A tire force estimation method based on physical information fusion recurrent neural network

Abstract

This invention belongs to the field of tire force estimation, and specifically relates to a tire force estimation method based on a recurrent neural network fused with physical information. The method includes the following steps: 1) Designing a loss function based on data labels and using a recurrent neural network algorithm to capture the intrinsic relationship between tire force and vehicle state sequence; 2) Combining a vehicle dynamics model to develop a loss function based on physical information, ensuring that the estimation process follows physical laws; 3) Integrating the above two loss functions to form a tire force estimation method that comprehensively considers data-driven and physical constraints. By deeply exploring the relationship between tire force and vehicle state through recurrent neural networks and incorporating a loss function based on vehicle dynamics, useful information is extracted from prior knowledge, improving the estimation effect under varying road conditions. This invention combines the flexibility of data-driven approaches with the rigor of physical models, not only improving the accuracy of tire force estimation but also enhancing the model's generalization ability and interpretability.

Description

Technical Field

[0001] This invention belongs to the field of tire force estimation, and specifically relates to a tire force estimation method based on physical information fusion recurrent neural network. Background Technology

[0002] In recent years, with the rapid development of the social economy, automobiles, as a modern tool to ensure people's daily needs and improve production efficiency, have seen a continuous increase in their ownership. The development of the automotive industry reflects the progress of my country's industrial technology and plays an indispensable and positive role in improving people's quality of life. However, with the increase in car ownership, traffic congestion and road accidents caused by driver fatigue, drunk driving, and other reasons occur frequently, increasing the risks to the road traffic system. With the rapid development of vehicle networking and artificial intelligence technologies, research on autonomous vehicles has attracted the attention of scholars worldwide in order to reduce traffic accidents caused by driver error.

[0003] Autonomous driving systems mainly consist of key modules such as an environmental perception module, a planning and decision-making module, and a vehicle control execution module. Autonomous vehicles use visual sensors and LiDAR sensors to collect information about their surrounding environment and make judgments about specific driving conditions. The sensor fusion and control module, based on the information output by the environmental perception module, completes global and local path planning. Finally, the control execution module, based on the results of the path planning module and considering the vehicle's longitudinal and lateral stability and safety constraints, solves the control strategy for the vehicle at various times, thereby achieving safe driving of the autonomous vehicle. Among these, vehicle motion control is at the end of the autonomous driving technology chain, carrying the main behavior of the intelligent system. It serves as a bridge for communication between the vehicle and other modules, supporting and implementing the algorithm architecture of the autonomous driving system.

[0004] Accurate dynamic models are fundamental to the design and optimization of control algorithms for intelligent driving vehicles. In vehicle models, tire-road friction is the primary source of vehicle dynamics, especially lateral tire force, which directly determines the vehicle's braking, steering, and other handling performance. Therefore, accurate estimation of lateral tire force is crucial for the safe operation of autonomous vehicles. However, because tires exhibit characteristics that are neither rigid nor elastic, and because friction is related to various factors such as ground conditions, lateral slip angle, longitudinal slip, tire load, tire pressure, temperature, and wear, it is difficult to accurately describe this parameter through mechanistic modeling. With the development of artificial intelligence, some studies have attempted to establish tire force models using machine learning methods. However, traditional tire force models based on pure data-driven machine learning methods often perform poorly in practical applications. Recent research has focused on enhancing the performance of machine learning models by incorporating prior physical knowledge, such as integrating prior knowledge into machine learning through algebraic or differential equations, to help the model better understand the data. For the tire force modeling problem, using a modeling method that fuses physical information with neural networks to improve model interpretability and generalization ability is of great significance for the application of models in real-world scenarios.

[0005] This invention addresses the strong nonlinearity of tire forces under complex and variable road conditions. To reduce the impact of unknown road conditions on vehicle tracking and safety performance, it introduces a vehicle tire force model based on a physical information fusion neural network, improving the estimation effect under variable road conditions and enhancing the model's generalization ability and interpretability. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a tire force estimation method based on physical information fusion recurrent neural network.

[0007] The technical solution of this invention is as follows:

[0008] A tire force estimation method based on physical information fusion recurrent neural network includes the following steps:

[0009] Step 1: Collect vehicle state data and tire force data under different driving conditions based on the Carsim simulation platform;

[0010] Step 2: From the vehicle state data in Step 1, select the following vehicle states related to tire force as network input parameters:

[0011]

[0012] Where δ is the front wheel steering angle, It is the yaw rate, v x ,v y These are the longitudinal velocity and the lateral velocity, a x ,ay It is the longitudinal acceleration and the lateral acceleration, σ fl ,,σ fr ,ω rl ,σ rr These are the rotational speeds of the left front wheel, right front wheel, left rear wheel, and right rear wheel, respectively.

[0013] Step 3: Perform mean-variance standardization on the vehicle status data from Step 2, and divide it into training and test sets;

[0014] Step 4, design the network structure; In this invention, the gated recurrent unit (GRU) is selected as the main recurrent neural network structure. In order to enhance the estimation effect of the model, a fully connected layer is added before the input layer as the input projection layer, and a fully connected layer is added after the output layer as the output regression layer.

[0015] Step 5: Process the vehicle state data training set sequence using the network from step 4 to obtain the estimated value of tire force f(x) at time i. i ,θ),i=T+1,…,N,T represents predicting the current tire force using the state at the past T time points, andN is the number of training data;

[0016] Step 6: Obtain the tire force estimate f(x) from the training set after step 5. i After considering the parameters θ, a loss function for physical information fusion is designed to update the network parameters. The design of the physical information fusion loss function consists of two parts: designing a data label-based loss function based on a recurrent neural network algorithm; and designing a physical information-based loss function based on a simplified vehicle dynamics model. The two loss functions are then superimposed to realize a tire force estimation method based on a physical information fusion recurrent neural network.

[0017] Step 7: Based on the loss function, update the weight parameters in the network according to the backpropagation algorithm and the chain rule;

[0018] Step 8: Train the network parameters, set the maximum number of training epochs M, and repeat steps 5 to 7 within the training epochs until the network meets the error requirements or reaches the maximum number of training epochs, then exit the loop.

[0019] Step 9: Test the network estimation effect; Based on the state data of the test set and the forward propagation process in step 4, obtain the tire force estimation value of the test set. At this time, the tire force estimation value is still under the standardized scale. After destandardization, obtain the tire force estimation value under the original scale, and realize the tire force estimation based on physical information fusion recurrent neural network.

[0020] Compared with the prior art, the present invention has the following beneficial effects:

[0021] In vehicle models, tire-road friction is a primary source of vehicle dynamics, especially lateral tire force, which directly determines the vehicle's braking, steering, and other handling performance. Therefore, accurate estimation of lateral tire force is crucial for the safe operation of autonomous vehicles. Traditional data-driven deep learning typically requires massive amounts of data to train a high-performance and sufficiently general model. Furthermore, purely data-driven models may not satisfy physical laws or external constraints. Therefore, this paper considers using a neural network that fuses physical information to estimate lateral tire force. In the network structure, to accurately capture the nonlinear characteristics of tire force, in addition to considering the static mechanical properties of the tire-ground contact area, dynamic characteristics caused by operations such as automatic transmission should also be considered. Therefore, a simple and high-performance recurrent neural network—the gated recurrent unit (GRU)—is chosen as the main network structure, and input projection layers and output regression layers are constructed to enhance the model's fitting ability. Simultaneously, an additional loss function based on the vehicle's dynamics model is constructed and fused into the recurrent neural network to improve estimation performance under varying road conditions, enhancing the model's generalization ability and interpretability. Attached Figure Description

[0022] Figure 1 This is a flowchart of the tire force estimation method based on physical information fusion neural network proposed in this invention;

[0023] Figure 2 This is a structural diagram of the gated loop unit according to an embodiment of the present invention;

[0024] Figure 3 This is a simplified eight-degree-of-freedom model diagram of a vehicle according to an embodiment of the present invention;

[0025] Figure 4 This is a structural diagram of the tire force model based on a physical information fusion recurrent neural network according to an embodiment of the present invention;

[0026] Figure 5 This is a comparison chart of the estimation results of PGRU in this embodiment of the invention with those of other pure data machine learning methods;

[0027] Figure 6 This is an estimation result diagram of the algorithm of the embodiment of the present invention with a tire-road friction coefficient of 0.75;

[0028] Figure 7 This is an estimation result diagram of the algorithm of the embodiment of the present invention when the tire-road friction coefficient is 0.85;

[0029] Figure 8 This is an estimation result diagram of the algorithm of the embodiment of the present invention with a tire-road friction coefficient of 0.9. Detailed Implementation

[0030] The technical solutions provided in this application will be further described below with reference to specific embodiments and accompanying drawings. The advantages and features of this application will become clearer from the following description.

[0031] like Figure 1 A tire force estimation method based on physical information fusion recurrent neural network includes the following steps:

[0032] Step 1: Collect vehicle status data and tire force data under different driving conditions based on the Carsim simulation platform.

[0033] Step 2: From the vehicle state data in Step 1, select the following vehicle states related to tire force as network input parameters:

[0034]

[0035] Where δ is the front wheel steering angle, It is the yaw rate, v x ,v y These are the longitudinal velocity and the lateral velocity, a x ,a y It is the longitudinal acceleration and the lateral acceleration, σ fl ,,σ fr ,ω rl ,σ rr These are the rotational speeds of the left front wheel, right front wheel, left rear wheel, and right rear wheel, respectively.

[0036] Step 3: Perform mean-variance standardization on the vehicle status data from Step 2, and divide it into training set and test set.

[0037] Step 4, design the network structure; In this invention, the gated recurrent unit (GRU) is selected as the main recurrent neural network structure. In order to enhance the estimation effect of the model, a fully connected layer is added before the input layer as the input projection layer, and a fully connected layer is added after the output layer as the output regression layer.

[0038] The overall network structure is as follows Figure 4 As shown, a gated recurrent unit (GRU) is selected as the main recurrent neural network structure (e.g., Figure 2 (As shown). To enhance the model's estimation performance, a fully connected layer is added before the input layer as an input projection layer. Each neuron in the input projection layer has a length x... il It can be represented as:

[0039] x il =f i (W i x t +b i )

[0040] Where, x t=[x t-T ,x t-T+1 ,…,x t-1 [W] represents the network input, where T represents the prediction of the current tire force using the states from the past T time points, and W... i It is the weight matrix, b i It is a deviation, f i It is the activation function. In this embodiment, the activation function of the input projection layer is selected as the linear rectified unit (ReLU).

[0041] The mathematical description of a gated loop unit is:

[0042] z t =σ(W z x il，t +U z h t-1 +b z )

[0043] r t =α(W r x il，t +U r h t-1 +b r )

[0044]

[0045]

[0046] Among them, z t and r t From input x il,t and bias b z b r W is computed through a fully connected layer. z W r W h and U z U r U h These are weight parameters; reset gate r t Controlling the hidden layer state h from the previous moment t-1 Update to the current candidate hidden layer state To what extent, update gate z t Controlling state h t-1 Update to the current hidden layer state h t The degree of sigmoid activation function σ limits the output to between 0 and 1, and is treated as a logical switch. Candidate states It is also calculated through a fully connected layer, using the tanh activation function to ensure that the values ​​in the candidate hidden states remain between -1 and 1. The symbol ⊙ represents the Hadamard product, which is element-wise multiplication.

[0047] To enhance model estimation performance, a fully connected layer is added to the output layer of the gated recurrent unit. The mathematical representation of the output layer is as follows:

[0048] y = W o h o,t +b o

[0049] Among them, h o,t It is the hidden state output by the GRU, W o It is the weight matrix, b o y represents the bias, and y is the tire force estimation data obtained after the GRU hidden layer state is regressed through the output layer.

[0050] For the training set, the forward propagation of the neural network is completed based on the above network structure. In this embodiment, T is set to 10, that is, the current tire force is predicted using the states of the past 10 time steps, and the network input is x. t =[x t-10 ,x t-9 ,…,x t-1 ]. x t The hidden state x of the input projection layer is obtained through the forward fully connected layer operation. il,t The hidden state x of the input projection layer. il,t The input is further fed into the gated loop unit, where long-term dependencies in the sequence are captured through reset gates, update gates, etc. Finally, the hidden layer state h is output by the gated loop unit. o,t After the output fully connected layer operation, h o,t This is converted into estimated tire force data. The output layer contains four neurons, representing the estimated tire forces f(x) for the left front tire, right front tire, left rear tire, and right rear tire. t ,θ).

[0051] Step 5: Process the vehicle state data training set sequence using the network from step 4 to obtain the estimated value of tire force f(x) at time i. i ,θ),i=T+1,…,N, where N is the number of training set data;

[0052] Step 6: Obtain the tire force estimate f(x) from the training set after step 5. i After (θ), a loss function for physical information fusion is designed to update the network parameters. The design of the loss function for physical information fusion mainly includes two parts: designing a loss function based on data labels based on the recurrent neural network algorithm; and designing a loss function based on physical information based on a simplified vehicle dynamics model. The two loss functions are then superimposed to realize a tire force estimation method based on physical information fusion recurrent neural network.

[0053] The specific implementation method of the loss function design is as follows:

[0054] (1) Vehicle dynamics model:

[0055] Consider using a widely used simplified eight-degree-of-freedom vehicle model to establish the dynamic equations, such as Figure 3 As shown. The coordinate system has the vehicle's center of mass (G) as its origin, with longitudinal motion as the X-axis and lateral motion as the Y-axis. Based on this simplified vehicle model, the following dynamic relationships exist:

[0056]

[0057]

[0058]

[0059] Where δ is the steering angle of the front wheel, v x and v y These are longitudinal velocity and lateral velocity, and These are longitudinal and lateral accelerations, l f and l r These are the distances from the vehicle's center of gravity (G) to the front and rear axles, respectively. It's the yaw angle. It is the yaw rate, l w It's the wheelbase, F xf and F xr F represents the resultant longitudinal tire force of the front and rear wheels, respectively. yf and F yr These represent the resultant lateral tire forces of the front and rear wheels, respectively, and their mathematical description is as follows:

[0060] F xf =F xfl +F xfr F yf =F yfl +F yfr

[0061] F xr =F xrl +F xrr F yr =F yrl +F yrr

[0062] ΔF xf =F xfl -F xfr ΔF yf =F yfl -F yfr

[0063] ΔF xr =F xrl -F xrr

[0064] Among them, F xj and F yj (j:= fl, fr, fl, rr) represent the longitudinal and lateral tire forces of the front left wheel, front right wheel, rear left wheel, and rear right wheel, respectively.

[0065] (2) Tire force estimation model based on physical information fusion recurrent neural network

[0066] The loss function of the network structure proposed in this invention, which combines a vehicle physics model with a deep neural network, mainly consists of two parts: a loss function based on data labels and a loss function based on physical information constraints, which are constructed as follows:

[0067] a) Loss function based on data labels

[0068] A data label-based loss function is constructed based on tire force estimation data and reference tire force data:

[0069]

[0070] Where, F = [F yfl ,F yfr ,F yrl ,F yrr ] T These are the tire force data labels from the training set. These are tire force estimates obtained from the network with weight θ.

[0071] b) Loss function based on physical information

[0072] To enable the network to acquire more information from prior physical knowledge and ensure the model satisfies fundamental physical constraints, this invention integrates dynamic constraints into the loss function. Considering the complexity, we transform the dynamic model into the following form:

[0073]

[0074] According to dynamic constraints, f PHY If the value is 0, then the following loss function based on physical information constraints can be constructed:

[0075]

[0076] The estimated tire forces and related states should satisfy the above loss function constraints, meaning the recurrent neural network structure must satisfy the dynamic physical constraints. Therefore, the network's loss function can be written as:

[0077] loss = λ l loss label +λ ploss phy +λ r R(θ)

[0078] Where, λ l ,λ p ,λ r It is a weight scalar, where λ p This can be intuitively understood as the proportion of physical information to be added to the network. It is a regularization function used to avoid overfitting.

[0079] Preferably, λ l ,λ p ,λ r Set them to 0.5, 0.5 and 0.001 respectively.

[0080] In the loss function above, the estimated value of tire force f(x) is calculated based on the data label loss function. i ,θ) and the true value F i The sum of squared errors in the training set is expected to bring the estimated value closer to the true value; the loss function based on physical information will estimate the tire force f(x). i Substituting θ into the zero-equality physical constraint f of the vehicle dynamics model PHY The expected training set f PHY The parameters tend to be 0 to satisfy physical constraints, thus enabling the training of network parameters under the joint constraints of physics and data.

[0081] Step 7: Based on the loss function, update the weight parameters in the network according to the backpropagation algorithm and the chain rule.

[0082] This step is mainly implemented based on the deep learning automatic differentiation framework in PyTorch.

[0083] Step 8: Train the network parameters and set the maximum number of training epochs M. Within the training epochs, repeat steps 5 to 7 until the network meets the error requirements or reaches the maximum number of training epochs, then exit the loop.

[0084] Step 9: Test the network estimation performance. Based on the state data of the test set and the forward propagation process in Step 4, obtain the tire force estimates for the test set. At this point, the tire force estimates are still under the standardized scale. After destandardization, obtain the tire force estimates under the original scale, thus realizing tire force estimation based on physical information fusion recurrent neural network.

[0085] To verify the superiority of this invention, a comparative test platform was built as follows:

[0086] Test environment: Windows 10 system; Carsim simulation environment, NVIDIA GTX 1080Ti GPU; Python 3.7; PyTorch framework.

[0087] Test dataset: Vehicle driving data is collected under different driving conditions, such as straight driving, curved driving, and two-lane lane changing.

[0088] Test metric: This invention uses MSE as the performance evaluation metric. In this invention, MSE refers to the mean square error between the estimated tire force and the data label.

[0089] Network parameter settings: For both PGRU and GRU, the input projection layer was set to a fully connected layer with 40 units, and the Rectified Linear Unit (ReLU) was selected as the activation function. The GRU had 20 hidden layers. Unlike PGRU, the GRU control group only considered a loss function based on data labels. A three-layer fully connected neural network was selected for the BP neural network, with 100 neurons in the hidden layer; the vehicle mass was 1500 kg.

[0090] Testing Methods: To verify the effectiveness of the algorithm of this invention, the following two sets of comparative experiments were designed:

[0091] The design is compared with pure data machine learning methods. The proposed PGRU method is compared with GRU networks and BP neural networks without physical constraints. The experimental results are as follows: Figure 5 As shown in the figure. The yellow dashed line represents the estimation result based on the BP neural network, the red dashed line represents the estimation result of the PGRU method proposed in this invention, and the green dashed line represents the estimation result of the GRU recurrent neural network.

[0092] The experiment compared the estimation performance under different tire-road friction coefficients. To further verify the generalization ability of PGRU in estimating tire forces under different road conditions, the proportion of the training set with tire-road friction coefficients of 0.75 and 0.85 was reduced to 50%, and the estimation performance under a tire-road friction coefficient of 0.9 was tested. The division of the training and test sets is shown in Table 2, and the test results are as follows: Figures 6 to 8 As shown in the figure. The purple dashed line represents the tire force data predicted by PGRU, and the black solid line represents the reference tire force data.

[0093] In the four sub-graphs, FY_FL, FY_FR, FY_RL, and FY_RR represent the lateral tire forces of the left front wheel, right front wheel, left rear wheel, and right rear wheel, respectively.

[0094] The test results are as follows:

[0095] Table 1. Comparison of the mean squared error (MSE) of the proposed algorithm with GRU and BP networks on the test set.

[0096]

[0097]

[0098] Table 2. Comparison of estimation results under different tire-road friction coefficients. Training and test set divisions in the experiment.

[0099]

[0100] Table 3. Comparison of Mean Square Error (MSE) of Estimation Results under Different Tire-Road Friction Coefficients

[0101] MSE <![CDATA[F yfl ]]> <![CDATA[F yfr ]]> <![CDATA[F yrl ]]> <![CDATA[F yrr ]]> Mu = 0.75 0.0113 0.0040 0.0102 0.0026 Mu = 0.85 0.0027 0.0016 0.0041 0.0016 Mu = 0.9 0.0029 0.0016 0.0028 0.0030

[0102] As shown in Table 1, PGRU's estimation errors for the right front and right rear tires are 0.0019 and 0.0018, respectively, significantly outperforming the other two networks. Under the same training epochs, the BP neural network's estimation errors for the left front and right rear tires reach 0.01 and 0.0101, respectively, significantly higher than the GRU network. Furthermore, the overall estimation error of the BP neural network is also greater than the other two GRU-based deep neural networks, suggesting that considering the long-term dependencies of the sequence yields better results in modeling lateral tire forces. Table 3 shows that even with a reduced training set ratio, PGRU can still estimate tire forces well when the tire-road friction coefficient is 0.85. Outside the training domain, with a friction coefficient of 0.9, PGRU's estimation of lateral tire forces also exhibits small errors. This demonstrates that PGRU can capture tire force characteristics well and possesses a certain degree of generalization ability.

[0103] It should be emphasized that the examples described in this invention are illustrative, intended to enable those skilled in the art to understand the content of this invention and implement it accordingly. This invention includes, but is not limited to, the examples described in the specific embodiments. All equivalent changes or modifications made in accordance with the spirit and essence of this invention should be covered within the protection scope of this invention.

Claims

1. A tire force estimation method based on physical information fusion recurrent neural network, characterized in that, Includes the following steps: Step 1: Collect vehicle state data and tire force data under different driving conditions based on the Carsim simulation platform; Step 2: From the vehicle state data in Step 1, select the following vehicle states related to tire force as network input parameters: in, It is the front wheel steering angle. It's the yaw rate. These are longitudinal velocity and lateral velocity, It is longitudinal acceleration and lateral acceleration. These are the rotational speeds of the left front wheel, right front wheel, left rear wheel, and right rear wheel, respectively. Step 3: Perform mean-variance standardization on the vehicle status data from Step 2, and divide it into training and test sets; Step 4, design the network structure: select gated recurrent units as the main recurrent neural network structure. In order to enhance the estimation effect of the model, add a fully connected layer before the input layer as the input projection layer and add a fully connected layer after the output layer as the output regression layer. Step 5: Perform step 4 on the vehicle state data sequence to obtain the tire force in the [number]th [phase]. The estimated value at each time point T represents the prediction of the current tire force based on the state at T past moments. It refers to the amount of training data; Step 6: Obtain the tire force estimates for the training set after step 5. Then, a loss function for physical information fusion is designed to update the network parameters. The design of the loss function for physical information fusion consists of two parts: a loss function based on data labels is designed based on the recurrent neural network algorithm; a loss function based on physical information is designed based on the simplified vehicle dynamics model; the two loss functions are superimposed to realize a tire force estimation method based on physical information fusion recurrent neural network. Step 7, based on the loss function The weight parameters in the network are updated according to the backpropagation algorithm and the chain rule. Step 8: Train network parameters and set the maximum number of training epochs. During each training round, repeat steps 5 to 7 until the network meets the error requirement or reaches the maximum number of training rounds, then exit the loop. Step 9, test the network estimation effect; obtain the tire force estimate of the test set based on the state data of the test set and the forward propagation process in step 4. At this time, the tire force estimate is still under the standardized scale. After destandardization, the tire force estimate under the original scale is obtained, realizing the tire force estimation based on physical information fusion recurrent neural network. A data label-based loss function is constructed based on tire force estimation data and reference tire force data: in, These are the tire force data labels from the training set. In weight The tire force estimate obtained from the network below; Based on a simplified vehicle dynamics model, tire forces and vehicle state are constructed. constraint functions between And based on this constraint function, a loss function based on physical information constraints is designed: 。 2. The tire force estimation method based on a physical information fusion recurrent neural network according to claim 1, characterized in that, The loss function of the network in step 7 is: in, To construct a data label-based loss function based on tire force estimation data and reference tire force data, The loss function is based on physical information constraints; It is a weighted scalar. It is a regularization function used to avoid overfitting.

3. The tire force estimation method based on a physical information fusion recurrent neural network according to claim 2, characterized in that, Set them to 0.5, 0.5 and 0.001 respectively.

Images