A vehicle motion state calculation method
By combining a three-layer neural network model with sensor data and an equivalent physical model, the problems of slow speed and low accuracy in calculating vehicle motion state are solved, achieving high-precision and fast vehicle motion state calculation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- TONGJI UNIV
- Filing Date
- 2023-03-22
- Publication Date
- 2026-07-21
AI Technical Summary
Existing methods for calculating vehicle motion states are slow in on-board conditions and have low accuracy in complex conditions. Furthermore, they are mostly single-threaded and rely on equivalence assumptions, resulting in insufficient adaptability.
A three-layer neural network model is adopted, which combines sensor measurement data. The first layer of the neural network calculates the wheel load, the second layer calculates the real-time vehicle speed, and the third layer combines the equivalent physical model to calculate the vehicle's kinematic parameters. Fully connected and dropout connection networks are used, and the Leaky ReLU activation function is used to construct the vehicle motion state calculation model.
It achieves high-precision vehicle motion state calculation under complex working conditions, simplifies on-board computing time and computing power consumption, and improves the practicality and calculation speed of the model.
Smart Images
Figure CN116373882B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of vehicle motion state calculation, and in particular to a method for calculating vehicle motion state. Background Technology
[0002] Determining a vehicle's motion state is a crucial aspect of its intelligent environmental perception. This perception includes both external shape perception (visual, lidar, ultrasonic radar, etc.) and dynamic perception (the aerodynamic and tire forces acting on the vehicle). External shape perception cannot serve as a direct input for vehicle control because it does not involve direct forces acting on the vehicle. Dynamic perception, however, involves external forces directly acting on the vehicle, directly guiding active vehicle control and significantly impacting safety, handling, and comfort. Compared to aerodynamic forces, tire forces have a greater influence on a vehicle's motion state, making the perception of tire forces increasingly important.
[0003] Existing methods for sensing vehicle dynamics include various approaches, such as directly using known vehicle dynamic models to calculate tire loads based on vehicle attitude and acceleration; research is being conducted on developing bearings that can measure loads for calculation; and other research involves analyzing the forces acting on the vehicle suspension to calculate tire loads using mechanical methods, etc.
[0004] Most of the aforementioned studies can achieve fast offline computation. However, in in-vehicle environments, limited computing power necessitates reasonable fitting and simplification of the algorithms to meet the basic requirements of high accuracy and speed. Furthermore, most of these methods operate on a single thread, and each model makes certain equivalence assumptions, which may limit their adaptability to complex operating conditions. Summary of the Invention
[0005] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a method for calculating vehicle motion state.
[0006] The objective of this invention can be achieved through the following technical solutions:
[0007] A method for calculating the motion state of a vehicle includes the following steps:
[0008] Use a sensor array to measure basic vehicle kinematics data;
[0009] The basic vehicle kinematics data are input into the vehicle motion state calculation model;
[0010] Output vehicle kinematic parameters;
[0011] The vehicle motion state calculation model includes a first layer neural network, a second layer neural network, and a third layer neural network.
[0012] The input to the first layer of the neural network is the basic vehicle kinematics data measured using a set of sensors, and the output is the calculated wheel load.
[0013] The input to the second layer of the neural network is the basic vehicle kinematics data measured by the sensor array and the wheel load calculated by the first layer of the neural network, and the output is the real-time vehicle speed.
[0014] The inputs to the third layer of the neural network are the basic vehicle kinematics data measured by the sensor group, the real-time vehicle speed, the wheel load calculated by the first layer of the neural network, and the wheel load reference value obtained through the equivalent physical model; the output is the vehicle kinematic parameters.
[0015] Furthermore, a fully connected method is used in the first layer of the neural network; and dropout is used in the second and third layers of the neural network.
[0016] Furthermore, the wheel load reference values obtained through the equivalent physical model are specifically calculated as follows: based on the known mass and acceleration information of the whole vehicle, the load transfer of the vehicle is calculated, thereby calculating the load reference values of each wheel.
[0017] Furthermore, the aforementioned basic vehicle kinematic data includes vehicle speed, vehicle acceleration, steering wheel angle, wheel speed, suspension compression, suspension force, brake fluid pressure, and throttle opening.
[0018] Furthermore, the sensor group includes a vehicle speed sensor, a vehicle acceleration sensor, a steering wheel angle sensor, a wheel speed sensor, a suspension compression line displacement sensor, a suspension force sensor, a brake oil pressure sensor, and a throttle opening sensor.
[0019] Furthermore, the vehicle kinematic parameters include angular acceleration, sideslip angle, and tire slip ratio.
[0020] Furthermore, the construction of the vehicle motion state calculation model includes the following steps:
[0021] S1. While the vehicle is in motion, use a sensor array to measure basic vehicle kinematic data;
[0022] S2. Calculate wheel loads based on the vehicle's basic kinematic data;
[0023] S3. Repeat steps S1 and S2 to construct the training and test sets for the first layer of the neural network;
[0024] S4. Construct the first layer of neural network. Based on the training set of the first layer of neural network, obtain the relationship between the basic kinematic data of the vehicle and the wheel load through nonlinear fitting of the neural network.
[0025] S5. Based on the vehicle's basic kinematic data and the wheel load, calculate the vehicle's real-time speed and construct the training and testing sets for the second-layer neural network.
[0026] S6. Construct a second-layer neural network. Based on the training set of the second-layer neural network, fit the basic kinematic data of the vehicle and the relationship between wheel load and real-time vehicle speed.
[0027] S7. Based on the vehicle's basic kinematic data, wheel load, and real-time vehicle speed, and combined with the wheel load reference value obtained through the equivalent physical model, calculate the vehicle's kinematic parameters and construct the training and testing sets for the third-layer neural network.
[0028] S8. Construct a third-layer neural network. Based on the training set of the third-layer neural network, fit the relationship between the basic vehicle kinematics data, wheel load, real-time vehicle speed, wheel load reference value and vehicle kinematic parameters to complete the model construction.
[0029] Furthermore, the construction of the vehicle motion state calculation model also includes the following steps:
[0030] The first layer of the neural network is tested using the test set of the first layer, the second layer is tested using the test set of the second layer, and the third layer is tested using the test set of the third layer. If all three layers meet the preset standards, the model is successfully built; otherwise, the model is retrained.
[0031] Further, in step S5, the method for calculating the real-time vehicle speed is as follows: the wheel rotation speed is calculated based on the data measured by the wheel speed sensor; during acceleration, the wheel rotation speed with the lowest rotation speed is taken as the equivalent real-time vehicle speed; during braking, the wheel rotation speed with the highest rotation speed is taken as the equivalent real-time vehicle speed.
[0032] Furthermore, the activation function used in the vehicle motion state calculation model is the Leaky ReLU function.
[0033] Compared with the prior art, the present invention has the following beneficial effects:
[0034] (1) This invention uses a sensor array to collect basic kinematic data of a car, takes the basic principles of car kinematics and dynamics as the guide for neural network training, and uses the neural network to fit special cases. It can still output high-precision calculation results for complex working conditions, thus achieving higher practicality of the model. It integrates the advantages of physical principles and data analysis.
[0035] (2) The present invention uses a fitted neural network to simplify complex vehicle calculations. The fitted network has a fast calculation speed, which greatly saves vehicle calculation time and computing power, and provides a wider range of potential uses for multi-layer processing of sensor data. Attached Figure Description
[0036] Figure 1 This is a schematic diagram of the model structure of the present invention;
[0037] Figure 2 This is a schematic diagram of the Dropout network structure used in the embodiments of the present invention;
[0038] Figure 3 This is a schematic diagram of the activation function (Leaky ReLU) used in the embodiments of the present invention;
[0039] Figure 4 This is a comparison chart of the model's predicted values and reference values in an embodiment of the present invention. Detailed Implementation
[0040] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.
[0041] To address the shortcomings of existing technologies in calculating vehicle kinematics states, such as slow on-board speed and low accuracy under complex conditions, this invention integrates multiple information sources to fully leverage the advantages of both dynamic and data-driven methods, thereby obtaining the most accurate vehicle operating parameters possible. The invention relates primarily to vehicle dynamics, neural networks and deep learning, and automated control. Specifically, it relates to a method for processing vehicle information and calculating vehicle motion states using a fusion of dynamics and neural networks.
[0042] This method can be applied not only to vehicle dynamics calculations but also to numerous other fields, demonstrating significant practical value. Analysis based on dynamic theory provides crucial reference and guidance for data-driven methods (neural networks), enhancing the reliability and accuracy of learning. Conversely, utilizing data-driven methods to learn and fit mechanistic theory analysis (here, dynamic methods) fully considers the limitations of mechanistic theory under different operating conditions and increases computational speed. The fusion of these two approaches can be used to study a variety of problems, solving issues that are difficult to address using a single method.
[0043] Specifically, this invention proposes a vehicle motion state calculation model, with a core algorithm being a neural network algorithm. The fitted network, obtained after extensive training on existing vehicle data, can quickly process new input data, thus eliminating the need for complex equation solving. For example... Figure 1 As shown, the vehicle motion state calculation model proposed in this invention includes a first layer neural network, a second layer neural network, and a third layer neural network;
[0044] The input to the first layer of the neural network is the basic vehicle kinematics data measured using a set of sensors, and the output is the calculated wheel load.
[0045] In one preferred embodiment, the basic vehicle kinematic data includes vehicle speed, vehicle acceleration, steering wheel angle, wheel speed, suspension compression, suspension force, brake fluid pressure, and throttle opening.
[0046] The input to the second layer of the neural network is the basic vehicle kinematics data measured by the sensor array and the wheel load calculated by the first layer of the neural network. The output is the real-time vehicle speed.
[0047] The inputs to the third layer of the neural network are the basic vehicle kinematics data measured by the sensor array, the real-time vehicle speed, the wheel load calculated by the first layer of the neural network, and the wheel load reference value obtained through the equivalent physical model; the outputs are some vehicle kinematic parameters that cannot be obtained directly and accurately, including angular acceleration, center of gravity sideslip angle, and tire slip ratio.
[0048] Specifically, the wheel load reference values obtained through the equivalent physical model are calculated by determining the vehicle's load transfer based on the vehicle's known mass and acceleration information, thereby calculating the load reference values for each wheel. Although this method has lower computational accuracy, the trend of its data closely matches the vehicle's kinematic characteristics. Therefore, it is used as the independent variable of the third-layer network to guide the algorithm to perform calculations along the correct trend.
[0049] The built-in function structure of the model is explained in detail below:
[0050] In the first layer of the neural network, a fully connected approach is used. Since the purpose of using a neural network in this problem is to fit a complex physical model, a fully connected approach helps to fully account for the influence of all dependent variables related to the model results. Because it is a fitting operation, the loss function is directly defined as the mean square error between the fitted result and the reference result. The output data at this point is the tire load. Since this output data has a fixed calculation formula, there is no need to worry about overfitting; on the contrary, a sufficient fit is the property required by this layer of the network.
[0051] In the second and third layers of the neural network, dropout is used to connect the networks. Since the output data of the second and third layers do not have sufficiently precise fixed formulas for calculation, extensive training is required to prevent overfitting. Dropout networks effectively address this need. Specifically, each neuron is retained with a probability of P and dropped with a probability of 1-P, thereby accelerating training and increasing randomness. The structure of a Dropout network is as follows: Figure 2 As shown.
[0052] Secondly, since the output of this dataset theoretically has no upper or lower bounds, it requires extensive training, and the choice of activation function must also conform to the characteristics of the original physical model. Due to the unbounded and non-linear nature of this dataset, the Leaky ReLU activation function is used in this embodiment, and its graph is shown below. Figure 3 As shown.
[0053] Because this algorithm operates in the harsh environment of an in-vehicle environment, the network fitting results need to realistically reproduce the signal vibrations during vehicle operation, such as the vibrations of tire loads when the vehicle travels over bumpy roads, rather than automatically smoothing them out; otherwise, important microscopic data information will be lost. Verification has shown that this network architecture has good dynamic performance, such as... Figure 4 The image shows a segment of data used to demonstrate the results. It is evident that the predicted values fit the reference values well, and the predicted values accurately reflect the fluctuations of the reference values.
[0054] In practical implementation, the onboard hardware requires a combination of sensors and controllers capable of measuring basic vehicle kinematic data. The sensor set includes vehicle speed sensors, vehicle acceleration sensors, steering wheel angle sensors, wheel speed sensors, suspension compression line displacement sensors, suspension force sensors, brake fluid pressure sensors, and throttle opening sensors. The controller includes a circuit board capable of running algorithms (or directly integrated into the vehicle's electronic control unit (ECU), a processor capable of amplifying and reducing noise from the sensor signals, and related wiring harnesses.
[0055] The algorithm for calculating wheel load using data collected by the aforementioned sensors has been described in detail in Chinese patent application CN2022103883415, and is presented here as background information.
[0056] To address the aforementioned model, this invention proposes a method for constructing a vehicle motion state calculation model, comprising the following steps:
[0057] S1. While the vehicle is in motion, use a sensor array to measure basic vehicle kinematic data;
[0058] S2. Calculate wheel loads based on basic vehicle kinematics data;
[0059] S3. Repeat steps S1 and S2 to construct the training and test sets for the first layer of the neural network;
[0060] S4. Construct the first layer of neural network. Based on the training set of the first layer of neural network, obtain the relationship between the basic kinematic data of the vehicle and the wheel load through the nonlinear fitting of the neural network.
[0061] S5. Based on the vehicle's basic kinematic data and wheel load, calculate the vehicle's real-time speed and construct the training and testing sets for the second-layer neural network.
[0062] S6. Construct a second-layer neural network. Based on the training set of the second-layer neural network, fit the basic kinematic data of the vehicle and the relationship between wheel load and real-time vehicle speed.
[0063] S7. Based on the basic vehicle kinematics data, wheel load and real-time vehicle speed, combined with the wheel load reference value obtained through the equivalent physical model, calculate the vehicle kinematic parameters and construct the training set and test set of the third layer neural network.
[0064] S8. Construct a third-layer neural network. Based on the training set of the third-layer neural network, fit the relationship between the basic vehicle kinematics data, wheel load, real-time vehicle speed, wheel load reference value and vehicle kinematic parameters to complete the model construction.
[0065] In step S5, the real-time vehicle speed is calculated as follows: based on data measured by wheel speed sensors, the wheel rotation speed is calculated. During acceleration, the wheel rotation speed with the lowest rotation speed is taken as the equivalent real-time vehicle speed; during braking, the wheel rotation speed with the highest rotation speed is taken as the equivalent real-time vehicle speed. Since the above calculation method involves equivalence, a neural network model is needed to establish the relationship between vehicle speed and physical parameters such as throttle opening and brake fluid pressure, correcting the results of the above physical calculation method to obtain a more accurate vehicle speed.
[0066] In step S8, the physical calculation method for vehicle kinematic parameters is as follows: the data obtained is estimated by on-board sensors or calculated by vehicle dynamics, and the data is used as the estimation result.
[0067] The model construction method also includes the following steps:
[0068] The first layer of the neural network is tested using the test set. If the test meets the preset standard, the training of the first layer is complete. This trained network will be used as a means for subsequent fast computation.
[0069] The second-layer neural network is tested using a test set. If the test meets the preset standard, the training of this layer is complete. After training, a fast fitting network that can reflect the actual vehicle speed can be obtained, providing a basis for the specific calculation of the third layer.
[0070] The third-layer neural network is tested using a test set. If the test results meet the preset criteria, the model is successfully built; otherwise, the model is retrained.
[0071] The third-layer neural network, combined with the network fitted by the first two layers, can serve as fast code for in-vehicle computing. By using data-driven methods, it avoids the complex calculations of physical principles while ensuring a certain level of accuracy, thus integrating the advantages of many methods.
[0072] It should be noted that this invention, as an analytical framework for vehicle dynamics, is not limited to the case of measuring wheel loads. Because a car is a complex object, its kinematic and dynamic models contain many assumptions and simplifications. Often, calculations based on physical models will result in some results that do not meet specific conditions, thus amplifying the shortcomings of the equivalent model; while results obtained through pure data analysis methods lack physical principles and cannot be reasonably interpreted. Integrating physical models and data analysis models can complement the advantages of both, thus proving highly useful in the field of automotive kinematics and dynamics.
[0073] The main innovation of this method compared to other single-threaded or multi-layered methods is:
[0074] (1) This invention uses a sensor array to collect basic kinematic data of a car, takes the basic principles of car kinematics and dynamics as the guide for neural network training, and uses the neural network to fit special cases. It can still output high-precision calculation results for complex working conditions, thus achieving higher practicality of the model. It integrates the advantages of physical principles and data analysis.
[0075] (2) The present invention uses a fitted neural network to simplify complex vehicle calculations. The fitted network has a fast calculation speed, which greatly saves vehicle calculation time and computing power, and provides a wider range of potential uses for multi-layer processing of sensor data.
[0076] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.
Claims
1. A method for calculating the motion state of a vehicle, characterized in that, Includes the following steps: Use a sensor array to measure basic vehicle kinematics data; The basic vehicle kinematics data are input into the vehicle motion state calculation model; Output vehicle kinematic parameters; The vehicle motion state calculation model includes a first layer neural network, a second layer neural network, and a third layer neural network. The input to the first layer of the neural network is the basic vehicle kinematics data measured using a set of sensors, and the output is the calculated wheel load. The input to the second layer of the neural network is the basic vehicle kinematics data measured by the sensor array and the wheel load calculated by the first layer of the neural network, and the output is the real-time vehicle speed. The inputs to the third layer of the neural network are the basic vehicle kinematics data measured by the sensor group, the real-time vehicle speed, the wheel load calculated by the first layer of the neural network, and the wheel load reference value obtained through the equivalent physical model; the output is the vehicle kinematic parameters.
2. The method for calculating the motion state of a vehicle according to claim 1, characterized in that, The first layer of the neural network uses a fully connected method; the second and third layers use dropout connections.
3. The method for calculating the motion state of a vehicle according to claim 1, characterized in that, The wheel load reference values obtained through the equivalent physical model are specifically calculated as follows: based on the known mass and acceleration information of the whole vehicle, the load transfer of the vehicle is calculated, thereby calculating the load reference values of each wheel.
4. The method for calculating the motion state of a vehicle according to claim 1, characterized in that, The basic vehicle kinematics data include vehicle speed, vehicle acceleration, steering wheel angle, wheel speed, suspension compression, suspension force, brake fluid pressure, and throttle opening.
5. The method for calculating the motion state of a vehicle according to claim 1, characterized in that, The sensor group includes a vehicle speed sensor, a vehicle acceleration sensor, a steering wheel angle sensor, a wheel speed sensor, a suspension compression line displacement sensor, a suspension force sensor, a brake oil pressure sensor, and a throttle opening sensor.
6. The method for calculating the motion state of a vehicle according to claim 5, characterized in that, The vehicle kinematic parameters include angular acceleration, sideslip angle, and tire slip ratio.
7. The method for calculating the motion state of a vehicle according to claim 6, characterized in that, The construction of the vehicle motion state calculation model includes the following steps: S1. While the vehicle is in motion, use a sensor array to measure basic vehicle kinematic data; S2. Calculate wheel loads based on the vehicle's basic kinematic data; S3. Repeat steps S1 and S2 to construct the training and test sets for the first layer of the neural network; S4. Construct the first layer of neural network. Based on the training set of the first layer of neural network, obtain the relationship between the basic kinematic data of the vehicle and the wheel load through nonlinear fitting of the neural network. S5. Based on the vehicle's basic kinematic data and the wheel load, calculate the vehicle's real-time speed and construct the training and testing sets for the second-layer neural network. S6. Construct a second-layer neural network. Based on the training set of the second-layer neural network, fit the basic kinematic data of the vehicle and the relationship between wheel load and real-time vehicle speed. S7. Based on the vehicle's basic kinematic data, wheel load, and real-time vehicle speed, and combined with the wheel load reference value obtained through the equivalent physical model, calculate the vehicle's kinematic parameters and construct the training and testing sets for the third-layer neural network. S8. Construct a third-layer neural network. Based on the training set of the third-layer neural network, fit the relationship between the basic vehicle kinematics data, wheel load, real-time vehicle speed, wheel load reference value and vehicle kinematic parameters to complete the model construction.
8. The method for calculating the motion state of a vehicle according to claim 7, characterized in that, The construction of the vehicle motion state calculation model also includes the following steps: The first layer of the neural network is tested using the test set of the first layer, the second layer is tested using the test set of the second layer, and the third layer is tested using the test set of the third layer. If all three layers meet the preset standards, the model is successfully built; otherwise, the model is retrained.
9. The method for calculating the motion state of a vehicle according to claim 8, characterized in that, In step S5, the method for calculating the real-time vehicle speed is as follows: the wheel rotation speed is calculated based on the data measured by the wheel speed sensor. During acceleration, the wheel rotation speed with the lowest rotation speed is taken as the equivalent real-time vehicle speed, and during braking, the wheel rotation speed with the highest rotation speed is taken as the equivalent real-time vehicle speed.
10. The method for calculating the motion state of a vehicle according to claim 1, characterized in that, The activation function used in the vehicle motion state calculation model is the Leaky ReLU function.