Vehicle state observation method, vehicle control method, and vehicle

Abstract

The present application relates to a kind of vehicle state observation method, vehicle control method and vehicle;Wherein, vehicle state observation method includes: based on first state data and first control data, first unmodeled dynamics is determined using first module;First module represents unmodeled dynamic prediction neural network;Based on first state data, first unmodeled dynamics and first control data, first estimated state is determined using second module;Wherein, second module is used to determine first estimated state using target mechanism model and first fusion coefficient;First fusion coefficient represents the fusion coefficient between target mechanism model and first unmodeled dynamics;According to first state data and first estimated state, second fusion coefficient is determined using third module;First fusion coefficient is updated using second fusion coefficient.The present application uses first module and second module to perform vehicle state estimation task, uses third module to perform fusion coefficient update or correction task, so, adopt bootstrap mechanism two-stage working mode, by adjusting fusion coefficient online, different vehicle platforms of different car types, different hardware configurations can be adapted, without retraining vehicle state observer, to improve the robustness of the vehicle state observation method.

Description

Technical Field

[0001] This invention relates to, but is not limited to, the field of vehicle dynamics control, and particularly to a vehicle state observation method, a vehicle control method, and a vehicle. Background Technology

[0002] As vehicle motion control technology develops towards applications across all scenarios, distributed drive-by-wire vehicles, with their overdrive structure featuring independent four-wheel steering, can fully tap into the vehicle's dynamic potential through multi-objective collaborative optimization, achieving high-precision trajectory tracking and driving stability control under extreme conditions.

[0003] The accuracy of vehicle dynamics models and the precision of motion state observations directly determine the control performance of distributed drive-by-wire vehicles. In related technologies, classic vehicle dynamics models used for real-time vehicle control and simulation are mostly based on physical mechanisms and belong to mechanistic dynamics models. However, in practical applications, the dynamic mechanistic models of distributed drive-by-wire vehicles suffer from significant model mismatch problems due to various factors: On the one hand, to ensure real-time control, mechanistic dynamics models often simplify longitudinal-lateral-yaw coupling motion characteristics, neglect the attenuation laws of longitudinal and lateral force coupling of tires, and actuator response hysteresis, leading to structural modeling errors; on the other hand, environmental disturbances and parameter perturbations such as changes in road adhesion coefficients, vehicle load transfer, and tire wear further exacerbate the model mismatch problem, causing a significant decrease in the accuracy of state observation methods based on traditional mechanistic dynamics models, and even triggering vehicle control instability.

[0004] Therefore, improving the accuracy of vehicle status observation has become an urgent problem to be solved. Summary of the Invention

[0005] One objective of this invention is to provide a vehicle state observation method to address the problem of low accuracy in vehicle state observation in the prior art; a second objective is to provide a vehicle control method; and a third objective is to provide a vehicle.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: On one hand, the present invention provides a vehicle state observation method, the method comprising: Based on the first state data and the first control data, the first unmodeled dynamic is determined using the first module; the first module represents the unmodeled dynamic prediction neural network. Based on the first state data, the first unmodeled dynamics, and the first control data, a first estimated state is determined using a second module; wherein, the second module is used to determine the first estimated state using a target mechanism model and a first fusion coefficient; the first fusion coefficient represents the fusion coefficient between the target mechanism model and the first unmodeled dynamics; Based on the first state data and the first estimated state, the second fusion coefficient is determined using the third module; The first fusion coefficient is updated using the second fusion coefficient.

[0007] Based on the aforementioned technical means, on the one hand, by fusing the first unmodeled dynamics to determine the first estimated state corresponding to the first state data, the model mismatch problem caused by factors such as model simplification and coupling time delay characteristics can be reduced. At the same time, the first estimated state is provided to the vehicle controller so that the controller can perform vehicle control based on the first estimated state, which can achieve accurate observation and unbiased control in the long time domain. On the other hand, the first and second modules are used to perform the vehicle state estimation task, and the third module is used to perform the fusion coefficient update or correction task. In this way, by adopting a boot-strapping two-stage working mode, the fusion coefficients can be adjusted online to adapt to heterogeneous vehicle platforms with different vehicle models and different hardware configurations without retraining the PINN-based vehicle observer, which solves the problem of cross-platform prediction failure of pure data-driven observation methods, thereby improving the robustness of the vehicle state observation method.

[0008] Furthermore, the step of determining the first estimated state using the second module based on the first state data, the first unmodeled dynamics, and the first control data includes: Based on the first state data, the first control data, and the first unmodeled dynamics, the fourth module is used to determine the observation gain corresponding to the first state data. Based on the first state data, the first unmodeled dynamics, the first control data, and the observation gain corresponding to the first state data, the second module is used to determine the first estimated state.

[0009] Based on the above technical means, the first estimated state is determined by fusing the observation gain corresponding to the first state data. This realizes the automatic determination of the correction magnitude of the target mechanism model based on the uncertainty of the target mechanism model and the vehicle sensor, so that the first estimated state can take into account both physical laws and measured accuracy, thereby improving the accuracy of state estimation.

[0010] Further, the step of determining the first estimated state using the second module based on the first state data, the first unmodeled dynamics, the first control data, and the observation gain corresponding to the first state data includes: Based on the second state data, the second control data, the first unmodeled dynamics and the target mechanism model, the second module is used to determine the first prior estimate of the fusion state corresponding to the first state data and the first unmodeled dynamics; wherein, the second state data and the second control data represent the state data and control data collected in the previous data acquisition cycle of the first state data and the first control data, respectively. Based on the first prior estimate, the first state data and the observation gain corresponding to the first state data, the second module is used to determine the first posterior estimate of the fusion state corresponding to the first unmodeled dynamics of the first state data. Based on the first posterior estimate and the target mechanism model, the first estimated state is determined using the second module.

[0011] Based on the above technical means, by fusing the observation gain corresponding to the first state data and the first unmodeled dynamics to determine the first estimated state, the effect of automatically determining the correction magnitude of the target mechanism model based on the uncertainty of the target mechanism model and the vehicle sensor, and compensating for the mismatch of the target mechanism model is achieved, thereby improving the estimation accuracy and reliability of the first estimated state and achieving unbiased estimation effect.

[0012] Furthermore, the first fusion coefficient represents the update coefficient corresponding to the state transition matrix of the target mechanism model.

[0013] According to the above technical means, since the first estimated state corresponding to the first state data is determined based on the first prior estimate, and the first prior estimate is determined based on the first unmodeled dynamic and the target mechanism model (including the state transition matrix corresponding to the target mechanism model), the fusion method of the unmodeled dynamic and the target mechanism model can be adjusted by using the fusion coefficient to update the state transition matrix corresponding to the target mechanism model.

[0014] Furthermore, the third module is a parametric extended Kalman filter.

[0015] Based on the aforementioned technical means, by implementing the third module as parameter EKF, independent and online correction of the fusion coefficients can be achieved. Furthermore, the method also includes: Acquire sample data; wherein the sample data includes sample state data and sample control data, and each sample state data has corresponding unmodeled dynamic and measured state data of the sample; Based on the sample state data and the sample control data, the first prediction unmodeled dynamic is determined using the first initial module; Based on the sample state data, the sample control data, and the predicted unmodeled dynamics, the prediction observation gain corresponding to the sample state data is determined using the fourth initialization module; Based on the sample state data, the unmodeled dynamics of the prediction, the sample control data, and the sample prediction observation gain, the second module is used to determine the prediction estimation state. The target loss is determined based on the predicted unmodeled dynamics, the predicted observed gain, the predicted estimated state, the sample unmodeled dynamics, and the measured state data. The first initial module and the fourth initial module are updated using the target loss to obtain the first module and the fourth module.

[0016] Based on the above technical means, the first initial module and the fourth initial module in the vehicle state observer are trained and updated using sample data, so as to obtain the first module and the fourth module that can accurately predict the unmodeled dynamics and observation gain, thereby effectively improving the estimation accuracy of vehicle state.

[0017] Furthermore, determining the target loss based on the unmodeled predictive dynamics, the predicted observed gain, the predicted estimated state, the unmodeled sample dynamics, and the measured state data includes: Based on the predicted unmodeled dynamics and the sample unmodeled dynamics, a first loss is determined; Based on the predicted estimated state and the measured state data, a second loss is determined; Based on the predicted observation gain corresponding to the sample state data, the third loss is determined; The target loss is determined based on the first loss, the second loss, and the third loss.

[0018] Based on the above technical means, this application can avoid problems such as oscillation and instability during the state observation process by embedding physical constraints and Kalman gain stability constraints during the training process of the vehicle state observer, thereby enabling the vehicle state observer to operate stably under extreme conditions for a long time.

[0019] A vehicle control method, comprising: Based on the first state data and the first control data, an unbiased state observer is used to determine the first unmodeled dynamic and the first estimated state corresponding to the first state data; the unbiased state observer is used to perform some or all of the steps in the above vehicle state observation method. Based on the first unmodeled dynamics and the first estimated state, the vehicle controller is used to determine the target control command for the target vehicle.

[0020] A vehicle condition monitoring device, comprising: An unmodeled dynamic determination module is used to determine a first unmodeled dynamic based on first state data and first control data using a first module; the first module represents an unmodeled dynamic prediction neural network. A state estimation module is used to determine a first estimated state based on the first state data, the first unmodeled dynamics, and the first control data using a second module; wherein, the second module is used to determine the first estimated state using a target mechanism model and a first fusion coefficient; the first fusion coefficient represents the fusion coefficient between the target mechanism model and the first unmodeled dynamics; The coefficient correction module is used to determine the second fusion coefficient based on the first state data and the first estimated state using the third module; The coefficient update module is used to update the first fusion coefficient using the second fusion coefficient.

[0021] A vehicle includes: a memory and a processor; wherein, The memory stores a computer program that can run on a processor; when the processor executes the program, it implements some or all of the steps in the above method.

[0022] The beneficial effects of this invention are: (1) By combining the generalization of the fusion mechanism model with the nonlinear fitting ability of PINN, and by using PINN to compensate for structural modeling errors, the dynamic characteristics of vehicles under extreme conditions can be accurately captured. (2) Embedding physical constraints and Kalman gain stability constraints into the training process of the vehicle state observer can avoid problems such as oscillation and instability during the observation process, and can run stably under extreme conditions for a long time. (3) By adopting the boot-strapping dual-stage working mode, the fusion coefficient can be adjusted online. It can adapt to heterogeneous vehicle platforms with different models and hardware configurations without retraining the PINN-based vehicle state observer, thus solving the problem of cross-platform prediction failure of pure data-driven observation methods. (4) This method is implemented using a modular system architecture with clear division of labor. When deployed on the vehicle, it does not require large-scale modification of the existing vehicle hardware. Therefore, it can be quickly integrated into the mainstream autonomous driving computing platform and the deployment cost is low. Attached Figure Description

[0023] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present invention and, together with the specification, serve to explain the technical solutions of the present invention.

[0024] Figure 1 A schematic diagram illustrating the implementation process of a vehicle state observation method provided by the present invention; Figure 2 A schematic diagram of the current process of an embodiment of the present invention is provided; Figure 3 A schematic diagram of the current process of an embodiment of the present invention is provided; Figure 4 A schematic diagram illustrating the implementation process of a vehicle control method provided by the present invention; Figure 5 A schematic diagram of a test result provided by the present invention; Figure 6 A schematic diagram of a test result provided by the present invention; Figure 7 A schematic diagram of a test result provided by the present invention; Figure 8 A schematic diagram of a test result provided by the present invention; Figure 9 A schematic diagram of a test result provided by the present invention; Figure 10 A schematic diagram of a test result provided by the present invention; Figure 11 This is a schematic diagram of the composition structure of a vehicle condition monitoring device provided by the present invention; Figure 12 This is a schematic diagram of the hardware entity of a vehicle provided by the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. The described embodiments should not be regarded as limitations on the present invention. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0027] The terms “first / second / third” are used merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that “first / second / third” may be interchanged in a specific order or sequence where permitted, so that the embodiments of the invention described herein can be implemented in an order other than that illustrated or described herein.

[0028] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein is for descriptive purposes only and is not intended to limit the scope of the invention.

[0029] To improve the accuracy of vehicle status observation, the following two solutions have been proposed in related technologies: Option 1: Traditional mechanistic observation methods, such as the classic Extended Kalman Filter (EKF) and Unscented Kalman Filter. Because they use a fixed covariance matrix to describe model mismatch, traditional mechanistic observation methods can only fit random noise and cannot effectively compensate for structural modeling errors. Especially under extreme conditions such as high-speed steering and emergency obstacle avoidance, the observation accuracy decreases sharply, making it difficult to meet high-precision control requirements. Option 2: Pure data-driven observation method based on deep neural networks. Because it relies on a large number of training samples covering the entire scenario, the generalization ability of the pure data-driven observation method is limited by the sample coverage. When transferred to heterogeneous vehicle platforms with different vehicle models and hardware configurations, it is prone to problems such as prediction failure and observation oscillation, resulting in poor engineering adaptability.

[0030] The technical solutions of the present invention will now be clearly and completely described with reference to the accompanying drawings in the embodiments of the present invention.

[0031] Figure 1 This is a schematic diagram illustrating the implementation process of a vehicle state observation method provided by the present invention, as shown below. Figure 1 As shown, the method includes the following steps S11 to S14: Step S11: Based on the first state data and the first control data, the first unmodeled dynamic is determined using the first module; the first module represents the unmodeled dynamic prediction neural network.

[0032] Here, the first state data refers to the vehicle state data collected for the target vehicle. In some embodiments, the first state data can be any type of state data corresponding to the target vehicle. In some embodiments, the first state data can be the longitudinal velocity, lateral velocity, yaw rate, wheel angle, steering wheel angle, etc., of the target vehicle. In some embodiments, the first state data can be sensor data collected using multiple sensors installed in the target vehicle, or it can be state data collected based on a vehicle simulation system. In some embodiments, the first state data can be raw sensor data collected by sensors, or it can be data obtained after preprocessing the raw sensor data.

[0033] The first control data refers to the control data corresponding to the first state data; that is, the first state data is the vehicle state data generated under the excitation of the first control data. In some embodiments, the first control data may include control data of any type. In some embodiments, the first control data may include front wheel steering angle, rear wheel steering angle, longitudinal acceleration, and / or additional yaw moment, etc. In some embodiments, the first control data may be control data read from the controller of the target vehicle, or it may be control data obtained by the controller used to execute the control algorithm in the vehicle simulation system.

[0034] Unmodeled dynamics refer to the inherent dynamic behaviors of a vehicle that are not included in the model equations during the modeling process due to simplification assumptions or other reasons. Examples of unmodeled dynamics include: tire nonlinear transient characteristics, body roll and pitch effects, suspension friction, bushing elastic deformation, and dynamic axle load transfer. It is evident that unmodeled dynamics cannot be eliminated simply by adjusting model parameters; they manifest as dynamic residuals between the vehicle dynamics and the output of the mechanistic model.

[0035] The first module is a neural network used to predict unmodeled dynamics based on input vehicle state data and control data. In some implementations, the first module can be any type of neural network capable of performing unmodeled dynamic prediction. For example, the first module can be implemented as a multilayer perceptron (MLP), a residual MLP, a long short-term memory (LSTM) network, a transformer network, etc. In some implementations, the first initial module can be trained using multiple sets of sample data so that the first initial module can predict the corresponding unmodeled dynamics based on the input state data and control data, thereby obtaining the first module.

[0036] In this way, the first state data and first control data of the target vehicle are input into the first module, so that the first module can perform inference to obtain the first unmodeled dynamics.

[0037] Step S12: Based on the first state data, the first unmodeled dynamics, and the first control data, a first estimated state is determined using a second module; wherein, the second module is used to determine the first estimated state using a target mechanism model and a first fusion coefficient; the first fusion coefficient represents the fusion coefficient between the target mechanism model and the first unmodeled dynamics.

[0038] A target mechanism model is a mechanistic dynamics model used to estimate the vehicle's state. In some implementations, the target mechanism model can be a two-degree-of-freedom vehicle dynamics model, a three-degree-of-freedom vehicle dynamics model, or a four-degree-of-freedom vehicle dynamics model, etc. The target mechanism model can output the observed values ​​corresponding to the vehicle state data based on the vehicle state data and the corresponding control data.

[0039] In some implementations, the second module may be a module built based on the target mechanism model and the Kalman filter correction mechanism.

[0040] The first fusion coefficient characterizes the fusion coefficient between the target mechanism model and the first unmodeled dynamics. That is, the first fusion coefficient is used to determine the fusion mode between the target mechanism model and the first unmodeled dynamics.

[0041] In some implementations, when fusing the observed values ​​output by the target mechanism model with the first unmodeled dynamic, the first fusion coefficient can be used as the fusion weight corresponding to the observed value or the first unmodeled dynamic. For example, when the first fusion coefficient is the fusion weight corresponding to the observed value, firstly, based on the first state data and the first control data, the observed value corresponding to the first state data is output using the target mechanism model; simultaneously, the fusion weight corresponding to the first unmodeled dynamic is determined according to the first fusion coefficient (for example, its value is the difference between the value 1 and the first fusion coefficient); then, the observed value and the first unmodeled dynamic are weighted and summed to obtain the first estimated state corresponding to the first state data.

[0042] In some implementations, firstly, the target mechanism model is updated using a first fusion coefficient, and augmented state data corresponding to the first unmodeled dynamic is determined based on the first unmodeled dynamic; then, based on the augmented state data corresponding to the first state data and the first control data, the updated target mechanism model is used to output observations, and the output observations are used as the first estimated state.

[0043] Step S13: Based on the first state data and the first estimated state, determine the second fusion coefficient using the third module.

[0044] Here, after determining the first estimated state corresponding to the first state data, the first fusion coefficient is corrected based on the residual between the first state data and the first estimated state to obtain the second fusion coefficient.

[0045] In some implementations, the third module can be implemented as a parametric extended Kalman filter (EKF) within a dual extended Kalman filter (DEKF). When determining the second fusion coefficient using the third module, the first fusion coefficient can be used as a priori estimate of the second fusion coefficient, and then the posterior estimate of the second fusion coefficient can be determined by calculating the Kalman gain. Thus, by implementing the third module as a parametric EKF, independent and online correction of the fusion coefficients can be achieved.

[0046] Step S14: Update the first fusion coefficient corresponding to the second module using the second fusion coefficient. Here, the first fusion coefficient corresponding to the second module is updated using the second fusion coefficient, so that the updated second fusion coefficient can more accurately reflect the fusion relationship between the unmodeled dynamics and the target mechanism model.

[0047] In some implementations, the first and second modules are referred to as vehicle state observers based on Physics-Informed Neural Networks (PINN), and the third module is referred to as a fusion coefficient estimator.

[0048] In the vehicle state observation method provided in this application, firstly, based on first state data and first control data, a first unmodeled dynamic is determined using a first module, wherein the first module represents an unmodeled dynamic prediction neural network; then, based on the first state data, the first unmodeled dynamic, and the first control data, a first estimated state is determined using a second module, wherein the second module is used to determine the first estimated state using a target mechanism model and a first fusion coefficient, wherein the first fusion coefficient represents the fusion coefficient between the target mechanism model and the first unmodeled dynamic; subsequently, based on the first state data and the first estimated state, a second fusion coefficient is determined using a third module; finally, the first fusion coefficient is updated using the second fusion coefficient. In this way, on the one hand, by fusing the first unmodeled dynamics to determine the first estimated state corresponding to the first state data, the model mismatch problem caused by factors such as model simplification and coupling time delay characteristics can be reduced. At the same time, providing the first estimated state to the vehicle controller so that the controller can perform vehicle control based on the first estimated state can achieve accurate observation and unbiased control in the long time domain. On the other hand, by using the first and second modules to perform the vehicle state estimation task and the third module to perform the fusion coefficient update or correction task, a boot-strapping two-stage working mode is adopted. By adjusting the fusion coefficients online, it can be adapted to heterogeneous vehicle platforms with different vehicle models and different hardware configurations without retraining the PINN-based vehicle state observer. This solves the problem of cross-platform prediction failure of pure data-driven observation methods, thereby improving the robustness of the vehicle state observation method.

[0049] In some implementations, the step of determining the first estimated state using the second module based on the first state data, the first unmodeled dynamics, and the first control data, i.e., step S12 above, can be implemented as steps S121 to S122: Step S121: Based on the first state data, the first control data, and the first unmodeled dynamics, the fourth module is used to determine the observation gain corresponding to the first state data.

[0050] In some implementations, the observation gain corresponding to the first state data characterizes the Kalman gain. Thus, the fourth module can be implemented as a module for outputting the Kalman gain.

[0051] In some implementations, the fourth module can determine the observation gain corresponding to the first state data based on the first state data, the first control data, and the first unmodeled dynamics, using the Kalman gain calculation formula corresponding to the conventional Kalman filter. The conventional Kalman filter can be a Kalman filter (EK), EKF, or DEKF.

[0052] In some implementations, the fourth module can be a deep neural network. In this way, after the first state data, first control data, and first unmodeled dynamic data are input into the fourth module, the fourth module outputs the observation gain corresponding to the first state data through network inference.

[0053] Step S122: Based on the first state data, the first unmodeled dynamics, and the observation gain corresponding to the first state data, the second module is used to determine the first estimated state.

[0054] In Kalman filtering, observation gain is used to dynamically measure the accuracy of the observed values ​​output by the dynamic model compared to the sensor measurements.

[0055] Here, after determining the observation gain corresponding to the first state, this observation gain is used to allocate the confidence levels of the prior estimated state corresponding to the first state data (i.e., the observed values ​​output by the target mechanism model based on the first state data, the first control data, and the first unmodeled dynamics) and the sensor measured values ​​(i.e., the first state data). Then, based on the observed values ​​and their corresponding confidence levels, and the measured values ​​and their corresponding confidence levels, the first estimated state of the first state data is determined. In some implementations, the observation gain corresponding to the first state can be used as a weighting parameter for the residual between the first state data and the observed values ​​output by the target mechanism model. Then, the observed values ​​output by the target mechanism model and this residual are weighted and summed to obtain the first estimated state.

[0056] In the embodiments provided in this application, the first estimated state is determined by fusing the observation gain corresponding to the first state data. This realizes the automatic determination of the correction magnitude of the target mechanism model based on the uncertainty of the target mechanism model and the vehicle sensor, so that the first estimated state can take into account both physical laws and measured accuracy, thereby improving the accuracy of state estimation.

[0057] In some implementations, the step of determining the first estimated state using the second module based on the first state data, the first unmodeled dynamics, the first control data, and the observation gain corresponding to the second state data, i.e., step S122 above, can be implemented as steps S1221 to S1223: Step S1221: Based on the second state data, the second control data, the first unmodeled dynamics and the target mechanism model, the second module is used to determine the first prior estimate of the fusion state corresponding to the first state data and the first unmodeled dynamics; wherein, the second state data and the second control data represent the state data and control data collected in the previous data acquisition cycle of the first state data and the first control data, respectively.

[0058] The second state data and the second control data refer to the state data and control data collected in the previous data acquisition cycle of the first state data and the first control data. At the same time, the second state data is the state data generated under the excitation of the second control data.

[0059] Here, after inputting the second state data, second control data, and first unmodeled dynamics into the second module, the second module first fuses the second state data and the first unmodeled dynamics to obtain the fused state corresponding to the second state data, constructing an augmented state space. Then, based on the fused state corresponding to the second state data and the second control data, the second module uses the state transition matrix and control matrix corresponding to the target mechanism model to calculate the estimated state corresponding to the fused state. Finally, the estimated state corresponding to the fused state is used as the first prior estimate of the fused state corresponding to the first state data and the first unmodeled dynamics. The target mechanism model is a three-degree-of-freedom vehicle dynamics model; the state transition matrix and control matrix are the state transition matrix and discrete control input matrix obtained after sequentially performing approximate linearization and discretization processing on the three-degree-of-freedom vehicle dynamics model, respectively.

[0060] Step S1222: Based on the first prior estimate, the first state data, and the observation gain corresponding to the first state data, the second module is used to determine the first posterior estimate of the fused state corresponding to the first unmodeled dynamic.

[0061] Here, the first prior estimate is corrected using the observation gain corresponding to the first state data and the first state data, thereby obtaining the posterior estimate of the fused state, i.e., the first posterior estimate.

[0062] Thus, after inputting the first prior estimate, the first state data, and the observation gain corresponding to the first state data into the second module, the second module first calculates the residual based on the first state data, the first prior estimate, and the state output matrix corresponding to the target mechanism model. Then, using the observation gain corresponding to the first state data as the weight parameter of the residual, the first prior estimate and the residual are weighted and summed to obtain the posterior estimate of the fused state, i.e., the first posterior estimate. Here, the target mechanism model is a three-degree-of-freedom vehicle dynamics model; the state output matrix is ​​the matrix obtained after sequentially performing approximate linearization and discretization processing on the three-degree-of-freedom vehicle dynamics model.

[0063] Step S1223: Based on the first posterior estimate and the target mechanism model, the second module is used to determine the first estimated state.

[0064] After determining the first posterior estimate, the second module calculates the first estimated state corresponding to the first state data based on the first posterior estimate and the state output matrix corresponding to the target mechanism model.

[0065] In the embodiments provided in this application, the first estimated state is determined by fusing the observation gain corresponding to the first state data and the first unmodeled dynamics. This achieves the effect of automatically determining the correction magnitude of the target mechanism model based on the uncertainty of the target mechanism model and the vehicle sensor, and compensating for the mismatch of the target mechanism model, thereby improving the estimation accuracy and reliability of the first estimated state and achieving unbiased estimation.

[0066] In some implementations, the first fusion coefficient represents the update coefficient corresponding to the state transition matrix of the target mechanism model.

[0067] As mentioned above, the state transition matrix is ​​obtained by sequentially performing approximate linearization and discretization on the target mechanism model (e.g., a three-degree-of-freedom vehicle dynamics model).

[0068] In the target mechanism model, the state transition matrix is ​​used to infer the state at the next moment based on the current state. Thus, by using the first fusion coefficient to update the state transition matrix corresponding to the target mechanism model, the process of inferring the state at the next moment in the target mechanism model can be adjusted.

[0069] In some implementations, the state transition matrix corresponding to the target mechanism model can be updated using the fusion coefficients based on the following formula (1): (1); Where A represents the nonlinear state transition matrix corresponding to the target mechanism model; diag( ), representing the function for calculating diagonal matrices; , indicating based on the first i The fusion coefficient is determined by +1 state data.

[0070] As can be seen from the above, in the embodiments provided by this application, since the first estimated state corresponding to the first state data is determined based on the first prior estimate, and the first prior estimate is determined based on the first unmodeled dynamic and the target mechanism model (including the state transition matrix corresponding to the target mechanism model), the state transition matrix corresponding to the target mechanism model can be updated by using the fusion coefficient, thereby adjusting the fusion method of the unmodeled dynamic and the target mechanism model.

[0071] In some embodiments, the vehicle state observation method provided in this application further includes a training method for the first module and the second module. Specifically, the method includes the following steps S15 to S110: Step S15: Obtain sample data; wherein the sample data includes sample state data and sample control data, and the sample state data has corresponding unmodeled dynamic and measured state data of the sample.

[0072] Here, sample data refers to the vehicle state data and control data used to train the vehicle state observer; the vehicle state observer includes the first module, the fourth module, and the second module mentioned above. It should be noted that the second module is constructed based on the target mechanism model and the Kalman filter mechanism; therefore, the training process for the vehicle state observer does not include updating the second module.

[0073] In some implementations, vehicle status data and control data can be acquired during the operation of a real vehicle to serve as sample status data and sample control data.

[0074] In some implementations, vehicle state data and control data can be obtained from a vehicle simulation system as sample state data and sample control data.

[0075] In some implementations, the sample state data may include the vehicle's longitudinal speed, lateral speed, yaw rate, wheel angle, steering wheel angle, etc.

[0076] In some implementations, sample control data may include the vehicle's front wheel steering angle, rear wheel steering angle, longitudinal acceleration, and / or additional yaw moment.

[0077] Unmodeled dynamics refer to the ground truth labels assigned to sample state data, which are then used to calculate the loss. In some implementations, unmodeled dynamics can be manually labeled for the sample state data; alternatively, they can be generated by a pre-trained model based on the sample state data and sample control data.

[0078] Measured state data refers to the state data actually measured by sensors installed in the vehicle after applying sample state data and sample control data to vehicle control. Therefore, the measured state can be used as the true value label for the estimated state corresponding to the sample state data.

[0079] Step S16: Based on the sample state data and the sample control data, the first initial module is used to determine the unmodeled dynamics for prediction.

[0080] Here, the first initial model refers to a deep neural network structure with an input layer, hidden layer, and output layer, used to predict unmodeled dynamics.

[0081] In this way, the sample state data and sample control data are input into the first initial module to perform inference using the first initial model and output the predicted unmodeled dynamics.

[0082] Step S17: Based on the sample state data, the sample control data, and the predicted unmodeled dynamics, the fourth initialization module is used to determine the predicted observation gain corresponding to the sample state data.

[0083] Here, the fourth initial module refers to a deep neural network structure with an input layer, hidden layer, and output layer, used to predict observation gain.

[0084] In this way, the sample state data, sample control data, and predicted unmodeled dynamics are input into the fourth initialization module to perform inference and output the predicted observation gain corresponding to the sample state data.

[0085] Step S18: Based on the sample state data, the predicted unmodeled dynamics, the sample control data, and the predicted observation gain, the second module is used to determine the predicted estimated state.

[0086] Here, based on sample state data, predicted unmodeled dynamics, sample control data, and predicted observation gain, the second module is used to determine the specific implementation method of the predicted estimated state. This is the same as the specific implementation method of using the second module to determine the first estimated state described in steps S121 to S122 above, and will not be repeated here.

[0087] Step S19: Determine the target loss based on the predicted unmodeled dynamics, the predicted observation gain, the predicted estimated state, the sample unmodeled dynamics, and the measured state data.

[0088] Here, after determining the unmodeled dynamics and the observed gain, the target loss for network training is determined based on the unmodeled dynamics, the observed gain, the estimated state, the unmodeled dynamics of the samples, and the measured state data.

[0089] In some implementations, the target loss can be determined based on the loss of the predicted unmodeled dynamics relative to the sample unmodeled dynamics.

[0090] In some implementations, the target loss can be determined based on the loss between the predicted estimated state and the measured state data.

[0091] In some implementations, the target loss can be determined based on the predicted observation gain, so that the observer can converge stably during training and ensure physical feasibility.

[0092] Step S110: Update the first initial module and the fourth initial module using the target loss to obtain the first module and the fourth module.

[0093] Here, after determining the target loss, the first and fourth initial modules are updated using the target loss to improve the accuracy of the updated vehicle state observer. Thus, the vehicle state observer is trained using multiple sample data sets, and the first and fourth initial modules are iteratively updated using the target loss. In some implementations, when the target loss is less than a loss threshold, the training epochs reach their maximum limit, or other training termination conditions are met, the last updated first initial module is used as the first model, and the last updated fourth initial module is used as the fourth module.

[0094] It should be noted that the second module is built based on the target mechanism model and the Kalman filter mechanism. Therefore, the training process for the vehicle state observer does not include the update process for the second module.

[0095] In some implementations, the sample data used for training vehicle state observation is divided into a training set and a validation set. The training set is used to update the network weights and parameters of the first initial module and the fourth initial module, while the validation set is used to monitor the training effect and adjust the hyperparameters.

[0096] In some implementations, during the training of the vehicle state observer, an Adaptive Moment Estimation (Adam) optimizer is used to update the network parameters, with the initial learning rate set to 0.01, the first-order gradient decay coefficient set to 0.99, and the second-order gradient decay coefficient set to 0.999. The learning rate decays linearly with each training epoch.

[0097] In some implementations, the training process for the vehicle state observer consists of 200 training epochs, and an early stopping strategy is introduced. In this strategy, training is stopped when the validation set loss does not show a significant decrease for 10 consecutive epochs to avoid overfitting.

[0098] In the embodiments provided in this application, the first initial module and the fourth initial module in the vehicle state observer are trained and updated using sample data to obtain the first module and the fourth module that can accurately predict unmodeled dynamics and observation gains, thereby effectively improving the accuracy of vehicle state estimation.

[0099] In some implementations, determining the target loss based on the unmodeled predictive dynamics, the predicted observed gain, the predicted estimated state, the unmodeled sample dynamics, and the measured state data, i.e., step S19 above, can be implemented as steps S191 to S194: Step S191: Determine the first loss based on the predicted unmodeled dynamics and the sample unmodeled dynamics.

[0100] Here, the first loss for network training is determined based on the difference between the predicted unmodeled dynamics and the sample unmodeled dynamics.

[0101] In some implementations, the first loss characterizes the mean square error between the predicted unmodeled dynamics and the sample unmodeled dynamics. Thus, the first loss L can be determined using the following formula (2). L : (2); in, N B This indicates the batch size of training sample data in one operation; in some implementations, it can be... N B Set to 128; r , indicates the sample data index; , indicating that the index is r The sample data corresponds to samples whose dynamics are not modeled; , indicating that the index is r The predictions corresponding to the sample data do not model dynamics; Q ω , indicating loss L L The corresponding weights.

[0102] Step S192: Determine the second loss based on the predicted estimated state and the measured state data.

[0103] Here, the second loss is determined based on the difference between the predicted estimated state and the measured state data.

[0104] In some implementations, the mean square error between the predicted estimated state and the measured state data over a time domain can be used as the second loss. Thus, the second loss L can be calculated using the following formula (3). P : (3); in, N S , representing the total amount of sample data; N S , which represents the batch size of training sample data in one session; r , indicating the first to the last N S Between the first sample data r Data index for each sample data; r + i , indicating the firstr The 1st sample data after the 1st sample data i Data index for each sample data; , indicating the first r + i The predicted observation gain corresponding to each sample data; , indicating that after fusing the predicted observation gain, the th r + i The predicted estimated state corresponding to the nth sample data; in some implementations, it can be based on the nth sample data. r + i The posterior estimate of the fusion state corresponding to each sample data point, and the state output equation corresponding to the target mechanism model, are used to determine... ; Q x , indicating the second loss L P The corresponding weights.

[0105] Step S193: Determine the third loss based on the predicted observation gain corresponding to the sample state data.

[0106] Here, the third loss is determined based on the predicted observation gain corresponding to the sample state data, and this third loss is used to ensure the stable convergence of the observation gain.

[0107] In some implementations, the third loss L can be determined by the following formula (4). K : (4); in, , representing observation gain The correlation spectral radius; when the value of the correlation spectral radius is less than 1, the vehicle state observer converges stably. ι , representing the boundary scaling factor; in some implementations, ι The value is 0.01; Q K : Indicates the third loss L K The corresponding weights.

[0108] In some implementations, the observation gain can be determined using the following formula (4). Correlation spectral radius : (4); in, , representing the predicted observation gain; I , represents the identity matrix; , represents the state transition matrix corresponding to the target mechanism model; , representing the state output matrix corresponding to the target mechanism model; , representing the predicted observation gain matrix output by the fourth initial module; , representing a matrix eigenvalues; This represents the largest eigenvalue.

[0109] Step S194: Determine the target loss based on the first loss, the second loss, and the third loss.

[0110] Thus, after determining the first loss, the second loss, and the third loss, the target loss L can be calculated using the following formula (5): (5).

[0111] As can be seen from the above embodiments, by embedding physical constraints and Kalman gain stability constraints during the training process of the vehicle state observer, this application can avoid problems such as oscillation and instability during the state observation process, thereby enabling the vehicle state observer to operate stably under extreme conditions for a long time.

[0112] Below, in conjunction with Figure 2 The illustrated embodiment explains the training process of a PINN-based vehicle state observer. Figure 2 As shown, this embodiment includes the following steps S21 to S27: Step S21: Build the vehicle simulation system; then, proceed to step S22. In some implementations, a vehicle simulation system is built using the automotive dynamics simulation software CarSim and the simulation modeling tool Simulink.

[0113] Step S22: Based on the vehicle simulation system, sample multiple sets of data; then, proceed to step S23. In some implementations, the vehicle state can be excited to the steady-state boundary excitation condition through the cooperative controller in Simulink, as well as coupled motion conditions including longitudinal-lateral coupling and extreme-normal composite motion conditions such as high-speed steering and emergency braking, so as to collect sample data under a variety of different motion conditions.

[0114] In some implementations, the sampling frequency of the sample data can be any frequency, such as 100Hz, 200Hz, etc.

[0115] In some implementations, the sample data includes motion state data and control data; wherein, the motion state data includes longitudinal velocity. v x (Speed ​​range: 0 to 35 m / s), lateral velocity v y (Speed ​​range: -5 to 5 m / s, with the left side of the vehicle being positive), yaw rate γ (Angular velocity range is -1.5 to 5 rad / s, with counter-clockwise leftward rotation of the vehicle defined as positive); control data includes front wheel steering angle. δ f (Steering angle range is -0.3 to 0.3 rad, with turning towards the left side of the vehicle being positive), Rear wheel steering angle δ r (Turn angle range is -0.1 to 0.1 rad, with turning towards the left side of the vehicle being positive), longitudinal acceleration a x (Acceleration range is -8 to 4 rad / s) 2 (where positive is the direction facing forward of the vehicle) and additional yaw moment. M z (The torque range is -3000 to 3000 Nm, where the torque that causes the vehicle to rotate counterclockwise to the left is positive).

[0116] In some implementations, after collecting multiple sets of sample data, data cleaning is performed on these sets of data. In some implementations, outliers in the multiple sets of sample data can be removed using the 3σ criterion to avoid the impact of outliers on the network training and observation accuracy.

[0117] In some implementations, multiple sets of sample data can be normalized to map sample data with different dimensions and numerical ranges to the same numerical interval (e.g., [0,1] or [-1,1], etc.), eliminating the influence of differences in physical dimensions and magnitudes, accelerating neural network training convergence, improving training stability and the generalization ability of the observer obtained after training.

[0118] In some implementations, multiple sets of sample data are timestamped to ensure time synchronization of sample data from different sources.

[0119] In some implementations, the preprocessed sample data is divided into training, validation, and test sets according to a specified ratio for use as the training and validation sets for the observer. In some implementations, the sample data can be divided into training, validation, and test sets in a 7:2:1 ratio.

[0120] In some implementations, to increase the realism of the sample data, the moving average filtering algorithm shown in formula (6) can be used to extract sensor noise. : (6); in, i This represents the data index of the current sample data, and it is located at the center of the sliding window, i.e. x ( i () represents the sample data at the center of the sliding window; The sliding window length is 2 N +1, where, N It is a natural number; sample x = ( v x , v y , γ ) T ; w , represents the weight of the corresponding sample data, and the weight can be calculated using the following formula (7). w ( j ): (7); in, j , represents the index of the sample data within the sliding window; σ 2 , where represents the Gaussian kernel parameter, corresponding to the variance of the noise distribution.

[0121] In this way, after determining the sensor noise data corresponding to each set of sample data, the sensor noise data is added to the sample data to improve the realism of the sample data.

[0122] Each sample state data has corresponding unmodeled dynamics and measured state data. The unmodeled dynamics can be manually labeled or generated by a pre-trained model based on the sample state data and sample control data; the measured state data is sensor data collected by the vehicle's sensors.

[0123] Step S23: Input multiple sets of sample data into the first initial module to obtain multiple predicted unmodeled dynamics; then, execute step S24. In some implementations, the input layer dimension of the first initial module is 7, to receive 7 input data items for each set of sample data, namely, 3 state data items, namely: longitudinal velocity. v x Lateral velocity v y yaw rate γ The four control data points are: front wheel steering angle. δ f Rear wheel steering angle δ r Longitudinal acceleration a x and additional yaw moment M z .

[0124] In some implementations, the output layer dimension of the first initial module is 3 to output 3 predicted unmodeled dynamics, namely, predicted longitudinal velocity unmodeled dynamics, predicted lateral velocity unmodeled dynamics, and predicted yaw rate unmodeled dynamics.

[0125] In some implementations, the first initial module is provided with four hidden layers, with the number of neurons in each hidden layer being 16, 32, 64, and 32, respectively.

[0126] In some implementations, the forward propagation formula corresponding to the first initial module is shown in formula (8): (8); in, x ( i ), indicating the first i Each sample state data, and x ( i ) = ( v x , v y , γ) ; u ( i ), indicating the first i Individual sample control data, and u ( i ) = ( δ f , δ r , a x , M z ); , represents the network hyperparameters of the first initial module, and = ( w a , b a ) T ;in: w a = ( w a,in ,w a,h,1 , ..., w a,h,n , w a,out ), representing the weight parameters of each network layer; w a,in , representing the weight parameters of the input layer; w a,h,1 to w a,h,n , representing the weight parameters of each hidden layer. n Indicates the number of hidden layers, and n It is a natural number greater than 1; w a,out , represents the weight parameters of the output layer; b a = ( b a,in , b a,h,1 , ..., b a,h,n , b a,out ), representing the bias parameters of each network layer; b a,in , representing the bias parameters of the input layer; b a,h,1 to b a,h,n , representing the bias parameters of each hidden layer. n Indicates the number of hidden layers, and n It is a natural number greater than 1; b a,out , represents the bias parameter of the output layer; tanh( ), representing the activation function of the first initial module; ( z a,in , z a,h,1 , ..., z a,h,n , z a,out ) T , representing the output of each network layer; = z a,out ( i ), indicating that the prediction is for unmodeled dynamics.

[0127] Step S24: Input multiple sets of sample data and their corresponding unmodeled predicted dynamics into the fourth initial module to obtain multiple predicted observation gains; then, execute step S25. In some implementations, the input layer dimension of the fourth initial module is 10, to receive 3 state data items, 4 control data items, and 3 predicted unmodeled dynamics for each set of sample data. Specifically, the 3 state data items are: longitudinal velocity... v x Lateral velocity v y yaw rate γ The four control data points are: front wheel steering angle. δ f Rear wheel steering angle δ r Longitudinal acceleration a x and additional yaw moment M z Three predicted dynamics were not modeled: predicted longitudinal velocity, predicted lateral velocity, and predicted yaw rate.

[0128] In some implementations, the output layer dimension of the fourth initial module is 18 to output an 18-dimensional prediction observation gain matrix.

[0129] In some implementations, the fourth initial module is provided with four hidden layers, with the number of neurons in each hidden layer being 16, 32, 64, and 32, respectively.

[0130] In some implementations, the forward propagation formula corresponding to the fourth initial module is shown in the following formula (9): (9); Among them, among them, x ( i ), indicating the first i One sample data, and x ( i ) = ( v x , v y , γ) ; u ( i ), indicating the first i one sample, and u ( i ) = ( δ f , δ r , a x , M z ); , indicating the first iThe predictions for each sample data point do not have dynamic modeling. , represents the network hyperparameters of the fourth initial module, and = ( w b , b b ) T ;in: w b = ( w b,in , w b,h,1 , ..., w b,h,n , w b,out ), representing the weight parameters of each network layer; w b,in , representing the weight parameters of the input layer; w b,h,1 to w b,h,n , representing the weight parameters of each hidden layer. n Indicates the number of hidden layers, and n It is a natural number greater than 1; w b,out , represents the weight parameters of the output layer; b b = ( b b,in , b b,h,1 , ..., b b,h,n , b b,out ), representing the bias parameters of each network layer; b b,in , representing the bias parameters of the input layer; b b,h,1 to b b,h,n , representing the bias parameters of each hidden layer. n Indicates the number of hidden layers, and n It is a natural number greater than 1; b b,out , represents the bias parameter of the output layer; tanh'( ), representing the activation function of the fourth initial module; ( z b,in , z b,h,1 , ..., z b,h,n , zb,out ) T , representing the output of each network layer; = z b,out ( i ), representing the predicted observation gain output.

[0131] Step S25: Input multiple sets of sample data, multiple unmodeled dynamics of prediction, and multiple predicted observation gains into the second module to obtain the predicted estimated state corresponding to each sample data; then, execute step S26. In some implementations, the following formula (10) can be used to determine the first... i The second prior estimate of the fusion state corresponding to each sample state data and its predicted unmodeled dynamics: (10); in, , indicating the first i The second prior estimate of the fusion state corresponding to each sample state data and its unmodeled dynamics; , representing the discrete time of the vehicle state observer; , respectively, represent the state transition matrix, control matrix, and state output matrix corresponding to the target mechanism model; I , represents the identity matrix; x ( i- 1), indicating the first i -1 sample state data; u ( i -1), indicating the first i -1 sample control data; , indicating the first i The predictions corresponding to the state data of each sample do not model the dynamics.

[0132] In some implementations, the following formula (11) can be used to determine the first... i The second posterior estimate of the fusion state corresponding to each sample state data and its predicted unmodeled dynamics: (11); in, , indicating the first i The second posterior estimate of the fusion state corresponding to the sample state data and the predicted unmodeled dynamics; , indicating the first i The predicted observation gain corresponding to each sample state data; , representing the state output matrix corresponding to the target mechanism model.

[0133] In some implementations, the following formula (12) can be used to determine the first... i Predicted estimated state corresponding to each sample state data: (12); in, , indicating the first i The predicted estimated state corresponding to each sample data; , indicating the first i The second posterior estimate of the fusion state corresponding to each sample data and its predicted unmodeled dynamics.

[0134] Step S26: Determine the target loss based on the predicted unmodeled dynamics, predicted observation gain, predicted estimated state, sample unmodeled dynamics, and measured state data; then, proceed to step S27. In some implementations, the target loss can be calculated using the loss calculation methods shown in formulas (2) to (5), which will not be elaborated here.

[0135] Step S27: Update the first initial module and the fourth initial module using the target loss to obtain the vehicle state observer.

[0136] In some implementations, the Adam optimizer can be used to update the first initial module and the fourth initial module.

[0137] Below, in conjunction with Figure 3 The following describes an embodiment of vehicle state observation provided by the present invention. Figure 3 As shown, this embodiment includes the following steps S31 to S34: Step S31: Obtain the third state data and the third control data; then, execute step S32. Here, the third state data represents the state data generated under the excitation of the third control data.

[0138] Step S32: Based on the third state data and the third control data, determine the third estimated state and the third unmodeled dynamic using a PINN-based vehicle state observer; then, execute step S33. Here, the PINN-based vehicle state observer can determine the third estimated state and the third unmodeled dynamics through the following steps: First, the third state data and the third control data are input into the first module of the vehicle state observer to output the third unmodeled dynamics; then, the third state data, the third control data, and the third unmodeled dynamics are input into the fourth module of the vehicle state observer to output the observation gain corresponding to the third state data; after that, the third state data, the third control data, the third unmodeled dynamics, and the observation gain are input into the second module of the vehicle state observer to output the third estimated state.

[0139] Step S33: Based on the third estimated state, the third state data, and the third control data, determine the corrected fusion coefficient using the fusion coefficient estimator; then, proceed to step S34. Here, after determining the third estimated state and the third unmodeled dynamics, the third estimated state and the third unmodeled dynamics can be output to the vehicle controller to achieve vehicle control. At the same time, the third estimated state and the third unmodeled dynamics are input to the fusion coefficient estimator (i.e., the third module mentioned above) to achieve fusion coefficient correction.

[0140] In some implementations, the corrected fusion coefficient can be calculated using the following formula (13). : (13); in, , characterizing the i Prior estimates of the corrected fusion coefficients corresponding to +1 state data; and ; , indicating the first i The coefficients corresponding to +1 state data are used to estimate the Kalman gain; , indicating the first i +1 measured state data corresponding to the state data; , representing the estimated state corresponding to the (i+1)th state data.

[0141] As can be seen, in this embodiment, the fusion coefficient is corrected by the state estimation residual.

[0142] In some implementations, the coefficients can be calculated using the following formula (14) to estimate the Kalman gain. : (14); in, , indicating the first i The prior estimate of the covariance matrix corresponding to +1 state data; where... ,in, , indicating the first i Posterior estimate of the coefficient estimation covariance matrix corresponding to each state data; , represents the state covariance matrix in coefficient Kalman filtering; , represents the Jacobian matrix of the observation equation corresponding to the target mechanism model with respect to the fusion coefficients; , where represents the measurement covariance matrix in parametric Kalman filtering.

[0143] In some implementations, the coefficient estimate covariance matrix can be calculated using the following formula (15). Corresponding posterior estimate : (15); in, I , represents the identity matrix.

[0144] Step S34: Update the state transition matrix in the vehicle state observer using the corrected fusion coefficients.

[0145] Here, the state transition matrix in the target mechanism model can be updated using the above formula (1).

[0146] As can be seen from the above, the vehicle state observation method provided in this application, on the one hand, by fusing the generalization of the fusion mechanism model and the nonlinear fitting capability of PINN, combined with PINN compensation for structural modeling errors, can accurately capture the dynamic characteristics of vehicles under extreme conditions; on the other hand, by embedding physical constraints and Kalman gain stability constraints into the training process of the vehicle state observer, problems such as oscillation and instability during the observation process can be avoided, and it can operate stably under extreme conditions for a long time; furthermore, by adopting a boot-strapping two-stage working mode, the fusion coefficient can be adjusted online, and it can adapt to heterogeneous vehicle platforms with different vehicle models and different hardware configurations without retraining the PINN-based vehicle state observer, thus solving the problem of cross-platform prediction failure of pure data-driven observation methods; finally, this method is implemented using a modular system architecture with clear division of labor, and when deployed on the vehicle side, there is no need to carry out large-scale modifications to the existing vehicle hardware, so it can be quickly integrated into mainstream autonomous driving computing platforms, and the deployment cost is low.

[0147] On the other hand, the present invention also provides a vehicle control method. For example... Figure 4 As shown, the method includes the following steps S41 to S42: Step S41: Based on the first state data and the first control data, an unbiased state observer is used to determine the first unmodeled dynamic and the first estimated state corresponding to the first state data; the unbiased state observer is used to perform some or all of the steps in the vehicle state observation method. Step S42: Based on the first unmodeled dynamics and the first estimated state, determine the target control command for the target vehicle using the vehicle controller.

[0148] Here, the first unmodeled dynamic and the first estimated state output by the unbiased state observer are input into the vehicle controller; the vehicle controller uses the first estimated state as the state input and the first unmodeled dynamic as the disturbance term to embed into the optimization objective function, thereby correcting the optimization objective and achieving online compensation for model mismatch.

[0149] In some implementations, the first unmodeled dynamics and the first estimated state are output to the vehicle controller via a Controller Area Network (CAN) bus to meet real-time control requirements.

[0150] This invention verifies the vehicle state observation performance of the invention by building a hardware-in-the-loop test platform. The test platform includes two control hosts, a PXIe-1088 chassis, and a Speedgoat real-time controller. Vehicle models and test scenarios are built using Carsim and Simulink, and unbiased state observers are deployed. The test instance employs a combined longitudinal and lateral motion coupling and conventional extreme conditions on a racetrack, while simultaneously testing the transfer capabilities on heterogeneous vehicle platforms.

[0151] Based on the aforementioned test platform, the state observation accuracy test results are as follows: In the longitudinal-lateral motion coupling scenario, the root mean square error (RMSE) of the longitudinal velocity observation error of this invention is 0.11 km / h, the RMSE of the lateral velocity observation error is 0.43 km / h, and the RMSE of the yaw rate observation error is 0.29 deg / s. Compared with the traditional EKF (RMSE of longitudinal velocity observation error is 0.28 km / h, RMSE of lateral velocity observation error is 1.11 km / h, and RMSE of yaw rate observation error is 0.57 deg / s), these represent improvements of 60.7%, 61.3%, and 49.1%, respectively, indicating a significant improvement in prediction accuracy. Related data comparisons are attached. Figures 5 to 7 As shown. In the appendix Figure 5 In the diagram, solid line 510 represents the longitudinal velocity observation error observed using the unbiased state observer provided by this invention, and dashed line 520 represents the longitudinal velocity observation error observed using a conventional EKF observer. (See appendix...) Figure 6In the diagram, solid line 610 represents the lateral velocity observation error observed using the unbiased state observer provided by this invention, and dashed line 620 represents the lateral velocity observation error observed using a conventional EKF observer. (See appendix...) Figure 7 In the diagram, solid line 710 represents the yaw rate observation error observed using the unbiased state observer provided by this invention, and dashed line 720 represents the yaw rate observation error observed using a conventional EKF observer. (See attached diagram.) Figures 5 to 7 It can be seen that the unbiased state observer provided by this invention has smaller and more accurate observation results for longitudinal velocity, lateral velocity and yaw rate.

[0152] Furthermore, based on the aforementioned test platform, the cross-platform migration test results are as follows: After migrating the unbiased state observer trained using the CarSim Class-Hatchback model to a full-size SUV, no retraining is required. Only through online adjustment of the fusion coefficients, the RMSE of the longitudinal velocity observation error is 0.21 km / h, the RMSE of the lateral velocity observation error is 0.47 km / h, and the RMSE of the yaw rate observation error is 0.39 deg / s. Compared to the observer without the boot-strapping structure provided by this invention (RMSE of 0.23 km / h for longitudinal velocity observation error, 0.66 km / h for lateral velocity observation error, and 0.49 deg / s for yaw rate observation error), these represent improvements of 9.5%, 28.8%, and 20.4%, respectively. It is evident that the migration adaptability of the unbiased state observer provided by this invention is significantly improved. Related data comparisons are attached. Figures 8 to 10 As shown. In the appendix Figure 8 In the diagram, solid line 810 represents the longitudinal velocity observation error observed using the unbiased state observer provided by this invention, and dashed line 820 represents the longitudinal velocity observation error observed using a conventional PINN observer. (See appendix...) Figure 9 In the diagram, solid line 910 represents the lateral velocity observation error observed using the unbiased state observer provided by this invention, and dashed line 920 represents the lateral velocity observation error observed using a conventional PINN observer. (See attached diagram.) Figure 10 In the diagram, solid line 1010 represents the yaw rate observation error observed using the unbiased state observer provided by this invention, and dashed line 1020 represents the yaw rate observation error observed using a conventional PINN observer. (See attached diagram.) Figures 8 to 10 The cross-platform migration test results show that the unbiased state observer provided by this invention has significantly improved migration adaptability compared to the traditional PINN observer.

[0153] Based on the foregoing method embodiments, the present invention also provides a vehicle status monitoring device. For example... Figure 11As shown, the vehicle state observation device 1100 includes: an unmodeled dynamic determination module 1110, a state estimation module 1120, a coefficient correction module 1130, and a coefficient update module 1140; wherein, Unmodeled dynamic determination module 1110 is used to determine a first unmodeled dynamic based on first state data and first control data using a first module; the first module represents an unmodeled dynamic prediction neural network. The state estimation module 1120 is used to determine a first estimated state based on the first state data, the first unmodeled dynamics, and the first control data using a second module; wherein, the second module is used to determine the first estimated state using a target mechanism model and a first fusion coefficient; the first fusion coefficient represents the fusion coefficient between the target mechanism model and the first unmodeled dynamics; The coefficient correction module 1130 is used to determine the second fusion coefficient based on the first state data and the first estimated state using the third module; The coefficient update module 1140 is used to update the first fusion coefficient using the second fusion coefficient.

[0154] In some embodiments, the state estimation module 1120 is used for: Based on the first state data, the first control data, and the first unmodeled dynamics, the fourth module is used to determine the observation gain corresponding to the first state data. Based on the first state data, the first unmodeled dynamics, the first control data, and the observation gain corresponding to the first state data, the second module is used to determine the first estimated state.

[0155] In some embodiments, the state estimation module 1120 is used for: Based on the second state data, the second control data, the first unmodeled dynamics and the target mechanism model, the second module is used to determine the first prior estimate of the fusion state corresponding to the first state data and the first unmodeled dynamics; wherein, the second state data and the second control data represent the state data and control data collected in the previous data acquisition cycle of the first state data and the first control data, respectively. Based on the first prior estimate, the first state data and the observation gain corresponding to the first state data, the second module is used to determine the first posterior estimate of the fusion state corresponding to the first unmodeled dynamics of the first state data. Based on the first posterior estimate and the target mechanism model, the first estimated state is determined using the second module.

[0156] In some implementations, the first fusion coefficient represents the update coefficient corresponding to the state transition matrix of the target mechanism model.

[0157] In some implementations, the third module is a parametric extended Kalman filter.

[0158] In some embodiments, the vehicle state observation device further includes an observer training module; wherein the training module is used for: Acquire sample data; wherein the sample data includes sample state data and sample control data, and each sample state data has corresponding unmodeled dynamic and measured state data of the sample; Based on the sample state data and the sample control data, the first prediction unmodeled dynamic is determined using the first initial module; Based on the sample state data, the sample control data, and the predicted unmodeled dynamics, the prediction observation gain corresponding to the sample state data is determined using the fourth initialization module; Based on the sample state data, the unmodeled dynamics of the prediction, the sample control data, and the sample prediction observation gain, the second module is used to determine the prediction estimation state. The target loss is determined based on the predicted unmodeled dynamics, the predicted observed gain, the predicted estimated state, the sample unmodeled dynamics, and the measured state data. The first initial module and the fourth initial module are updated using the target loss to obtain the first module and the fourth module.

[0159] In some implementations, the observer training module is used for: Based on the predicted unmodeled dynamics and the sample unmodeled dynamics, a first loss is determined; Based on the predicted estimated state and the measured state data, a second loss is determined; Based on the predicted observation gain corresponding to the sample state data, the third loss is determined; The target loss is determined based on the first loss, the second loss, and the third loss.

[0160] On the other hand, the present invention also provides a vehicle. For example... Figure 12 As shown, the vehicle 1200 includes: a memory 1210 and a processor 1220; wherein, The memory 1210 stores a computer program that can run on the processor 1220; when the processor 1220 executes the program, it implements some or all of the steps in the above-described vehicle state observation method embodiments and / or vehicle control method embodiments.

[0161] The above embodiments are merely preferred embodiments provided to fully illustrate the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are all within the scope of protection of the present invention.

Claims

1. A vehicle condition observation method, comprising: Based on the first state data and the first control data, the first unmodeled dynamic is determined using the first module; The first module represents an unmodeled dynamic prediction neural network; Based on the first state data, the first unmodeled dynamics, and the first control data, a first estimated state is determined using a second module; wherein, the second module is used to determine the first estimated state using a target mechanism model and a first fusion coefficient; the first fusion coefficient represents the fusion coefficient between the target mechanism model and the first unmodeled dynamics; Based on the first state data and the first estimated state, the second fusion coefficient is determined using the third module; The first fusion coefficient is updated using the second fusion coefficient.

2. The method according to claim 1, wherein determining the first estimated state based on the first state data, the first unmodeled dynamics, and the first control data using the second module comprises: Based on the first state data, the first control data, and the first unmodeled dynamics, the fourth module is used to determine the observation gain corresponding to the first state data. Based on the first state data, the first unmodeled dynamics, the first control data, and the observation gain corresponding to the first state data, the second module is used to determine the first estimated state.

3. The method according to claim 2, wherein determining the first estimated state using the second module based on the first state data, the first unmodeled dynamics, the first control data, and the observation gain corresponding to the first state data includes: Based on the second state data, the second control data, the first unmodeled dynamics and the target mechanism model, the second module is used to determine the first prior estimate of the fusion state corresponding to the first state data and the first unmodeled dynamics; wherein, the second state data and the second control data represent the state data and control data collected in the previous data acquisition cycle of the first state data and the first control data, respectively. Based on the first prior estimate, the first state data and the observation gain corresponding to the first state data, the second module is used to determine the first posterior estimate of the fusion state corresponding to the first unmodeled dynamics of the first state data. Based on the first posterior estimate and the target mechanism model, the first estimated state is determined using the second module.

4. The method according to claim 3, wherein the first fusion coefficient represents the update coefficient corresponding to the state transition matrix of the target mechanism model.

5. The method according to any one of claims 1 to 4, wherein the third module is a parametric extended Kalman filter.

6. The method according to claim 2, further comprising: Acquire sample data; wherein the sample data includes sample state data and sample control data, and each sample state data has corresponding unmodeled dynamic and measured state data of the sample; Based on the sample state data and the sample control data, the first prediction unmodeled dynamic is determined using the first initial module; Based on the sample state data, the sample control data, and the predicted unmodeled dynamics, the prediction observation gain corresponding to the sample state data is determined using the fourth initialization module; Based on the sample state data, the unmodeled dynamics of the prediction, the sample control data, and the sample prediction observation gain, the second module is used to determine the prediction estimation state. The target loss is determined based on the predicted unmodeled dynamics, the predicted observed gain, the predicted estimated state, the sample unmodeled dynamics, and the measured state data. The first initial module and the fourth initial module are updated using the target loss to obtain the first module and the fourth module.

7. The method according to claim 6, wherein determining the target loss based on the predicted unmodeled dynamics, the predicted observed gain, the predicted estimated state, the sample unmodeled dynamics, and the measured state data comprises: Based on the predicted unmodeled dynamics and the sample unmodeled dynamics, a first loss is determined; Based on the predicted estimated state and the measured state data, a second loss is determined; Based on the predicted observation gain corresponding to the sample state data, the third loss is determined; The target loss is determined based on the first loss, the second loss, and the third loss.

8. A vehicle control method, comprising: Based on the first state data and the first control data, the first unmodeled dynamic and the first estimated state corresponding to the first state data are determined using an unbiased state observer. The unbiased state observer is used to perform the method of any one of claims 1 to 7; Based on the first unmodeled dynamics and the first estimated state, the vehicle controller is used to determine the target control command for the target vehicle.

9. A vehicle condition monitoring device, comprising: The unmodeled dynamic determination module is used to determine the first unmodeled dynamic based on the first state data and the first control data using the first module. The first module represents an unmodeled dynamic prediction neural network; A state estimation module is used to determine a first estimated state based on the first state data, the first unmodeled dynamics, and the first control data using a second module; wherein, the second module is used to determine the first estimated state using a target mechanism model and a first fusion coefficient; the first fusion coefficient represents the fusion coefficient between the target mechanism model and the first unmodeled dynamics; The coefficient correction module is used to determine the second fusion coefficient based on the first state data and the first estimated state using the third module; The coefficient update module is used to update the first fusion coefficient using the second fusion coefficient.

10. A vehicle comprising: Memory and processor; among which, The memory stores a computer program that can run on a processor; when the processor executes the program, it implements the steps of the method according to any one of claims 1 to 8.

Images