A Sequential Neural Network Construction Method for Vehicle State Estimation Fusion with Normative Model
By integrating specification model and deep learning model in vehicle state estimation and combining vehicle dynamic model for output constraints, the problems of unreasonable output and large calculation volume in vehicle state prediction technology are solved, and the stability of output within a reasonable range is achieved and improve computing efficiency.
Patent Information
- Application Number
- CN202210015519.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-01-07
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2042-01-07
AI Technical Summary
In the existing vehicle state prediction technology, neural network-based estimation calculation method cannot ensure that the output estimation value is always reasonable, and the method combining deep learning and traditional models will take into account the error values of deep learning, which is huge in calculation.
Using a fusion specification model to construct a neural network for vehicle state estimation, constrain the output by setting a reasonable range, combining deep learning models and formal boundary models, ensure that the output is within a reasonable range, and is calculated through vehicle dynamics models. The rate of change and direction vectors are standardized to verify.
It ensures that the output of vehicle state estimation is always within a reasonable range, improves the calculation speed and simplicity of the model, and avoids the uncertainty of the deep learning model and the excessive calculation amount.
Smart Images

Figure CN114357623B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of vehicle state prediction, and in particular to a method for constructing a vehicle state estimation sequence neural network integrating a specification model. Background Art
[0002] Vehicle state estimation is a fundamental problem in dynamic control. Traditional vehicle state estimation mostly relies on low-cost sensor data and Kalman filtering techniques. These methods have low estimation accuracy and the results obtained in the nonlinear domain of the vehicle are almost unusable.
[0003] In recent years, artificial intelligence technology has been widely used, among which deep learning technology performs well in fitting strong nonlinear systems, making it possible to use deep learning technology to estimate vehicle dynamics. Studies have shown that the use of deep learning to estimate vehicle state is more accurate than traditional methods and performs well in nonlinear domains.
[0004] However, the security of deep learning models is also an open issue. Their black box characteristics make them unexplainable. Insufficient training data may lead to poor generalization performance and cannot guarantee that their output is always within a reasonable range. This is fatal in the task of vehicle state estimation. The results of state prediction are often used as input information for vehicle dynamics control. Occasional prediction values with large deviations may cause large changes in the control amount, and in severe cases may cause vehicle instability. The method of combining deep learning with traditional model estimation will take into account the error values of deep learning and require huge calculations. Summary of the invention
[0005] 1. Technical issues to be resolved
[0006] In view of the shortcomings of the prior art, the present invention provides a method for constructing a vehicle state estimation sequence neural network that integrates a specification model, which solves the problem in the existing vehicle state prediction technology that the vehicle state estimation algorithm based on the neural network cannot guarantee that the output estimate value is always reasonable, and the method of combining deep learning with traditional model estimation will take the error value of deep learning into account and the amount of calculation is huge.
[0007] (II) Technical solution
[0008] In order to achieve the above-mentioned purpose, the present invention specifically adopts the following technical solutions:
[0009] A method for constructing a sequential neural network for vehicle state estimation integrating a specification model comprises the following steps:
[0010] S1, set the reasonable range of each estimator to be estimated π i, forming a set π;
[0011] The estimated quantity refers to the vehicle state quantity that the overall model needs to estimate;
[0012] S2, obtain sensor information, input it into the sequence deep learning model for prediction, and obtain the predicted vehicle state value vector Spt (where the element is Spti);
[0013] S3, obtain sensor information and the state information output at the previous moment, input the formal boundary model to calculate the limit vector Ωt of the change rate (where the elements are Ωti), and the direction vector Vt (where the elements are Vti) composed of the directions of the change rates of each value;
[0014] S4. Get the output St-1 of the previous moment, and use the following formula to calculate the rate of change vector αt of each state value, where ρ is the time step of the model prediction
[0015] ;
[0016] S5, taking π, Spt, Ωt, Vt, αt as input, performing standard verification, obtaining the final output St, and feeding it back to the sequence deep learning model;
[0017] S6. Return to step S2 to perform output estimation at the next moment.
[0018] Specifically, in step S3, the method for obtaining the restriction vector Ωt and the direction vector Vt specifically includes the following steps:
[0019] S31, obtaining the sensor input and the vehicle state quantity output at the last moment;
[0020] S32, calculating the tire force on each tire of the vehicle based on the above information and a simple tire model;
[0021] S33, calculating the direction and magnitude of the force and moment acting on the center of mass of the vehicle from the forces of each tire;
[0022] S34, calculating the rate of change of the estimated quantity based on the force and the torque;
[0023] S35, splitting the above change rate into a direction vector Vt and a limit vector Ωt.
[0024] Specifically, in step S5, the final output St acquisition and feedback specifically include the following steps:
[0025] S51, check whether the directions of each item αti in αt are consistent with the directions of each item in Vt; check whether the absolute values of each item αti in αt are less than or equal to the corresponding items Ωti in Ωt; check whether each value in Spt is within a preset range π;
[0026] S52, check the correct estimated item and assign sti = spti;
[0027] S53, adding the estimated amount that has been checked for error to the error set E;
[0028] S54, pop up an estimate in E, if E is empty, go to step S57;
[0029] S55. If the estimated value is wrong due to the direction of the rate of change or the value is greater than the limit value, the output of the estimated value at this moment is estimated according to the following formula; then return to step S54
[0030] ;
[0031] S56. If the error of the estimation is caused by exceeding the value range, the output is limited to the boundary of the value range; return to step S54
[0032] Sti = πi;
[0033] S57. Take the estimated vector St finally obtained as output.
[0034] (III) Beneficial effects
[0035] Compared with the prior art, the present invention provides a method for constructing a vehicle state estimation sequence neural network integrating a standard model, which has the following beneficial effects:
[0036] 1. The present invention adds a specification model based on vehicle dynamics to the feedback channel of the traditional sequential neural network, uses the model to constrain unreasonable output values, and then uses the model as feedback for the sequential neural network. Using formalized boundaries to constrain the output of the deep learning network can ensure that its output is always within the reasonable range of the formalized boundary constraints.
[0037] 2. In the present invention, the specification model is designed based on vehicle dynamics. It only needs to calculate the force or torque acting on the center of mass of the vehicle according to the current state, and then divide it by the corresponding mass or moment of inertia to obtain the direction and rate of change of the corresponding state. The parameters of the tire model are taken under the high friction coefficient of the dry road surface. The longitudinal and lateral tire models are simplified, which improves the calculation speed. The model is simple and has a wide range. The output of the neural network will not be strictly limited, and it will only be restricted when it is unreasonable. BRIEF DESCRIPTION OF THE DRAWINGS
[0038] FIG1 is a schematic diagram of the overall principle of the present invention;
[0039] FIG2 is a diagram of a standard model of the present invention based on vehicle dynamics;
[0040] FIG3 is a schematic diagram of the overall solution flow of the present invention;
[0041] FIG4 is a schematic diagram of the sub-process flow of step S3 of the present invention;
[0042] FIG5 is a schematic diagram of the sub-process flow of step S5 of the present invention;
[0043] FIG. 6 is a schematic diagram of the structural scheme principle of an embodiment of the present invention. DETAILED DESCRIPTION
[0044] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0045] Example
[0046] As shown in Figures 1-6, the overall scheme is shown in Figure 1, and the overall flow chart is shown in Figure 3. An embodiment of the present invention proposes a method for constructing a vehicle state estimation sequence neural network that integrates a standard model. This embodiment takes the estimation of a simple vehicle longitudinal velocity V, lateral velocity v and yaw angular velocity r as an example to illustrate the working mode of the overall model. At the same time, this embodiment is also a practical method for estimating these two state quantities. Example 1 specifies the sequence neural network used in the model, which is a GRU model; and specifies the standard model, using a simple tire model to calculate the tire force, based on which the direction vector and the restriction vector are obtained.
[0047] The estimated vector of the model is:
[0048] ;
[0049] The input vector of the sensor is:
[0050] ;
[0051] Wherein: δf is the steering wheel angle, ωfl, ωfr, ωrl, ωrr are the wheel speeds of the four wheels, φ is the accelerator pedal opening, ay is the lateral acceleration obtained by ESP, and r(~) is the yaw angular velocity.
[0052] Example 1 Scheme diagram Figure 6 shown.
[0053] S1. Set the reasonable range of the quantity to be estimated, π i , to form a set π . The reasonable range of the lateral velocity v is [Vlower, Vupper , the reasonable range of lateral velocity v is [vlower, vupper , the reasonable range of yaw angular velocity r is rlower, rupper .
[0054] S2. Obtain sensor information and input it into the sequence deep learning model for prediction to obtain the predicted vehicle state value vector Spt (where the elements are Vt, vt, rt).
[0055] S3. Obtain sensor information and the state information output at the previous moment, input the formal boundary model to calculate the restriction vector Ωt (where the elements are Ωti) of the change rate, and the direction vector Vt (where the elements are Vti) composed of the directions of the change rates of each value; see the sub-process for specific steps, and the specification model is shown in Figure 2.
[0056] S4. Get the output St-1 of the previous moment, and use the following formula to calculate the change rate vector αt of each state value, where ρ is the time step of the model output, and the calculation formula is as follows:
[0057] ;
[0058] S5. Take π, Spt, Ωt, Vt, αt as input, perform standard verification, obtain the final output St,s and feed it back to the GRU model; see the sub-process for standard verification.
[0059] S6. Return to step 2 to perform output estimation at the next moment.
[0060] The sub-process flow of step S3 is shown in FIG4 . The method for obtaining the restriction vector Ωt and the direction vector Vt specifically includes the following steps:
[0061] Step 1: Obtain the input steering wheel angle δf of the sensor and the vehicle state quantity Vt−1, vt−1, rt−1 output at the previous moment;
[0062] Step 2: Based on the above information and a simple tire model, calculate the tire force on each tire of the vehicle. The calculation formula is as follows:
[0063] ;
[0064] ;
[0065] ;
[0066] ;
[0067] Where Fyf is the resultant lateral force of the two tires on the front axle, Fyr is the resultant lateral force of the two tires on the rear axle, C αf is the front wheel cornering stiffness, C αr is the rear wheel cornering stiffness, a is the distance from the center of mass to the front axle, b is the distance from the center of mass to the rear axle, s is the slip rate of the tire, and Fxi is the longitudinal force of each tire.
[0068] Step 3: Calculate the direction and magnitude of the force and moment acting on the vehicle's center of mass from the forces of each tire.
[0069] ;
[0070] ;
[0071] ;
[0072] Where Σ Fy is the resultant lateral force and Σ Mz is the moment about the Z axis.
[0073] Step 4: Calculate the rate of change of the estimated quantity based on the force and moment.
[0074] ;
[0075] ;
[0076] ;
[0077] Where v. is the rate of change of the lateral velocity, and r. is the rate of change of the yaw angular velocity.
[0078] Step 5: Split the above rate of change into direction vector Vt and limit vector Ωt
[0079] ;
[0080] ;
[0081] in Represents the sign function.
[0082] The sub-process flow of step S5 is shown in FIG5 . The final output St acquisition and feedback specifically include the following steps:
[0083] Step 1: Check whether the directions of each item αti in αt are consistent with the directions of each item in Vt; check whether the absolute values of each item αti in αt are less than or equal to the corresponding items Ωti in Ωt; check whether each value in Spt is within the preset range π.
[0084] Step 2: Check for correct estimates to assign values
[0085] Sti = Spti;
[0086] Step 3: Add the estimated amount that has been checked for error to the error set E.
[0087] Step 4: Pop an estimator in E. If E is empty, go to step 7.
[0088] Step 5: If the estimated value is wrong due to the direction of the rate of change or the value is greater than the limit value, estimate the output of the estimated value at this moment according to the following formula; then return to step 4
[0089] ;
[0090] Step 6: If the error of the estimator is caused by exceeding the range, limit its output to the boundary of the range; return to step 4
[0091] Sti = πi;
[0092] Step 7: The final estimated vector St is taken as output.
[0093] In one aspect of the present invention, it is ensured that the output predicted by the model always remains within a reasonable range. This advantage comes from the addition of a specification model based on vehicle dynamics to the feedback channel of the traditional sequential neural network, which is used to constrain unreasonable output values and then used as feedback for the sequential neural network. The deep learning network has a black box characteristic. Even after training with a large amount of data, it is still possible to encounter untrained data in practical applications, resulting in uncertainty in its output. In addition, accidental jitter of the input data may also cause uncertainty in the output. Using formalized boundaries to constrain the output of the deep learning network can ensure that its output is always within a reasonable range of formalized boundary constraints.
[0094] In another aspect of the present invention, the specification model is designed based on vehicle dynamics. The model is simple and has a wide range. It will not strictly limit the output of the neural network, but will only limit it when it is unreasonable. This advantage comes from the design method of the specification model. The design principle of the specification model is: to ensure that the output value of the predicted quantity is within a reasonable range; the directions of each predicted quantity must be coordinated and consistent, and no physically impossible postures can appear; the change rate of each predicted quantity is within a reasonable range, and no sudden jumps can occur. Based on this designed model, it is only necessary to calculate the force or torque exerted on the center of mass of the vehicle according to the current state, and then divide it by the corresponding mass or moment of inertia to obtain the direction and change rate of the corresponding state. The parameters of the tire model take the values under the high friction coefficient of the dry road surface, and the longitudinal and lateral tire models are simplified to improve the calculation speed.
[0095] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.
Claims
1. A method for constructing a sequential neural network for vehicle state estimation that integrates a normative model. Its characteristics are In: including the following steps: S1. Set a reasonable range of each estimated quantity that needs to be estimated π i to form a set π ; the estimated quantity refers to the vehicle state quantity that needs to be estimated by the overall model; S2, obtain sensor information, input it into the sequence deep learning model for prediction, and obtain the predicted vehicle state value vector Spt, where the element is Spti; S3, obtain sensor information and the state information output at the previous moment, input the formal boundary model to calculate the change rate limit vector Ωt, where the elements Ωti, and the direction vector Vt composed of the directions of the change rates of each value, where the elements are Vti; S4. Get the output St-1 of the previous moment, and use the following formula to calculate the change rate vector α t of each state value, where ρ is the time step of the model prediction: ; S5, taking π, Spt, Ωt, Vt, αt as input, performs standard verification, obtains the final output St, and feeds it back to the sequence deep learning model; S6, return to step S2 to perform output estimation at the next moment; In step S3, the method for obtaining the restriction vector Ωt and the direction vector Vt specifically includes the following steps: S31, obtaining the sensor input and the vehicle state quantity output at the last moment; S32, calculating the tire force on each tire of the vehicle based on the above information and a simple tire model; S33, calculating the direction and magnitude of the force and moment acting on the center of mass of the vehicle from the forces of each tire; S34, calculating the rate of change of the estimated quantity based on the force and the torque; S35, splitting the above change rate into a direction vector Vt and a limit vector Ωt; In step S5, the final output St acquisition and feedback specifically includes the following steps: S51, check whether the directions of each item αti in αt are consistent with the directions of each item in Vt; check whether the absolute values of each item αti in αt are less than or equal to the corresponding items Ωti in Ωt; check whether each value in Spt is within a preset range π; S52. Check the correct estimated items and assign value sti = spti; S53, adding the estimated amount that has been checked for error to the error set E; S54, pop up an estimate in E, if E is empty, go to step S57; S55. If the estimated value is wrong due to the direction of the rate of change or the value is greater than the limit value, the output of the estimated value at this moment is estimated according to the following formula; then return to step S54 ; S56. If the error of the estimation is caused by exceeding the value range, the output is limited to the boundary of the value range; Return to step S54 Sti = πi; S57. Take the estimated vector St finally obtained as output.
