Automobile intelligent road condition monitoring and automatic adjustment system

Abstract

The present invention relates to an intelligent road surface state monitoring and automatic adjustment system for automobiles, which belongs to the field of automobile control and includes: a camera, a sensor, a memory, a processor and an electronic control unit; the camera is used to shoot road surface images; the sensor is used to obtain sensor data; the memory is used to store road surface images and sensor data; the processor is used to identify road surface conditions according to road surface images and sensor data through a road surface state intelligent recognition model based on a convolutional neural network; the electronic control unit is used to adjust vehicle traction and braking force according to the road surface conditions through an intelligent control model based on deep reinforcement learning. The present invention can capture complex textures and illumination changes in road surface images through a convolutional neural network, and introduce additional sensor data in addition to road surface images, which can improve the road surface state recognition effect and the accuracy of road surface state recognition.

Description

Technical Field

[0001] The invention belongs to the field of automobile control, and in particular relates to an automobile intelligent road surface state monitoring and automatic adjustment system. Background Art

[0002] At present, when driving, drivers mainly rely on visual observation and experience to judge the road conditions. However, in bad weather or complex road conditions, this method is often not accurate and timely. In addition, most existing automobile safety systems focus on the status monitoring of the vehicle itself, but have limited monitoring and response capabilities for road conditions.

[0003] With the rapid development of science and technology, more and more modern technologies are being applied to intelligent driving. Intelligent identification of road conditions and adaptive control of vehicle conditions can effectively improve driving safety, driving comfort and energy efficiency, while promoting the popularization and development of intelligent driving technology. This capability is of vital importance to the realization and comprehensive optimization of future autonomous driving systems.

[0004] At present, there are mainly road surface recognition methods based on support vector machines, which take road surface images and then use support vector machines to perform road surface recognition. However, support vector machines are essentially binary classification models, and road surface conditions are diverse. Extending support vector machines to multi-classification problems requires "one-to-one" or "one-to-many" methods, resulting in high model complexity. At the same time, support vector machines mainly rely on features such as color, texture, and edges in images. These primary features are not sufficient to capture the complex texture and lighting changes in road surface images, resulting in limited recognition effects and affecting the accuracy of road surface condition recognition. Summary of the invention

[0005] In view of the above deficiencies in the prior art, the purpose of the invention is to provide an intelligent road condition monitoring and automatic adjustment system for automobiles, which can identify the road condition according to the road image and sensor data through a road condition intelligent recognition model based on a convolutional neural network, thereby reducing the complexity of the model, capturing the complex texture and lighting changes in the road image through the convolutional neural network, and introducing additional sensor data in addition to the road image, which can improve the road surface condition recognition effect and improve the accuracy of road surface condition recognition.

[0006] The present invention proposes an intelligent road condition monitoring and automatic adjustment system for automobiles, comprising: a camera, a sensor, a memory, a processor and an electronic control unit;

[0007] The camera is used to capture road surface images;

[0008] The sensor is used to obtain sensor data;

[0009] The memory is used to store the road surface image and the sensor data;

[0010] The processor is used to identify the road surface state according to the road surface image and the sensor data through a road surface state intelligent recognition model based on a convolutional neural network;

[0011] The electronic control unit is used to adjust the vehicle traction and braking force according to the road surface condition through an intelligent control model based on deep reinforcement learning.

[0012] The beneficial effects of the present invention are as follows:

[0013] (1) The method and system of the present invention can identify the road surface state according to the road surface image and sensor data through a road surface state intelligent recognition model based on a convolutional neural network, thereby reducing the complexity of the model. The convolutional neural network can capture the complex texture and illumination changes in the road surface image, and the additional sensor data introduced in addition to the road surface image can improve the road surface state recognition effect and improve the accuracy of road surface state recognition.

[0014] (2) The method and system of the present invention can adjust the vehicle traction and braking force according to the road surface conditions through an intelligent control model based on deep reinforcement learning. This can not only improve the driving safety of the vehicle under different road conditions, but also reduce unnecessary waste of braking force and optimize energy utilization efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The accompanying drawings are only used to illustrate specific embodiments and are not considered to limit the present invention. In the entire drawings, the same reference symbols represent the same components. Obviously, the drawings described below are only some embodiments recorded in the embodiments of the present invention. For those of ordinary skill in the art, other drawings can also be obtained based on these drawings.

[0016] Figure 1 A schematic diagram of the structure of an intelligent road condition monitoring and automatic adjustment system for an automobile provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0017] In order to enable those skilled in the art to better understand the technical solutions in the embodiments of the present invention, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. It should be understood that these descriptions are only exemplary and are not intended to limit the scope of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in the field without making creative work should fall within the scope of protection of the present invention.

[0018] Furthermore, in the following description, descriptions of well-known structures and technologies are omitted to avoid unnecessarily obscuring the concepts disclosed in the present invention.

[0019] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the orientation or position relationship indicated by the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inside", "outside", etc. is based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present invention and simplifying the description, and does not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation of the present invention. In addition, the terms "first", "second", and "third" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance. The terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be a connection between the two elements. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0020] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Instead, they are merely examples of methods and systems consistent with some aspects of the present invention as detailed in the appended claims.

[0021] The present invention proposes an intelligent road condition monitoring and automatic adjustment system for automobiles, which solves the road recognition method based on support vector machines in the prior art. However, support vector machines are essentially binary classification models, while road conditions are diverse. Extending support vector machines to multi-classification problems requires the use of "one-to-one" or "one-to-many" methods, resulting in high model complexity. At the same time, support vector machines mainly rely on features such as color, texture, and edges in images. These primary features are not sufficient to capture the complex texture and illumination changes in road images, resulting in limited recognition effects and affecting the accuracy of road condition recognition.

[0022] Reference Manual Attached Figure 1 , showing a structural schematic diagram of an automobile intelligent road condition monitoring and automatic adjustment system provided by an embodiment of the present invention.

[0023] An embodiment of the present invention provides an intelligent road condition monitoring and automatic adjustment system for an automobile, comprising: a camera 1, a sensor 2, a memory 3, a processor 4 and an electronic control unit (ECU) 5.

[0024] Camera 1 is used to capture road surface images.

[0025] Sensor 2 is used to obtain sensor 2 data.

[0026] Optionally, the sensor 2 includes a vibration sensor 2 and an acoustic sensor 2 .

[0027] The vibration sensor 2 is used to obtain vehicle body vibration data.

[0028] The acoustic sensor 2 is used to obtain tire road noise data.

[0029] The memory 3 is used to store the road surface image and the sensor 2 data.

[0030] Processor 4 is used to identify the road surface state according to the road surface image and the sensor 2 data through a road surface state intelligent recognition model based on a convolutional neural network.

[0031] In a possible implementation, the road surface state intelligent recognition model based on a convolutional neural network is used to recognize the road surface state, and the road surface state is specifically identified by:

[0032] The color features and texture features of the road surface image are extracted to form a color feature map and a texture feature map.

[0033] The color features include the first-order color moment, the second-order color moment and the third-order color moment of each color channel.

[0034] It should be noted that color moment is a statistic used to describe the color characteristics of an image and is widely used in image analysis and classification. According to the order of the moment, it is divided into first-order color moment (mean), second-order color moment (variance) and third-order color moment (skewness).

[0035] Among them, texture features include energy value, entropy value, contrast and correlation.

[0036] It should be noted that the energy value describes the uniformity of the grayscale distribution in the image. The larger the value, the smoother the image texture.

[0037] The entropy value indicates the amount of image information, and the higher the value, the more complex the texture.

[0038] Contrast measures the local change of grayscale value and reflects the clarity of texture.

[0039] Correlation measures the correlation between the grayscale values ​​of pixels in an image. A higher value indicates a more regular grayscale distribution in the image.

[0040] The vibration characteristics of the vehicle body vibration data are extracted to form a vibration characteristic graph.

[0041] Among them, the vibration characteristics include acceleration characteristics.

[0042] The acoustic characteristics of tire-road noise data are extracted to form an acoustic characteristic map.

[0043] Among them, the acoustic features include MFCC features.

[0044] It should be noted that MFCC (Mel Frequency Cepstral Coefficient) feature is a commonly used audio signal feature, which is widely used in speech recognition, audio analysis, road condition recognition and other fields. It can effectively capture the spectral characteristics of audio signals and compress them into low-dimensional features to facilitate subsequent analysis and classification.

[0045] The color feature map, texture feature map, vibration feature map, acoustic feature map and original road surface image are respectively input as different channels into the road surface condition intelligent recognition model based on convolutional neural network.

[0046] In the present invention, by combining images (color and texture features) with non-image data (vibration features, acoustic features), the model can analyze the road surface conditions from multiple dimensions, making up for the shortcomings of a single feature. At the same time, information from different channels can complement each other. For example, the sensitivity of acoustic features to slippery roads can make up for the difficulty of image recognition caused by changes in lighting conditions, thereby improving the accuracy of road surface condition recognition.

[0047] The road surface condition is identified through an intelligent road surface condition recognition model based on a convolutional neural network.

[0048] Among them, Convolutional Neural Network (CNN) is a deep learning model that is particularly suitable for processing image and video data and is widely used in computer vision, speech processing, natural language processing and other fields. CNN extracts local features from data and gradually forms a global representation by simulating the working principle of biological vision systems.

[0049] The road condition intelligent recognition model based on convolutional neural network includes: convolution layer, pooling layer, flattening layer, discard layer, channel attention layer, fully connected layer and classification layer.

[0050] The convolution layer is used to extract the deep feature map from the feature map of each channel through the convolution operation. The convolution layer is the core module of the convolutional neural network, which is used to extract features from local areas by sliding the convolution kernel on the input feature map. The convolution layer can effectively capture the spatial pattern and local characteristics of the input data (such as images or feature maps), such as edges, textures, and shapes. By stacking multiple convolution layers, the network can gradually extract low-level to high-level deep features, thereby realizing the expression of complex features.

[0051] The pooling layer is used to reduce the dimension of the deep feature map through pooling operations. The pooling layer reduces the dimension of the feature map by downsampling the feature map, reduces the computational complexity, and retains the key feature information. Pooling operations (such as maximum pooling or average pooling) can improve the model's translation invariance and robustness to the input data, and help the network ignore subtle local changes or noise. This process also plays a role in compressing features and reducing overfitting, providing a more refined feature representation for subsequent layer processing.

[0052] The flattening layer is used to convert the multi-dimensional feature map after pooling into a one-dimensional feature vector, which is convenient for the subsequent fully connected layer to process. The flattening layer expands the multi-dimensional feature map processed by the pooling layer into a one-dimensional feature vector, which is convenient for the subsequent fully connected layer to process. This process converts the spatial information of the feature map into a linear form, making the data transfer between the convolutional layer and the fully connected layer more efficient. The flattening layer is an important bridge connecting the feature extraction stage and the classification stage.

[0053] The dropout layer is used to randomly drop some neurons to prevent overfitting. The dropout layer prevents the model from overfitting the training data by randomly "dropping" some neurons during the training process. The dropout layer introduces a regularization mechanism that improves the generalization ability of the model by reducing the interdependence between neurons. The random dropout operation also enhances the robustness of the network, making it more stable when facing unknown data.

[0054] The channel attention layer is used to determine the weights of features extracted from different channels. The channel attention layer is used to assign weights to feature maps extracted from different channels to help the model focus on more meaningful features. By compressing the feature information of each channel and calculating the weights, the channel attention layer can dynamically adjust the priority of features, strengthen the expression of important features, and weaken the influence of irrelevant features. This mechanism improves the efficiency of feature fusion and enables the model to better understand multi-channel input data.

[0055] Optionally, the channel attention layer is specifically used to compress the feature maps extracted from each channel:

[0056]

[0057] Among them, Y k represents the feature map after the k-th channel is compressed, S represents the compression function, X k represents the feature map extracted by the kth channel, (i, j) represents the pixel coordinates in the feature map, H represents the height of the feature map, and W represents the width of the feature map.

[0058] Reconstruct the compressed feature map and determine the weights of the feature maps extracted from each channel:

[0059]

[0060] Among them, W k represents the weight of the feature map extracted by the kth channel, Sigmoid represents the Sigmoid activation function, ReLU represents the ReLU activation function, and W s represents the reconstruction of the fully connected layer, W r represents a compressed fully connected layer.

[0061] In the present invention, by calculating the global mean of the feature map (i.e., compression operation), the channel attention layer can capture the global representation of each channel, thereby quantifying the importance of the channel. By assigning weights to different channels, the model can dynamically enhance the focus on important features, while weakening irrelevant or redundant features, thereby improving the model's efficiency in utilizing key features. This is particularly important for processing multimodal data (such as color features, texture features, vibration features, acoustic features, etc.), because it can combine the advantages of different modalities to generate more discriminative feature representations.

[0062] The fully connected layer is used to summarize the features extracted from each channel. The fully connected layer aggregates the extracted features and generates a high-dimensional feature representation by connecting each neuron to all neurons in the previous layer. It maps the features extracted by the convolution and pooling layers to inputs that can be used by the classifier through linear transformation and nonlinear activation. The fully connected layer is an important part of the model decision-making, serving as a bridge from features to categories or numerical outputs.

[0063] The classification layer is used to generate the probability distribution of each road condition category through the activation function, and the road condition category with the largest probability value is used as the final recognition result. The classification layer is used to generate the final classification result. Through Softmax or other activation functions, the classification layer converts the output of the fully connected layer into the probability distribution of each category, and uses the category with the largest probability as the prediction result of the model. The classification layer is responsible for mapping features to specific road condition categories, such as dry asphalt, wet asphalt or snowy road surface, and is the final output stage of the intelligent recognition model.

[0064] Among them, road conditions include: dry asphalt, slightly wet asphalt, wet asphalt, heavily wet asphalt, dry cement, cobblestones, snow and ice.

[0065] Optionally, the classification layer is specifically used to generate a probability distribution of each road surface condition category through a Softmax activation function, and take the road surface condition category with the largest probability value as the final recognition result:

[0066]

[0067] Where P represents the probability distribution of road surface conditions, p urepresents the probability value of belonging to the u-th road condition category, U represents the total number of road condition categories, Softmax represents the Softmax activation function, and b k represents the bias term of the kth channel, and K represents the total number of channels.

[0068] In the present invention, the probability distribution is generated by the Softmax activation function and the category with the maximum probability is selected as the recognition result, which can intuitively reflect the classification confidence of the model, fully integrate multi-channel feature information, and provide stable, efficient, and easy-to-interpret classification decisions. This method not only improves the accuracy of road state recognition, but also ensures the stability and applicability of the model.

[0069] The electronic control unit 5 is used to adjust the vehicle traction and braking force according to the road conditions through an intelligent control model based on deep reinforcement learning.

[0070] Among them, Deep Reinforcement Learning (DRL) is a machine learning technology that combines deep learning and reinforcement learning, focusing on enabling intelligent agents to learn optimal decision-making strategies in dynamic environments through trial and error mechanisms. By using deep neural networks as function approximators, DRL can handle high-dimensional state spaces and is suitable for complex control problems, such as vehicle traction and braking force adjustment in intelligent driving.

[0071] Optionally, the environment space of the intelligent control model based on deep reinforcement learning is specifically:

[0072]

[0073] Among them, s represents the environment, v represents the current vehicle speed, represents the current acceleration, e represents the current tire slip rate, F r Indicates the current traction, F a represents the current braking force, z represents the current vehicle body vibration, and y represents the current road surface condition.

[0074] Optionally, the action space of the intelligent control model based on deep reinforcement learning is specifically:

[0075]

[0076] Among them, a represents action, ΔF r Indicates the traction adjustment range, ΔF a Indicates the braking force adjustment range.

[0077] Optionally, the reward function of the intelligent control model based on deep reinforcement learning is specifically:

[0078]

[0079] Among them, r represents the reward function, r 1 represents the driving smoothness reward function, r 2 represents the slip rate reward function, r 3 represents the fuel consumption penalty function, r 4 represents the engine efficiency reward function, λ 1 represents the driving comfort reward coefficient, λ 2 represents the slip rate reward coefficient, λ 3 represents the fuel consumption penalty coefficient, λ 4 Represents the engine efficiency bonus coefficient.

[0080] Among them, those skilled in the art can set the driving smoothness reward coefficient λ according to actual conditions. 1 , slip rate reward coefficient λ 2 , fuel consumption penalty coefficient λ 3 and the engine efficiency bonus factor λ 4 The present invention does not limit the size.

[0081] In the present invention, by clearly defining the environment space, action space and reward function, the intelligent control model based on deep reinforcement learning can respond to complex driving environments in real time and dynamically adjust the vehicle state, which not only improves safety and comfort, but also optimizes energy efficiency, and ultimately achieves the global optimal control strategy.

[0082] Optionally, the driving smoothness reward function is specifically:

[0083]

[0084] Wherein, I represents the driving comfort index.

[0085] It should be noted that the driving smoothness reward function ensures that the worse the smoothness, the smaller the reward; the better the smoothness, the larger the reward, thereby driving the model to learn the optimal smoothness control strategy.

[0086] Optionally, the driving comfort index is calculated based on the current vehicle body vibration:

[0087]

[0088] Where I represents the driving comfort index, V represents the vibration amplitude measurement value, and V bi represents the vibration amplitude measurement value of the ith suspension-body connection, V c Represents the vibration amplitude measurement value at the seat, λ b represents the weight coefficient of the suspension-body connection, λ c represents the weight coefficient at the seat, represents the acceleration at the ith suspension-body connection, represents the acceleration of the i-th suspension-body connection at time t, d represents the differential sign, and T represents the length of the time window.

[0089] It should be noted that the vibration data at the suspension body connection and the seat can be obtained through the vibration sensor 2.

[0090] Furthermore, by calculating the vibration amplitude at the suspension-body connection and the vibration amplitude at the seat, the vibration conditions of each key part of the vehicle are comprehensively evaluated, which can fully reflect the driving smoothness. At the same time, the introduction of the time window makes the vibration evaluation not only focus on the instantaneous value, but also reflect the overall smoothness during driving, which is particularly suitable for analyzing the comfort of long-term driving.

[0091] Optionally, the slip rate reward function is specifically:

[0092]

[0093] Where, e represents the current tire slip rate, e g Indicates the target slip ratio.

[0094] It should be noted that the slip rate reward function clarifies the optimization goal of slip rate control, which is to minimize the deviation between the current slip rate and the target slip rate. With the continuous adjustment of the control strategy, the model can gradually optimize the dynamic control of the slip rate, maximize the reward function, and quickly reach a stable state.

[0095] Optionally, the target slip ratio is specifically:

[0096]

[0097] Among them, C 1 , C 2 , C 3 It represents the fitting coefficient, which is related to the road surface condition, and ln represents an exponential function with a natural constant as the base.

[0098] When the road surface is dry asphalt, C 1 =1.281, C 2 =23.993, C 3 =0.520.

[0099] When the road surface is slightly wet asphalt, C 1 =0.628, C 2 =33.765, C 3 =0.200.

[0100] When the road surface is wet asphalt, C 1 =0.856, C 2=33.821, C 3 =0.345.

[0101] When the road surface is wet asphalt, C 1 =1.072, C 2 =29.494, C 3 =0.422.

[0102] When the pavement is in dry cement condition, C 1 =1.196, C 2 =25.166, C 3 =0.539.

[0103] When the road surface is cobblestone, C 1 =0.400, C 2 =60.010, C 3 =0.120.

[0104] When the road surface is covered with snow, C 1 =0.195, C 2 =94.129, C 3 =0.065.

[0105] When the road surface is icy, C 1 =0.050, C 2 =306.390, C 3 =0.001.

[0106] Among them, the fitting coefficients cover common road types (such as dry asphalt, wet asphalt, snow, ice, etc.), and the model can flexibly adapt to various road conditions.

[0107] It should be noted that the target slip rate is dynamically adjusted according to different road conditions (such as dry asphalt, wet road surface, and snowy road surface) to ensure that the vehicle maintains optimal traction and stability in various complex environments.

[0108] Optionally, the fuel consumption penalty function is specifically:

[0109]

[0110] Wherein, F represents the instantaneous fuel consumption.

[0111] It should be noted that the fuel consumption penalty function provides negative rewards through instantaneous fuel consumption, driving the model to reduce excessive fuel use and optimize the vehicle's fuel economy.

[0112] Optionally, the engine efficiency reward function is specifically:

[0113]

[0114] Among them, η m Indicates engine efficiency.

[0115] It should be noted that the engine efficiency reward function provides positive rewards according to the efficiency value of the engine, and the drive system keeps the engine's operating point in the high-efficiency area to reduce energy waste.

[0116] Optionally, the specific working method of the intelligent control model based on deep reinforcement learning includes:

[0117] In the online network, the ε-greedy strategy is used in the environment s t Next select an action a t ,s t represents the environment at the tth iteration, a t represents the action selected at the tth iteration.

[0118] Among them, the ε-greedy strategy is a common action selection strategy in reinforcement learning, which is used to find a balance between exploration and exploitation. Through this strategy, the agent can fully explore the environment in the early stage of learning, and gradually tend to use the existing optimal strategy to make decisions in the later stage of learning. Select a random action (exploration) with probability ε to ensure that the agent tries different actions and understands the various possibilities of the environment. Select the action that is currently considered to be the best (exploitation) with probability 1-ε, and use the learned experience to maximize the current reward.

[0119] Execute action a t , get reward r t , r t Represents the reward obtained at the tth iteration.

[0120] When this Stored in the experience pool as the parameter at time t.

[0121] The Q function is introduced to calculate the Q value under various strategies. The Q value represents the expectation of the long-term return of all actions:

[0122]

[0123] Among them, Q represents the Q value, π represents the strategy, and Q π represents the Q value when adopting strategy π, E represents the expected operation, r t+1 represents the reward obtained at the t+1th iteration, and γ represents the discount factor.

[0124] It should be noted that by using the Q function to estimate the long-term returns of all strategies and combining it with the discount factor, the model can weigh short-term rewards against long-term goals and optimize vehicle performance overall.

[0125] In the target network, a deep neural network is used to fit the optimal Q value function, which is specifically:

[0126]

[0127] in, represents the optimal Q value, max represents the maximum value, r t+2 Represents the reward obtained at the t+2th iteration.

[0128] Randomly extract small batches of samples from the experience pool. Storing the data of the agent's interaction with the environment in the experience pool and randomly extracting small batches of samples to train the model can break the time correlation of the data and improve training stability and efficiency.

[0129] Construct the mean square error loss function:

[0130]

[0131] Among them, L represents the mean square error loss function, θ t represents the target network parameters at the tth iteration, represents the online network parameters at the tth iteration, D represents the experience pool, and M represents the number of samples drawn from the experience pool.

[0132] It should be noted that the mean square error loss function minimizes the difference between the true Q value and the predicted Q value, driving the model to more accurately estimate the long-term reward of each action, thereby optimizing the decision quality.

[0133] Update the target network parameters using gradient descent:

[0134]

[0135] Among them, θ t+1 represents the target network parameter at the t+1th iteration, ρ represents the learning rate, Represents a gradient operation.

[0136] According to the optimal Q value function, the vehicle traction and braking force are adjusted.

[0137] In the present invention, the intelligent and dynamic adjustment of vehicle traction and braking force is achieved through the exploration and learning mechanism of reinforcement learning, the efficient sample utilization method and the powerful feature extraction ability of deep neural network. This method not only improves driving safety and energy efficiency, but also enhances the stability and adaptability of the model, providing a comprehensive and reliable solution for intelligent driving control.

[0138] In a possible implementation, after the braking force adjustment range is determined by the intelligent control model based on deep reinforcement learning, the front and rear wheel braking force distribution method specifically includes:

[0139] Get the current braking force adjustment range.

[0140] Convert this braking force adjustment range into braking intensity:

[0141]

[0142] Where q represents the braking intensity, F a Indicates the current braking force, ΔF a Indicates the braking force adjustment range, F am Indicates the maximum braking force.

[0143] When the braking strength meets When the car is in a light braking state, braking is performed only by the front wheel motor:

[0144]

[0145] Among them, F fa Indicates the front wheel braking force, F ba represents the rear wheel braking force, and G represents the vehicle's gravity.

[0146] It should be noted that during light braking, the front wheels bear the braking force alone, reducing rear wheel intervention and lowering energy consumption.

[0147] When the braking strength meets When the brake is in the medium state, the front and rear wheel electric motor power is distributed as follows:

[0148] Among them, b is the distance from the center of mass of the vehicle to the center line of the rear axle, h is the height of the center of mass of the vehicle, and L is the wheelbase.

[0149] It should be noted that during moderate braking, the front and rear axle distribution ratio is accurately calculated to ensure that the front axle bears a larger proportion of the braking force, thereby maximizing energy recovery efficiency.

[0150] When the braking strength meets When the emergency brake is in effect, the front and rear wheel electric motor power is distributed as follows:

[0151] in, It represents the road adhesion coefficient, which is related to the road condition. a represents the distance from the center of mass of the vehicle to the center line of the front axle.

[0152] It should be noted that in an emergency braking state, the braking force is dynamically adjusted according to the adhesion coefficient of different road surfaces (dry, wet, snowy, etc.) to ensure the braking stability of the vehicle under complex road conditions.

[0153] In the present invention, based on the segmented distribution of braking intensity, physical parameters (such as center of mass position, adhesion coefficient) are combined with dynamic adjustment formulas, which not only improves the safety and efficiency of the braking system, but also takes into account the needs of energy recovery, smoothness and intelligent control. This design enables the vehicle to adapt to a variety of driving scenarios and complex road conditions, and is an efficient way to implement an intelligent braking system.

[0154] The beneficial effects of the present invention are as follows:

[0155] (1) The method and system of the present invention can identify the road surface state according to the road surface image and the data of sensor 2 through a road surface state intelligent recognition model based on a convolutional neural network, thereby reducing the complexity of the model. The convolutional neural network can capture the complex texture and illumination changes in the road surface image. In addition to the road surface image, the additional introduction of sensor 2 data can improve the road surface state recognition effect and improve the accuracy of road surface state recognition.

[0156] (2) The method and system of the present invention can adjust the vehicle traction and braking force according to the road surface conditions through an intelligent control model based on deep reinforcement learning. This can not only improve the driving safety of the vehicle under different road conditions, but also reduce unnecessary waste of braking force and optimize energy utilization efficiency.

[0157] The applicant of the present invention has made a detailed explanation and description of the implementation examples of the present invention in conjunction with the drawings in the specification. However, those skilled in the art should understand that the above implementation examples are only preferred implementation schemes of the present invention, and the detailed description is only to help readers better understand the spirit of the present invention, but not to limit the scope of protection of the present invention. On the contrary, any improvements or modifications based on the inventive spirit of the present invention should fall within the scope of protection of the present invention.

[0158] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the embodiments of the present invention, rather than to limit them. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the above embodiments, or replace some of the technical features therein by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention. Any changes or replacements that can be easily thought of by those skilled in the art within the technical scope disclosed in the present invention should be covered within the protection scope of the present invention.

Claims

1. An intelligent road condition monitoring and automatic adjustment system for automobiles, characterized in that: include: Cameras, sensors, memory, processors, and electronic control units; The camera is used to capture road surface images; The sensor is used to obtain sensor data; The memory is used to store the road surface image and the sensor data; The processor is used to identify the road surface state according to the road surface image and the sensor data through a road surface state intelligent recognition model based on a convolutional neural network; The electronic control unit is used to adjust the vehicle traction and braking force according to the road surface state through an intelligent control model based on deep reinforcement learning; The sensor includes a vibration sensor and an acoustic sensor; the vibration sensor is used to obtain vehicle body vibration data; the acoustic sensor is used to obtain tire road noise data; The specific environment space of the intelligent control model based on deep reinforcement learning is: ; in, s Indicates the environment, v Indicates the current vehicle speed. Indicates the current acceleration, e Indicates the current tire slip rate, F r Indicates the current traction force, F a Indicates the current braking force. z Indicates the current vehicle body vibration condition. y Indicates the current road condition; The action space of the intelligent control model based on deep reinforcement learning is as follows: ;in, a Indicates action, Δ F r Indicates the traction adjustment range, Δ F a Represents the adjustment range of the braking force; the reward function of the intelligent control model based on deep reinforcement learning is: ; in, r represents the reward function, r 1 represents the driving smoothness reward function, r 2 represents the slip rate reward function, r 3 represents the fuel consumption penalty function, r 4 represents the engine efficiency reward function, λ 1 represents the driving smoothness reward coefficient, λ 2 represents the slip rate bonus coefficient, λ 3 represents the fuel consumption penalty coefficient, λ 4 represents the engine efficiency bonus factor; The driving comfort reward function is specifically: ; in, I Indicates the driving comfort index; The driving comfort index is calculated based on the current vehicle body vibration condition: ; in, I Indicates the driving comfort index. V Indicates the vibration amplitude measurement value, V bi Indicates i A measure of the vibration amplitude at the suspension-body connection, V c Indicates the vibration amplitude measurement value at the seat, λ b represents the weight coefficient of the suspension-body connection, λ c represents the weight coefficient at the seat, Indicates i The acceleration at the suspension-body connection, Indicates i Suspension-body connection t The acceleration at the moment, d represents the differential sign, T Indicates the length of the time window; The slip rate reward function is specifically: ; in, e Indicates the current tire slip rate, e g represents the target slip ratio; The fuel consumption penalty function is specifically: ; in, F Indicates instantaneous fuel consumption; The engine efficiency reward function is specifically: ; in, η m Indicates engine efficiency; The specific working methods of the intelligent control model based on deep reinforcement learning include: In the online network, using the ε-greedy strategy in the environment s t Next select an action a t , s t Indicates t The environment at the time of the iteration, a t Indicates t The action selected at the iteration; Execute an action a t , get reward r t , r t Indicates t The reward obtained at the iteration; When this As t The parameters at each moment are stored in the experience pool. s t+1 Indicates t +1 environment at iteration time; The Q function is introduced to calculate the Q value under various strategies. The Q value represents the expectation of the long-term return of all actions: ； in, Q represents the Q value, π Indicates strategy, Q π Indicates the adoption strategy π of Q value, E represents the expectation operation, r t+1 Indicates t +1 reward for the iteration, γ represents the discount factor; In the target network, a deep neural network is used to fit the optimal Q value function, which is specifically: ； in, Q * indicates the optimal Q value, and max indicates the maximum value. r t+2 Indicates t +Rewards obtained in 2 iterations; According to the optimal Q value function, the vehicle traction and braking force are adjusted.

2. The automobile intelligent road condition monitoring and automatic adjustment system according to claim 1 is characterized in that: The road surface state intelligent recognition model based on the convolutional neural network specifically includes: Extracting color features and texture features of the road surface image to form a color feature map and a texture feature map; Extracting vibration characteristics of the vehicle body vibration data to form a vibration characteristic graph; Extracting acoustic features of the tire-pavement noise data to form an acoustic feature map; The color feature map, the texture feature map, the vibration feature map, the acoustic feature map and the original road surface image are respectively input as different channels into a road surface state intelligent recognition model based on a convolutional neural network; Identify the road surface condition through the road surface condition intelligent recognition model based on convolutional neural network; The road surface state intelligent recognition model based on convolutional neural network includes: a convolution layer, a pooling layer, a flattening layer, a discarding layer, a channel attention layer, a fully connected layer and a classification layer; The convolution layer is used to extract the depth feature map in the feature map of each channel through a convolution operation; The pooling layer is used to reduce the dimension of the depth feature map through a pooling operation; The flattening layer is used to convert the pooled multi-dimensional feature map into a one-dimensional feature vector to facilitate the processing of the subsequent fully connected layer; The dropout layer is used to randomly drop a portion of neurons to prevent overfitting; The channel attention layer is used to determine the weights of features extracted from different channels; The fully connected layer is used to summarize the features extracted from each channel; The classification layer is used to generate the probability distribution of each road surface condition category through an activation function, and take the road surface condition category with the largest probability value as the final recognition result.

3. The automobile intelligent road condition monitoring and automatic adjustment system according to claim 2 is characterized in that: The color features include first-order color moments, second-order color moments, and third-order color moments of each color channel, the texture features include energy values, entropy values, contrast, and correlation, the vibration features include acceleration features, and the acoustic features include MFCC features.

4. The automobile intelligent road condition monitoring and automatic adjustment system according to claim 1 is characterized in that: After the braking force adjustment range is determined by the intelligent control model based on deep reinforcement learning, the front and rear wheel braking force distribution methods include: Get the braking force adjustment range of this time; Convert this braking force adjustment range into braking intensity: ; in, q Indicates the braking strength. F a Indicates the current braking force, Δ F a Indicates the braking force adjustment range. F am Indicates the maximum braking force; When the braking intensity meets When the car is in a light braking state, braking is performed only by the front wheel motor: ; in, F fa Indicates the front wheel braking force, F ba Represents the rear wheel braking force, G Indicates the weight of the whole vehicle; When the braking intensity meets When the brake is in the medium state, the front and rear wheel electric motor power is distributed as follows: ; in, b It represents the distance from the center of mass of the vehicle to the centerline of the rear axle. h represents the height of the vehicle's center of mass, L Wheelbase; When the braking intensity meets When the emergency brake is in effect, the front and rear wheel electric motor power is distributed as follows: ; in, Represents the road adhesion coefficient, which is related to the road surface condition. a Indicates the distance from the vehicle's center of mass to the centerline of the front axle.

Images