Vehicle dynamic parameter real-time estimation method and system based on hybrid network

Abstract

The invention discloses a vehicle dynamic parameter real-time estimation method and system based on a hybrid network, and the method comprises the steps: building an implicit mapping function between a vehicle-mounted sensor signal parameter and a tire-road surface friction coefficient and a side slip angle based on a vehicle nonlinear kinetic equation; on the basis of an implicit mapping function, a hybrid network architecture is constructed, and the hybrid network architecture comprises a set of lightweight Inception V3 pavement classification networks and two sets of deep regression networks with multiple layers of long and short term memory units; constructing a working condition high-coverage data set of hybrid network offline training and an intermediate parameter space-time synchronization strategy; and executing off-line training of the hybrid network based on the working condition high-coverage training data set, and completing model deployment on a real vehicle. According to the method, future road surface type priori knowledge obtained based on machine vision is comprehensively considered, and the method has great significance in improving the estimation precision and real-time performance of the tire-road surface friction coefficient and the side slip angle.

Description

Technical Field

[0001] The invention belongs to the field of automobile control, and in particular relates to a method and system for real-time estimation of vehicle dynamic parameters based on a hybrid network. Background Art

[0002] The effective implementation of advanced driver assistance systems and even autonomous driving depends on the vehicle's accurate perception of the environment and real-time feedback of its own dynamic state, among which the key information is the tire-road friction coefficient and the center of mass sideslip angle. In some active safety control algorithms, the accurate acquisition of the values ​​of the two will really affect the stability and safety of the vehicle. Similarly, these two key parameters cannot be directly measured by the sensors currently configured in mass-produced vehicles unless expensive optical sensors or high-precision motion positioning global navigation satellite systems are used. Therefore, studying virtual sensor technology to estimate the two is the main research direction in the industry. However, due to the uncertainty of the external environment and the high nonlinearity of the tire, it is still challenging to accurately estimate the tire-road friction coefficient and the center of mass sideslip angle.

[0003] Obtaining an accurate tire-road friction coefficient is a prerequisite for achieving accurate estimation of the center of mass sideslip angle. The tire-road friction coefficient estimation methods that are often discussed in the literature can be divided into two categories based on cause and effect. Among them, the cause-based method generally requires the support of additional sensors (such as cameras, acoustic sensors, etc.) to obtain road characteristics associated with the tire-road friction coefficient, and combines prior knowledge to obtain potential tire-road friction coefficient estimates. The effect-based method uses the vehicle dynamic response collected by the sensors configured in mass-produced vehicles, and uses a model-based or data-based solution to observe or map to the continuous value of the tire-road friction coefficient. In contrast, the effect-based method has an inherent "lag", and the cause-based method can obtain predictive information on future road types.

[0004] After obtaining the tire-road friction coefficient, the estimation of the center of mass sideslip angle has a priori constraints. The widely discussed center of mass sideslip angle estimation methods can be divided into two categories: model-based and data-based. Among them, the effective implementation of the model-based solution depends on whether the vehicle has sufficient dynamic excitation. Unfortunately, maneuvers with sufficient excitation are rare, and the vehicle dynamics under normal working conditions do not meet this requirement. Data-based solutions, such as artificial neural networks, do not require prior knowledge of physical models and can directly learn the nonlinear mapping relationship between sensor signal input and center of mass sideslip angle output. The ability of data-based solutions to solve the problem of center of mass sideslip angle estimation is unquestionable. I believe that with the support of a sufficiently rich training data set, it will bring amazing performance. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a method and system for real-time estimation of vehicle dynamic parameters based on a hybrid network, which solves the problem in the prior art of the difficulty in accurately estimating the tire-road friction coefficient and the center of mass sideslip angle.

[0006] The present invention adopts the following technical solutions to solve the above technical problems:

[0007] The method for real-time estimation of vehicle dynamic parameters based on hybrid network includes the following steps:

[0008] S1. Based on the seven-degree-of-freedom nonlinear dynamic equation, an implicit mapping function between the vehicle sensor signal parameters and the tire-road friction coefficient and the center of mass sideslip angle is established to clarify the input and output vectors of the hybrid network;

[0009] S2. Based on the implicit mapping function in S1, a hybrid network structure is constructed including a set of lightweight Inception V3 road classification networks and two sets of deep regression networks with multi-layer long short-term memory units;

[0010] S3, based on the hybrid network input and output vectors in S1 and the hybrid network structure in S2, construct a high-coverage dataset of working conditions and a time synchronization strategy for intermediate parameters for offline training of the hybrid network;

[0011] S4: Perform offline training of the hybrid network in S2 based on the high-coverage training data set of working conditions in S3, and complete model deployment and testing on the actual vehicle to achieve high-precision and real-time estimation of the tire-road friction coefficient and the sideslip angle of the center of mass.

[0012] Step S1 includes the following sub-steps:

[0013] S101. According to Newton's second law, a seven-degree-of-freedom vehicle dynamics model is established, including three vehicle planar motion degrees of freedom and four wheel rotation degrees of freedom;

[0014] S102, based on the need to accurately describe the tire force generation mechanism under the longitudinal and lateral coupling conditions, establish a fitting UniTire model to analyze the mechanical properties of the tire;

[0015] S103. Based on the nonlinear dynamic equations established in S101 and S102, the relationship between the vehicle sensor signal parameters, the intermediate signal parameters and the signal parameters to be estimated is obtained, an implicit mapping function between the parameters is constructed, and the input and output vectors of the hybrid network are clarified.

[0016] In step S103, the implicit mapping function between the parameters , the expressions of the input and output vectors of the hybrid network are:

[0017]

[0018] In the formula, , is the longitudinal and lateral acceleration of the vehicle, , is the longitudinal and lateral speed, is the yaw angular velocity, , and are the longitudinal, lateral and vertical forces of the tire, respectively. , are tire slip rate and sideslip angle, is the wheel rotation angular velocity, , is the wheel drive and brake torque, is the wheel angle, is the tire-road friction coefficient.

[0019] Step S2 includes the following sub-steps:

[0020] S201. Based on the characteristics of the road surface classification task and considering the real-time requirements of real vehicle deployment reasoning, a lightweight Inception V3 road surface classification network is established;

[0021] S202. Based on the tire-road friction coefficient range output of the lightweight Inception V3 road classification network and the implicit mapping function in step S1, the input and output vectors of the hybrid network, two sets of deep regression networks with multi-layer LSTM are established.

[0022] In step S202, in order to prevent the regression value of the tire-road friction coefficient from exceeding its prior value range, the mean square error loss function of the first group of LSTM deep regression networks is modified to impose a penalty on the tire-road friction coefficient exceeding the value range. The expression is:

[0023]

[0024] In the formula, and are the true and predicted values ​​of the tire-road friction coefficient, , The maximum and minimum values ​​of the tire-road friction coefficient output by the lightweight Inception V3. is the number of training samples.

[0025] Step S3 includes the following sub-steps:

[0026] S301. Design a three-dimensional active excitation curve according to the dynamic response law of the vehicle, automatically acquire a large amount of simulation data sets and real vehicle road test data sets in batches, and form a hybrid network offline training data set;

[0027] S302, based on the hybrid network offline training data set established in S301, formulate a hybrid training scheme of simulation data pre-training and real vehicle data fine-tuning;

[0028] S303. According to the time domain characteristics of the road surface classification network and the deep regression network, a spatiotemporal synchronization strategy of intermediate parameters is designed to align the output of the classification network and the input of the regression network.

[0029] In step S303, the intermediate parameter spatiotemporal synchronization strategy design process is:

[0030] The camera is installed on the top of the test vehicle, the sampling frequency is set to 10 Hz, the effective detection range is 5-50m in front of the vehicle, and the current time is defined as , the center of mass of the vehicle is the origin O, the camera has completed sampling, then the distance near the vehicle is , far end distance , based on the simplified longitudinal kinematics of the vehicle, its longitudinal displacement can be expressed as:

[0031]

[0032] Define the time when the vehicle reaches the near and far end distances as , , then the relationship between vehicle travel time and space is established through the following formula:

[0033]

[0034] In the formula, is the output feedback of the hybrid network. In the corresponding interval The maximum and minimum values ​​of the tire-road friction coefficient of the first group of LSTM deep regression networks are input as delays at the time interval Output value of road surface classification network under time label, here is the network inference time.

[0035] In step S4, based on the high-coverage training data set of the working conditions in S3 and combined with the spatiotemporal synchronization strategy of the intermediate parameters, the offline training of the hybrid network in S2 is performed, and the model deployment and testing are completed on the actual vehicle.

[0036] A vehicle dynamic parameter real-time estimation system based on a hybrid network includes a test vehicle, a camera arranged on the top of the test vehicle, and a control module arranged inside the test vehicle. The control module, based on the camera on the top of the test vehicle and the on-board sensor, applies the method to perform real-time estimation of the tire-road friction coefficient and the center of mass sideslip angle of the test vehicle.

[0037] A computer-readable storage medium stores computer-readable instructions, which, when executed by a processor, call all or part of the steps of the method.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] 1. The present invention first establishes an implicit mapping function between the vehicle-mounted sensor signal parameters and the tire-road friction coefficient and the center of mass sideslip angle, and clarifies the input and output vectors of the hybrid network. Subsequently, the present invention constructs a hybrid network architecture, which includes a set of lightweight Inception V3 road classification networks and two sets of deep regression networks with multi-layer LSTM. After that, the present invention designs and constructs a high-coverage data set of working conditions and an intermediate parameter time synchronization strategy for offline training of the hybrid network. Finally, the present invention performs offline training of the hybrid network based on the hybrid training scheme, and completes model deployment and testing on a real vehicle, achieving high-precision and real-time estimation of the tire-road friction coefficient and the center of mass sideslip angle.

[0040] 2. Taking into account the prior knowledge of future road types obtained based on machine vision, the estimation method for the tire-road friction coefficient and the sideslip angle of the center of mass is reasonable, feasible and effective, which is of great significance for improving the accuracy of estimation while ensuring real-time performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 The present invention is a block diagram of a method for real-time estimation of vehicle dynamic parameters based on a hybrid network.

[0042] Figure 2 This is a vehicle state parameter mapping relationship diagram in an embodiment of the present invention.

[0043] Figure 3 Schematic diagram of the intermediate parameter spatiotemporal synchronization strategy in an embodiment of the present invention.

[0044] Figure 4 It is the estimation result of the hybrid network real vehicle road test in the embodiment of the present invention. DETAILED DESCRIPTION

[0045] The structure and working process of the present invention will be further described below in conjunction with the accompanying drawings.

[0046] In view of some of the problems mentioned in the background technology, this scheme designs and develops a real-time estimation method for vehicle dynamic parameters based on a hybrid network. It combines the future road type and its corresponding tire-road friction coefficient prior range obtained based on machine vision, and makes full use of the nonlinear feature fitting ability of the deep regression network. It is of great significance to improve the estimation accuracy and real-time performance of the tire-road friction coefficient and the sideslip angle of the center of mass.

[0047] The method for real-time estimation of vehicle dynamic parameters based on hybrid network includes the following steps:

[0048] S1. Based on the seven-degree-of-freedom nonlinear dynamic equation, an implicit mapping function between the vehicle sensor signal parameters and the tire-road friction coefficient and the center of mass sideslip angle is established to clarify the input and output vectors of the hybrid network;

[0049] S2. Based on the implicit mapping function in S1, a hybrid network structure is constructed including a set of lightweight Inception V3 road classification networks and two sets of deep regression networks with multi-layer long short-term memory units;

[0050] S3, based on the hybrid network input and output vectors in S1 and the hybrid network structure in S2, construct a high-coverage dataset of working conditions and a time synchronization strategy for intermediate parameters for offline training of the hybrid network;

[0051] S4: Perform offline training of the hybrid network in S2 based on the high-coverage training data set of working conditions in S3, and complete model deployment and testing on the actual vehicle to achieve high-precision and real-time estimation of the tire-road friction coefficient and the sideslip angle of the center of mass.

[0052] Specific embodiments, such as Figures 1 to 4 As shown,

[0053] A method for real-time estimation of vehicle dynamic parameters based on a hybrid network comprises the following steps:

[0054] 1) Based on the seven-degree-of-freedom nonlinear dynamic equation, an implicit mapping function between the vehicle sensor signal parameters and the tire-road friction coefficient and the center of mass sideslip angle is established to clarify the input and output vectors of the hybrid network;

[0055] 11) According to Newton's second law, a seven-degree-of-freedom vehicle dynamics model is established, including three vehicle planar motion degrees of freedom (longitudinal, lateral and yaw motion) and four wheel rotation degrees of freedom;

[0056] 12) Based on the need to accurately describe the tire force generation mechanism under longitudinal and lateral coupling conditions, a UniTire model with high fitting accuracy is established to analyze the mechanical properties of the tire;

[0057] 13) Clarify the relationship between the on-board sensor signal parameters, intermediate signal parameters and signal parameters to be estimated, construct implicit mapping functions between parameters, and clarify the input and output vectors of the hybrid network.

[0058] 2) According to the implicit mapping function in step 1), a hybrid network architecture is constructed, which includes a set of lightweight Inception V3 road classification networks and two sets of deep regression networks with multi-layer LSTM;

[0059] 21) Based on the characteristics of the road surface classification task and considering the real-time requirements of real vehicle deployment reasoning, a lightweight Inception V3 road surface classification network was established;

[0060] 22) Based on the tire-road friction coefficient range output of the lightweight Inception V3 road classification network and the implicit mapping function in step 1), the input and output vectors of the hybrid network are established, two sets of deep regression networks with multi-layer LSTM.

[0061] 3) Based on the hybrid network input and output vectors in step 1) and the hybrid network structure in step 2), a high-coverage dataset of working conditions and a spatiotemporal synchronization strategy for intermediate parameters for offline training of the hybrid network are constructed;

[0062] 31) Design three-dimensional active excitation curves (vehicle speed, steering wheel angle, road adhesion coefficient) according to the dynamic response law of the vehicle, and automatically obtain massive simulation data sets and real vehicle road test data sets in batches;

[0063] 32) According to the time domain characteristics of the road surface classification network and the deep regression network, a spatiotemporal synchronization strategy of intermediate parameters is designed to align the output of the classification network and the input of the regression network.

[0064] 4) Based on the high-coverage training data set of working conditions in step 3), perform offline training of the hybrid network in step 2), and complete model deployment and testing on the actual vehicle;

[0065] 41) Based on the high-coverage training dataset of the working conditions in step 3), combined with the intermediate parameter spatiotemporal synchronization strategy, perform offline training of the hybrid network in step 2), and complete model deployment and testing on the actual vehicle.

[0066] In this embodiment, the specific steps of realizing high-precision real-time estimation of tire-road friction coefficient and center of mass sideslip angle according to the above method are:

[0067] Step 1: In order to more accurately characterize the nonlinear characteristics of the system, a seven-degree-of-freedom vehicle dynamics model was established. According to Newton's second law, the dynamic equation of the vehicle's planar motion can be expressed as:

[0068]

[0069] In the formula, the vehicle state parameters include , , , and The basic parameters of the vehicle include , , , and They are vehicle mass, distance from center of mass to front and rear axles, yaw moment of inertia and wheelbase.

[0070] Furthermore, in order to more accurately describe the vehicle's motion state, the rotational freedom of the four wheels is also considered, and its rotational dynamics equation can be expressed as:

[0071]

[0072] In the formula, is the wheel moment of inertia, is the effective radius of the wheel.

[0073] Furthermore, in order to accurately describe the generation mechanism of tire force under longitudinal and lateral coupling conditions, a UniTire model with high fitting accuracy is established to analyze the mechanical properties of the tire. In the general UniTire tire model, the longitudinal force and lateral force can be expressed as:

[0074]

[0075] In the formula,

[0076]

[0077] In the formula, , are the longitudinal and lateral tire-road friction coefficients, , are the relative longitudinal and lateral slip rates, is the relative total slip rate.

[0078] In the above formula, the tire longitudinal slip rate, sideslip angle and vertical force can be expressed as:

[0079]

[0080]

[0081]

[0082] In the formula, is the wheel center speed, is the acceleration due to gravity, is the vehicle wheelbase, Represents the centroid height.

[0083] Furthermore, based on the above-mentioned seven-degree-of-freedom vehicle dynamics model, wheel rotation dynamics model and tire dynamics model, an implicit mapping function between parameters is constructed: ,like Figure 2 As shown, the expressions of the input and output vectors of the hybrid network are:

[0084]

[0085]

[0086] Step 2: The main part of the hybrid network mainly includes a set of lightweight Inception V3 classification networks based on machine vision and two sets of deep regression networks with multi-layer LSTM, such as Figure 1 shown.

[0087] Specifically, the lightweight Inception V3 road classification network is first used to obtain the prediction information of the future tire-road friction coefficient value range. The output is the maximum and minimum values ​​of the tire-road friction coefficient under the road state. Then, after being processed by the spatiotemporal synchronization strategy, it is used as the augmented input to enter the first group of LSTM deep regression networks to output the continuous value of the tire-road friction coefficient. Subsequently, the continuous value of the tire-road friction coefficient is again used as the augmented input to enter the second group of LSTM deep regression networks to output state information such as the sideslip angle of the center of mass.

[0088] The parameter transfer of the hybrid network is cascaded, and the estimated values ​​of the tire-road friction coefficient and the center of mass sideslip angle can be easily obtained from the two groups of LSTM deep regression networks. Except for the augmented input, the other inputs of the two groups of LSTM deep regression networks are consistent, and are all signals that are easily obtained by the vehicle sensors. In addition, in order to prevent the regression value of the tire-road friction coefficient from exceeding its prior value range, the mean square error loss function of the first group of LSTM deep regression networks is modified to impose a penalty on it if it exceeds the value range, which can be expressed as:

[0089]

[0090] Step 3: Based on the hybrid network input and output vectors in step 1 and the hybrid network structure in step 2, a high-coverage dataset of working conditions and a spatiotemporal synchronization strategy for intermediate parameters for offline training of the hybrid network are constructed.

[0091] For the lightweight Inception V3 classification network based on machine vision in the hybrid network architecture, a suitable open source dataset is selected for training. Among them, the road conditions are divided into four common categories: ice and snow, gravel, wet asphalt and dry asphalt. The number of samples in each category and the corresponding tire-road friction coefficient range are shown in Table 1. For the training of the two groups of LSTM deep regression networks, the dynamic signals of the vehicle itself are required. The three dimensions of target vehicle speed, steering wheel angle and road friction coefficient are considered comprehensively. Through the orderly combination of the three input dimensions, the active excitation conditions that may occur in the vehicle are simulated to obtain an offline training dataset with high coverage of working conditions.

[0092] Table 1

[0093]

[0094] Furthermore, the road surface classification based on machine vision obtains the predicted information of the tire-road friction coefficient range of the road ahead, while the input of the LSTM deep regression network is based on the dynamic response of the vehicle at the current moment. Therefore, it is necessary to design a spatiotemporal synchronization strategy to align the output of the classification network and the input of the regression network, such as Figure 3 shown.

[0095] Specifically, the camera is installed on the top of the test vehicle, and its sampling frequency is set to 10 Hz, that is, the sampling time The effective detection range is 5-50m in front of the vehicle. Define the current time as , the center of mass of the vehicle is the origin O, the camera has completed sampling, then the distance near the vehicle is , far end distance Based on the simplified longitudinal kinematics of the vehicle, its longitudinal displacement can be expressed as:

[0096]

[0097] Define the time when the vehicle reaches the near and far end distances as , , then:

[0098]

[0099] The above formula establishes the connection between vehicle travel time and space. When the vehicle travels to In the corresponding interval The maximum and minimum values ​​of the tire-road friction coefficient of the first group of LSTM deep regression networks are input as delays at the time interval Output value of road surface classification network under time label, here is the network inference time.

[0100] Step 4: Based on the high-coverage training data set of working conditions in step 3, perform offline training of the hybrid network in step 2, and complete model deployment and testing on the actual vehicle.

[0101] The actual vehicle deployment and testing were carried out on an A0-class, fully wire-controlled four-wheel drive test vehicle. The test vehicle was equipped with an RT3000 inertial navigation system to obtain the center of mass sideslip angle label value. Additional conventional sensors including steering wheel angle, brake master cylinder oil pressure sensor, etc. were used to obtain corresponding signals. The signals of these sensors were uniformly collected and recorded by the vehicle control unit Speedgoat via the CAN bus, and the sampling frequency was set to 100Hz.

[0102] The algorithm of the present invention is verified by real vehicle road test in the alternating dry and wet asphalt driving scenario:

[0103] Figure 4 Preliminary results show that the estimation accuracy of tire-road friction coefficient and center of mass sideslip angle based on hybrid network is significantly higher than that based on physical model in road tests, especially in the process of turning when the vehicle dynamics change sharply. In addition, the task execution time results in the vehicle control unit Speedgoat indicate that the hybrid network can fully guarantee the real-time performance of operation while improving the estimation accuracy.

[0104] It can be seen from the above actual vehicle road test results that the present invention can effectively achieve high-precision real-time estimation of the tire-road friction coefficient and the center of mass sideslip angle.

[0105] A vehicle dynamic parameter real-time estimation system based on a hybrid network includes a test vehicle, a camera arranged on the top of the test vehicle, and a control module arranged inside the test vehicle. The control module, based on the camera on the top of the test vehicle and the on-board sensor, applies the method to perform real-time estimation of the tire-road friction coefficient and the center of mass sideslip angle of the test vehicle.

[0106] A computer-readable storage medium stores computer-readable instructions, which, when executed by a processor, call all or part of the steps of the method.

[0107] If the above functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc., which can store program codes.

[0108] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A real-time estimation method of vehicle dynamic parameters based on a hybrid network, characterized by: The steps include: S1. Based on the seven-degree-of-freedom nonlinear dynamic equation, an implicit mapping function between the vehicle sensor signal parameters and the tire-road friction coefficient and the center of mass sideslip angle is established to clarify the input and output vectors of the hybrid network; S2. Based on the implicit mapping function in S1, a hybrid network structure is constructed including a set of lightweight Inception V3 road classification networks and two sets of deep regression networks with multi-layer long short-term memory units; S3, based on the hybrid network input and output vectors in S1 and the hybrid network structure in S2, construct a high-coverage dataset of working conditions and a time synchronization strategy for intermediate parameters for offline training of the hybrid network; S4: Perform offline training of the hybrid network in S2 based on the high-coverage training data set of working conditions in S3, and complete model deployment and testing on the actual vehicle to achieve high-precision and real-time estimation of the tire-road friction coefficient and the sideslip angle of the center of mass.

2. The method for real-time estimation of vehicle dynamic parameters based on hybrid network according to claim 1, characterized in that: Step S1 includes the following sub-steps: S101. According to Newton's second law, a seven-degree-of-freedom vehicle dynamics model is established, including three vehicle planar motion degrees of freedom and four wheel rotation degrees of freedom; S102, based on the need to accurately describe the tire force generation mechanism under the longitudinal and lateral coupling conditions, establish a fitting UniTire model to analyze the mechanical properties of the tire; S103. Based on the nonlinear dynamic equations established in S101 and S102, the relationship between the vehicle sensor signal parameters, the intermediate signal parameters and the signal parameters to be estimated is obtained, an implicit mapping function between the parameters is constructed, and the input and output vectors of the hybrid network are clarified.

3. The method for real-time estimation of vehicle dynamic parameters based on hybrid network according to claim 2 is characterized in that: In step S103, the implicit mapping function between the parameters , the expressions of the input and output vectors of the hybrid network are: In the formula, , is the longitudinal and lateral acceleration of the vehicle, , is the longitudinal and lateral speed, is the yaw angular velocity, , and are the longitudinal, lateral and vertical forces of the tire, respectively. , are tire slip rate and sideslip angle, is the wheel rotation angular velocity, , is the wheel drive and brake torque, is the wheel angle, is the tire-road friction coefficient.

4. The method for real-time estimation of vehicle dynamic parameters based on hybrid network according to claim 1, characterized in that: Step S2 includes the following sub-steps: S201. Based on the characteristics of the road surface classification task and considering the real-time requirements of real vehicle deployment reasoning, a lightweight Inception V3 road surface classification network is established; S202. Based on the tire-road friction coefficient range output of the lightweight Inception V3 road classification network and the implicit mapping function in step S1, the input and output vectors of the hybrid network, two sets of deep regression networks with multi-layer LSTM are established.

5. The method for real-time estimation of vehicle dynamic parameters based on hybrid network according to claim 4 is characterized in that: In step S202, in order to prevent the regression value of the tire-road friction coefficient from exceeding its prior value range, the mean square error loss function of the first group of LSTM deep regression networks is modified to impose a penalty on the tire-road friction coefficient exceeding the value range. The expression is: In the formula, and are the true and predicted values ​​of the tire-road friction coefficient, , The maximum and minimum values ​​of the tire-road friction coefficient output by the lightweight Inception V3. is the number of training samples.

6. The method for real-time estimation of vehicle dynamic parameters based on hybrid network according to claim 1, characterized in that: Step S3 includes the following sub-steps: S301. Design a three-dimensional active excitation curve according to the dynamic response law of the vehicle, automatically acquire a large amount of simulation data sets and real vehicle road test data sets in batches, and form a hybrid network offline training data set; S302, based on the hybrid network offline training data set established in S301, formulate a hybrid training scheme of simulation data pre-training and real vehicle data fine-tuning; S303. According to the time domain characteristics of the road surface classification network and the deep regression network, a spatiotemporal synchronization strategy of intermediate parameters is designed to align the output of the classification network and the input of the regression network.

7. The method for real-time estimation of vehicle dynamic parameters based on hybrid network according to claim 6 is characterized in that: In step S303, the intermediate parameter spatiotemporal synchronization strategy design process is: The camera is installed on the top of the test vehicle, the sampling frequency is set to 10 Hz, the effective detection range is 5-50m in front of the vehicle, and the current time is defined as , the center of mass of the vehicle is the origin O, the camera has completed sampling, then the distance near the vehicle is , far end distance , based on the simplified longitudinal kinematics of the vehicle, its longitudinal displacement can be expressed as: Define the time when the vehicle reaches the near and far end distances as , , then the relationship between vehicle travel time and space is established through the following formula: In the formula, is the output feedback of the hybrid network. In the corresponding interval The maximum and minimum values ​​of the tire-road friction coefficient of the first group of LSTM deep regression networks are input as delays at the time interval Output value of road surface classification network under time label, here is the network inference time.

8. The method for real-time estimation of vehicle dynamic parameters based on hybrid network according to claim 1, characterized in that: In step S4, based on the high-coverage training data set of the working conditions in S3 and combined with the spatiotemporal synchronization strategy of the intermediate parameters, the offline training of the hybrid network in S2 is performed, and the model deployment and testing are completed on the actual vehicle.

9. A vehicle dynamic parameter real-time estimation system based on a hybrid network, characterized by: The invention comprises a test vehicle, a camera arranged on the top of the test vehicle, and a control module arranged inside the test vehicle. The control module applies any one of the methods of claims 1 to 8 based on the camera on the top of the test vehicle and the on-board sensor to estimate the tire-road friction coefficient and the sideslip angle of the center of mass of the test vehicle in real time.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-readable instructions, and when the computer-readable instructions are executed by a processor, all or part of the steps of the method described in any one of claims 1 to 8 are called.

Images