Road surface friction force estimation method and device based on deep learning and medium

Abstract

The invention provides a pavement friction force estimation method and device based on deep learning and a medium, and belongs to the technical field of deep learning. The method comprises the following steps: acquiring a data set which comprises a pavement crack type and a corresponding friction coefficient, preprocessing the data set, and dividing the data set into a test set and a training set; establishing a pavement friction coefficient prediction model, wherein the pavement friction coefficient prediction model is a BP neural network optimized by adopting an improved eagle optimization algorithm; inputting the training set data into the road surface friction coefficient prediction model, and training the optimal weight and bias; and predicting the friction coefficient of the pavement crack by using the trained pavement friction coefficient prediction model. According to the invention, the deep learning technology is adopted to realize automatic detection, the dependence on manual operation is reduced, and the detection efficiency of the pavement friction coefficient is greatly improved.

Description

Technical Field

[0001] The present invention relates to a road friction estimation method, device and medium based on deep learning, belonging to the technical field of deep learning. Background Art

[0002] With the rapid development of expressways in my country, asphalt pavement has been widely used in my country's high-grade highway pavement due to its advantages such as high strength and good durability. The surface layer of asphalt pavement is exposed to the natural environment. Under the repeated action of driving load and environmental factors, the anti-skid performance declines and the friction coefficient decreases, which makes the vehicle easy to lose control during braking, thus causing traffic accidents. Therefore, timely estimation of the friction coefficient during driving helps to ensure driving safety and reduce the occurrence of traffic accidents. Commonly used road friction coefficient test methods include the use of a pendulum friction coefficient meter. The pendulum friction coefficient meter is simple to operate and easy to use. It can be used for on-site road testing and indoor testing, but it is easily affected by human factors and cannot be measured continuously. The dynamic friction coefficient meter can test the friction coefficient of the road at a speed of 0~80 km / h, but it can only be tested at a single point and requires closed traffic. The lateral force coefficient test equipment sets the test wheel to a certain angle with the driving direction. When driving, the tire contacts the road surface to generate a lateral friction resistance, thereby calculating the lateral force coefficient SFC, but the equipment is expensive and not suitable for large-scale promotion. Therefore, the research on a fast, efficient and economical friction coefficient estimation method has great application value and practical needs. In recent years, machine learning has become one of the most intelligent and cutting-edge research fields in artificial intelligence. Using computer simulation to achieve human-like learning activities has been widely used in image recognition, automatic speech recognition, data prediction, website search and other fields. Many experts and scholars predict the friction coefficient through machine learning. Zhangyou et al. introduced a new deep residual network (ResNets), called Friction-ResNets, which is designed for road friction prediction based on surface texture datasets. The Friction-ResNets architecture consists of 11 convolutional layers, 1 average pooling layer, and 1 fully connected layer with millions of neurons. Pedro Marcelino used the Python machine learning library scikit-learn to predict asphalt pavement friction. Using data from the Long-Term Pavement Performance (LTPP) database, asphalt concrete pavements in 113 different sections across the United States were selected. Based on these data, two machine learning models were established to predict the friction coefficient. The skid resistance of the pavement is closely related to pavement cracks. The present invention aims to identify and detect cracks while being able to feedback the friction coefficient of the cracks, so as to remind the driver and achieve the application purpose of intelligent driving. Therefore, the intrinsic relationship between the number of pavement cracks and the skid resistance of the pavement is targeted. The BP neural network algorithm has strong nonlinear mapping capabilities, but the traditional BP network algorithm not only has a slow convergence speed, but also the selection of the initial neural network weights and thresholds is random, which will lead to problems such as poor prediction accuracy and unstable prediction. Summary of the invention

[0003] The purpose of the present invention is to provide a road friction estimation method, device and medium based on deep learning, so as to improve the detection efficiency and accuracy of the road friction coefficient.

[0004] In order to achieve the above object, the present invention is implemented through the following technical solutions: A road friction estimation method based on deep learning, comprising: Collect data sets, including pavement crack types and corresponding friction coefficients, and divide the data sets into test sets and training sets after preprocessing; A road friction coefficient prediction model is established, wherein the road friction coefficient prediction model is a BP neural network optimized by an improved Osprey optimization algorithm, and the optimal weight and bias of the model are obtained by the improved Osprey optimization algorithm; the improved Osprey optimization algorithm initializes weights and biases based on a Sobol sequence, a random number of a Weibull distribution is used as a step factor in the first-stage position update of the improved Osprey optimization algorithm, and a firefly algorithm is used to perturb the position of the second-stage update of the improved Osprey optimization algorithm; Input the training set data into the road friction coefficient prediction model and train the optimal weights and biases; The friction coefficient of pavement cracks is predicted using the trained pavement friction coefficient prediction model.

[0005] Preferably, the optimal weight and bias of the model are obtained by using the improved Osprey optimization algorithm, specifically in the following manner: Initialize weights and biases based on the Sobol sequence and calculate the initial fitness value; In the first phase of the Osprey optimization algorithm, the weights and bias positions are updated and new fitness values ​​are calculated; random numbers from the Weibull distribution are introduced as step size factors to improve the first phase position update process of the Osprey optimization algorithm; The positions of weights and biases are updated through the second phase of the Osprey optimization algorithm; The weights and biases after the second stage are used as the initial population, and the perturbation strategy of the firefly algorithm is used to update the position in the second stage; Iterate the first and second phases of the Osprey optimization algorithm until the termination condition is met and output the optimal weight and bias parameters.

[0006] Preferably, the activation function between neurons of the road friction coefficient prediction model adopts a nonlinear transformation Sigmoid function.

[0007] Preferably, weights and biases are initialized based on the Sobol sequence, specifically in the following manner: Generate an N×D Sobol matrix, each element represents a set of weights and biases, traverse each parameter, and convert each value of the Sobol matrix Map from [0,1] to the upper and lower bounds of the corresponding parameters In the BP neural network, the weight and bias population distribution are obtained. The specific formula is as follows: , in, is the dimension of the search space, For the population size.

[0008] Preferably, the first-stage position update formula after introducing Weibull distributed random numbers as step factors is as follows.

[0009] , In the formula, is a random step size factor affected by the Weibull distribution, with a scale parameter of 1 and a shape parameter of 0.5; are the new weights and biases for the first stage, For the group weights and biases, are the selected weights and biases, is a random number, ranging from .

[0010] Preferably, the second-stage position update of the Osprey optimization algorithm improved by the firefly algorithm is specifically performed as follows: The weights and biases updated by the second stage of the Osprey algorithm are used as the initial population, and the fitness values ​​of the weights and biases are used as the initial brightness, and the relative brightness and attractiveness of each pair of weights and biases are calculated; The positions are moved according to the relative brightness of any two weights and biases in the initial population. When the relative brightness of one set of weights and biases is smaller than that of another set of weights and biases, the set with smaller relative brightness of weights and biases moves toward the set with larger relative brightness of weights and biases according to the formula. The formula is as follows: , in, is the position of a set of weights and biases with relatively small brightness after passing through the second stage, is the position of a set of weights and biases with relatively large brightness after passing through the second stage, is the position after the firefly interference, belongs to [0,1], is a random number that follows a uniform distribution [0,1]. is the attraction between the two sets of weights; The boundary is processed by the following formula: , If the fitness value of the updated position is higher, the position is replaced and the position update of the weight and bias at this stage is obtained. The formula is as follows: , in, It's location The fitness value of It is The position of the group weights and biases after the perturbation, After the second stage Fitness values ​​for group weights and biases.

[0011] A road friction estimation device based on deep learning comprises a processor and a memory storing program instructions, wherein the processor is configured to execute the road friction estimation method based on deep learning when running the program instructions.

[0012] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the road friction estimation method based on deep learning.

[0013] The advantages of the present invention are: the present invention adopts deep learning technology to realize automatic detection, reduces the dependence on manual operation, and greatly improves the detection efficiency of the road friction coefficient. Compared with the traditional manual detection method, the present invention can realize real-time and continuous road friction coefficient detection, avoiding data delays and inaccuracies caused by manual operation. By utilizing the improved optimization of the weights and thresholds of the BP neural network, the convergence speed and prediction accuracy of the model are improved. Compared with the high-cost dynamic friction coefficient meter and lateral force coefficient test equipment, the method of the present invention relies on image acquisition and processing, the equipment cost is relatively low, and it is convenient for large-scale promotion and application. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0015] Figure 1 It is a schematic diagram of the process of the present invention.

[0016] Figure 2 This is a flow chart of model training of the present invention.

[0017] Figure 3 This is a comparison diagram of the effects of the embodiments of the present invention. DETAILED DESCRIPTION

[0018] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0019] like Figure 1 As shown, the current road friction coefficient test method mainly relies on manual operation, which is time-consuming and labor-intensive, and is easily affected by the subjective factors of the testers, resulting in inaccurate data. In addition, manual testing requires closed traffic, which has an adverse effect on driving safety. High equipment cost: Although the existing dynamic friction coefficient meter and lateral force coefficient test equipment can provide relatively accurate measurement data, the equipment is expensive and not suitable for large-scale promotion, which limits its popularity in practical applications. Measurement discontinuity: The pendulum friction coefficient meter can only perform single-point testing and cannot achieve continuous measurement, resulting in limited data coverage and inability to reflect the friction conditions of the entire road surface in real time. The present invention combines deep learning technology and defect detection technology to propose a road friction estimation method based on deep learning, including: S1: Collect data sets, including pavement crack types and corresponding friction coefficients, and divide the data sets into test sets and training sets after preprocessing; S2: Establishing a road friction coefficient prediction model, wherein the road friction coefficient prediction model is a BP neural network optimized by an improved Osprey optimization algorithm, and obtaining the optimal weight and bias of the model by the improved Osprey optimization algorithm; the improved Osprey optimization algorithm initializes weights and biases based on a Sobol sequence, and in the first-stage position update of the improved Osprey optimization algorithm, a random number of a Weibull distribution is used as a step factor, and the firefly algorithm is used to perturb the position of the second-stage update of the improved Osprey optimization algorithm; S3: input the training set data into the road friction coefficient prediction model, and train the optimal weights and biases; S4: Use the trained pavement friction coefficient prediction model to predict the friction coefficient of pavement cracks.

[0020] As a refinement of the above embodiment, the specific method of collecting the data set in step S1 is as follows: Construct a road crack detection model and obtain road defect images. The original road defect image dataset used in this embodiment is the public dataset GRDD2022, which consists of road images from multiple countries. The damage categories considered are longitudinal cracks D00, transverse cracks D10, mesh cracks D20, and potholes D40. Objects that do not contain the detection targets of this embodiment are eliminated. After processing, there are 17,630 pictures remaining, including 3,853 Chinese road images, 1,072 Czech road images, 7,900 Japanese road images, and 4,806 American road images. The label format of this dataset is .xml format. Use the python program to convert it into a .txt format label that can be used by the YOLO model, and divide it into a training set and a validation set in a ratio of 7:3, with 12,339 training sets and 5,291 validation sets. Configure the YOLOv8 operating environment and hyperparameters; set the learning rate = 0.01, input image size = 640 × 640, batch size = 64, Epoch = 200, learning rate momentum = 0.937, optimizer = SGD, weight decay = 0.0005.

[0021] Input the preprocessed training set and validation set into the improved YOLOv8 detection model, and start model training according to the set hyperparameters. The specific training process includes: The data-augmented training set images are input into the backbone network of the improved YOLOv8 model for feature extraction, and the input images are converted into multi-level feature maps. The obtained feature maps are passed into the Neck network to fuse the feature maps from different levels and generate multi-scale feature representations. This enables the model to detect objects of different sizes at the same time and pass these features to the head layer. The head is responsible for generating the final detection results. It includes multiple convolutional layers to predict the target category (Class), bounding box (Bounding Box) and objectness score (Objectness Score) in each feature map. For each grid cell (Grid Cell), the detection head outputs multiple bounding box predictions, each of which contains category probability, bounding box regression parameters and objectness score. According to the prediction results and the true annotations, the classification loss, bounding box regression loss and objectness loss are calculated, the gradient is calculated using the backpropagation algorithm, and the optimizer (SGD) is used to update the model parameters. This process will be repeated multiple times on the entire training dataset until the set epoch (200) is reached, that is, the training is completed.

[0022] After the training is completed, the model is tested, and the weight file generated by the training is loaded into the test program to obtain the test results of the model, including defect type, defect category, confidence level and defect quantity.

[0023] Since the road friction coefficient is affected by the road material, vehicle tires, ambient temperature and the degree of humidity in different weather, the present invention aims to achieve effective identification and detection of road cracks during driving, and simultaneously obtain the friction coefficient of the cracks, as the ultimate goal, so the collection and measurement of the friction coefficient adopts a unified standard, and does not consider other external factors. This embodiment only uses a pendulum instrument to perform statistical measurement of cracks on dry ground. The road section is divided into 60 road section measurement points with a spacing of 30 meters, and the number of cracks in the road section is counted respectively, and the pendulum friction coefficient measuring instrument is used to collect the friction coefficient of each road section crack. Due to the large number of different values, the friction coefficient in each road section measurement point is counted and the average value is taken as the friction coefficient value of the measurement point. The pendulum friction measuring instrument is used for data collection, and the BPN value is used as the friction coefficient as an evaluation index between road cracks and skid resistance. The road crack type and the corresponding friction coefficient data set are established.

[0024] As a refinement of the above embodiment, in order to realize the application scenario of intelligent driving in step S2, the friction coefficient taken by the pendulum friction coefficient measuring instrument is predicted and analyzed, so that in the process of detecting cracks, the overall number of longitudinal cracks, transverse cracks and block cracks in the detected section can be identified to predict the friction coefficient of the section. The BP neural network is selected to predict and analyze the friction coefficient. As a multi-layer feedforward neural network for prediction training, the BP neural network can effectively meet the data training and prediction functions of the present invention, and the neural network model consists of an input layer, a hidden layer and an output layer. In order to reduce the error value during training, the BP neural network adopts a combination of sample forward propagation and error back propagation. The input sample can be progressively input to the hidden layer and the output layer through the input layer, and the number of samples is weighted to obtain the output result; and the back propagation is to compare the result of the forward propagation output with the expected value, back propagate the error to each layer, and use the gradient descent method to calculate and update the weight and bias of each layer to minimize the training error. Taking the three-layer perceptron as an example, when there is a large difference between the output value and the expected value, there is an output error E: , For the hidden layer, the formula can be expanded as follows: , The formula can be further extended to the output layer as follows: , in, , Represents the weight of each layer, is expressed as an error and can be adjusted by weights, represents the ideal output, that is, the correct result expected from the neural network. represents the actual output, that is, the output value given by the neural network based on the current weights and inputs. Represents a general layer, used to index a specific layer number, Represents the net input of a neuron, that is, the weighted sum input received by the neuron represents the output of the hidden layer neurons, represents the number of neurons, The input layer Input, that is, the network receives input data, represents the activation function, Represents the total number of neurons in the input layer. The purpose of adjusting the weights is to minimize the error. Therefore, the weights should be proportional to the gradient descent of the error, as shown below: , , in, Represents the learning rate, which is a hyperparameter used to control the step size of each weight update.

[0025] BP neural algorithm belongs to The learning rule class continuously adjusts the weights and biases of the model through back propagation to minimize the sum of squared errors. Since the BP neural network algorithm has the characteristics of random selection of initial weights and biases, it always leads to problems such as slow network convergence and unstable output during the training process. Therefore, this embodiment improves the BP neural network to improve network accuracy.

[0026] First, the activation function between neurons adopts the nonlinear transformation Sigmoid function, which is characterized by the fact that the function itself and its derivatives are continuous, which is convenient for data processing and can greatly improve the computing efficiency of the computer; the nonlinear transformation Sigmoid function is as follows: , in: represents the function input value, is the negative exponential term of the natural exponential function, indicating of Power.

[0027] Secondly, an improved Osprey optimization algorithm (Osprey) is proposed to optimize the BP neural network. The Osprey optimization algorithm consists of two stages: global exploration (first stage) and local development (second stage). The specific implementation process includes:

[0028] 1. Population initialization. Each osprey is regarded as a candidate solution to the problem, that is, the weight and bias of the BP neural network. The N×D-dimensional matrix X composed of N groups of BP neural network weights and biases is used as the initial population. Each group of weights and biases is randomly initialized according to the following formula: , in, It is The group weights and biases are The location of the dimension, and They are Dimensional lower and upper bounds, is a random number in [0,1], is the number of weights and biases, is the dimension of the solution.

[0029] Each set of weights and biases is a candidate solution, and the quality of the candidate solution can be evaluated based on the fitness value of the objective function F. The fitness value is calculated according to the following formula.

[0030] , in, No. The fitness values ​​of the group weights and biases, It is Group weights and biases.

[0031] 2. Positioning and fishing (first stage). For each set of weights and biases, other weights and biases with better fitness values ​​in the search space are considered as fish under water. The position calculation formula for each set of weights and biases is shown below.

[0032] , in, It is group weights and biases, N is the number of weights and biases, It is Group weights and bias fitness values. are the optimal weights and biases. In the search space, the osprey randomly locates a fish and attacks it. The following formula is used to simulate the position update process when the osprey moves to the fish: , in are the new weights and biases in the first stage, It is group weights and biases, are the selected weights and biases, is a random number in [0,1], Pick a random value from the set {1, 2}.

[0033] If the updated position is out of bounds, the boundary is processed according to the following formula. If the new position is less than the lower bound of the problem, it is set to the lower bound value. If the new position is greater than the upper bound, it is set to the upper bound value.

[0034] , If the fitness value of the updated position is better, it replaces the previous position. The operation is as follows: , in, After the first stage Fitness values ​​for group weights and biases. are the weights and biases after the first stage, Indicates the first stage The position after the group weights and biases are updated.

[0035] 3. Bring the fish to the correct location (second stage). After the osprey catches the fish, it will take it to a location that it thinks is safe for consumption. This process will cause the osprey's position in the search space to change slightly, that is, the position of the weights and biases will change slightly. This is the local development stage of the osprey optimization algorithm. The position of the weights and biases is updated according to the following formula. Similar to the global exploration stage, boundary processing operations should also be performed in this stage.

[0036] , , in, are the new weights and biases in the second stage, is the current iteration number of the algorithm; is the maximum number of iterations. Similar to the global exploration phase, if the fitness value of the updated position is better, the previous position is replaced: , in, Indicates the second stage The position after the group weights and biases are updated. Indicates the second stage Location of group weights and biases. After the second stage Fitness values ​​for group weights and biases.

[0037] The Osprey optimization algorithm is inspired by the hunting behavior of osprey and has strong global search capabilities. It can avoid the problem that traditional BP neural networks are prone to fall into local optimal solutions, and improve the robustness and stability of the optimization process. The Osprey optimization algorithm uses the hunting strategy of the osprey to accelerate the convergence speed during the optimization process and can significantly reduce the training time. At the same time, due to the small number of data sets, the Osprey algorithm will still have the problem of overfitting, which will affect the generalization ability during the training process and thus reduce the evaluation criteria of the entire model. The Osprey optimization algorithm is used to optimize the hyperparameters of the back propagation neural network (BPNN). However, the Osprey algorithm has the disadvantages of imbalance between global exploration ability and local development ability, and is prone to fall into local optimality and has weak global exploration ability. In order to enhance the global search ability of the Osprey algorithm and avoid BPNN overfitting, the Osprey optimization algorithm is improved. The optimal weights and biases of the model are obtained by the improved Osprey optimization algorithm, as follows: S201: Initialize weights and biases based on the Sobol sequence and calculate the initial fitness value; S202: updating the positions of weights and biases in the first phase of the Osprey optimization algorithm and calculating new fitness values; introducing random numbers from a Weibull distribution as step size factors to improve the first phase position update process of the Osprey optimization algorithm; S203: updating the positions of weights and biases through the second stage of the Osprey optimization algorithm; S204: Using the weights and biases after the second stage as the initial population, and using the perturbation strategy of the firefly algorithm to perform the second stage position update; S205: Iterate the first and second stages of the Osprey optimization algorithm until the termination condition is met, and output the optimal weight and bias parameters.

[0038] Optionally, weights and biases are initialized based on the Sobol sequence as follows: Generate an N×D Sobol matrix, each element represents a set of weights and biases, traverse each parameter, and convert each value of the Sobol matrix Map from [0,1] to the upper and lower bounds of the corresponding parameters In, we get the BP neural network weight and bias population distribution, the specific formula is as follows: , in, is the dimension of the search space, For the population size.

[0039] Optionally, the first-stage position update formula after introducing Weibull distributed random numbers as step factors is as follows.

[0040] , In the formula, is a random step size factor affected by the Weibull distribution, with a scale parameter of 1 and a shape parameter of 0.5; are the new weights and biases for the first stage, For the group weights and biases, are the selected weights and biases, is a random number, ranging from .

[0041] The probability density function of the Weibull distribution is shown below: , In the formula, is a random variable, >0 is the scale parameter, >0 is the shape parameter. Specific implementation: Generate random numbers according to Weibull distribution, use them as the step factor in position update, and modify the position update formula in the standard OOA to include the Weibull step factor.

[0042] Optionally, the second-stage position update of the Osprey optimization algorithm improved by the firefly algorithm is specifically performed as follows: S2041: Use the weights and biases updated in the second stage of the Osprey algorithm as the initial population, and the fitness values ​​of the weights and biases as the initial brightness, and calculate the relative brightness and attractiveness of each pair of weights and biases; in the Firefly algorithm, the light intensity With distance Monotonous exponential change, its relative light intensity I is: , In the formula, is the initial light intensity, is the absorption coefficient of light intensity, which is 0.01 in this embodiment. Represents two sets of weights and biases and The attraction between fireflies is proportional to the light intensity seen by neighboring fireflies. Defined as: , in, It is the distance In this embodiment, the attraction is 1.

[0043] S2042: Move the positions according to the relative brightness of any two weights and biases in the initial population. When the relative brightness of one set of weights and biases is smaller than that of another set of weights and biases, the set with smaller relative brightness of weights and biases moves toward the set with larger relative brightness of weights and biases according to a formula. The formula is as follows: , in, is the position of a set of weights and biases with relatively small brightness after passing through the second stage, is the position of a set of weights and biases with relatively large brightness after passing through the second stage, is the position after the firefly interference, belongs to [0,1], is a random number that follows a uniform distribution [0,1]. is the attraction between the two sets of weights; S2043: The boundary is processed by the following formula: , S2044: If the fitness value of the updated position is higher, the position is replaced and the position update of the weight and bias of this stage is obtained. The formula is as follows: , in, It's location The fitness value of It is The position of the group weights and biases after the perturbation, After the second stage Fitness values ​​for group weights and biases.

[0044] As a refinement of the above embodiment, a total of 60 sets of sample data are collected in step S3, including crack types as input and corresponding friction coefficients as output. The data are divided into training sets and test sets in a ratio of 7:3. In order to reduce the error in the training process and improve the training efficiency, the normalization method is used to train the samples of the experimental data set, and the normalization formula is shown in Formula 7.

[0045] in, is the normalized data, For normalization, the previous data, is the maximum value of the elements in the vector, is the minimum value among the elements of the vector.

[0046] At the same time, the mean square error (MSE) in the training sample is used as a measure of the fitness function. The mean square error is specifically used to measure the square value of the difference between the predicted value and the true value during the training process. Its main function is to evaluate the size of the error. The specific formula is as follows.

[0047] , in, and Respectively represent The predicted and actual values ​​of training samples; Represents the number of training samples.

[0048] The advantages of this method are analyzed in combination with experimental data: The comparison of different numbers of hidden layer nodes shows that the mean square error (MSE) of the model is the smallest when the number of nodes is 7. Whether the number of nodes is increased or decreased, the MSE will increase. The comparison results show that the MSE of the improved BP neural network is lower than 2 during training, while the MSE of the original BP neural network remains at around 3. This proves that the error of the original BP neural network is large and is not suitable for actual prediction.

[0049] Table 1 MSE values ​​of different nodes in the hidden layer In order to evaluate the improved BP neural network's high accuracy in crack prediction, the improved BP neural network was compared with the original BP neural network, and the two networks were trained and tested using the same training set and test set data. The evaluation was performed by comparing the mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE), and determination coefficient (R-squared, R2) of the two networks on the test set. The comparison table is shown in Table 2.

[0050] Table 2 Comparison of evaluation indicators According to the comparison table of evaluation indicators, the improved BP neural network is lower than the original BP neural network in MSE, RMES, and MAE. R2 represents the degree of fit of the entire network model. The closer the value is to 1, the higher the degree of fit of the model to the data, and the closer the predicted value is to the true value. From Table 2, it can be seen that the R2 of the improved BP neural network is 0.9410, which is 0.1020 higher than the original BP neural network. The real values ​​and test values ​​in the test set of the improved BP neural network and the original BP neural network are compared. The experimental results are as follows Figure 3 shown.

[0051] In order to verify the effectiveness of the proposed method, 5 test points were selected from a certain road section for testing. The predicted values ​​of the original BP neural network and the improved BP neural network were compared with the true values. The results are shown in Table 3.

[0052] Table 3 Friction coefficient of test points It can be seen from Table 3 that the prediction results of the improved BP neural network are closer to the true value and have a smaller error. It also proves that the prediction of the friction coefficient by the number of cracks proposed in this embodiment has certain practical research significance and can meet the actual research scenario.

[0053] The embodiment of the present disclosure also provides a road friction estimation device based on deep learning, including a processor and a memory. Optionally, the device may also include a communication interface and a bus. The processor, the communication interface, and the memory may communicate with each other through the bus. The communication interface may be used for information transmission. The processor may call the logic instructions in the memory to execute the road friction estimation method based on deep learning of the above embodiment.

[0054] In addition, the logic instructions in the above-mentioned memory can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product.

[0055] The memory, as a computer-readable storage medium, can be used to store software programs and computer executable programs, such as program instructions / modules corresponding to the method in the embodiment of the present disclosure. The processor executes the function application and data processing by running the program instructions / modules stored in the memory, that is, realizing the road friction estimation method based on deep learning in the above embodiment.

[0056] The memory may include a program storage area and a data storage area, wherein the program storage area may store an operating system and an application required for at least one function; the data storage area may store data created according to the use of the terminal device, etc. In addition, the memory may include a high-speed random access memory and may also include a non-volatile memory.

[0057] An embodiment of the present disclosure provides a computer-readable storage medium storing computer-executable instructions, wherein the computer-executable instructions are configured to execute the above-mentioned road friction estimation method based on deep learning.

[0058] The computer-readable storage medium mentioned above may be a transient computer-readable storage medium or a non-transitory computer-readable storage medium.

[0059] The technical solution of the embodiment of the present disclosure can be embodied in the form of a software product, which is stored in a storage medium and includes one or more instructions for enabling a computer device (which may be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in the embodiment of the present disclosure. The aforementioned storage medium may be a non-transient storage medium, including: a USB flash drive, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a disk or an optical disk, and other media that can store program codes, or a transient storage medium.

[0060] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A road friction estimation method based on deep learning, characterized in that: include: Collect data sets, including pavement crack types and corresponding friction coefficients, and divide the data sets into test sets and training sets after preprocessing; A road friction coefficient prediction model is established, wherein the road friction coefficient prediction model is a BP neural network optimized by an improved Osprey optimization algorithm, and the optimal weight and bias of the model are obtained by the improved Osprey optimization algorithm; the improved Osprey optimization algorithm initializes weights and biases based on a Sobol sequence, a random number of a Weibull distribution is used as a step factor in the first-stage position update of the improved Osprey optimization algorithm, and a firefly algorithm is used to perturb the position of the second-stage update of the improved Osprey optimization algorithm; Input the training set data into the road friction coefficient prediction model and train the optimal weights and biases; The friction coefficient of pavement cracks is predicted using the trained pavement friction coefficient prediction model.

2. The road friction estimation method based on deep learning according to claim 1, characterized in that: The optimal weight and bias of the model are obtained by the improved Osprey optimization algorithm, and the specific method is as follows: Initialize weights and biases based on the Sobol sequence and calculate the initial fitness value; In the first phase of the Osprey optimization algorithm, the weights and bias positions are updated and new fitness values ​​are calculated; random numbers from the Weibull distribution are introduced as step size factors to improve the first phase position update process of the Osprey optimization algorithm; The positions of weights and biases are updated through the second phase of the Osprey optimization algorithm; The weights and biases after the second stage are used as the initial population, and the perturbation strategy of the firefly algorithm is used to update the position in the second stage; Iterate the first and second phases of the Osprey optimization algorithm until the termination condition is met and output the optimal weight and bias parameters.

3. The road friction estimation method based on deep learning according to claim 1, characterized in that: The activation function between neurons of the road friction coefficient prediction model adopts the nonlinear transformation Sigmoid function.

4. The road friction estimation method based on deep learning according to claim 1, characterized in that: Initialize weights and biases based on the Sobol sequence as follows: Generate an N×D Sobol matrix, each element represents a set of weights and biases, traverse each parameter, and convert each value of the Sobol matrix Map from [0,1] to the upper and lower bounds of the corresponding parameters In, we get the BP neural network weight and bias population distribution, the specific formula is as follows: ， in, is the dimension of the search space, For the population size.

5. The road friction estimation method based on deep learning according to claim 4, characterized in that: After introducing Weibull distributed random numbers as step size factors, the first-stage position update formula is as follows: ， In the formula, is a random step size factor affected by the Weibull distribution, with a scale parameter of 1 and a shape parameter of 0.5; are the new weights and biases for the first stage, For the group weights and biases, are the selected weights and biases, is a random number, ranging from .

6. The road friction estimation method based on deep learning according to claim 5, characterized in that: The second stage position update of the Osprey optimization algorithm improved by the firefly algorithm is specifically performed as follows: The weights and biases updated by the second stage of the Osprey algorithm are used as the initial population, and the fitness values ​​of the weights and biases are used as the initial brightness, and the relative brightness and attractiveness of each pair of weights and biases are calculated; The positions are moved according to the relative brightness of any two weights and biases in the initial population. When the relative brightness of one set of weights and biases is smaller than that of another set of weights and biases, the set with smaller relative brightness of weights and biases moves toward the set with larger relative brightness of weights and biases according to the formula. The formula is as follows: ， in, is the position of a set of weights and biases with relatively small brightness after passing through the second stage, is the position of a set of weights and biases with relatively large brightness after passing through the second stage, is the position after the firefly interference, belongs to [0,1], is a random number that follows a uniform distribution [0,1]. is the attraction between the two sets of weights; The boundary is processed by the following formula: ， If the fitness value of the updated position is higher, the position is replaced and the position update of the weight and bias at this stage is obtained. The formula is as follows: ， in, It's location The fitness value of It is The position of the group weights and biases after the perturbation, After the second stage Fitness values ​​for group weights and biases.

7. A road friction estimation device based on deep learning, comprising a processor and a memory storing program instructions, characterized in that: The processor is configured to execute the road friction estimation method based on deep learning as described in any one of claims 1-6 when running the program instructions.

8. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and when the program is executed by a processor, the method described in any one of claims 1 to 6 is implemented.

Images