WO2022156182A1 - 车辆动力学模型的构建及车辆状态信息预测方法、装置 - Google Patents

车辆动力学模型的构建及车辆状态信息预测方法、装置 Download PDF

Info

Publication number
WO2022156182A1
WO2022156182A1 PCT/CN2021/109534 CN2021109534W WO2022156182A1 WO 2022156182 A1 WO2022156182 A1 WO 2022156182A1 CN 2021109534 W CN2021109534 W CN 2021109534W WO 2022156182 A1 WO2022156182 A1 WO 2022156182A1
Authority
WO
WIPO (PCT)
Prior art keywords
time
state information
sample
vehicle
dynamics model
Prior art date
Application number
PCT/CN2021/109534
Other languages
English (en)
French (fr)
Inventor
肖卓越
徐蒙恩
徐梓哲
梁健楠
周莫凡
付志强
严瑞辉
Original Assignee
魔门塔(苏州)科技有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 魔门塔(苏州)科技有限公司 filed Critical 魔门塔(苏州)科技有限公司
Priority to EP21920563.0A priority Critical patent/EP4216098A1/en
Publication of WO2022156182A1 publication Critical patent/WO2022156182A1/zh
Priority to US18/358,169 priority patent/US20230367934A1/en

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
    • B—PERFORMING OPERATIONS; TRANSPORTING
    • B60—VEHICLES IN GENERAL
    • B60W—CONJOINT CONTROL OF VEHICLE SUB-UNITS OF DIFFERENT TYPE OR DIFFERENT FUNCTION; CONTROL SYSTEMS SPECIALLY ADAPTED FOR HYBRID VEHICLES; ROAD VEHICLE DRIVE CONTROL SYSTEMS FOR PURPOSES NOT RELATED TO THE CONTROL OF A PARTICULAR SUB-UNIT
    • B60W50/00—Details of control systems for road vehicle drive control not related to the control of a particular sub-unit, e.g. process diagnostic or vehicle driver interfaces
    • B—PERFORMING OPERATIONS; TRANSPORTING
    • B60—VEHICLES IN GENERAL
    • B60W—CONJOINT CONTROL OF VEHICLE SUB-UNITS OF DIFFERENT TYPE OR DIFFERENT FUNCTION; CONTROL SYSTEMS SPECIALLY ADAPTED FOR HYBRID VEHICLES; ROAD VEHICLE DRIVE CONTROL SYSTEMS FOR PURPOSES NOT RELATED TO THE CONTROL OF A PARTICULAR SUB-UNIT
    • B60W50/00—Details of control systems for road vehicle drive control not related to the control of a particular sub-unit, e.g. process diagnostic or vehicle driver interfaces
    • B60W50/0097—Predicting future conditions
    • B—PERFORMING OPERATIONS; TRANSPORTING
    • B60—VEHICLES IN GENERAL
    • B60W—CONJOINT CONTROL OF VEHICLE SUB-UNITS OF DIFFERENT TYPE OR DIFFERENT FUNCTION; CONTROL SYSTEMS SPECIALLY ADAPTED FOR HYBRID VEHICLES; ROAD VEHICLE DRIVE CONTROL SYSTEMS FOR PURPOSES NOT RELATED TO THE CONTROL OF A PARTICULAR SUB-UNIT
    • B60W60/00—Drive control systems specially adapted for autonomous road vehicles
    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B13/00—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion
    • G05B13/02—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric
    • G05B13/04—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric involving the use of models or simulators
    • G05B13/048—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric involving the use of models or simulators using a predictor
    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B17/00—Systems involving the use of models or simulators of said systems
    • G05B17/02—Systems involving the use of models or simulators of said systems electric
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/10—Geometric CAD
    • G06F30/15—Vehicle, aircraft or watercraft design
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/004—Artificial life, i.e. computing arrangements simulating life
    • G06N3/006—Artificial life, i.e. computing arrangements simulating life based on simulated virtual individual or collective life forms, e.g. social simulations or particle swarm optimisation [PSO]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • B—PERFORMING OPERATIONS; TRANSPORTING
    • B60—VEHICLES IN GENERAL
    • B60W—CONJOINT CONTROL OF VEHICLE SUB-UNITS OF DIFFERENT TYPE OR DIFFERENT FUNCTION; CONTROL SYSTEMS SPECIALLY ADAPTED FOR HYBRID VEHICLES; ROAD VEHICLE DRIVE CONTROL SYSTEMS FOR PURPOSES NOT RELATED TO THE CONTROL OF A PARTICULAR SUB-UNIT
    • B60W50/00—Details of control systems for road vehicle drive control not related to the control of a particular sub-unit, e.g. process diagnostic or vehicle driver interfaces
    • B60W2050/0001—Details of the control system
    • B60W2050/0019—Control system elements or transfer functions
    • B60W2050/0028—Mathematical models, e.g. for simulation
    • B60W2050/0037—Mathematical models of vehicle sub-units
    • B60W2050/0039—Mathematical models of vehicle sub-units of the propulsion unit
    • B—PERFORMING OPERATIONS; TRANSPORTING
    • B60—VEHICLES IN GENERAL
    • B60W—CONJOINT CONTROL OF VEHICLE SUB-UNITS OF DIFFERENT TYPE OR DIFFERENT FUNCTION; CONTROL SYSTEMS SPECIALLY ADAPTED FOR HYBRID VEHICLES; ROAD VEHICLE DRIVE CONTROL SYSTEMS FOR PURPOSES NOT RELATED TO THE CONTROL OF A PARTICULAR SUB-UNIT
    • B60W2556/00—Input parameters relating to data
    • B60W2556/10—Historical data

Definitions

  • the present invention relates to the technical field of automatic driving, and in particular, to a method and device for constructing a vehicle dynamics model, and a method and device for predicting vehicle state information.
  • autonomous vehicles In the field of autonomous driving, autonomous vehicles generally use pre-built vehicle dynamics models to predict vehicle state information, and then complete autonomous driving. Correspondingly, the accuracy of the prediction results of the pre-built vehicle dynamics model has a greater impact on the safety of autonomous vehicles.
  • the simulation software CarSim is generally used to simulate the vehicle dynamics model of the vehicle.
  • the user needs to have a very detailed understanding of the characteristic parameters and working conditions of each system of the vehicle.
  • the simulation industry in the field of autonomous driving, in most cases, it is impossible to obtain the specific characteristic parameters of each system of the vehicle from automobile manufacturers and parts suppliers. Errors in the true dynamics of the vehicle.
  • the invention provides a vehicle dynamics model construction method and device, and a vehicle state information prediction method and device, so as to realize the construction of a vehicle dynamics model more suitable for the vehicle, and then determine the accurate vehicle state information.
  • a method and device for constructing a vehicle dynamics model, and a method and device for predicting vehicle state information obtained by the embodiments of the present invention obtain sample historical state information corresponding to each sample time of a target vehicle, a sample control parameter sequence, and a label at each sample time Vehicle state information, wherein the sample control parameter sequence corresponding to each sample time includes: the sample time and the control parameters of each time in the first time period ahead; for each sample time, the sample historical state information corresponding to the sample time and Sample control parameter sequence, input the initial vehicle dynamics model, and determine the sample predicted state information corresponding to the sample time; for each sample time, use the sample predicted state information corresponding to the sample time and the label vehicle state information at the sample time, The current loss value corresponding to the initial vehicle dynamics model is determined; based on the current loss value, the model parameters of the
  • the sample historical state information corresponding to each sample time of the target vehicle, the sample control parameter sequence, and the label vehicle state information at each sample time can be used to train the initial vehicle dynamics model, so as to perform the vehicle dynamics model by using the vehicle dynamics model.
  • Supervised learning learn the relationship between the historical state information of each sample and the sample control parameter sequence of the target vehicle, and the state information of the labeled vehicle at the sample moment, to achieve end-to-end modeling of vehicle dynamics, and the modeling process does not require manual labor
  • the data of the training model are collected according to the real situation of the target vehicle.
  • the constructed vehicle dynamics model is more suitable for the vehicle's own characteristics, and the pre-built vehicle dynamics model can determine a more accurate vehicle state. information.
  • vehicle dynamics model for supervised learning, learn the relationship between the historical state information of each sample and the sample control parameter sequence of the target vehicle, and the labeled vehicle state information at the sample time, so as to realize the end-to-end analysis of vehicle dynamics.
  • Modeling the modeling process does not require manual participation, and the data of the training model is collected according to the real situation of the target vehicle. The model can determine more accurate vehicle state information.
  • the dynamic system is a delay system
  • the vehicle state information at the sample time is related to the state information at the time corresponding to the second time period before the sample time, and is related to the time corresponding to the second time period before the sample time and the previous time.
  • the sample control parameter sequence is related. Considering the randomness of the delay value of the dynamic system and the amount of calculation, set the control parameters between the sample time and the time corresponding to the first time duration that is further forward, as the sample control parameter sequence corresponding to the sample time, to ensure that The sample control parameter sequence must include control parameters related to the vehicle state information at the sample moment to ensure the effectiveness of the training model and take into account the computational burden. And the above setting can make the model implicitly learn the delay value of the dynamic system of the vehicle, so that the pre-built vehicle dynamic model obtained by training can better determine the accurate vehicle state information.
  • FIG. 1 is a schematic flowchart of a method for constructing a vehicle dynamics model according to an embodiment of the present invention
  • Fig. 2A is a kind of schematic diagram of the data flow of the state cycle prediction layer
  • 2B is a schematic structural diagram of a vehicle dynamics model
  • FIG. 3 is a schematic flowchart of a method for predicting vehicle state information provided by an embodiment of the present invention
  • FIG. 4 is a schematic structural diagram of an apparatus for constructing a vehicle dynamics model according to an embodiment of the present invention.
  • FIG. 5 is a schematic structural diagram of an apparatus for predicting vehicle state information provided by an embodiment of the present invention.
  • the invention provides a vehicle dynamics model construction method and device, and a vehicle state information prediction method and device, so as to realize the construction of a vehicle dynamics model more suitable for the vehicle, and then determine the accurate vehicle state information.
  • vehicle dynamics model construction method and device and a vehicle state information prediction method and device, so as to realize the construction of a vehicle dynamics model more suitable for the vehicle, and then determine the accurate vehicle state information.
  • FIG. 1 is a schematic flowchart of a method for constructing a vehicle dynamics model according to an embodiment of the present invention. The method may include the following steps:
  • S101 Obtain sample historical state information corresponding to each sample time of the target vehicle, a sample control parameter sequence, and tag vehicle state information at each sample time.
  • the sample control parameter sequence corresponding to each sample time includes: the sample time and the control parameters of each time in the first time period ahead.
  • the method for constructing a vehicle dynamics model provided by the embodiment of the present invention can be applied to any first electronic device with computing capability, and the first electronic device can be a terminal or a server.
  • the functional software for implementing the method may exist in the form of a separate client software, or may exist in the form of a plug-in of the currently related client software, for example, it may exist in the form of a functional module of a dynamic system , this is all possible.
  • the first electronic device may be an in-vehicle device disposed inside the target vehicle, or it may be a non-vehicle device, which may obtain relevant information of the target vehicle.
  • the first electronic device can obtain sample historical state information corresponding to each sample time of the target vehicle, a sample control parameter sequence, and label vehicle state information at each sample time, wherein, the labeled vehicle state information at the sample time is the real vehicle state information of the target vehicle at the sample time.
  • the sample historical state information corresponding to the sample time is: the historical state information corresponding to the time before the sample time.
  • the sample control parameter sequence corresponding to the sample time includes: the sample time and the control parameters of each time in the first time period ahead.
  • the first duration is set according to experience. Each time corresponds to multiple types of control parameters, and the types of control parameters corresponding to each time are the same.
  • the vehicle state information may include, but is not limited to, the speed, acceleration, yaw rate, and attitude angle of the vehicle.
  • the types of control parameters corresponding to each moment include, but are not limited to, brake control quantities and accelerator control quantities.
  • the real dynamic system of the vehicle is a delay system
  • the change of the dynamic system from the vehicle state O(T-N) at the time T-N to the vehicle state O(T) at the time T is taken as the random variable Y; Y and
  • the control parameter sequence of (- ⁇ , T-X) is related to the vehicle state O(T-N) at time T-N. It is irrelevant to the control parameter sequence at each time after time T-N until time T.
  • the control parameter sequence between (T-N- ⁇ , T- ⁇ ), where ⁇ represents the time delay ⁇ of the dynamical system.
  • ⁇ represents the time delay ⁇ of the dynamical system.
  • the value range of the time delay ⁇ of the dynamic system is approximately between 1.5 frames and 2.5 frames, that is, between 30 ms and 50 ms.
  • the time delay ⁇ of the dynamic system is a random variable
  • the time corresponding to the beginning and the end of the control data sequence that can be used to determine the vehicle state at time T is extended to (T-N- ⁇ -A, T), that is, (T-K, T)
  • the control data sequence at time (T-K, T) is used as the input of the vehicle dynamics model for predicting the vehicle state at time T, so that the vehicle dynamics model can implicitly learn the random variable ⁇ , so that the subsequent
  • T, N, ⁇ , A, and K are all positive numbers, and K is greater than N.
  • N represents the first duration
  • K represents the second duration.
  • the sample historical state information corresponding to the sample time is the vehicle state information of the target vehicle at a time corresponding to a second time period before the sample time, and the second time period is smaller than the first time period.
  • S102 For each sample time, input the sample historical state information and the sample control parameter sequence corresponding to the sample time into the initial vehicle dynamics model, and determine the sample predicted state information corresponding to the sample time.
  • the first electronic device can use the single frame of data to train the initial vehicle dynamics model. , that is, for each sample time, the initial vehicle dynamics model is trained by using the sample historical state information and sample control parameter sequence corresponding to the sample time and the label vehicle state information at the sample time.
  • the first electronic device inputs the sample historical state information and sample control parameter sequence corresponding to the sample time into the initial vehicle dynamics model, and determines the sample predicted state information corresponding to the sample time.
  • the initial vehicle dynamics model is a deep learning-based recurrent neural network model, which may include a feature encoding layer, a state cycle prediction layer, and a feature decoding layer.
  • the feature encoding layer and the feature decoding layer can be implemented by a fully connected layer
  • the state cycle prediction layer can be implemented by a GRU (Gated Recurrent Unit, gated recurrent unit).
  • the first electronic device uses the sample predicted state information corresponding to the sample time and the labeled vehicle state information at the sample time to calculate the sample predicted state information and the labeled vehicle state information corresponding to the sample time.
  • the calculated distance is determined as the current loss value corresponding to the initial vehicle dynamics model.
  • the first electronic device uses the sample predicted state information corresponding to the sample time and the labeled vehicle state information at the sample time to calculate the sample predicted state information and the labeled vehicle state corresponding to the sample time The distance between the information; the average or sum of the distances between the sample predicted state information corresponding to the preset number of sample moments and the labeled vehicle state information, as the current loss value corresponding to the initial vehicle dynamics model.
  • the first electronic device may first determine whether the current loss value is greater than the preset loss threshold value, and if the current loss value is determined to be greater than the preset loss threshold value, it is determined that the initial vehicle dynamics model has not converged, and the preset optimization algorithm is used, Adjust the model parameters of the initial vehicle dynamics model, and return to re-execution for each sample moment, input the sample historical state information and sample control parameter sequence corresponding to the sample moment, input the initial vehicle dynamics model after adjusting the parameters, and determine the The sample prediction state information corresponding to the sample time, and then based on the sample prediction state information corresponding to the sample time and the label vehicle state information at the sample time, the loss value corresponding to the initial vehicle dynamics model is determined as the current loss value, and the current loss is re-judged Whether the value is greater than the preset loss threshold, if it is judged that the current loss value is greater than the preset loss threshold, continue to use the preset optimization algorithm to adjust the model parameters of the initial vehicle dynamics model after adjusting the parameters, and
  • the preset optimization algorithm may be a gradient descent method least square method, a genetic algorithm, or the like.
  • the current loss value it can be calculated by the LOSS function, for example, it can be calculated by the L2LOSS function.
  • the sample historical state information corresponding to each sample time of the target vehicle, the sample control parameter sequence, and the label vehicle state information at each sample time can be used to train the initial vehicle dynamics model, so as to use the vehicle dynamics model.
  • Supervised learning learn the relationship between the historical state information of each sample and the sample control parameter sequence of the target vehicle, and the state information of the labeled vehicle at the sample moment, to achieve end-to-end modeling of vehicle dynamics, and the modeling process does not require manual labor
  • the data of the training model are collected according to the real situation of the target vehicle.
  • the constructed vehicle dynamics model is more suitable for the vehicle's own characteristics, and the pre-built vehicle dynamics model can determine a more accurate vehicle state. information.
  • the S102 may include the following steps 011-013:
  • the historical state information corresponding to the sample time is the state information of the time corresponding to the second time period before the sample time, which is a single frame of data.
  • the first electronic device For each sample time, the sample historical state information corresponding to the sample time is input into the feature coding layer of the initial vehicle dynamics model to obtain the hidden vector corresponding to the sample historical state information corresponding to the sample time, and the sample corresponding to the sample time.
  • the hidden vector corresponding to the historical state information is used as the initial state vector and the sample control parameter sequence corresponding to the sample moment, and is input into the state cycle prediction layer of the initial vehicle dynamics model, so that the state cycle prediction layer is based on the initial state vector and the sample.
  • the sample control parameter sequence corresponding to the time, and the implicit vector corresponding to the vehicle state information at each time after the time corresponding to the initial state vector is determined cyclically in turn, until the implicit vector corresponding to the vehicle state information at the sample time is determined.
  • the first electronic device inputs the latent vector corresponding to the vehicle state information corresponding to the sample moment into the feature decoding layer of the initial vehicle dynamics model for decoding, and decodes to obtain the sample predicted state corresponding to the sample moment information, that is, the predicted vehicle state information corresponding to the sample moment.
  • the feature encoding layer of the initial vehicle dynamics model can be implemented by a 32-dimensional fully connected layer, and the state cycle prediction layer of the initial vehicle dynamics model can be implemented by a GRU gated recurrent unit.
  • the feature decoding layer of the dynamic model can be implemented by a 16-dimensional fully connected layer and a 3-dimensional fully connected layer.
  • FIG. 2A it is a schematic diagram of a data flow of the state cycle prediction layer, wherein z t and r t in FIG. 2A represent the value corresponding to the update gate and the value corresponding to the reset gate, respectively.
  • the update gate is used to control the degree to which the vehicle state information at time t-1 is brought into the vehicle state information at time t. The larger the value of the update gate, the more vehicle state information at time t-1 is brought in.
  • the reset gate controls how much of the vehicle state information at time t-1 is written into the vehicle state information at time t, and the smaller the reset gate, the less vehicle state information at time t-1 is written.
  • each processing process in the state cycle prediction layer can be expressed by the following formula:
  • W z , W r and W are the parameters of the state cycle prediction layer, obtained through training; h t-1 represents the vehicle state information at time t-1.
  • the initial value of h t-1 corresponds to the sample time
  • the initial value of h t-1 is the historical state information corresponding to the current time
  • h t represents the vehicle state information at time t
  • h t is the corresponding sample time
  • h t is the vehicle state information corresponding to the current moment
  • x t is the sample control parameter sequence corresponding to the sample moment
  • x t is the current control parameter sequence corresponding to the current moment.
  • [] means that two vectors are connected
  • * means the product of matrices.
  • the method may further include a process of testing the pre-established vehicle dynamics model to determine the pros and cons of the vehicle state determined by the pre-established vehicle dynamics model.
  • the method may further include the following steps 021-024:
  • the original test data includes: test history state information, test control parameter sequence and test vehicle state information corresponding to each test time generated during the driving of the target vehicle, and the test control parameter sequence corresponding to each test time includes: test time and For the control parameters at each time in the first time period, the test history state information corresponding to the test time is: vehicle state information at the time corresponding to the second time period before the test time.
  • the first test time includes: the earliest test time and each time before the time corresponding to the second duration after that.
  • the second test time is: the time in the test time except the first test time
  • the prediction history state information corresponding to the second test time is: the test prediction state information corresponding to the time corresponding to the second time length before the second test time .
  • the first electronic device may first obtain original test data of the target vehicle, where the original test data may include: The generated test history status information, test control parameter sequence, and test vehicle status information corresponding to each moment.
  • the moment corresponding to the data used in the test process is called the test moment, wherein the test corresponding to each test moment
  • the vehicle state information is the real vehicle state information of the target vehicle.
  • the first electronic device sequentially inputs the test history state information and the test control parameter sequence corresponding to the first test time into the pre-built vehicle dynamics model in order to pass the pre-built vehicle dynamics model.
  • the feature encoding layer, the state cycle prediction layer and the feature decoding layer of the model determine the test prediction state information corresponding to each first test time.
  • the vehicle state corresponding to time T-N needs to be used, that is, "O(T-N)" shown in Fig. 2B, and the sequence of control parameters from time T-K to time T is shown in Fig. "S(T)" shown in 2B, where N represents the first duration, K represents the second duration, and N is less than K.
  • time T is the earliest time in the test time
  • time T to T+N-1 is the first test time.
  • the test history state information and test control parameter sequence corresponding to the first test moment when determining the corresponding vehicle state information, that is, the corresponding test prediction state information, it is necessary to use In the original test data, the test history state information and test control parameter sequence corresponding to the first test moment.
  • O(T-N) and S(T) are used as the feature encoding layer of the pre-established vehicle dynamics model, that is, the 32-dimensional fully connected layer "FC(32)Initial” States” to encode to obtain the corresponding hidden vector;
  • the hidden vector is used as the initial state vector and the S(T) input state loop prediction layer, that is, the GRU gated loop unit, as shown in Figure 2B "GRU".
  • the initial state vector and S(T) that is, the sequence of control parameters from time T-K to time T, cyclically determine the hidden vector corresponding to the vehicle state information at each time after the time corresponding to the initial state vector, until the time corresponding to time T is determined.
  • the hidden vector corresponding to the vehicle state corresponding to the T moment is input into the feature decoding layer, that is, a 16-dimensional fully connected layer as shown in “FC(16)” in 2B and a 3-dimensional fully connected layer as shown in “FC(16)” in 2B FC(3)", decode the hidden vector corresponding to the vehicle state corresponding to time T, and obtain the vehicle state corresponding to time T, such as O(T) in Figure 2.
  • the vehicle state corresponding to time T includes: the state information of the test vehicle corresponding to time T in the original test data, and includes the test prediction state information output by the pre-built vehicle dynamics model.
  • test prediction state information corresponding to time T+N In order to test the accuracy of the prediction and determination results of the pre-built vehicle dynamics model, when determining the test prediction state information corresponding to time T+N, it is necessary to use the test prediction state information corresponding to time T as the test history corresponding to time T+N State information, correspondingly, input the test prediction state information corresponding to time T and the test control parameter sequence corresponding to time T+N, that is, the control parameter sequence from time T+N-K to time T+N in the original test data, into the pre-built
  • the vehicle dynamics model determines the test prediction state information corresponding to the second test moment.
  • the test prediction state information at each time after time T+N is calculated sequentially.
  • the time T+N and the subsequent test times are the above-mentioned second test times.
  • the first electronic device after determining the test prediction state information corresponding to each first test time, sequentially records the prediction history state information corresponding to the second test time and the test control parameter sequence according to the chronological order of each second test time. , and input the pre-built vehicle dynamics model to determine the test prediction state information corresponding to each second test time through the feature encoding layer, state cycle prediction layer and feature decoding layer of the pre-built vehicle dynamics model.
  • the prediction history state information corresponding to the second test time is the test prediction state information corresponding to the time corresponding to the second time period before the second test time. For example, when the second test time is T+N time, the predicted historical state information corresponding to the second test time is the test predicted state information corresponding to the T time output by the pre-built vehicle dynamics model.
  • test prediction state information corresponding to each test time After determining the test prediction state information corresponding to each test time, use the test prediction state information and test vehicle state information corresponding to each test time to calculate the distance between the test prediction state information and the test vehicle state information corresponding to each test time; The distance between the test prediction state information corresponding to each test time and the test vehicle state information determines the test result of the pre-built vehicle dynamics model.
  • the process of determining the test result of the pre-built vehicle dynamics model may be: counting the distance between the test predicted state information corresponding to the test moment and the test vehicle state information, the number of distances that do not exceed the preset distance threshold, If the ratio of the number of distances that do not exceed the preset distance threshold to the total number of distances exceeds the preset ratio, the test result of the pre-built vehicle dynamics model includes information indicating that the pre-built vehicle dynamics model has passed the test; If the ratio of the number of distances exceeding the preset distance threshold to the total number of distances does not exceed the preset ratio, the test result of the pre-built vehicle dynamics model includes information indicating that the pre-built vehicle dynamics model fails the test. etc., the embodiments of the present invention do not limit the specific process of determining the test result of the pre-built vehicle dynamics model.
  • the process of determining the vehicle state information of the target vehicle is characterized by the pre-built vehicle dynamics model; Retraining is required to obtain a pre-built vehicle dynamics model.
  • the data for training and testing the vehicle dynamics model includes the control parameters and vehicle state information recorded during the driving of the target vehicle, and the data for training and testing the vehicle dynamics model does not need to be manually marked. Save labor costs to a certain extent.
  • the vehicle dynamics model does not need to maintain any internal state.
  • the input vehicle state information is the vehicle state information of a certain frame in history, that is, a certain moment.
  • An independent sample corresponding to each frame can increase the number of sample data in the process of training and testing the vehicle dynamics model to a certain extent, and reduce the computational difficulty of the training process and subsequent testing and actual prediction and determination processes .
  • the model can start to predict at any time, and does not need to be warmed up in advance, that is, the process of predicting and determining the vehicle state information starts from the stationary state of the target vehicle.
  • the embodiments of the present invention provide a vehicle state information prediction method, which relies on the pre-established vehicle dynamics model constructed in the above method embodiments.
  • the method may include the following steps S301-S302:
  • S301 Obtain historical state information and a current control parameter sequence corresponding to the target vehicle at the current moment.
  • the current control parameter sequence includes: the current time and the control parameters at each time in the first time period ahead.
  • the vehicle state information prediction method provided in the embodiment of the present invention can be applied to any second electronic device with computing capability, and the second electronic device can be a terminal or a server.
  • the functional software for implementing the method may exist in the form of a separate client software, or may exist in the form of a plug-in of the currently related client software, for example, it may exist in the form of a functional module of a dynamic system , this is all possible.
  • the second electronic device may be the same physical device as the above-mentioned first electronic device, or may be a different physical device.
  • the second electronic device may be an in-vehicle device, which is provided inside the target vehicle, or may be an off-board device, which can obtain relevant information of the target vehicle.
  • the second electronic device can obtain the historical state information and the current control parameter sequence corresponding to the target vehicle at the current time, wherein the historical state information corresponding to the current time is: before the current time
  • the historical state information corresponding to the time is: before the current time
  • the current control parameter sequence corresponding to the current time includes: the current time and the control parameters at each time in the first time period ahead.
  • the first duration is set according to experience. Each time corresponds to multiple types of control parameters, and the types of control parameters corresponding to each time are the same.
  • the vehicle state information may include, but is not limited to, the speed, acceleration, yaw rate, and attitude angle of the vehicle.
  • the types of control parameters corresponding to each moment include, but are not limited to, brake control quantities and accelerator control quantities.
  • the historical state information is the vehicle state information of the target vehicle at the time corresponding to the second time period before the current time, and the second time period less than the first duration.
  • S302 Input the historical state information and the current control parameter sequence into a pre-built vehicle dynamics model to determine the vehicle state information of the target vehicle at the current moment.
  • the pre-built vehicle dynamics model is a recurrent neural network model trained based on the sample state information corresponding to each historical moment of the target vehicle and its corresponding sample control parameter sequence. It may include a feature encoding layer, a state loop prediction layer, and a feature decoding layer.
  • the feature encoding layer and the feature decoding layer can be implemented by a fully connected layer
  • the state cycle prediction layer can be implemented by a GRU (Gated Recurrent Unit, gated recurrent unit).
  • the second electronic device inputs the historical state information and the current control parameter sequence into the pre-built vehicle dynamics model, and the pre-built vehicle dynamics model, through its feature encoding layer, state cycle prediction layer, and feature decoding layer, is used for historical state information and
  • the current control parameter sequence is processed to determine the vehicle state information of the target vehicle at the current moment. Then, the vehicle state information of the target vehicle at the current time is output, so as to control the running state of the target vehicle by the vehicle state information of the target vehicle at the current time.
  • the cyclic neural network model of the relationship between the historical state information of each sample and the sample control parameter sequence of the target vehicle and the state information of the labeled vehicle at the sample moment is obtained by learning, that is, the pre-built vehicle dynamics model, By determining the vehicle status information of the target vehicle at the current moment, the vehicle status information with higher accuracy and more suitable for the performance of the target vehicle can be obtained, thereby ensuring the safety of the vehicle during driving.
  • the S302 may include the following steps 031-033:
  • the historical state information corresponding to the current time is the state information of the time corresponding to the second time period before the current time, which is a single frame of data.
  • the first The second electronic device inputs the historical state information corresponding to the current moment into the feature coding layer of the pre-established vehicle dynamics model, obtains the hidden vector corresponding to the historical state information corresponding to the current moment, and converts the corresponding historical state information to the current moment.
  • the hidden vector is used as the initial state vector and the current control parameter sequence to input the state cycle prediction layer of the pre-established vehicle dynamics model, so that the state cycle prediction layer is based on the initial state vector and the current control parameter sequence.
  • the latent vector corresponding to the vehicle state information at each time after the time corresponding to the state vector of , until the latent vector corresponding to the vehicle state information at the current time is determined.
  • the second electronic device inputs the latent vector corresponding to the vehicle state information corresponding to the current moment into the feature decoding layer of the pre-established vehicle dynamics model for decoding, and decodes to obtain the vehicle state information corresponding to the current moment.
  • the feature encoding layer of the pre-established vehicle dynamics model can be implemented by a 32-dimensional fully connected layer
  • the state cycle prediction layer of the pre-established vehicle dynamics model can be implemented by a GRU gated recurrent unit
  • the feature decoding layer of the pre-established vehicle dynamics model can be implemented by a 16-dimensional fully connected layer and a 3-dimensional fully connected layer.
  • the method may further include the following steps: 041-042:
  • the target control parameter sequence further includes: the current time and the control parameters at each time in the time before the time corresponding to the first forward time period.
  • the second electronic device may input the vehicle status information at the current moment into the preset control parameter determination model, so that the preset control parameter determination model is based on the current moment
  • the vehicle state information of the target vehicle is determined to determine the control parameters corresponding to the target vehicle at the next moment from the current moment.
  • the second electronic device obtains the control parameter corresponding to the next moment of the current moment; the control parameter corresponding to the next moment of the current moment is added to the target control parameter sequence as the control parameter corresponding to the next moment of the current moment. sequence.
  • the target control parameter sequence including the current control parameters and the historical state information corresponding to the next moment at the current moment are input into the pre-built vehicle dynamics model to determine the vehicle status information of the target vehicle at the current moment and the next moment.
  • the preset control parameter determination model may use any determination algorithm in the related art that can determine vehicle control parameters, and determine the control parameters corresponding to the next moment at the current moment based on the vehicle state information at the current moment, which is used to control the target vehicle of driving.
  • the embodiments of the present invention provide an apparatus for constructing a vehicle dynamics model.
  • the apparatus may include:
  • the first obtaining module 410 is configured to obtain sample historical state information corresponding to each sample time of the target vehicle, a sample control parameter sequence and label vehicle state information at each sample time, wherein the sample control parameter sequence corresponding to each sample time includes: The sample time and the control parameters of each time in the first time period ahead;
  • the first determination module 420 is configured to, for each sample time, input the sample historical state information and sample control parameter sequence corresponding to the sample time into the initial vehicle dynamics model, and determine the sample predicted state information corresponding to the sample time;
  • the second determination module 430 is configured to, for each sample time, use the sample predicted state information corresponding to the sample time and the labeled vehicle state information at the sample time to determine the current loss value corresponding to the initial vehicle dynamics model;
  • the adjustment module 440 is configured to adjust the model parameters of the initial vehicle dynamics model based on the current loss value until the initial vehicle dynamics model reaches a preset convergence state to obtain a pre-established vehicle dynamics model.
  • the sample historical state information corresponding to each sample time of the target vehicle, the sample control parameter sequence, and the label vehicle state information at each sample time can be used to train the initial vehicle dynamics model, so as to use the vehicle dynamics model.
  • Supervised learning learn the relationship between the historical state information of each sample and the sample control parameter sequence of the target vehicle, and the state information of the labeled vehicle at the sample moment, to achieve end-to-end modeling of vehicle dynamics, and the modeling process does not require manual labor
  • the data of the training model are collected according to the real situation of the target vehicle.
  • the constructed vehicle dynamics model is more suitable for the vehicle's own characteristics, and the pre-built vehicle dynamics model can determine a more accurate vehicle state. information.
  • the sample historical state information corresponding to the sample time is state information of the target vehicle at a time corresponding to a second time period before the sample time, and the second time period is shorter than the first time.
  • the first determination module 420 is specifically configured to, for each sample moment, input the sample historical state information corresponding to the sample moment into the feature encoding layer of the initial vehicle dynamics model , obtain the hidden vector corresponding to the historical state information of the sample corresponding to the sample moment;
  • the latent vector corresponding to the sample historical state information corresponding to the sample moment and the sample control parameter sequence corresponding to the sample moment are input into the state cycle prediction layer of the initial vehicle dynamics model to obtain the sample moment.
  • the latent vector corresponding to the vehicle state information corresponding to the sample time is input into the feature decoding layer of the initial vehicle dynamics model, and the sample prediction state information corresponding to the sample time is determined.
  • the apparatus further includes:
  • the second obtaining module (not shown in the figure) is configured to obtain the original test data of the target vehicle after the determining that the initial vehicle dynamics model converges to obtain the pre-built vehicle dynamics model, wherein,
  • the original test data includes: the test history state information, the test control parameter sequence and the test vehicle state information corresponding to each test time generated during the running of the target vehicle, and the test control parameter sequence corresponding to each test time includes: The test time and the control parameters at each time in the first time duration, the test history state information corresponding to the test time is: the vehicle state information at the time corresponding to the second time duration before the test time;
  • the fourth determination module (not shown in the figure) is configured to input the test history state information and test control parameter sequence corresponding to the first test time into the pre-built vehicle dynamics model, and determine the corresponding test time of the first test time.
  • the test prediction state information of , wherein, the first test time includes: the earliest test time and each time before the time corresponding to the second time length after the earliest test time;
  • the fifth determination module (not shown in the figure) is configured to input the predicted historical state information and the test control parameter sequence corresponding to the second test moment into the pre-built vehicle dynamics model to determine the second test moment Corresponding test prediction state information, wherein the second test time is: a time in the test time other than the first test time, and the prediction history state information corresponding to the second test time is: the first test time 2. Test prediction state information corresponding to the time corresponding to the first second duration of the test time;
  • the sixth determination module (not shown in the figure) is configured to determine the test result of the pre-built vehicle dynamics model by using the test prediction state information and the test vehicle state information corresponding to each test time.
  • an embodiment of the present invention provides a vehicle state information prediction device.
  • the device may include:
  • the third obtaining module 510 is configured to obtain the historical state information and the current control parameter sequence corresponding to the target vehicle at the current time, wherein the current control parameter sequence includes: the current time and each time in the first forward period of time. control parameter;
  • the seventh determination module 520 is configured to input the historical state information and the current control parameter sequence into a pre-built vehicle dynamics model to determine the vehicle state information of the target vehicle at the current moment, wherein the pre-built vehicle dynamics model
  • the vehicle dynamics model is: a recurrent neural network model obtained by training based on the sample state information corresponding to each historical moment of the target vehicle and its corresponding sample control parameter sequence.
  • the cyclic neural network model of the relationship between the historical state information of each sample and the sample control parameter sequence of the target vehicle and the state information of the labeled vehicle at the sample moment is obtained by learning, that is, the pre-built vehicle dynamics model, By determining the vehicle status information of the target vehicle at the current moment, the vehicle status information with higher accuracy and more suitable for the performance of the target vehicle can be obtained, thereby ensuring the safety of the vehicle during driving.
  • the historical state information is vehicle state information of the target vehicle at a time corresponding to a second time period before the current time, and the second time period is smaller than the first time period.
  • the seventh determination module 520 is specifically configured to input the historical state information into the feature encoding layer of the pre-built vehicle dynamics model, and obtain the corresponding historical state information the hidden vector;
  • the latent vector corresponding to the vehicle state information corresponding to the current moment is input into the feature decoding layer of the pre-built vehicle dynamics model to obtain the vehicle state information corresponding to the current moment.
  • the apparatus further includes:
  • a fourth obtaining module (not shown in the figure), configured to obtain the current control parameters determined by the preset control parameter determination model based on the vehicle state information at the current moment;
  • the eighth determination module (not shown in the figure) is configured to input a target control parameter sequence including the current control parameter and historical state information corresponding to the next moment of the current moment into the pre-built vehicle power
  • a learning model is used to determine the vehicle state information of the target vehicle at the next moment of the current moment, wherein the target control parameter sequence further includes: the current moment and the moment before the moment corresponding to the first forward time period. control parameters at each time.
  • the modules in the apparatus in the embodiment may be distributed in the apparatus in the embodiment according to the description of the embodiment, and may also be located in one or more apparatuses different from this embodiment with corresponding changes.
  • the modules in the foregoing embodiments may be combined into one module, or may be further split into multiple sub-modules.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • General Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • Software Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Data Mining & Analysis (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Mathematical Physics (AREA)
  • Biomedical Technology (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Automation & Control Theory (AREA)
  • Geometry (AREA)
  • Human Computer Interaction (AREA)
  • Mechanical Engineering (AREA)
  • Transportation (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Medical Informatics (AREA)
  • Computer Hardware Design (AREA)
  • Computational Mathematics (AREA)
  • Mathematical Analysis (AREA)
  • Mathematical Optimization (AREA)
  • Pure & Applied Mathematics (AREA)
  • Aviation & Aerospace Engineering (AREA)
  • Feedback Control In General (AREA)
  • Traffic Control Systems (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

本发明实施例公开一种车辆动力学模型的构建方法、装置及车辆状态信息预测方法、装置，该车辆动力学模型的构建方法包括: 获得目标车辆的各样本时刻对应的样本历史状态信息、样本控制参数序列及标签车辆状态信息; 针对每一样本时刻，将样本历史状态信息及样本控制参数序列，输入初始车辆动力学模型，确定样本预测状态信息; 利用样本预测状态信息及标签车辆状态信息，确定当前损失值; 基于当前损失值，调整初始车辆动力学模型的模型参数，直至初始车辆动力学模型达到预设收敛状态，得到预先建立的车辆动力学模型，以实现构建出更加贴合车辆的车辆动力学模型。

Description

车辆动力学模型的构建及车辆状态信息预测方法、装置 技术领域
本发明涉及自动驾驶技术领域，具体而言，涉及一种车辆动力学模型的构建方法、装置及车辆状态信息预测方法、装置。
背景技术
在自动驾驶领域中，自动驾驶车辆一般利用预先构建的车辆动力学模型预测得到车辆状态信息，进而完成自动驾驶。相应的，预先构建的车辆动力学模型的预测结果的准确性对自动驾驶车辆的安全性影响较大。
目前，一般使用仿真软件CarSim，仿真构建车辆的车辆动力学模型。在利用仿真软件CarSim针对车辆构建其车辆动力学模型的过程中，用户需要对车辆的各系统的特性参数以及工况有非常详细的了解。对于自动驾驶领域中仿真行业来说，多数情况下，无法从汽车制造商以及零部件供应商得到具体的车辆各系统的特性参数，只能通过不断通过启发式的参数调整，来不断减少仿真与车辆真实动力学的误差。
除此之外，即使在构建车辆的车辆动力学模型时获得该车辆的各系统的特性参数，由于车辆的制作工艺以及零件损耗等因素，利用仿真软件CarSim所构建的基于数学模型推衍的车辆动力学模型，与真实的车辆也会存在一定的误差。
发明内容
本发明提供了一种车辆动力学模型的构建方法、装置及车辆状态信息预测方法、装置，以实现构建出更加贴合车辆的车辆动力学模型，进而确定出准确的车辆状态信息。本发明实施例提供的一种车辆动力学模型的构建方法、装置及车辆状态信息预测方法、装置，获得目标车辆的各样本时刻对应的样本历史状态信息、样本控制参数序列以及各样本时刻的标签车辆状态信息，其中，各样本时刻对应的样本控制参数序列包括：该样本时刻及向前第一时长内各时刻的控制参数；针对每一样本时刻，将该样本时刻对应的样本历史状态信息及样本控制参数序列，输入初始车辆动力学模型，确定出该样本时刻对应的样本预测状态信息；针对每一样本时刻，利用该样本时刻对应的样本预测状态信息以及该样本时刻的标签车辆状态信息，确定初始车辆动力学模型对应的当前损失值；基于当前损失值，调整初始车辆动力学模型的模型参数，直至初始车辆动力学模型达到 预设收敛状态，得到预先建立的车辆动力学模型。
应用本发明实施例，可以利用目标车辆的各样本时刻对应的样本历史状态信息、样本控制参数序列以及各样本时刻的标签车辆状态信息，训练初始车辆动力学模型，以通过利用车辆动力学模型进行监督学习，学习得到目标车辆的各样本历史状态信息和样本控制参数序列，与样本时刻的标签车辆状态信息之间的关系，实现端到端的对车辆动力学进行建模，建模过程不需要人工参与，并且，训练模型的数据均为根据目标车辆的真实情况采集得到，所构建的车辆动力学模型更加贴合车辆自身特性，进而通过预先构建的车辆动力学模型可以确定出更加准确的车辆状态信息。当然，实施本发明的任一产品或方法并不一定需要同时达到以上所述的所有优点。
本发明实施例的创新点包括：
1、通过利用车辆动力学模型进行监督学习，学习得到目标车辆的各样本历史状态信息和样本控制参数序列，与样本时刻的标签车辆状态信息之间的关系，实现端到端的对车辆动力学进行建模，建模过程不需要人工参与，并且，训练模型的数据均为根据目标车辆的真实情况采集得到，所构建的车辆动力学模型更加贴合车辆自身特性，进而通过预先构建的车辆动力学模型可以确定出更加准确的车辆状态信息。
2、考虑到动力学系统为延时系统，即样本时刻的车辆状态信息与样本时刻的前第二时长所对应时刻的状态信息有关，且与样本时刻的前第二时长所对应时刻及之前时刻的样本控制参数序列有关。考虑到动力学系统的延时取值的随机性以及计算量，设置样本时刻及其更向前的第一时长所对应时刻之间的控制参数，作为样本时刻对应的样本控制参数序列，以保证该样本控制参数序列中必包含与样本时刻的车辆状态信息相关的控制参数，以保证训练模型的有效性且考虑到计算负担。并且上述设置可以使得模型隐式的学习车辆的动力学系统的延时取值，以使得训练得到的预先构建的车辆动力学模型可以更好的确定出准确的车辆状态信息。
3、利用学习得到目标车辆的各样本历史状态信息和样本控制参数序列，与样本时刻的标签车辆状态信息之间的关系的循环神经网络模型，即预先构建的车辆动力学模型，确定目标车辆在当前时刻的车辆状态信息，可以得到准确性更高且更贴合目标车辆自身性能的车辆状态信息，进而保证车辆行驶过程中的安全性。
附图说明
为了更清楚地说明本发明实施例或现有技术中的技术方案，下面将对实施例或现有技术描述中所需要使用的附图作简单介绍。显而易见地，下面描述中的附图仅仅是本发明的一些实施例。对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可 以根据这些附图获得其他的附图。
图1为本发明实施例提供的车辆动力学模型的构建方法的一种流程示意图；
图2A为状态循环预测层的数据流动的一种示意图；
图2B为一种车辆动力学模型的结构示意图；
图3为本发明实施例提供的车辆状态信息预测方法的一种流程示意图；
图4为本发明实施例提供的车辆动力学模型的构建装置的一种结构示意图；
图5为本发明实施例提供的车辆状态信息预测装置的一种结构示意图。
具体实施方式
下面将结合本发明实施例中的附图，对本发明实施例中的技术方案进行清楚、完整的描述。显然，所描述的实施例仅仅是本发明的一部分实施例，而不是全部的实施例。基于本发明中的实施例，本领域普通技术人员在没有付出创造性劳动前提下所获得的所有其他实施例，都属于本发明保护的范围。
需要说明的是，本发明实施例及附图中的术语“包括”和“具有”以及它们的任何变形，意图在于覆盖不排他的包含。例如包含的一系列步骤或单元的过程、方法、系统、产品或设备没有限定于已列出的步骤或单元，而是可选地还包括没有列出的步骤或单元，或可选地还包括对于这些过程、方法、产品或设备固有的其他步骤或单元。
本发明提供了一种车辆动力学模型的构建方法、装置及车辆状态信息预测方法、装置，以实现构建出更加贴合车辆的车辆动力学模型，进而确定出准确的车辆状态信息。下面对本发明实施例进行详细说明。
图1为本发明实施例提供的车辆动力学模型的构建方法的一种流程示意图。该方法可以包括如下步骤：
S101：获得目标车辆的各样本时刻对应的样本历史状态信息、样本控制参数序列以及各样本时刻的标签车辆状态信息。
其中，各样本时刻对应的样本控制参数序列包括：该样本时刻及向前第一时长内各时刻的控制参数。
本发明实施例所提供的车辆动力学模型的构建方法，可以应用于任一具有计算能力的第一电子设备，该第一电子设备可以为终端或者服务器。在一种实现中，实现该方法的功能软件可以以单独的客户端软件的形式存在，也可以以目前相关的客户端软件的插件的形式存在，例如可以以动力学系统的功能模块的形式存在，这都是可以的。
在一种情况中，该第一电子设备可以为车载设备，设置于目标车辆的内部，也可以为非车载设备，可以获得目标车辆的相关信息，这都是可以的。
在存在针对目标车辆构建其车辆动力学模型的需求的情况下，第一电子设备可以获得目标车辆的各样本时刻对应的样本历史状态信息、样本控制参数序列以及各样本时刻的标签车辆状态信息，其中，该样本时刻的标签车辆状态信息为目标车辆在样本时刻时的真实的车辆状态信息。样本时刻对应的样本历史状态信息为：样本时刻之前的时刻对应的历史状态信息。样本时刻对应的样本控制参数序列包括：该样本时刻及向前第一时长内各时刻的控制参数。该第一时长为根据经验设置的。每一时刻对应多种类型的控制参数，各时刻对应的控制参数的类型相同。
在一种情况中，车辆状态信息可以包括但不限于：车辆的速度、加速度、横摆角速度以及姿态角等。各时刻对应的控制参数的类型包括但不限于：刹车控制量以及油门控制量等。
考虑到车辆的真实的动力学系统为延时系统，假设将动力学系统从T-N时刻的车辆状态O(T-N)到T时刻的车辆状态O(T)的变化量，作为随机变量Y；Y与(-∞，T-X)的控制参数序列、T-N时刻的车辆状态O(T-N)相关。与T-N时刻的之后，直至T时刻之间的各时刻的控制参数序列不相关。相应的，若需要预测T-N时刻的车辆状态O(T-N)到T时刻的车辆状态O(T)的变化量，需要利用(T-N-σ，T-σ)之间的控制参数序列，其中，σ表示动力学系统的时延σ。一种情况中，动力学系统的时延σ的取值范围约在1.5帧～2.5帧之间即30ms～50ms之间。
考虑到动力学系统的时延σ为随机变量，相应的，可以用于确定T时刻的车辆状态的控制数据序列的首尾对应的时刻拓宽至(T-N-σ-A,T)，即(T-K,T)，以时刻(T-K,T)的控制数据序列作为用于预测时刻T的车辆状态时，车辆动力学模型的输入，以让车辆动力学模型隐式的去学习随机变量σ，以使得后续的训练所得的预先构建的车辆动力学模型的输出结果达到更好的效果。T、N、σ、A以及K均为正数，K大于N。相应的，N表示第一时长，K表示第二时长。
相应的，在本发明的一种实现方式中，样本时刻对应的样本历史状态信息为目标车辆在该样本时刻的前第二时长所对应的时刻的车辆状态信息，第二时长小于第一时长。
S102：针对每一样本时刻，将该样本时刻对应的样本历史状态信息及样本控制参数序列，输入初始车辆动力学模型，确定出该样本时刻对应的样本预测状态信息。
第一电子设备在利用各样本时刻对应的样本历史状态信息、样本控制参数序列以及各样本时刻的标签车辆状态信息训练初始车辆动力学模型的过程中，可以利用单帧数据训练初始车辆动力学模型，即针对每一样本时刻，利用该样本时刻对应的样本历史状态信息及样本控制参数序列以及样本时刻的标签车辆状态信息训练初始车辆动力学模型。相应的，第一电子设备针对每一样本时刻，将该样本时刻对应的样本历史状态信息及样 本控制参数序列，输入初始车辆动力学模型，确定出该样本时刻对应的样本预测状态信息。
一种实现中，初始车辆动力学模型为基于深度学习的循环神经网络模型，其可以包括特征编码层、状态循环预测层以及特征解码层。在一种情况中，特征编码层和特征解码层可以通过全连接层实现，状态循环预测层可以通过GRU(Gated Recurrent Unit，门控循环单元)实现。
S103：针对每一样本时刻，利用该样本时刻对应的样本预测状态信息以及该样本时刻的标签车辆状态信息，确定初始车辆动力学模型对应的当前损失值。
一种实现中，第一电子设备针对每一样本时刻，利用该样本时刻对应的样本预测状态信息以及该样本时刻的标签车辆状态信息，计算该样本时刻对应的样本预测状态信息及标签车辆状态信息之间的距离，将所计算的距离确定为初始车辆动力学模型对应的当前损失值。
另一种实现中，第一电子设备针对每一样本时刻，利用该样本时刻对应的样本预测状态信息以及该样本时刻的标签车辆状态信息，计算该样本时刻对应的样本预测状态信息及标签车辆状态信息之间的距离；将预设数量个样本时刻对应的样本预测状态信息及标签车辆状态信息之间的距离的平均值或和，作为初始车辆动力学模型对应的当前损失值。
S104：基于当前损失值，调整初始车辆动力学模型的模型参数，直至初始车辆动力学模型达到预设收敛状态，得到预先建立的车辆动力学模型。
本步骤中，第一电子设备可以首先判断当前损失值是否大于预设损失阈值，若判断当前损失值大于预设损失阈值，则确定该初始车辆动力学模型未收敛，则利用预设优化算法，调整初始车辆动力学模型的模型参数，并返回重新执行针对每一样本时刻，将该样本时刻对应的样本历史状态信息及样本控制参数序列，输入调整参数后的初始车辆动力学模型，确定出该样本时刻对应的样本预测状态信息，进而基于该样本时刻对应的样本预测状态信息与该样本时刻的标签车辆状态信息，确定初始车辆动力学模型对应的损失值，作为当前损失值，重新判断当前损失值是否大于预设损失阈值，若判断当前损失值大于预设损失阈值，则继续利用预设优化算法，调整调整参数后的初始车辆动力学模型的模型参数，以此类推，直至计算出的调整参数后的初始车辆动力学模型对应的损失值不大于预设损失阈值，则确定调整参数后的初始车辆动力学模型达到预设收敛状态，则确定当前的该参数后的初始车辆动力学模型，为预先建立的车辆动力学模型。
其中，预设优化算法可以为梯度下降法最小二乘法以及遗传算法等。计算当前损失值的过程中可以通过LOSS函数计算，例如可以通过L2LOSS函数计算。
应用本发明实施例，可以利用目标车辆的各样本时刻对应的样本历史状态信息、样本控制参数序列以及各样本时刻的标签车辆状态信息，训练初始车辆动力学模型，以通过利用车辆动力学模型进行监督学习，学习得到目标车辆的各样本历史状态信息和样本控制参数序列，与样本时刻的标签车辆状态信息之间的关系，实现端到端的对车辆动力学进行建模，建模过程不需要人工参与，并且，训练模型的数据均为根据目标车辆的真实情况采集得到，所构建的车辆动力学模型更加贴合车辆自身特性，进而通过预先构建的车辆动力学模型可以确定出更加准确的车辆状态信息。
在本发明的另一实施例中，所述S102，可以包括如下步骤011-013：
011：针对每一样本时刻，将该样本时刻对应的样本历史状态信息输入初始车辆动力学模型的特征编码层，得到该样本时刻所对应样本历史状态信息对应的隐向量。
012：针对每一样本时刻，将该样本时刻所对应样本历史状态信息对应的隐向量以及该样本时刻对应的样本控制参数序列，输入初始车辆动力学模型的状态循环预测层，得到该样本时刻所对应车辆状态信息对应的隐向量。
013：针对每一样本时刻，将该样本时刻所对应车辆状态信息对应的隐向量输入初始车辆动力学模型的特征解码层，确定出该样本时刻对应的样本预测状态信息。
本实现方式中，样本时刻对应的历史状态信息为样本时刻的前第二时长所对应的时刻的状态信息，为单帧数据，为了保证后续的初始车辆动力学模型的有效训练，第一电子设备针对每一样本时刻，将该样本时刻对应的样本历史状态信息输入初始车辆动力学模型的特征编码层，得到该样本时刻所对应样本历史状态信息对应的隐向量，并将该样本时刻所对应样本历史状态信息对应的隐向量作为初始的状态向量以及该样本时刻对应的样本控制参数序列，输入初始车辆动力学模型的状态循环预测层，以使得状态循环预测层基于该初始的状态向量以及该样本时刻对应的样本控制参数序列，依次循环确定该初始的状态向量所对应时刻之后的各时刻的车辆状态信息对应的隐向量，直至确定出该样本时刻的车辆状态信息对应的隐向量。
进而，第一电子设备针对每一样本时刻，将该样本时刻所对应车辆状态信息对应的隐向量输入初始车辆动力学模型的特征解码层，以进行解码，解码得到该样本时刻对应的样本预测状态信息，即所预测的该样本时刻对应的车辆状态信息。
在一种实现中，该初始车辆动力学模型的特征编码层可以通过一个32维的全连接层实现，该初始车辆动力学模型的状态循环预测层可以通过GRU门控循环单元实现，该初始车辆动力学模型的特征解码层可以通过一个16维的全连接层和一个3维的全连接层实现。
如图2A所示，为状态循环预测层的一种数据流动示意图，其中，图2A中的z t和r t 分别表示更新门对应的值和重置门对应的值。更新门用于控制t-1时刻的车辆状态信息被带入到t时刻的车辆状态信息中的程度，更新门的值越大说明t-1时刻的车辆状态信息被带入的越多。重置门控制t-1时刻的车辆状态信息有多少信息被写入到t时刻的车辆状态信息，重置门越小t-1时刻的车辆状态信息被写入的越少。
其中，状态循环预测层中各处理过程可以通过如下公式表示：
z t＝σ(W z·[h t-1,x t])；
r t＝σ(W r·[h t-1,x t])；
Figure PCTCN2021109534-appb-000001
Figure PCTCN2021109534-appb-000002
其中，W z、W r以及W为状态循环预测层的参数，通过训练得到；h t-1表示t-1时刻的车辆状态信息，训练过程中，h t-1的初始值为样本时刻对应的样本历史状态信息，后续的实际预测过程中，h t-1的初始值为当前时刻对应的历史状态信息，h t表示t时刻的车辆状态信息，训练过程中，h t的为样本时刻对应的样本预测状态信息，后续的实际预测过程中，h t的为当前时刻对应的车辆状态信息；训练过程中，x t为样本时刻对应的样本控制参数序列；后续的实际预测过程中，x t为当前时刻对应的当前控制参数序列。上述公式中[]表示两个向量相连，*表示矩阵的乘积。
在本发明的另一实施例中，所述方法还可以包括对预先建立的车辆动力学模型的测试过程，以确定预先建立的车辆动力学模型所确定的车辆状态的优劣。相应的，在S105之后，所述方法还可以包括如下步骤021-024：
021：获得目标车辆的原始测试数据。
其中，原始测试数据包括：目标车辆行驶过程中所生成的各测试时刻对应的测试历史状态信息、测试控制参数序列以及测试车辆状态信息，每一测试时刻对应的测试控制参数序列包括：测试时刻及向前所述第一时长内各时刻的控制参数，测试时刻对应的测试历史状态信息为：测试时刻的前第二时长所对应的时刻的车辆状态信息。
022：将第一测试时刻对应的测试历史状态信息和测试控制参数序列，输入预先构建的车辆动力学模型，确定出第一测试时刻对应的测试预测状态信息。
其中，第一测试时刻包括：最早的测试时刻及其之后的第二时长所对应时刻之前的各时刻。
023：将第二测试时刻对应的预测历史状态信息以及测试控制参数序列，输入预先构建的车辆动力学模型，确定出该第二测试时刻对应的测试预测状态信息。
其中，第二测试时刻为：测试时刻中除第一测试时刻外的时刻，第二测试时刻对应的预测历史状态信息为：第二测试时刻的前第二时长所对应时刻对应的测试预测状态信 息。
024：利用各测试时刻对应的测试预测状态信息和测试车辆状态信息，确定预先构建的车辆动力学模型的测试结果。
本实现方式中，为了对预先建立的车辆动力学模型的预测结果的准确性进行测试，第一电子设备可以首先获得目标车辆的原始测试数据，该原始测试数据可以包括：目标车辆行驶过程中所生成的各时刻对应的测试历史状态信息、测试控制参数序列以及测试车辆状态信息，为了描述清楚，将测试过程中所利用到的数据对应的时刻称为测试时刻，其中，各测试时刻对应的测试车辆状态信息为目标车辆真实的车辆状态信息。
第一电子设备按照各第一测试时刻的时间先后顺序，依次将第一测试时刻对应的测试历史状态信息和测试控制参数序列，输入预先构建的车辆动力学模型，以通过预先构建的车辆动力学模型的特征编码层、状态循环预测层以及特征解码层，确定出各第一测试时刻对应的测试预测状态信息。
如图2B所示，预测T时刻对应的车辆状态时，需要使用T-N时刻对应的车辆状态，即图2B中所示的“O(T-N)”，以及T-K时刻至T时刻的控制参数序列即图2B中所示的“S(T)”，其中，N表示第一时长，K表示第二时长，N小于K。在T时刻为测试时刻中最早的时刻的情况下，T到T+N-1时刻即为第一测试时刻，相应的，确定其对应的车辆状态信息即对应的测试预测状态信息时，需要利用原始测试数据中，该第一测试时刻对应的测试历史状态信息和测试控制参数序列。
如图2B所示，预测T时刻对应的车辆状态时，将O(T-N)和S(T)作为预先建立的车辆动力学模型的特征编码层即32维的全连接层“FC(32)Initial States”进行编码，得到相应的隐向量；将隐向量作为初始的状态向量以及S(T)输入状态循环预测层即GRU门控循环单元，如图2B所示的“GRU”，以通过基于该初始的状态向量以及S(T)即T-K时刻至T时刻的控制参数序列，依次循环确定该初始的状态向量所对应时刻之后的各时刻的车辆状态信息对应的隐向量，直至确定出T时刻对应的车辆状态对应的隐向量。将该T时刻对应的车辆状态对应的隐向量输入特征解码层，即一个16维的全连接层如图2B中的“FC(16)”和一个3维的全连接层如图2B中的“FC(3)”，对T时刻对应的车辆状态对应的隐向量进行解码，得到T时刻对应的车辆状态，如图2中的O(T)。
相应的，T+N时刻开始，计算该T+N时刻对应的车辆状态信息即对应的测试预测状态信息时，需要使用T时刻对应的车辆状态，以及T+N-K时刻至T+N时刻的控制参数序列。此时，T时刻对应的车辆状态包括：原始测试数据中T时刻对应的测试车辆状态信息，且包括预先构建的车辆动力学模型所输出的测试预测状态信息。
为了测试预先构建的车辆动力学模型的预测确定结果的准确性，确定T+N时刻对应的测试预测状态信息时，需要将T时刻对应的测试预测状态信息，作为T+N时刻对应的测试历史状态信息，相应的，将T时刻对应的测试预测状态信息以及T+N时刻对应的测试控制参数序列，即原始测试数据中T+N-K时刻至T+N时刻的控制参数序列，输入预先构建的车辆动力学模型，确定出该第二测试时刻对应的测试预测状态信息。依次类推，依次计算T+N时刻之后的各时刻的测试预测状态信息。相应的，该T+N时刻及其之后的各测试时刻为上述的第二测试时刻。
相应的，确定出各第一测试时刻对应的测试预测状态信息之后，第一电子设备按照各第二测试时刻的时间先后顺序，依次将第二测试时刻对应的预测历史状态信息和测试控制参数序列，输入预先构建的车辆动力学模型，以通过预先构建的车辆动力学模型的特征编码层、状态循环预测层以及特征解码层，确定出各第二测试时刻对应的测试预测状态信息。该第二测试时刻对应的预测历史状态信息为第二测试时刻的前第二时长所对应时刻对应的测试预测状态信息。举例而言，第二测试时刻为T+N时刻的情况下，第二测试时刻对应的预测历史状态信息，为预先构建的车辆动力学模型输出的T时刻对应的测试预测状态信息。
确定出各测试时刻对应的测试预测状态信息之后，利用各测试时刻对应的测试预测状态信息和测试车辆状态信息，计算各测试时刻对应的测试预测状态信息和测试车辆状态信息之间的距离；基于各测试时刻对应的测试预测状态信息和测试车辆状态信息之间的距离，确定预先构建的车辆动力学模型的测试结果。
其中，确定预先构建的车辆动力学模型的测试结果的过程，可以是：统计测试时刻对应的测试预测状态信息和测试车辆状态信息之间的距离中，未超过预设距离阈值的距离的数量，若未超过预设距离阈值的距离的数量与距离总数的比值超过预设比值，则预先构建的车辆动力学模型的测试结果包括表征预先构建的车辆动力学模型测试通过的信息；反之，若未超过预设距离阈值的距离的数量与距离总数的比值不超过预设比值，则预先构建的车辆动力学模型的测试结果包括表征预先构建的车辆动力学模型测试未通过的信息。等，本发明实施例并不对确定预先构建的车辆动力学模型的测试结果的具体过程进行限定。
在确定预先构建的车辆动力学模型测试通过之后，则表征该预先构建的车辆动力学模型可用于目标车辆的车辆状态信息的确定过程；反之，确定预先构建的车辆动力学模型测试未通过，则需要重新训练获得预先构建的车辆动力学模型。
在本发明实施例中，用于训练和测试车辆动力学模型的数据包括在目标车辆的行驶过程中所记录的控制参数以及车辆状态信息，训练和测试车辆动力学模型的数据无需人 工标注，在一定程度上节省了人工成本。车辆动力学模型不需要维护任何内部状态，模型的单次预测车辆状态的过程中，输入的车辆状态信息为历史某一帧即某一时刻的车辆状态信息。每一帧对应的一个独立的样本，在一定程度上可以增加在训练和测试车辆动力学模型的过程中的样本数据的数量，且降低了训练过程以及后续的测试和实际预测确定过程的计算难度。模型可以在任一时刻开始预测，不需要提前预热，即从目标车辆静止状态下开始执行车辆状态信息的预测确定过程。
相应于上述方法实施例，本发明实施例提供了一种车辆状态信息预测方法，该方法依赖于上述方法实施例中所构建的预先建立的车辆动力学模型。如图3所示，所述方法可以包括如下步骤S301-S302：
S301：获得目标车辆在当前时刻对应的历史状态信息以及当前控制参数序列。
其中，当前控制参数序列包括：当前时刻及向前第一时长内各时刻的控制参数。
本发明实施例所提供的车辆状态信息预测方法，可以应用于任一具有计算能力的第二电子设备，该第二电子设备可以为终端或者服务器。在一种实现中，实现该方法的功能软件可以以单独的客户端软件的形式存在，也可以以目前相关的客户端软件的插件的形式存在，例如可以以动力学系统的功能模块的形式存在，这都是可以的。该第二电子设备可以与上述第一电子设备为同一物理设备，也可以为不同物理设备。
在一种情况中，该第二电子设备可以为车载设备，设置于目标车辆的内部，也可以为非车载设备，可以获得目标车辆的相关信息，这都是可以的。
第二电子设备在对目标车辆的车辆状态信息预测确定的过程中，可以获得目标车辆在当前时刻对应的历史状态信息以及当前控制参数序列，其中，当前时刻对应的历史状态信息为：当前时刻之前的时刻对应的历史状态信息。当前时刻对应的当前控制参数序列包括：该当前时刻及向前第一时长内各时刻的控制参数。该第一时长为根据经验设置的。每一时刻对应多种类型的控制参数，各时刻对应的控制参数的类型相同。
在一种情况中，车辆状态信息可以包括但不限于：车辆的速度、加速度、横摆角速度以及姿态角等。各时刻对应的控制参数的类型包括但不限于：刹车控制量以及油门控制量等。
考虑到车辆的真实的动力学系统为延时系统，在本发明的一种实现方式中，历史状态信息为目标车辆在当前时刻的前第二时长所对应的时刻的车辆状态信息，第二时长小于第一时长。
S302：将历史状态信息以及当前控制参数序列，输入预先构建的车辆动力学模型，确定出目标车辆在当前时刻的车辆状态信息。
其中，预先构建的车辆动力学模型为：基于目标车辆的各历史时刻对应的样本状态 信息及其对应的样本控制参数序列训练所得的循环神经网络模型。可以包括特征编码层、状态循环预测层以及特征解码层。在一种情况中，特征编码层和特征解码层可以通过全连接层实现，状态循环预测层可以通过GRU(Gated Recurrent Unit，门控循环单元)实现。
第二电子设备将历史状态信息以及当前控制参数序列，输入预先构建的车辆动力学模型，预先构建的车辆动力学模型通过其特征编码层、状态循环预测层以及特征解码层，对历史状态信息以及当前控制参数序列进行处理，确定出目标车辆在当前时刻的车辆状态信息。进而输出目标车辆在当前时刻的车辆状态信息，以通过目标车辆在当前时刻的车辆状态信息控制目标车辆的行驶状态。
应用本发明实施例，利用学习得到目标车辆的各样本历史状态信息和样本控制参数序列，与样本时刻的标签车辆状态信息之间的关系的循环神经网络模型，即预先构建的车辆动力学模型，确定目标车辆在当前时刻的车辆状态信息，可以得到准确性更高且更贴合目标车辆自身性能的车辆状态信息，进而保证车辆行驶过程中的安全性。
在本发明的另一实施例中，所述S302，可以包括如下步骤031-033：
031：将历史状态信息输入预先构建的车辆动力学模型的特征编码层，得到历史状态信息对应的隐向量。
032：将历史状态信息对应的隐向量以及当前控制参数序列，输入预先构建的车辆动力学模型的状态循环预测层，得到当前时刻所对应车辆状态信息对应的隐向量。
033：将当前时刻所对应车辆状态信息对应的隐向量输入预先构建的车辆动力学模型的特征解码层，得到当前时刻对应的车辆状态信息。
本实现方式中，当前时刻对应的历史状态信息为当前时刻的前第二时长所对应的时刻的状态信息，为单帧数据，为了保证后续的预先建立的车辆动力学模型的有效预测确定，第二电子设备将该当前时刻对应的历史状态信息输入预先建立的车辆动力学模型的特征编码层，得到当前时刻所对应历史状态信息对应的隐向量，并将该当前时刻所对应历史状态信息对应的隐向量作为初始的状态向量以及当前控制参数序列，输入预先建立的车辆动力学模型的状态循环预测层，以使得状态循环预测层基于该初始的状态向量以及当前控制参数序列，依次循环确定该初始的状态向量所对应时刻之后的各时刻的车辆状态信息对应的隐向量，直至确定出当前时刻的车辆状态信息对应的隐向量。
进而，第二电子设备将当前时刻所对应车辆状态信息对应的隐向量输入预先建立的车辆动力学模型的特征解码层，以进行解码，解码得到当前时刻对应的车辆状态信息。
在一种实现中，该预先建立的车辆动力学模型的特征编码层可以通过一个32维的全连接层实现，该预先建立的车辆动力学模型的状态循环预测层可以通过GRU门控循环单 元实现，该预先建立的车辆动力学模型的特征解码层可以通过一个16维的全连接层和一个3维的全连接层实现。
在本发明的另一实施例中，所述方法还可以包括如下步骤：041-042：
041：获得预设控制参数确定模型基于当前时刻的车辆状态信息，确定的当前控制参数。
042：将包含当前控制参数的目标控制参数序列以及当前时刻的下一时刻对应的历史状态信息，输入预先构建的车辆动力学模型，确定出目标车辆在当前时刻的下一时刻的车辆状态信息。
其中，目标控制参数序列还包括：当前时刻及向前第一时长所对应时刻的前一时刻中的各时刻的控制参数。
本实现方式中，第二电子设备在确定出当前时刻的车辆状态信息之后，可以将该当前时刻的车辆状态信息输入至预设控制参数确定模型，以通过该预设控制参数确定模型基于当前时刻的车辆状态信息，确定出目标车辆在当前时刻的下一时刻对应的控制参数。相应的，第二电子设备获得该当前时刻的下一时刻对应的控制参数；将当前时刻的下一时刻对应的控制参数，添加至目标控制参数序列，作为当前时刻的下一时刻对应的控制参数序列。将包含当前控制参数的目标控制参数序列以及当前时刻的下一时刻对应的历史状态信息，输入预先构建的车辆动力学模型，确定出目标车辆在当前时刻的下一时刻的车辆状态信息。
其中，该预设控制参数确定模型可以利用相关技术中任一可以确定车辆控制参数的确定算法，基于该当前时刻的车辆状态信息确定当前时刻的下一时刻对应的控制参数，用于控制目标车辆的行驶。
相应于上述方法实施例，本发明实施例提供了一种车辆动力学模型的构建装置，如图4所示，所述装置可以包括：
第一获得模块410，被配置为获得目标车辆的各样本时刻对应的样本历史状态信息、样本控制参数序列以及各样本时刻的标签车辆状态信息，其中，各样本时刻对应的样本控制参数序列包括：该样本时刻及向前第一时长内各时刻的控制参数；
第一确定模块420，被配置为针对每一样本时刻，将该样本时刻对应的样本历史状态信息及样本控制参数序列，输入初始车辆动力学模型，确定出该样本时刻对应的样本预测状态信息；
第二确定模块430，被配置为针对每一样本时刻，利用该样本时刻对应的样本预测状态信息以及该样本时刻的标签车辆状态信息，确定所述初始车辆动力学模型对应的当前损失值；
调整模块440，被配置为基于所述当前损失值，调整所述初始车辆动力学模型的模型参数，直至所述初始车辆动力学模型达到预设收敛状态，得到预先建立的车辆动力学模型。
应用本发明实施例，可以利用目标车辆的各样本时刻对应的样本历史状态信息、样本控制参数序列以及各样本时刻的标签车辆状态信息，训练初始车辆动力学模型，以通过利用车辆动力学模型进行监督学习，学习得到目标车辆的各样本历史状态信息和样本控制参数序列，与样本时刻的标签车辆状态信息之间的关系，实现端到端的对车辆动力学进行建模，建模过程不需要人工参与，并且，训练模型的数据均为根据目标车辆的真实情况采集得到，所构建的车辆动力学模型更加贴合车辆自身特性，进而通过预先构建的车辆动力学模型可以确定出更加准确的车辆状态信息。
在本发明的另一实施例中，所述样本时刻对应的样本历史状态信息为所述目标车辆在该样本时刻的前第二时长所对应的时刻的状态信息，所述第二时长小于所述第一时长。
在本发明的另一实施例中，所述第一确定模块420，被具体配置为针对每一样本时刻，将该样本时刻对应的样本历史状态信息输入所述初始车辆动力学模型的特征编码层，得到该样本时刻所对应样本历史状态信息对应的隐向量；
针对每一样本时刻，将该样本时刻所对应样本历史状态信息对应的隐向量以及该样本时刻对应的样本控制参数序列，输入所述初始车辆动力学模型的状态循环预测层，得到该样本时刻所对应车辆状态信息对应的隐向量；
针对每一样本时刻，将该样本时刻所对应车辆状态信息对应的隐向量输入所述初始车辆动力学模型的特征解码层，确定出该样本时刻对应的样本预测状态信息。
在本发明的另一实施例中，所述装置还包括：
第二获得模块(图中未示出)，被配置为在所述确定所述初始车辆动力学模型收敛，得到预先构建的车辆动力学模型之后，获得所述目标车辆的原始测试数据，其中，所述原始测试数据包括：所述目标车辆行驶过程中所生成的各测试时刻对应的测试历史状态信息、测试控制参数序列以及测试车辆状态信息，每一测试时刻对应的测试控制参数序列包括：所述测试时刻及向前所述第一时长内各时刻的控制参数，所述测试时刻对应的测试历史状态信息为：所述测试时刻的前第二时长所对应的时刻的车辆状态信息；
第四确定模块(图中未示出)，被配置为将第一测试时刻对应的测试历史状态信息和测试控制参数序列，输入所述预先构建的车辆动力学模型，确定出第一测试时刻对应的测试预测状态信息，其中，第一测试时刻包括：最早的测试时刻及其之后的第二时长所对应时刻之前的各时刻；
第五确定模块(图中未示出)，被配置为将第二测试时刻对应的预测历史状态信息 以及测试控制参数序列，输入所述预先构建的车辆动力学模型，确定出该第二测试时刻对应的测试预测状态信息，其中，所述第二测试时刻为：所述测试时刻中除所述第一测试时刻外的时刻，所述第二测试时刻对应的预测历史状态信息为：所述第二测试时刻的前第二时长所对应时刻对应的测试预测状态信息；
第六确定模块(图中未示出)，被配置为利用各测试时刻对应的测试预测状态信息和测试车辆状态信息，确定所述预先构建的车辆动力学模型的测试结果。
相应于上述方法实施例，本发明实施例提供了一种车辆状态信息预测装置，如图5所示，所述装置可以包括：
第三获得模块510，被配置为获得目标车辆在当前时刻对应的历史状态信息以及当前控制参数序列，其中，所述当前控制参数序列包括：所述当前时刻及向前第一时长内各时刻的控制参数；
第七确定模块520，被配置为将所述历史状态信息以及当前控制参数序列，输入预先构建的车辆动力学模型，确定出所述目标车辆在当前时刻的车辆状态信息，其中，所述预先构建的车辆动力学模型为：基于所述目标车辆的各历史时刻对应的样本状态信息及其对应的样本控制参数序列训练所得的循环神经网络模型。
应用本发明实施例，利用学习得到目标车辆的各样本历史状态信息和样本控制参数序列，与样本时刻的标签车辆状态信息之间的关系的循环神经网络模型，即预先构建的车辆动力学模型，确定目标车辆在当前时刻的车辆状态信息，可以得到准确性更高且更贴合目标车辆自身性能的车辆状态信息，进而保证车辆行驶过程中的安全性。
在本发明的另一实施例中，所述历史状态信息为所述目标车辆在当前时刻的前第二时长所对应的时刻的车辆状态信息，所述第二时长小于所述第一时长。
在本发明的另一实施例中，所述第七确定模块520，被具体配置为将所述历史状态信息输入所述预先构建的车辆动力学模型的特征编码层，得到所述历史状态信息对应的隐向量；
将所述历史状态信息对应的隐向量以及所述当前控制参数序列，输入所述预先构建的车辆动力学模型的状态循环预测层，得到所述当前时刻所对应车辆状态信息对应的隐向量；
将所述当前时刻所对应车辆状态信息对应的隐向量输入所述预先构建的车辆动力学模型的特征解码层，得到当前时刻对应的车辆状态信息。
在本发明的另一实施例中，所述装置还包括：
第四获得模块(图中未示出)，被配置为获得预设控制参数确定模型基于所述当前时刻的车辆状态信息，确定的当前控制参数；
第八确定模块(图中未示出)，被配置为将包含所述当前控制参数的目标控制参数序列以及所述当前时刻的下一时刻对应的历史状态信息，输入所述预先构建的车辆动力学模型，确定出所述目标车辆在当前时刻的下一时刻的车辆状态信息，其中，所述目标控制参数序列还包括：所述当前时刻及向前第一时长所对应时刻的前一时刻中的各时刻的控制参数。
上述系统、装置实施例与系统实施例相对应，与该方法实施例具有同样的技术效果，具体说明参见方法实施例。装置实施例是基于方法实施例得到的，具体的说明可以参见方法实施例部分，此处不再赘述。本领域普通技术人员可以理解：附图只是一个实施例的示意图，附图中的模块或流程并不一定是实施本发明所必须的。
本领域普通技术人员可以理解：实施例中的装置中的模块可以按照实施例描述分布于实施例的装置中，也可以进行相应变化位于不同于本实施例的一个或多个装置中。上述实施例的模块可以合并为一个模块，也可以进一步拆分成多个子模块。
最后应说明的是：以上实施例仅用以说明本发明的技术方案，而非对其限制；尽管参照前述实施例对本发明进行了详细的说明，本领域的普通技术人员应当理解：其依然可以对前述实施例所记载的技术方案进行修改，或者对其中部分技术特征进行等同替换；而这些修改或者替换，并不使相应技术方案的本质脱离本发明实施例技术方案的精神和范围。

Claims (10)

  1. 一种车辆动力学模型的构建方法，其特征在于，所述方法包括：
    获得目标车辆的各样本时刻对应的样本历史状态信息、样本控制参数序列以及各样本时刻的标签车辆状态信息，其中，各样本时刻对应的样本控制参数序列包括：该样本时刻及向前第一时长内各时刻的控制参数；
    针对每一样本时刻，将该样本时刻对应的样本历史状态信息及样本控制参数序列，输入初始车辆动力学模型，确定出该样本时刻对应的样本预测状态信息；
    针对每一样本时刻，利用该样本时刻对应的样本预测状态信息以及该样本时刻的标签车辆状态信息，确定所述初始车辆动力学模型对应的当前损失值；
    基于所述当前损失值，调整所述初始车辆动力学模型的模型参数，直至所述初始车辆动力学模型达到预设收敛状态，得到预先建立的车辆动力学模型。
  2. 如权利要求1所述的方法，其特征在于，所述样本时刻对应的样本历史状态信息为所述目标车辆在该样本时刻的前第二时长所对应的时刻的车辆状态信息，所述第二时长小于所述第一时长。
  3. 如权利要求1所述的方法，其特征在于，所述针对每一样本时刻，将该样本时刻对应的样本历史状态信息及样本控制参数序列，输入初始车辆动力学模型，确定出该样本时刻对应的样本预测状态信息的步骤，包括：
    针对每一样本时刻，将该样本时刻对应的样本历史状态信息输入所述初始车辆动力学模型的特征编码层，得到该样本时刻所对应样本历史状态信息对应的隐向量；
    针对每一样本时刻，将该样本时刻所对应样本历史状态信息对应的隐向量以及该样本时刻对应的样本控制参数序列，输入所述初始车辆动力学模型的状态循环预测层，得到该样本时刻所对应车辆状态信息对应的隐向量；
    针对每一样本时刻，将该样本时刻所对应车辆状态信息对应的隐向量输入所述初始车辆动力学模型的特征解码层，确定出该样本时刻对应的样本预测状态信息。
  4. 如权利要求1所述的方法，其特征在于，在所述确定所述初始车辆动力学模型收敛，得到预先构建的车辆动力学模型的步骤之后，所述方法还包括：
    获得所述目标车辆的原始测试数据，其中，所述原始测试数据包括：所述目标车辆行驶过程中所生成的各测试时刻对应的测试历史状态信息、测试控制参数序列以及测试车辆状态信息，每一测试时刻对应的测试控制参数序列包括：所述测试时刻及向前所述第一时长内各时刻的控制参数，所述测试时刻对应的测试历史状态信息为：所述测试时 刻的前第二时长所对应的时刻的车辆状态信息；
    将第一测试时刻对应的测试历史状态信息和测试控制参数序列，输入所述预先构建的车辆动力学模型，确定出第一测试时刻对应的测试预测状态信息，其中，第一测试时刻包括：最早的测试时刻及其之后的第二时长所对应时刻之前的各时刻；
    将第二测试时刻对应的预测历史状态信息以及测试控制参数序列，输入所述预先构建的车辆动力学模型，确定出该第二测试时刻对应的测试预测状态信息，其中，所述第二测试时刻为：所述测试时刻中除所述第一测试时刻外的时刻，所述第二测试时刻对应的预测历史状态信息为：所述第二测试时刻的前第二时长所对应时刻对应的测试预测状态信息；
    利用各测试时刻对应的测试预测状态信息和测试车辆状态信息，确定所述预先构建的车辆动力学模型的测试结果。
  5. 一种基于权利要求1-4任一项所述的方法所构建车辆动力学模型的车辆状态信息预测方法，其特征在于，所述方法包括：
    获得目标车辆在当前时刻对应的历史状态信息以及当前控制参数序列，其中，所述当前控制参数序列包括：所述当前时刻及向前第一时长内各时刻的控制参数；
    将所述历史状态信息以及当前控制参数序列，输入预先构建的车辆动力学模型，确定出所述目标车辆在当前时刻的车辆状态信息，其中，所述预先构建的车辆动力学模型为：基于所述目标车辆的各历史时刻对应的样本状态信息及其对应的样本控制参数序列训练所得的循环神经网络模型。
  6. 如权利要求5所述的方法，其特征在于，所述历史状态信息为所述目标车辆在当前时刻的前第二时长所对应的时刻的车辆状态信息，所述第二时长小于所述第一时长。
  7. 如权利要求5或6所述的方法，其特征在于，所述将所述历史状态信息以及当前控制参数序列，输入预先构建的车辆动力学模型，确定出所述目标车辆在当前时刻的车辆状态信息的步骤，包括：
    将所述历史状态信息输入所述预先构建的车辆动力学模型的特征编码层，得到所述历史状态信息对应的隐向量；
    将所述历史状态信息对应的隐向量以及所述当前控制参数序列，输入所述预先构建的车辆动力学模型的状态循环预测层，得到所述当前时刻所对应车辆状态信息对应的隐向量；
    将所述当前时刻所对应车辆状态信息对应的隐向量输入所述预先构建的车辆动力学模型的特征解码层，得到当前时刻对应的车辆状态信息。
  8. 如权利要求5或6所述的方法，其特征在于，所述方法还包括：
    获得预设控制参数确定模型基于所述当前时刻的车辆状态信息，确定的当前控制参数；
    将包含所述当前控制参数的目标控制参数序列以及所述当前时刻的下一时刻对应的历史状态信息，输入所述预先构建的车辆动力学模型，确定出所述目标车辆在当前时刻的下一时刻的车辆状态信息，其中，所述目标控制参数序列还包括：所述当前时刻及向前第一时长所对应时刻的前一时刻中的各时刻的控制参数。
  9. 一种车辆动力学模型的构建装置，其特征在于，所述装置包括：
    第一获得模块，被配置为获得目标车辆的各样本时刻对应的样本历史状态信息、样本控制参数序列以及各样本时刻的标签车辆状态信息，其中，各样本时刻对应的样本控制参数序列包括：该样本时刻及向前第一时长内各时刻的控制参数；
    第一确定模块，被配置为针对每一样本时刻，将该样本时刻对应的样本历史状态信息及样本控制参数序列，输入初始车辆动力学模型，确定出该样本时刻对应的样本预测状态信息；
    第二确定模块，被配置为针对每一样本时刻，利用该样本时刻对应的样本预测状态信息以及该样本时刻的标签车辆状态信息，确定所述初始车辆动力学模型对应的当前损失值；
    调整模块，被配置为基于所述当前损失值，调整所述初始车辆动力学模型的模型参数，直至所述初始车辆动力学模型达到预设收敛状态，得到预先建立的车辆动力学模型。
  10. 一种基于权利要求9所述的装置所构建车辆动力学模型的车辆状态信息预测装置，其特征在于，所述装置包括：
    第三获得模块，被配置为获得目标车辆在当前时刻对应的历史状态信息以及当前控制参数序列，其中，所述当前控制参数序列包括：所述当前时刻及向前第一时长内各时刻的控制参数；
    第七确定模块，被配置为将所述历史状态信息以及当前控制参数序列，输入预先构建的车辆动力学模型，确定出所述目标车辆在当前时刻的车辆状态信息，其中，所述预先构建的车辆动力学模型为：基于所述目标车辆的各历史时刻对应的样本状态信息及其对应的样本控制参数序列训练所得的循环神经网络模型。
PCT/CN2021/109534 2021-01-25 2021-07-30 车辆动力学模型的构建及车辆状态信息预测方法、装置 WO2022156182A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
EP21920563.0A EP4216098A1 (en) 2021-01-25 2021-07-30 Methods and apparatuses for constructing vehicle dynamics model and for predicting vehicle state information
US18/358,169 US20230367934A1 (en) 2021-01-25 2023-07-25 Method and apparatus for constructing vehicle dynamics model and method and apparatus for predicting vehicle state information

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202110092643.3A CN112464577B (zh) 2021-01-25 2021-01-25 车辆动力学模型的构建及车辆状态信息预测方法、装置
CN202110092643.3 2021-01-25

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US18/358,169 Continuation US20230367934A1 (en) 2021-01-25 2023-07-25 Method and apparatus for constructing vehicle dynamics model and method and apparatus for predicting vehicle state information

Publications (1)

Publication Number Publication Date
WO2022156182A1 true WO2022156182A1 (zh) 2022-07-28

Family

ID=74802705

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/109534 WO2022156182A1 (zh) 2021-01-25 2021-07-30 车辆动力学模型的构建及车辆状态信息预测方法、装置

Country Status (4)

Country Link
US (1) US20230367934A1 (zh)
EP (1) EP4216098A1 (zh)
CN (1) CN112464577B (zh)
WO (1) WO2022156182A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115826544A (zh) * 2023-02-17 2023-03-21 江苏御传新能源科技有限公司 一种汽车配件的生产调参系统

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112464577B (zh) * 2021-01-25 2021-04-20 魔门塔(苏州)科技有限公司 车辆动力学模型的构建及车辆状态信息预测方法、装置
CN113492827A (zh) * 2021-06-23 2021-10-12 东风柳州汽车有限公司 一种混合动力汽车能量管理方法及装置

Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20200114926A1 (en) * 2018-10-16 2020-04-16 Toyota Motor Engineering & Manufacturing North America, Inc. Vehicle velocity predictor using neural networks based on v2x data augmentation to enable predictive optimal control of connected and automated vehicles
CN111125854A (zh) * 2018-10-31 2020-05-08 百度在线网络技术（北京）有限公司 车辆动力学模型的优化方法、装置、存储介质和终端设备
CN111152796A (zh) * 2020-04-07 2020-05-15 北京三快在线科技有限公司 一种车辆运动状态预测方法及装置
CN111626219A (zh) * 2020-05-28 2020-09-04 深圳地平线机器人科技有限公司 轨迹预测模型生成方法、装置、可读存储介质及电子设备
CN111931286A (zh) * 2020-06-29 2020-11-13 北京百度网讯科技有限公司 纵向动力学模型的训练方法、装置及设备
CN111930015A (zh) * 2020-09-16 2020-11-13 北京三快在线科技有限公司 一种无人车控制方法及装置
CN112238857A (zh) * 2020-09-03 2021-01-19 北京新能源汽车技术创新中心有限公司 自动驾驶车辆的控制方法
CN112464577A (zh) * 2021-01-25 2021-03-09 魔门塔(苏州)科技有限公司 车辆动力学模型的构建及车辆状态信息预测方法、装置

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP3803103B2 (ja) * 2004-05-21 2006-08-02 シャープ株式会社 イオン濃度分布の予測方法、解析装置およびイオン濃度分布の予測プログラム
CN106671985A (zh) * 2016-10-25 2017-05-17 长春工业大学 电动汽车动力学系统建模方法

Patent Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20200114926A1 (en) * 2018-10-16 2020-04-16 Toyota Motor Engineering & Manufacturing North America, Inc. Vehicle velocity predictor using neural networks based on v2x data augmentation to enable predictive optimal control of connected and automated vehicles
CN111125854A (zh) * 2018-10-31 2020-05-08 百度在线网络技术（北京）有限公司 车辆动力学模型的优化方法、装置、存储介质和终端设备
CN111152796A (zh) * 2020-04-07 2020-05-15 北京三快在线科技有限公司 一种车辆运动状态预测方法及装置
CN111626219A (zh) * 2020-05-28 2020-09-04 深圳地平线机器人科技有限公司 轨迹预测模型生成方法、装置、可读存储介质及电子设备
CN111931286A (zh) * 2020-06-29 2020-11-13 北京百度网讯科技有限公司 纵向动力学模型的训练方法、装置及设备
CN112238857A (zh) * 2020-09-03 2021-01-19 北京新能源汽车技术创新中心有限公司 自动驾驶车辆的控制方法
CN111930015A (zh) * 2020-09-16 2020-11-13 北京三快在线科技有限公司 一种无人车控制方法及装置
CN112464577A (zh) * 2021-01-25 2021-03-09 魔门塔(苏州)科技有限公司 车辆动力学模型的构建及车辆状态信息预测方法、装置

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115826544A (zh) * 2023-02-17 2023-03-21 江苏御传新能源科技有限公司 一种汽车配件的生产调参系统

Also Published As

Publication number Publication date
CN112464577B (zh) 2021-04-20
US20230367934A1 (en) 2023-11-16
EP4216098A1 (en) 2023-07-26
CN112464577A (zh) 2021-03-09

Similar Documents

Publication Publication Date Title
WO2022156182A1 (zh) 车辆动力学模型的构建及车辆状态信息预测方法、装置
CN107203134B (zh) 一种基于深度卷积神经网络的前车跟随方法
CN110235148B (zh) 训练动作选择神经网络
EP3593288B1 (en) Training action selection neural networks using look-ahead search
WO2021103834A1 (zh) 换道决策模型生成方法和无人车换道决策方法及装置
CN112119404A (zh) 样本高效的强化学习
EP3610418A1 (en) Distributional reinforcement learning
WO2020052480A1 (zh) 无人驾驶行为决策及模型训练
EP3926420A1 (en) Making time-series predictions of a computer-controlled system
CN112907970A (zh) 一种基于车辆排队长度变化率的可变车道转向控制方法
Miller et al. 23 Real-Time Simulation of Physical Systems Using Simscape™
CN110103960B (zh) 车辆自适应巡航控制方法、系统及车辆
CN115311860A (zh) 一种交通流量预测模型的在线联邦学习方法
CN113419424B (zh) 减少过估计的模型化强化学习机器人控制方法及系统
CN111625457A (zh) 基于改进的dqn算法的虚拟自动驾驶测试优化方法
CN109752952B (zh) 一种获取多维随机分布及强化控制器的方法和装置
CN113033902A (zh) 一种基于改进深度学习的自动驾驶换道轨迹规划方法
CN110826695B (zh) 数据处理方法、装置和计算机可读存储介质
CN109725639B (zh) 巡航系统的线性控制方法及装置
CN115086375A (zh) 网联车辆运动状态信息延时补偿方法、装置、系统和介质
Shamsudin et al. Recursive Gauss-Newton based training algorithm for neural network modelling of an unmanned rotorcraft dynamics
CN112198794A (zh) 一种基于类人驾驶规则与改进深度确定性策略梯度的无人驾驶方法
CN111290118A (zh) 一种变形镜的解耦控制方法及装置
CN115489572B (zh) 基于强化学习的列车ato控制方法、设备及存储介质
CN114639255B (zh) 一种交通信号控制方法、装置、设备和介质

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 21920563

Country of ref document: EP

Kind code of ref document: A1

ENP Entry into the national phase

Ref document number: 2021920563

Country of ref document: EP

Effective date: 20230418

NENP Non-entry into the national phase

Ref country code: DE