Electronic hydraulic brake system modeling method based on machine learning

Abstract

The invention is suitable for the technical field of electronic hydraulic braking system modeling, and provides an electronic hydraulic braking system modeling method based on machine learning, and the method comprises the steps: designing a braking working condition covering a plurality of road surfaces, vehicle speeds, braking strengths and steering; based on the braking working condition, input and output data are collected through joint simulation of the whole vehicle dynamics model and the braking control strategy black box model; dividing the data set into a training set and a verification set; constructing a feedforward neural network consistent with the input and output dimensions of the black box model; training the network by using a training set, and establishing nonlinear mapping between input and output; a training effect is checked through a loss function convergence condition; and finally, verifying the consistency of the output of the feedforward neural network model and the output of the black box model by using the verification set to obtain a feedforward neural network control strategy model. According to the method, the complexity of traditional modeling is avoided, and the brake control strategy can be efficiently and accurately reproduced.

Description

Technical Field

[0001] This invention belongs to the field of electro-hydraulic braking system modeling technology, and particularly relates to a machine learning-based electro-hydraulic braking system modeling method. Background Technology

[0002] Electro-hydraulic braking system control strategies can achieve dynamic distribution and precise adjustment of braking force under different braking conditions, thereby ensuring vehicle safety and stability. However, the braking control strategy algorithm itself is quite complex, and the modeling methods are not uniform.

[0003] Currently, mainstream brake controllers such as Bosch IPB and Continental MK C2 generally do not disclose their internal control strategies, making it difficult to effectively reproduce them using traditional mechanism-based modeling methods. Therefore, it is crucial to research a machine learning-based brake control strategy modeling method to achieve control performance equivalent to that of black-box models.

[0004] For modeling control strategies, existing technologies mostly adopt mechanism-based modeling methods. However, the modeling process of such methods is relatively cumbersome. Moreover, with the increasing complexity of vehicle control strategies and the introduction of more active safety functions, it is difficult to accurately reproduce the internal logic and control dynamics of the original vehicle strategy, resulting in certain differences between its control effect and the original braking control strategy model. Summary of the Invention

[0005] The purpose of this invention is to provide a machine learning-based modeling method for electro-hydraulic braking systems, aiming to solve the problems mentioned in the background art.

[0006] The present invention is implemented as follows: a machine learning-based modeling method for an electro-hydraulic braking system includes the following steps:

[0007] Step 1: Design braking conditions that cover various road surface conditions, initial vehicle speed, braking intensity, and steering wheel angle;

[0008] Step 2: Based on the braking conditions designed in Step 1, the vehicle dynamics model and the braking control strategy black box model are integrated in the simulation environment for joint simulation. The input and output data of the black box model under different conditions are collected to form a simulation dataset. The input data includes vehicle speed, four-wheel wheel speed, braking intensity, road adhesion coefficient, yaw rate, steering wheel angle and center of gravity sideslip angle. The output data includes four-wheel cylinder pressure.

[0009] Step 3: Divide the simulation dataset into a training set and a validation set;

[0010] Step 4: Construct a feedforward neural network model, with its input and output settings consistent with the black-box model in Step 2;

[0011] Step 5: Use the training set obtained in Step 3 to train the feedforward neural network model to establish a nonlinear mapping relationship between the input and output of the braking control strategy;

[0012] Step 6: Based on the convergence of the loss function during the training process, verify the training effect of the feedforward neural network model;

[0013] Step 7: Based on the validation set obtained in Step 3, verify the consistency between the output of the trained feedforward neural network model and the output of the black box model described in Step 2. If the consistency meets the preset requirements, a feedforward neural network control strategy model to replace the black box model is obtained.

[0014] A further technical solution involves setting the road surface adhesion coefficient in step 1 to 0.2, 0.3, 0.5, and 0.7, corresponding to four typical road conditions: icy and snowy road surface, wet and slippery road surface, damp cement road surface, and dry cement road surface, respectively. Additionally, a split road surface condition with an adhesion coefficient of 0.7 on the left and 0.3 on the right is simultaneously set, along with connecting road surface conditions ranging from 0.7 to 0.3 and from 0.3 to 0.7.

[0015] Regarding the initial vehicle speed settings, 30, 60, 90 and 120 km / h were selected respectively, covering low speed, medium speed, medium-high speed and high speed conditions;

[0016] Regarding the braking intensity settings, 0.2, 0.4, 0.6, 0.8 and 1.0 were used as the final values ​​of the step signal to simulate the condition of suddenly pressing the brake pedal; at the same time, a ramp signal that linearly increases from 0 to 0.8 braking intensity within 1 second or 2 seconds was used to simulate the condition of gently pressing the brake pedal.

[0017] Regarding the steering wheel angle setting, 30° to 270° was selected as the final value of the step signal to simulate the condition of sudden steering wheel turning; at the same time, sinusoidal signals with frequencies of 0.2Hz, 0.5Hz, and 1.0Hz, corresponding to amplitudes of 80°, 50°, and 20°, were applied to simulate the steering response of the vehicle during dynamic driving.

[0018] In a further technical solution, in step 2, the simulation environment is a Matlab / Simulink environment.

[0019] A further technical solution is that, in step 3, the simulation dataset is divided into a training set of 70% and a validation set of the remaining 30%.

[0020] In a further technical solution, in step 4, the parameters of the feedforward neural network model are set as follows:

[0021] The feedforward neural network has four layers: an input layer, two hidden layers, and an output layer. The input layer contains 10 neurons, each of the two hidden layers contains 64 neurons, and the output layer contains 4 neurons. Layers are connected via fully connected layers. The ReLU activation function is chosen to introduce non-linearity. The ReLU activation function is also used between hidden layers and between the second hidden layer and the output layer. The expression for the ReLU activation function is:

[0022] (1);

[0023] In the formula, This represents the sum of signals input to the activation function. This represents the output value after activation.

[0024] A further technical solution involves updating the neural network parameters in step 5 using stochastic gradient descent, with a learning rate set to 0.001 and a batch size of 6. The expression for stochastic gradient descent is:

[0025] (2);

[0026] in, For the current parameter, For loss function, For the updated parameters, This is the gradient operator.

[0027] A further technical solution is that, in step 6, the loss function used during training is the mean squared error (MSE), expressed as:

[0028] (3);

[0029] In the formula, To output the true value, To output the estimated value, This represents the total number of data points in the training set.

[0030] The present invention provides a machine learning-based modeling method for an electro-hydraulic braking system, the advantages of which are as follows:

[0031] (1) Solved the problem of black box model reproduction: In view of the problem that it is difficult to accurately reproduce the dynamic response of the black box model of the special control strategy in the electro-hydraulic braking system based on the traditional physical mechanism modeling method, the present invention adopts the data-driven machine learning method, directly learns its internal complex mapping through input and output data, effectively bypasses the difficulty of reverse analysis of the internal logic of the black box, and realizes high-precision replacement modeling of the original control strategy.

[0032] (2) High model accuracy and strong generalization ability: By designing comprehensive and reasonable braking conditions (covering various road surface adhesion conditions, vehicle speed range, braking intensity change mode and steering operation), a high-quality training dataset that can fully stimulate the dynamic characteristics of the system was generated. The feedforward neural network model trained based on this data can accurately capture the nonlinear response law of the braking control strategy under various complex conditions, ensuring that the model exhibits output characteristics that are highly consistent with the original black box model on both the training set and the validation set.

[0033] (3) The modeling process is highly efficient and automated: This method transforms the modeling process into a standard data acquisition and neural network training process. By utilizing mature simulation tools and machine learning frameworks, it can achieve automated or semi-automated execution, which significantly reduces the workload of manual analysis and derivation, improves modeling efficiency, and provides a convenient tool for the rapid verification, optimization and porting of braking control strategies.

[0034] (4) It has good scalability and applicability: This method is universal and is not only applicable to the modeling of specific braking control strategies. Its technical path based on working condition design, co-simulation, data acquisition, neural network training and verification can be extended to the alternative modeling of other difficult-to-analyze black box models of vehicle electronic control systems, and has a wide range of application prospects. Attached Figure Description

[0035] Figure 1 A flowchart illustrating a machine learning-based modeling method for an electro-hydraulic braking system, provided as an embodiment of the present invention;

[0036] Figure 2 This is a schematic diagram of the neural network vehicle speed training set.

[0037] Figure 3 A schematic diagram of the neural network braking intensity training set;

[0038] Figure 4 A schematic diagram of the training set for the neural network wheel cylinder pressure;

[0039] Figure 5 This is a schematic diagram of the training loss of a neural network.

[0040] Figure 6 A simulation comparison chart showing the actual and predicted values ​​of the verification set. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0042] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0043] like Figure 1 As shown, an embodiment of the present invention provides a machine learning-based modeling method for an electro-hydraulic braking system, comprising the following steps:

[0044] Step 1: Design reasonable and comprehensive braking conditions to generate a training dataset for learning the braking control strategy model. The braking conditions are shown in Table 1 below.

[0045] Table 1 Braking Conditions

[0046]

[0047] Regarding the setting of road surface adhesion coefficients, 0.2, 0.3, 0.5, and 0.7 were selected, corresponding to four common typical road conditions: icy and snowy road surface, wet and slippery road surface, damp cement road surface, and dry cement road surface, respectively. In addition, a split road surface condition with an adhesion coefficient of 0.7 on the left and 0.3 on the right was also set, as well as a connecting road surface condition from high adhesion (0.7) to low adhesion (0.3) and from low adhesion (0.3) to high adhesion (0.7).

[0048] Regarding the initial vehicle speed settings, 30, 60, 90, and 120 km / h were selected respectively, covering low-speed, medium-speed, medium-high-speed, and high-speed operating conditions.

[0049] Regarding the braking intensity settings, 0.2, 0.4, 0.6, 0.8 and 1.0 were used as the final values ​​of the step signal to simulate the condition of suddenly pressing the brake pedal; at the same time, a ramp signal that linearly increases from 0 to 0.8 braking intensity within 1 second or 2 seconds was used to simulate the condition of gently pressing the brake pedal.

[0050] Regarding the steering wheel angle setting, 30° to 270° was selected as the final value of the step signal to simulate the condition of sudden steering wheel turning; at the same time, sinusoidal signals with frequencies of 0.2Hz, 0.5Hz, and 1.0Hz, corresponding to amplitudes of 80°, 50°, and 20°, were applied to simulate the steering response of the vehicle during dynamic driving.

[0051] Step 2: Based on the braking conditions constructed in Step 1, obtain the input and output simulation dataset of the braking control strategy black box model. By traversing all preset braking conditions, the vehicle dynamics model and the braking control strategy black box model are integrated in the Matlab / Simulink environment for joint simulation. The input and output data of the black box model under different conditions are collected to form the simulation dataset. The input data includes vehicle speed, four-wheel wheel speed, braking intensity, road adhesion coefficient, yaw rate, steering wheel angle, and center of gravity sideslip angle. The output data includes four-wheel cylinder pressure.

[0052] Step 3: Divide the simulation dataset into a 70% training set and a 30% validation set. Taking an initial speed of 60 km / h, braking intensity of 0.6, and road adhesion coefficient of 0.4 as an example, the training set would be as follows: Figures 2 to 4 As shown.

[0053] Step 4: Employ a feedforward neural network as the algorithm for modeling the machine learning control strategy. The input and output settings of this network are consistent with the black-box model described in Step 2. Set the feedforward neural network model parameters: The feedforward neural network has four layers: an input layer, two hidden layers, and an output layer. The input layer contains 10 neurons, each of the two hidden layers contains 64 neurons, and the output layer contains 4 neurons. Layers are connected via fully connected layers. The ReLU activation function is selected to introduce non-linearity. The ReLU activation function is used between hidden layers and between the second hidden layer and the output layer. The ReLU activation function expression is:

[0054] (1);

[0055] In the formula, This represents the sum of signals input to the activation function. This represents the output value after activation.

[0056] Step 5: Based on the training set obtained in Step 3, train the feedforward neural network model to establish a complex nonlinear mapping relationship between the input and output of the braking control strategy, thus realizing the construction of the braking control strategy model based on the feedforward neural network. Set the training parameters for the feedforward neural network: learning rate. The value is 0.001, and the batch size is 6. Stochastic gradient descent is used to update the neural network parameters. The expression for stochastic gradient descent is:

[0057] (2);

[0058] in, For the current parameter, For loss function, For the updated parameters, This is the gradient operator.

[0059] Step 6: Validate the training effect of the braking control strategy model built based on the feedforward neural network. As the training cycle increases, such as... Figure 5 As shown, if the training loss gradually converges, it proves that the model's learning performance gradually improves. The loss function uses the mean squared error (MSE), and its expression is:

[0060] (3);

[0061] In the formula, To output the true value, To output the estimated value, This represents the total number of data points in the training set.

[0062] Step 7: Based on the validation set obtained in Step 3, and under the condition that the input data is the same, verify the consistency between the braking control strategy model built based on the feedforward neural network and the output data of the original black-box model. For example... Figure 6 To verify the simulation comparison between the actual and predicted values ​​under a certain working condition, the model outputs the wheel cylinder pressure. The actual and predicted values ​​show a high degree of consistency. This step verifies the effectiveness of the method.

[0063] The above method effectively solves the problem that mechanism-based modeling methods struggle to accurately reproduce the control effects of black-box braking control strategies. By employing neural network technology from machine learning, the complex traditional modeling process can be bypassed, and a neural network control strategy model that can directly replace the original black-box model can be constructed.

[0064] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A machine learning based electronic hydraulic brake system modeling method, characterized by, The method comprises the following steps: Step 1: design braking conditions covering various road conditions, initial vehicle speeds, braking intensities and steering wheel angles; Step 2: based on the braking conditions designed in step 1, integrate a vehicle dynamics model and a black-box braking control strategy model in a simulation environment to perform joint simulation, collect input and output data of the black-box model under different conditions to form a simulation data set, wherein the input data comprises vehicle speed, four-wheel speed, braking intensity, road adhesion coefficient, yaw rate, steering wheel angle and center of mass side slip angle, and the output data comprises four-wheel cylinder pressure; Step 3: divide the simulation data set into a training set and a verification set; Step 4: construct a feedforward neural network model, and set the input and output of the feedforward neural network model to be consistent with the black-box model in step 2; Step 5: use the training set obtained in step 3 to train the feedforward neural network model to establish a nonlinear mapping relationship between the input and output of the braking control strategy; Step 6: based on the loss function convergence during the training process, verify the training effect of the feedforward neural network model; Step 7: based on the verification set obtained in step 3, verify the consistency between the output of the trained feedforward neural network model and the output of the black-box model in step 2, and if the consistency meets the preset requirement, a feedforward neural network control strategy model is obtained to replace the black-box model.

2. The machine learning based electronic hydraulic brake system modeling method of claim 1, wherein, In step 1, in terms of road adhesion coefficient setting, 0.2, 0.3, 0.5 and 0.7 are selected to correspond to ice and snow road, wet road, wet cement road and dry cement road respectively; in addition, a split road condition with left adhesion coefficient of 0.7 and right adhesion coefficient of 0.3 is set synchronously, and a butt joint road condition from road adhesion coefficient 0.7 to road adhesion coefficient 0.3 and from road adhesion coefficient 0.3 to road adhesion coefficient 0.7 is set synchronously; In terms of initial vehicle speed setting, 30, 60, 90 and 120 km / h are selected to cover low speed, medium speed, medium-high speed and high speed conditions; In terms of braking intensity setting, 0.2, 0.4, 0.6, 0.8 and 1.0 are used as the final value of the step signal to simulate the condition of stepping on the brake pedal; at the same time, a ramp signal linearly increasing from 0 to 0.8 braking intensity within 1 second or 2 seconds is used to simulate the condition of slowly stepping on the brake pedal; In terms of steering wheel angle setting, 30° to 270° is selected as the final value of the step signal to simulate the condition of stepping on the steering wheel; at the same time, a sine signal with a frequency of 0.2 Hz, 0.5 Hz and 1.0 Hz and an amplitude of 80°, 50° and 20° is applied to simulate the steering response condition of the vehicle in dynamic driving.

3. The machine learning based electronic hydraulic brake system modeling method of claim 1, wherein, In step 2, the simulation environment is Matlab / Simulink environment.

4. The machine learning based electronic hydraulic brake system modeling method of claim 1, wherein, In step 3, the simulation data set is divided into a training set according to a proportion of 70%, and the remaining 30% is used as a verification set.

5. The machine learning based electronic hydraulic brake system modeling method of claim 2, wherein, In step 4, the parameters of the feedforward neural network model are set as follows: The feedforward neural network has four layers, which are an input layer, two hidden layers and an output layer. The input layer includes 10 neurons, each of the two hidden layers includes 64 neurons, and the output layer includes 4 neurons. The layers are connected through full connection. A ReLU activation function is selected to introduce a nonlinear factor. The ReLU activation function is used between the hidden layers and between the second hidden layer and the output layer. The expression of the ReLU activation function is: （1）； In the formula, represents the sum of signals input to the activation function, represents the output value after activation.

6. The machine learning based electronic hydraulic brake system modeling method of claim 1, wherein, In the step 5, the neural network parameters are updated by using a stochastic gradient descent method, the learning rate is set to 0.001, and the update batch size is set to 6. The expression of the stochastic gradient descent method is: （2）； wherein, is the current parameter, is the loss function, is the updated parameter, is the gradient operator.

7. The machine learning based electronic hydraulic brake system modeling method of claim 1, wherein, In the step 6, a mean square error (MSE) is used as the loss function in the training process, and the expression is: （3）； wherein is the output true value, is the output estimated value, is the total number of training set data.

Images