The components of the specific embodiments of the present invention described and shown in the drawings herein can be arranged and designed in various different configurations, and the present invention can also have other embodiments.

[0140] Therefore, the following detailed description of the specific embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents the selected specific embodiments of the present invention. Based on the specific embodiments of the present invention, all other specific embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present invention.

[0141] In order to further understand the content, features and effects of the present invention, the following specific implementation methods are given as examples, and the attached Figure 1 The detailed instructions are as follows:

[0142] Embodiment 1:

[0143] A method for constructing a road surface wave rapid identification model comprises the following steps:

[0144] S1. Using a vehicle-mounted three-dimensional structured light camera to collect road images, and then compressing the collected road images to obtain a compressed road image;

[0145] Furthermore, the specific implementation method of step S1 includes the following steps:

[0146] S1.1. Use a vehicle-mounted 3D structured light camera to collect road images, and control the vehicle speed within 70 km / h during the collection process;

[0147] S1.2. compressing the collected road surface image by using Huffman coding to obtain a compressed road surface image;

[0148] S1.3. Perform a plane projection on the compressed road surface image, and obtain a projected road surface image which is a compressed road surface image.

[0149] S2. The compressed road surface image obtained in step S1 is annotated to obtain a data set for a road surface wave rapid recognition model;

[0150] Furthermore, the specific implementation method of step S2 includes the following steps:

[0151] S2.1. By manually labeling, the compressed pavement image obtained in step S1 is framed to select waves and three types of defects, namely, peeling, rutting, and bulging. The waves are labeled as Wave, and the other three types of defects are labeled as other1, other2, and other3, respectively, to obtain a data set for the pavement wave rapid recognition model.

[0152] S2.2. Divide the data set of the road wave rapid recognition model into a training set and a test set in a ratio of 8:2;

[0153] S3. Construct the basic model, the first auxiliary model, the second auxiliary model and the enhanced model for rapid identification of road surface waves;

[0154] Furthermore, the specific implementation method of step S3 includes the following steps:

[0155] S3.1. Constructing a basic model for rapid identification of road surface waves:

[0156] S3.1.1. The basic model A consists of 1 input layer, 9 convolutional layers, 9 activation layers, 3 pooling layers, 1 merging layer, 1 upsampling layer, 1 path aggregation network layer PANet, 1 pooling splicing layer, 1 spatial pyramid pooling layer SPP, and 1 output layer. The specific network structure is the following structure connected in sequence:

[0157] Input layer: image size 227×227×3;

[0158] Convolutional layer 1: convolution kernel size 3×3; number of convolution kernels 32; stride: 1; batch normalization;

[0159] Activation layer 1: The activation function is LeakyReLU;

[0160] Convolutional layer 2: kernel size 3×3; number of kernels 64; Stride: 1; batch normalization;

[0161] Activation layer 2: The activation function is LeakyReLU;

[0162] Convolution layer 3: convolution kernel size 3×3; number of convolution kernels 128; stride: 1; batch normalization;

[0163] Activation layer 3: The activation function is LeakyReLU;

[0164] Convolutional layer 4: convolution kernel size 3×3; number of convolution kernels 256; stride: 1; batch normalization;

[0165] Activation layer 4: The activation function is LeakyReLU;

[0166] SPP layer: extracts features of different scales and fuses them to increase the receptive field of the model;

[0167] Pooling layer 1: Type is global average pooling; pooling size: 5×5; Stride: 1;

[0168] Pooling layer 2: Type is global average pooling; pooling size: 9×9; Stride: 1;

[0169] Pooling layer 3: Type is global average pooling; pooling size: 13×13; Stride: 1;

[0170] Pooling concatenation layer: connect the outputs of the three pooling layers;

[0171] PANet layer: aggregates features at different levels and implements feature pyramids through upsampling and downsampling operations;

[0172] Convolution layer 5: kernel size 3×3; number of kernels 768; stride: 2; batch normalization;

[0173] Activation layer 5: The activation function is LeakyReLU;

[0174] Convolutional layer 6: kernel size 3×3; number of kernels 768; stride: 2; batch normalization;

[0175] Activation layer 6: The activation function is LeakyReLU;

[0176] Upsampling layer: uses the nearest neighbor interpolation method to upsample to the same size as the pooling splicing layer;

[0177] Merging layer: The output feature map of the downsampling layer is merged with the output feature map of the upsampling layer;

[0178] Convolutional layer 7: kernel size 3×3; number of kernels 256; stride: 1; batch normalization;

[0179] Activation layer 7: The activation function is LeakyReLU;

[0180] Convolutional layer 8: kernel size 3×3; number of kernels 128; Stride: 1; batch normalization;

[0181] Activation layer 8: The activation function is LeakyReLU;

[0182] Convolution layer 9: kernel size 3×3; number of kernels 64; stride: 1; batch normalization;

[0183] Activation layer 9: The activation function is LeakyReLU;

[0184] Output layer: type is normalized exponential function Softmax; number of neurons: 4;

[0185] S3.1.2. Input the training set obtained in step S2 into the basic model A, calculate the weights and biases of the neural network units through the error back propagation algorithm, complete the training of the basic model A, and obtain the trained basic model A;

[0186] The calculation formula of the loss function of the basic model A is:

[0187] ;

[0188] Among them, La is the loss function of the basic model A; Na is the total number of data samples in the basic model A; For the it The true labels of samples, Based on the prediction of model A it The probability that a sample is class 1;

[0189] S3.1.3. The node output values ​​of the four output units corresponding to the basic model A are a 1. a 2. a 3. a 4. Perform initial conversion on the node output values ​​of the four output units to obtain the output values The corresponding probability value The calculation formula is:

[0190] ;

[0191] Then the node output values ​​of the four output units are further optimized and converted to obtain the output value The corresponding enhancement probability value The calculation formula is:

[0192] ;

[0193] in, G is the enhancement coefficient, and then the actual disease corresponding to each node is marked as , specifically, for road waves 1, for other diseases is 0;

[0194] S3.1.4. Adoption The training set obtained in step S2 is combined with the error chain derivation rule to train the basic model A. After the training is completed, the output values ​​of the four types of diseases corresponding to the basic model A are obtained, which are recorded as , , , ;

[0195] S3.2. Constructing the first auxiliary model for rapid identification of road surface waves:

[0196] S3.2.1. The first auxiliary model B consists of 1 input layer, 9 convolutional layers, 9 activation layers, 5 pooling layers, 1 merging layer, 1 upsampling layer, 1 PANet layer, 1 pooling splicing layer, 1 SPP layer, and 1 output layer. The specific network structure is the following structure connected in sequence:

[0197] Input layer: image size 227×227×3;

[0198] Convolutional layer 1: convolution kernel size 3×3; number of convolution kernels 32; Stride: 1; batch normalization;

[0199] Activation layer 1: The activation function is LeakyReLU;

[0200] Convolutional layer 2: kernel size 3×3; number of kernels 64; Stride: 1; batch normalization;

[0201] Activation layer 2: The activation function is LeakyReLU;

[0202] Convolution layer 3: convolution kernel size 3×3; number of convolution kernels 128; stride: 1; batch normalization;

[0203] Activation layer 3: The activation function is LeakyReLU;

[0204] Convolutional layer 4: convolution kernel size 3×3; number of convolution kernels 256; stride: 1; batch normalization;

[0205] Activation layer 4: The activation function is LeakyReLU;

[0206] SPP layer: extracts features of different scales and fuses them to increase the receptive field of the model;

[0207] Pooling layer 1: Type is global average pooling; pooling size: 5×5; Stride: 1;

[0208] Pooling layer 2: Type is global average pooling; pooling size: 7×7; Stride: 1;

[0209] Pooling layer 3: Type is global average pooling; pooling size: 9×9; Stride: 1;

[0210] Pooling layer 4: Type is global average pooling; pooling size: 11×11; Stride: 1;

[0211] Pooling layer 5: Type is global average pooling; pooling size: 13×13; Stride: 1;

[0212] Pooling concatenation layer: connect the outputs of the five pooling layers;

[0213] PANet: Aggregates features at different levels and implements feature pyramids through upsampling and downsampling operations;

[0214] Convolution layer 5: kernel size 3×3; number of kernels 1280; stride: 2; batch normalization;

[0215] Activation layer 5: The activation function is LeakyReLU;

[0216] Convolutional layer 6: kernel size 3×3; number of kernels 1280; stride: 2; batch normalization;

[0217] Activation layer 6: The activation function is LeakyReLU;

[0218] Upsampling layer: uses the nearest neighbor interpolation method to upsample to the same size as the pooling splicing layer;

[0219] Merging layer: The output feature map of the downsampling layer is merged with the output feature map of the upsampling layer;

[0220] Convolutional layer 7: kernel size 3×3; number of kernels 512; stride: 1; batch normalization;

[0221] Activation layer 7: The activation function is LeakyReLU;

[0222] Convolutional layer 7: kernel size 3×3; number of kernels 256; stride: 1; batch normalization;

[0223] Activation layer 7: The activation function is LeakyReLU;

[0224] Convolutional layer 8: kernel size 3×3; number of kernels 128; Stride: 1; batch normalization;

[0225] Activation layer 8: The activation function is LeakyReLU;

[0226] Convolution layer 9: kernel size 3×3; number of kernels 64; stride: 1; batch normalization;

[0227] Activation layer 9: The activation function is LeakyReLU;

[0228] Output layer: type is Softmax; number of neurons: 4;

[0229] S3.2.2. Input the training set obtained in step S2 into the first auxiliary model B, calculate the weights and biases of the neural network units through the error back propagation algorithm, complete the training of the first auxiliary model B, and obtain the trained first auxiliary model B;

[0230] The calculation formula of the loss function of the first auxiliary model B is:

[0231] ;

[0232] Wherein, Lb is the loss function of the first auxiliary model B; Nb is the total number of data samples in the first auxiliary model B; For the ib The true labels of samples, The first auxiliary model B predicts ib The probability that a sample is class 1;

[0233] S3.2.3. The node output values ​​of the four output units corresponding to the first auxiliary model B are b 1. b 2. b 3. b 4. Perform initial conversion on the node output values ​​of the four output units to obtain the output values b i The corresponding probability value The calculation formula is:

[0234] ;

[0235] Then the node output values ​​of the four output units are further optimized and converted to obtain the output value b i The corresponding enhancement probability value The calculation formula is:

[0236] ;

[0237] S3.2.4. Adoption The training set obtained in step S2 is combined with the error chain derivation rule to train the first auxiliary model B. After the training is completed, the disease output value corresponding to the first auxiliary model B is obtained as , among the 4 nodes in the output layer, the actual disease corresponding to each node is marked as , i =1,2,3,4, specifically, for road surface waves 1, for other diseases is 0;

[0238] S3.3. Constructing the second auxiliary model for rapid identification of road surface waves:

[0239] S3.3.1. The second auxiliary model C consists of 1 input layer, 12 convolutional layers, 12 activation layers, 5 pooling layers, 1 merging layer, 1 upsampling layer, 1 PANet layer, 1 pooling splicing layer, 1 SPP layer, and 1 output layer. The specific network structure is the following structure connected in sequence:

[0240] Input layer: image size 227×227×3;

[0241] Convolutional layer 1: convolution kernel size 3×3; number of convolution kernels 32; Stride: 1; batch normalization;

[0242] Activation layer 1: The activation function is LeakyReLU;

[0243] Convolutional layer 2: kernel size 3×3; number of kernels 64; Stride: 1; batch normalization;

[0244] Activation layer 2: The activation function is LeakyReLU;

[0245] Convolution layer 3: convolution kernel size 3×3; number of convolution kernels 128; stride: 1; batch normalization;

[0246] Activation layer 3: The activation function is LeakyReLU;

[0247] Convolutional layer 4: convolution kernel size 3×3; number of convolution kernels 128; stride: 1; batch normalization;

[0248] Activation layer 4: The activation function is LeakyReLU;

[0249] Convolution layer 5: kernel size 3×3; number of kernels 256; stride: 1; batch normalization;

[0250] Activation layer 5: The activation function is LeakyReLU;

[0251] Convolutional layer 6: kernel size 3×3; number of kernels 256; stride: 1; batch normalization;

[0252] Activation layer 6: The activation function is LeakyReLU;

[0253] SPP layer: extracts features of different scales and fuses them to increase the receptive field of the model;

[0254] Pooling layer 1: Type is global average pooling; pooling size: 5×5; Stride: 1;

[0255] Pooling layer 2: Type is global average pooling; pooling size: 7×7; Stride: 1;

[0256] Pooling layer 3: Type is global average pooling; pooling size: 9×9; Stride: 1;

[0257] Pooling layer 4: Type is global average pooling; pooling size: 11×11; Stride: 1;

[0258] Pooling layer 5: Type is global average pooling; pooling size: 13×13; Stride: 1;

[0259] Pooling concatenation layer: connect the outputs of the three pooling layers;

[0260] PANet: Aggregates features at different levels and implements feature pyramids through upsampling and downsampling operations;

[0261] Convolutional layer 7: kernel size 3×3; number of kernels 1280; stride: 2; batch normalization;

[0262] Activation layer 7: The activation function is LeakyReLU;

[0263] Convolutional layer 8: kernel size 3×3; number of kernels 1280; stride: 2; batch normalization;

[0264] Activation layer 8: The activation function is LeakyReLU;

[0265] Upsampling layer: uses the nearest neighbor interpolation method to upsample to the same size as the pooling splicing layer;

[0266] Merging layer: The output feature map of the downsampling layer is merged with the output feature map of the upsampling layer;

[0267] Convolutional layer 9: kernel size 3×3; number of kernels 512; Stride: 1; batch normalization;

[0268] Activation layer 9: The activation function is LeakyReLU;

[0269] Convolution layer 10: kernel size 3×3; number of kernels 256; stride: 1; batch normalization;

[0270] Activation layer 10: The activation function is LeakyReLU;

[0271] Convolutional layer 11: kernel size 3×3; number of kernels 128; Stride: 1; batch normalization;

[0272] Activation layer 11: The activation function is LeakyReLU;

[0273] Convolution layer 12: kernel size 3×3; number of kernels 64; stride: 1; batch normalization;

[0274] Activation layer 12: The activation function is LeakyReLU;

[0275] Output layer: type is normalized exponential function Softmax; number of neurons: 4;

[0276] S3.3.2. Input the training set obtained in step S2 into the second auxiliary model C, calculate the weights and biases of the neural network units through the error back propagation algorithm, complete the training of the second auxiliary model C, and obtain the trained second auxiliary model C;

[0277] The calculation formula of the loss function of the second auxiliary model C is:

[0278] ;

[0279] Wherein, Lc is the loss function of the second auxiliary model C; Nc is the total number of data samples in the second auxiliary model C; For the ic The true labels of samples, The second auxiliary model C predicts ic The probability that a sample is class 1;

[0280] S3.3.3. The node output values ​​of the four output units corresponding to the second auxiliary model C are , the node output values ​​of the four output units are initially converted to obtain The corresponding probability value The calculation formula is:

[0281] ;

[0282] Then the node output values ​​of the four output units are further optimized and converted to obtain the output value The corresponding enhancement probability value The calculation formula is:

[0283] ;

[0284] in, G is the enhancement coefficient, and then the actual disease corresponding to each node is marked as , specifically, for road waves 1, for other diseases is 0;

[0285] S3.3.4. Adoption The training set obtained in step S2 is combined with the error chain derivation rule to train the second auxiliary model C. After the training is completed, the output values ​​of the four types of diseases corresponding to the second auxiliary model C are obtained as follows: ;

[0286] S3.4. Constructing an enhanced model for rapid identification of road waves:

[0287] S3.4.1. The enhanced model D consists of 1 input layer, 14 convolutional layers, 14 activation layers, 5 pooling layers, 1 merging layer, 1 upsampling layer, 1 PANet layer, 1 pooling splicing layer, 1 SPP layer, and 1 output layer. The specific network structure is the following structure connected in sequence:

[0288] Input layer: image size 227×227×3;

[0289] Convolutional layer 1: convolution kernel size 3×3; number of convolution kernels 32; Stride: 1; batch normalization;

[0290] Activation layer 1: The activation function is LeakyReLU;

[0291] Convolutional layer 2: convolution kernel size 3×3; number of convolution kernels 32; stride: 1; batch normalization;

[0292] Activation layer 2: The activation function is LeakyReLU;

[0293] Convolution layer 3: convolution kernel size 3×3; number of convolution kernels 64; stride: 1; batch normalization;

[0294] Activation layer 3: The activation function is LeakyReLU;

[0295] Convolutional layer 4: convolution kernel size 3×3; number of convolution kernels 64; stride: 1; batch normalization;

[0296] Activation layer 4: The activation function is LeakyReLU;

[0297] Convolution layer 5: kernel size 3×3; number of kernels 128; stride: 1; batch normalization;

[0298] Activation layer 5: The activation function is LeakyReLU;

[0299] Convolutional layer 6: kernel size 3×3; number of kernels 128; Stride: 1; batch normalization;

[0300] Activation layer 6: The activation function is LeakyReLU;

[0301] Convolutional layer 7: kernel size 3×3; number of kernels 256; stride: 1; batch normalization;

[0302] Activation layer 7: The activation function is LeakyReLU;

[0303] Convolutional layer 8: kernel size 3×3; number of kernels 256; stride: 1; batch normalization;

[0304] Activation layer 8: The activation function is LeakyReLU;

[0305] SPP layer: extracts features of different scales and fuses them to increase the receptive field of the model;

[0306] Pooling layer 1: Type is global average pooling; pooling size: 5×5; Stride: 1;

[0307] Pooling layer 2: Type is global average pooling; pooling size: 7×7; Stride: 1;

[0308] Pooling layer 3: Type is global average pooling; pooling size: 9×9; Stride: 1;

[0309] Pooling layer 4: Type is global average pooling; pooling size: 11×11; Stride: 1;

[0310] Pooling layer 5: Type is global average pooling; pooling size: 13×13; Stride: 1;

[0311] Pooling concatenation layer: connect the outputs of the three pooling layers;

[0312] PANet: Aggregates features at different levels and implements feature pyramids through upsampling and downsampling operations;

[0313] Convolution layer 9: kernel size 3×3; number of kernels 1280; stride: 2; batch normalization;

[0314] Activation layer 9: The activation function is LeakyReLU;

[0315] Convolution layer 10: kernel size 3×3; number of kernels 1280; stride: 2; batch normalization;

[0316] Activation layer 10: The activation function is LeakyReLU;

[0317] Upsampling layer: uses the nearest neighbor interpolation method to upsample to the same size as the pooling splicing layer;

[0318] Merging layer: The output feature map of the downsampling layer is merged with the output feature map of the upsampling layer;

[0319] Convolutional layer 11: kernel size 3×3; number of kernels 512; Stride: 1; batch normalization;

[0320] Activation layer 11: The activation function is LeakyReLU;

[0321] Convolutional layer 12: kernel size 3×3; number of kernels 256; stride: 1; batch normalization;

[0322] Activation layer 12: The activation function is LeakyReLU;

[0323] Convolution layer 13: kernel size 3×3; number of kernels 128; stride: 1; batch normalization;

[0324] Activation layer 13: The activation function is LeakyReLU;

[0325] Convolutional layer 14: kernel size 3×3; number of kernels 64; Stride: 1; batch normalization;

[0326] Activation layer 14: the activation function is LeakyReLU;

[0327] Output layer: type is normalized exponential function Softmax; number of neurons: 4;

[0328] S3.4.2. Input the training set obtained in step S2 into the enhanced model D, calculate the weights and biases of the neural network units through the error back propagation algorithm, complete the training of the enhanced model D, and obtain the trained enhanced model D;

[0329] The loss function of the enhanced model D is calculated as:

[0330] ;

[0331] Where Ld is the loss function of the enhanced model D; Nd is the total number of data samples in the enhanced model D; For the id The true labels of samples, To strengthen the model D to predict the id The probability that a sample is class 1;

[0332] S3.4.3. The node output values ​​of the four output units corresponding to the enhanced model D are , the node output values ​​of the four output units are initially converted to obtain The corresponding probability value The calculation formula is:

[0333] ;

[0334] Then the node output values ​​of the four output units are further optimized and converted to obtain the output value The corresponding enhancement probability value The calculation formula is:

[0335] ;

[0336] in, G is the enhancement coefficient, and then the actual disease corresponding to each node is marked as , specifically, for road waves 1, for other diseases is 0;

[0337] S3.4.4. Adoption The training set obtained in step S2 is combined with the error chain derivation rule to train the enhanced model D. After the training is completed, the output values ​​of the four types of diseases corresponding to the enhanced model D are obtained as follows: .

[0338] S4. Construct the loss function among the basic model, the first auxiliary model, the second auxiliary model and the output layer of the enhanced model for rapid identification of road surface waves;

[0339] Furthermore, the specific implementation method of step S4 includes the following steps:

[0340] S4.1. Based on and , establish the loss function between the first auxiliary model B and the output layer of the basic model A, and the calculation formula is:

[0341] ;

[0342] based on and , construct the output value The corresponding probability value The loss function of the base model A , the calculation formula is:

[0343] ;

[0344] based on and , establish the relationship loss function between the output layer of the first auxiliary model B and the basic model A , the calculation formula is:

[0345] ,

[0346] in, for The weight coefficient of for The weight coefficient of

[0347] S4.2. Establish the relationship loss function between the output layer of the second auxiliary model C and the first auxiliary model B , the expression is:

[0348] ;

[0349] in, for The weight coefficient of for The weight coefficient of Based on and The loss function of Based on and The loss function of

[0350] The calculation formula is:

[0351] ;

[0352] The calculation formula is:

[0353] ;

[0354] S4.3. Establish the relationship loss function between the output layer of the second auxiliary model C and the basic model A , the expression is:

[0355] ;

[0356] in, for The weight coefficient of for The weight coefficient of Based on and The loss function is Based on and The loss function of

[0357] The calculation formula is:

[0358] ;

[0359] The calculation formula is:

[0360] ;

[0361] S4.4. Establishing the relationship loss function between the output layer of the enhanced model D and the base model A , the expression is:

[0362] ;

[0363] in, for The weight coefficient of for The weight coefficient of Based on and The loss function of

[0364] The calculation formula is:

[0365] ;

[0366] S4.5. Establish the relationship loss function between the output layer of the reinforcement model D and the second auxiliary model C , the expression is:

[0367] ;

[0368] in, for The weight coefficient of for The weight coefficient of Based on and The loss function is Based on and The loss function of

[0369] The calculation formula is:

[0370] ;

[0371] The calculation formula is:

[0372] .

[0373] S5. construct an overall loss function based on the loss function between the basic model, the first auxiliary model, the second auxiliary model and the output layer of the enhanced model obtained in step S4, and then construct a model optimization method based on the overall loss function, and obtain an optimized basic model after training;

[0374] Furthermore, the specific implementation method of step S5 includes the following steps:

[0375] S5.1. Define the corresponding overall loss function between the basic model A, the first auxiliary model B, the second auxiliary model C, and the enhanced model D The calculation formula is:

[0376] ;

[0377] in, , , , , They are , , , , The weight coefficient of , , , , The value of is 0 or 1;

[0378] By changing , , , , The value of The overall loss function In the form of, thus obtaining the basic model A with 32 parameters;

[0379] S5.2. Using the data set of the road wave rapid recognition model obtained in step S2, the model is trained based on the chain rule and by error back propagation to obtain 32 basic models A;

[0380] Set the accuracy of 32 basic models A to be , i=1,2,3,…,32, the calculation formula is:

[0381] ;

[0382] in, Represents the first i The number of samples that a model predicts to be cracks and are actually cracks when identifying the data in the test set; Represents the first i The number of samples that a model predicts as cracks but is not actually cracks when identifying the data in the test set;

[0383] Comparison The maximum value in ;

[0384] The corresponding basic model A is the basic model after training and optimization.

[0385] S6. The optimized basic model obtained in step S5 is enhanced and improved to obtain a final road wave rapid recognition model;

[0386] Furthermore, the specific implementation method of step S6 includes the following steps:

[0387] S6.1. For the optimized basic model obtained in step S5, the enhancement coefficient G is set to be: G=1, 3, 5, 7, 9, 11, 15, 20, 30, 40, 50;

[0388] S6.2. Based on different enhancement coefficients G, 11 working conditions are obtained, and the optimized basic model is trained to obtain 11 models. The accuracy rates corresponding to the 11 optimized basic models are set as follows: , i=1,2,3,…,11;

[0389] Comparison The maximum value in ;

[0390] The corresponding optimized basic model is the final road wave rapid identification model.

[0391] S7. Construct a digital twin model based on the final pavement wave rapid identification model for updating pavement wave damage.

[0392] Furthermore, the specific implementation method of step S7 includes the following steps:

[0393] S7.1. Export the final road wave rapid identification model obtained in step S6 as a deployment file for independent services, including one of a pt file, a pth file, a json file, and a txt file;

[0394] S7.2. Construct a road information model;

[0395] S7.2.1. Collect information on the location, length, width, and layout of roads;

[0396] S7.2.2. The data obtained in step S7.2.1 is stored in a database, and a basic three-dimensional model of the road is constructed, wherein the basic three-dimensional model of the road includes geometric information of the road;

[0397] S7.2.3. Based on step S7.2.2, material properties, design specifications, and historical maintenance data are added to the basic three-dimensional model of the road to obtain a road information model;

[0398] S7.2.4. Export the road information model to a format file supported by the web application, including a json file and a js file;

[0399] S7.3. In the digital twin platform, import the deployment file of the final road surface wave rapid identification model obtained in step S7.1, and import the road information model format file obtained in step S7.2;

[0400] S7.4. Use Python to construct an API interface between the image data and the final road surface wave rapid recognition model for intelligent rutting identification. Use Python to construct an API interface between the final road surface wave rapid recognition model and the road information model to complete the construction of a digital twin model for updating the road surface rutting status based on the final road surface wave rapid recognition model.

[0401] Embodiment 2:

[0402] An electronic device includes a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of a method for constructing a road surface wave rapid identification model described in Example 1 when executing the computer program.

[0403] The computer device of the present invention may be a device including a processor and a memory, such as a single chip microcomputer including a central processing unit, etc. Moreover, the processor is used to implement the steps of the above-mentioned method for constructing a road wave rapid identification model when executing the computer program stored in the memory.

[0404] The processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc.

[0405] The memory may mainly include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function (such as a sound playback function, an image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, a phone book, etc.), etc. In addition, the memory may include a high-speed random access memory, and may also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0406] Embodiment 3:

[0407] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements a method for constructing a road surface wave rapid identification model as described in Example 1.

[0408] The computer-readable storage medium of the present invention can be any form of storage medium that can be read by a processor of a computer device, including but not limited to non-volatile memory, volatile memory, ferroelectric memory, etc. A computer program is stored on the computer-readable storage medium. When the processor of the computer device reads and executes the computer program stored in the memory, the steps of the above-mentioned method for constructing a road wave rapid identification model can be implemented.

[0409] The computer program includes computer program code, which may be in source code form, object code form, executable file or some intermediate form, etc. The computer readable medium may include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium, etc.

[0410] It should be noted that relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0411] Although the present application has been described above with reference to specific embodiments, various modifications may be made thereto and parts thereof may be replaced with equivalents without departing from the scope of the present application. In particular, as long as there is no structural conflict, the various features in the specific embodiments disclosed in the present application may be used in combination with each other in any manner, and the fact that these combinations are not exhaustively described in this specification is only for the sake of omitting space and saving resources. Therefore, the present application is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A method for constructing a road surface wave rapid identification model, characterized in that: The steps include: S1. Using a vehicle-mounted three-dimensional structured light camera to collect a road surface image, and then compressing the collected road surface image to obtain a compressed road surface image; S2. Annotate the compressed road surface image obtained in step S1 to obtain a data set for a road surface wave rapid recognition model; S3. Construct the basic model, the first auxiliary model, the second auxiliary model and the enhanced model for rapid identification of road surface waves; The specific implementation method of step S3 includes the following steps: S3.

1. Construct a basic model for rapid identification of road waves: S3.1.

1. The basic model A consists of 1 input layer, 9 convolutional layers, 9 activation layers, 3 pooling layers, 1 merging layer, 1 upsampling layer, 1 path aggregation network layer PANet, 1 pooling splicing layer, 1 spatial pyramid pooling layer SPP, and 1 output layer; S3.1.

2. Input the training set obtained in step S2 into the basic model A, calculate the weights and biases of the neural network units through the error back propagation algorithm, complete the training of the basic model A, and obtain the trained basic model A; The calculation formula of the loss function of the basic model A is: Among them, La is the loss function of the basic model A; Na is the total number of data samples in the basic model A; y ia is the true label of the iath sample, p ia Predict the probability that the iath sample is class 1 based on the basic model A; S3.1.

3. The node output values ​​of the four output units corresponding to the basic model A are a1, a2, a3, and a4. The node output values ​​of the four output units are initially converted to obtain the output value a i The corresponding probability value f ai The calculation formula is: Then the node output values ​​of the four output units are further optimized and converted to obtain the output value a i The corresponding enhancement probability value g ai The calculation formula is: Among them, G is the enhancement coefficient, and then the actual disease corresponding to each node is marked as h ai ; S3.1.

4. Use g ai The training set obtained in step S2 is combined with the error chain derivation rule to train the basic model A. After the training is completed, the output values ​​of the four types of diseases corresponding to the basic model A are obtained, which are recorded as j a1 , j a2 , j a3 , j a4 ; S3.

2. Construct the first auxiliary model for rapid identification of road waves: S3.2.

1. The first auxiliary model B consists of 1 input layer, 9 convolutional layers, 9 activation layers, 5 pooling layers, 1 merging layer, 1 upsampling layer, 1 PANet layer, 1 pooling splicing layer, 1 SPP layer, and 1 output layer; S3.2.

2. Input the training set obtained in step S2 into the first auxiliary model B, calculate the weights and biases of the neural network units through the error back propagation algorithm, complete the training of the first auxiliary model B, and obtain the trained first auxiliary model B; The calculation formula of the loss function of the first auxiliary model B is: Where Lb is the loss function of the first auxiliary model B; Nb is the total number of data samples in the first auxiliary model B; y ib is the true label of the ib-th sample, p ib Predict the probability that the ibth sample is category 1 for the first auxiliary model B; S3.2.

3. The node output values ​​of the four output units corresponding to the first auxiliary model B are b1, b2, b3, and b4. The node output values ​​of the four output units are initially converted to obtain the output value b i The corresponding probability value f bi The calculation formula is: Then the node output values ​​of the four output units are further optimized and converted to obtain the output value b i The corresponding enhancement probability value g bi The calculation formula is: S3.2.

4. Use g bi The training set obtained in step S2 is combined with the error chain derivation rule to train the first auxiliary model B. After the training is completed, the disease output value corresponding to the first auxiliary model B is j bi , among the 4 nodes in the output layer, the actual disease corresponding to each node is labeled as h bi , i=1,2,3,4; S3.

3. Construct the second auxiliary model for rapid identification of road waves: S3.3.

1. The second auxiliary model C consists of 1 input layer, 12 convolutional layers, 12 activation layers, 5 pooling layers, 1 merging layer, 1 upsampling layer, 1 PANet layer, 1 pooling splicing layer, 1 SPP layer, and 1 output layer; S3.3.

2. Input the training set obtained in step S2 into the second auxiliary model C, calculate the weights and biases of the neural network units through the error back propagation algorithm, complete the training of the second auxiliary model C, and obtain the trained second auxiliary model C; The calculation formula of the loss function of the second auxiliary model C is: Where Lc is the loss function of the second auxiliary model C; Nc is the total number of data samples in the second auxiliary model C; y ic is the true label of the icth sample, p ic Predict the probability that the icth sample is class 1 for the second auxiliary model C; S3.3.

3. The node output values ​​of the four output units corresponding to the second auxiliary model C are c i , the node output values ​​of the four output units are initially converted to obtain c i The corresponding probability value f ci The calculation formula is: Then the node output values ​​of the four output units are further optimized and converted to obtain the output value c i The corresponding enhancement probability value g ci The calculation formula is: Among them, G is the enhancement coefficient, and then the actual disease corresponding to each node is marked as h ci ; S3.3.

4. Use g ci The training set obtained in step S2 is combined with the error chain derivation rule to train the second auxiliary model C. After the training is completed, the output values ​​of the four types of diseases corresponding to the second auxiliary model C are j ci ; S3.

4. Construct an enhanced model for rapid identification of road waves: S3.4.

1. The enhanced model D consists of 1 input layer, 14 convolutional layers, 14 activation layers, 5 pooling layers, 1 merging layer, 1 upsampling layer, 1 PANet layer, 1 pooling splicing layer, 1 SPP layer, and 1 output layer; S3.4.

2. Input the training set obtained in step S2 into the enhanced model D, calculate the weights and biases of the neural network units through the error back propagation algorithm, complete the training of the enhanced model D, and obtain the trained enhanced model D; The loss function of the enhanced model D is calculated as: Where Ld is the loss function of the enhanced model D; Nd is the total number of data samples in the enhanced model D; y id is the true label of the idth sample, p id Predict the probability that the idth sample is class 1 for the enhanced model D; S3.4.

3. The node output values ​​of the four output units corresponding to the enhanced model D are d i , the node output values ​​of the four output units are initially converted to obtain d i The corresponding probability value f di The calculation formula is: Then the node output values ​​of the four output units are further optimized and converted to obtain the output value d i The corresponding enhancement probability value g di The calculation formula is: Among them, G is the enhancement coefficient, and then the actual disease corresponding to each node is marked as h di ; S3.4.

4. Use g di The training set obtained in step S2 is combined with the error chain derivation rule to train the enhanced model D. After the training is completed, the output values ​​of the four types of diseases corresponding to the enhanced model D are j di ; S4. Construct the loss function between the basic model for rapid identification of road surface waves, the first auxiliary model, the second auxiliary model and the output layer of the enhanced model; S5. construct an overall loss function based on the loss function between the basic model, the first auxiliary model, the second auxiliary model and the output layer of the enhanced model obtained in step S4, and then construct a model optimization method based on the overall loss function, and obtain an optimized basic model after training; S6. Enhance and improve the optimized basic model obtained in step S5 to obtain the final road wave rapid recognition model; S7. Construct a digital twin model based on the final pavement wave rapid identification model for updating pavement wave damage.

2. A method for constructing a road surface wave rapid identification model according to claim 1, characterized in that: The specific implementation method of step S1 includes the following steps: S1.

1. Use a vehicle-mounted 3D structured light camera to collect road images, and control the vehicle speed within 70 km / h during the collection process; S1.

2. compressing the collected road surface image by using Huffman coding to obtain a compressed road surface image; S1.

3. Perform a planar projection on the compressed road surface image, and obtain a projected road surface image which is a compressed road surface image.

3. A method for constructing a road surface wave rapid identification model according to claim 1 or 2, characterized in that: The specific implementation method of step S2 includes the following steps: S2.

1. By manually labeling, the compressed pavement image obtained in step S1 is framed to select waves and three types of defects, namely, peeling, rutting, and bulging. The waves are labeled as Wave, and the other three types of defects are labeled as other1, other2, and other3, respectively, to obtain a data set for the pavement wave rapid recognition model. S2.

2. The data set of the road wave rapid identification model is divided into a training set and a test set according to a ratio of 8:

2.

4. A method for constructing a road surface wave rapid identification model according to claim 3, characterized in that: The specific implementation method of step S4 includes the following steps: S4.

1. Based on j bi and g ai , establish the loss function between the first auxiliary model B and the output layer of the basic model A, and the calculation formula is: Based on f ai and h ai , construct the output value a i The corresponding probability value f ai The loss function L of the base model A AA , the calculation formula is: Based on L BA and L AA , establish the relationship loss function L between the output layer of the first auxiliary model B and the basic model A BAA , the calculation formula is: L BAA =c BA ·L BA +g AA ·L AA , Among them, γ BA For L BA The weight coefficient, γ AA For L AA The weight coefficient of S4.

2. Establish the relationship loss function L between the output layer of the second auxiliary model C and the first auxiliary model B CBB , the expression is: L CBB =c CB ·L CB +g BB ·L BB Among them, γ CB For L CB The weight coefficient, γ BB For L BB The weight coefficient, L CB Based on j ci and g bi The loss function of L BB Based on f bi and h bi The loss function of L CB The calculation formula is: L BB The calculation formula is: S4.

3. Establish the relationship loss function L between the output layer of the second auxiliary model C and the basic model A CAA , the expression is: L CAA =c CA ·L CA +g AA ·L AA Among them, γ CA For L CA The weight coefficient, γ AA For L AA The weight coefficient, L CA Based on j ci and g ai The loss function, L AA Based on f ai and h ai The loss function of L CA The calculation formula is: L AA The calculation formula is: S4.

4. Establish the relationship between the output layer of the enhanced model D and the basic model A. Loss function L DAA , the expression is: L DAA =c DA ·L DA +g AA ·L AA Among them, γ DA For L DA The weight coefficient, γ AA For L AA The weight coefficient, L DA Based on j di and g ai The loss function of L DA The calculation formula is: S4.

5. Establish the relationship loss function L between the output layer of the reinforcement model D and the second auxiliary model C DCC , the expression is: L DCC =c DC ·L DC +g CC ·L CC Among them, γ DC For L DC The weight coefficient, γ CC For L CC The weight coefficient, L DC Based on j di and g ci The loss function, L CC Based on f ci and h ci The loss function of L DC The calculation formula is: L CC The calculation formula is:

5. A method for constructing a road surface wave rapid identification model according to claim 4, characterized in that: The specific implementation method of step S5 includes the following steps: S5.

1. Define the overall loss function L corresponding to the basic model A, the first auxiliary model B, the second auxiliary model C, and the enhanced model D total The calculation formula is: L total =β1L DCC +β2L CBB +β3L BAA +β4L DAA +β5L CAA Among them, β1, β2, β3, β4, and β5 are L DCC , L CBB , L BAA , L DAA , L CAA The weight coefficient of , and the values ​​of β1, β2, β3, β4, and β5 are 0 or 1; By changing the values ​​of β1, β2, β3, β4, and β5, we get 2 5 = 32 overall loss functions L total In the form of, thus obtaining the basic model A with 32 parameters; S5.

2. Using the data set of the road wave rapid recognition model obtained in step S2, the model is trained based on the chain rule and the error back propagation method to obtain 32 basic models A; Set the accuracy of the 32 basic models A to be ACa i ,i=1,2,3,…,32, the calculation formula is: Among them, TP i It indicates the number of samples predicted as cracks and actually found to be cracks when the i-th model trained with the training set identifies the data in the test set; FP i It indicates the number of samples predicted as cracks but not actually cracks by the i-th model trained with the training set when identifying the data in the test set; Compare to get ACa i The maximum value ACa in max =max[ACa i ]; ACa max The corresponding basic model A is the basic model after training and optimization.

6. A method for constructing a road surface wave rapid identification model according to claim 5, characterized in that: The specific implementation method of step S6 includes the following steps: S6.

1. For the optimized basic model obtained in step S5, the enhancement coefficient G is set to be: G = 1, 3, 5, 7, 9, 11, 15, 20, 30, 40, 50; S6.

2. Based on different enhancement coefficients G, 11 working conditions are obtained, and the optimized basic model is trained to obtain 11 models. The accuracy rates corresponding to the 11 optimized basic models are set as AACa ni , i=1,2,3,…,11; Compare to AACa i The maximum value of AACa max =max[AACa i ]; AAC max The corresponding optimized basic model is the final road wave rapid identification model.

7. A method for constructing a road surface wave rapid identification model according to claim 6, characterized in that: The specific implementation method of step S7 includes the following steps: S7.

1. Export the final road wave rapid identification model obtained in step S6 as a deployment file for independent services, including one of a pt file, a pth file, a json file, and a txt file; S7.

2. Construct a road information model; S7.2.

1. Collect information on the location, length, width, and layout of roads; S7.2.

2. The data obtained in step S7.2.1 is stored in a database, and a basic three-dimensional model of the road is constructed, wherein the basic three-dimensional model of the road includes geometric information of the road; S7.2.

3. Based on step S7.2.2, material properties, design specifications, and historical maintenance data are added to the basic three-dimensional model of the road to obtain a road information model; S7.2.

4. Export the road information model to a format file supported by the Web application, including a json file and a js file; S7.

3. In the digital twin platform, import the deployment file of the final road surface wave rapid identification model obtained in step S7.1, and import the road information model format file obtained in step S7.2; S7.

4. Use Python to construct an API interface between the image data and the final road surface wave rapid recognition model for intelligent road surface wave recognition; use Python to construct an API interface between the final road surface wave rapid recognition model and the road information model to complete the construction of a digital twin model for road surface wave status update based on the final road surface wave rapid recognition model.

8. An electronic device, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of a method for constructing a road surface wave rapid identification model as described in any one of claims 1 to 7 when executing the computer program.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, a method for constructing a road surface wave rapid identification model as described in any one of claims 1 to 7 is implemented.

Images