JP7152367B2 - Vehicle performance prediction method - Google Patents

Description

The present invention relates to a vehicle body performance prediction method.

In vehicle development, research is being carried out to estimate the vehicle's behavior during running from the specifications of the vehicle body using computer aided engineering (CAE), and to predict the performance of the vehicle based on the estimation. ing.

Patent Literature 1 discloses an invention of a vehicle motion model generating apparatus and method for generating a vehicle motion model that contributes to vehicle body performance prediction using a recurrent neural network (RNN).

Japanese Patent Application Laid-Open No. 2004-249812

However, the RNN sometimes has missing input information in each layer of the multi-layered structure, and in such a case, there is a problem that the accuracy of vehicle performance prediction is lowered.

SUMMARY OF THE INVENTION An object of the present invention is to provide a vehicle body performance prediction method capable of accurately predicting vehicle body performance.

The method for predicting vehicle body performance according to claim 1 includes a step of inputting a true value at the beginning of learning, randomly inputting a true value and an estimated value during the middle of learning, and inputting an estimated value at the end of learning to train the system. a step of calculating an estimated value related to the behavior of the vehicle body based on the variable related to the behavior of the vehicle body to the system after learning, and feeding back the calculated estimated value to the input to calculate the estimated value of the next stage in time series. a step of repeating the steps of calculating an estimated value in the next stage in the time series a predetermined number of times; and outputting a response related to the behavior of the vehicle based on the estimated value calculated in the time series and the variable. and simultaneously optimizing the estimated loss and the response loss respectively calculated in the time series.

According to the vehicle body performance prediction method of claim 1, a true value is input at the beginning of learning, a true value and an estimated value are input at random during the middle of learning, and an estimated value is input at the end of learning to efficiently operate the system. can be taught to

Further, according to the vehicle body performance prediction method of claim 1, in the post-learning neural network system 100, the estimated value related to the behavior of the vehicle body calculated based on the variable related to the behavior of the vehicle body is fed back to the input. Optimized estimates can be generated step-by-step by repeating the process of calculating the next-stage estimates on the series.

Further, according to the vehicle body performance prediction method of claim 1, by outputting a response related to the behavior of the vehicle based on the estimated values and variables respectively calculated in time series, The obtained information can be supplemented with variables to obtain a more accurate response.

Further, according to the vehicle body performance prediction method of claim 1, the accuracy of the response can be improved by simultaneously optimizing the loss of the estimated value and the loss of the response calculated in chronological order.

As described above, according to the vehicle body performance prediction method according to the present invention, it is possible to provide a vehicle body performance prediction method capable of accurately predicting vehicle body performance.

1 is a block diagram showing an example of a specific configuration of an arithmetic processing device related to a vehicle body performance prediction method according to an embodiment of the present invention; FIG. 1 is a schematic diagram showing an example of a neural network system constructed by a vehicle body performance prediction program; FIG. 1 is a schematic diagram showing an example of calculated results of an NCAP fishhook test for testing vehicle roll stability; FIG. FIG. 4 is an explanatory diagram showing the effects of the vehicle body performance prediction method according to the embodiment of the present invention;

A vehicle body performance prediction method according to the present embodiment will be described below with reference to FIG. FIG. 1 is a block diagram showing an example of a specific configuration of an arithmetic processing unit 10 relating to a vehicle body performance prediction method according to an embodiment of the present invention.

The arithmetic processing unit 10 is configured including a computer 30 . The computer 30 has a CPU 32 , a ROM 34 , a RAM 36 and an input/output port 38 . As an example, the computer 30 is desirably a model capable of executing advanced arithmetic processing at high speed, such as an engineering workstation or a supercomputer.

In computer 30, CPU 32, ROM 34, RAM 36, and input/output port 38 are connected to each other via various buses such as an address bus, a data bus, and a control bus. The input/output port 38 includes various input/output devices such as a display 40, a mouse 42, a keyboard 44, a hard disk (HDD) 46, and various disks (for example, CD-ROM, DVD, etc.) 48 for reading information from the disk. A drive 50 is connected to each.

A network 52 is connected to the input/output port 38 so that information can be exchanged with various devices connected to the network 52 . In this embodiment, a data server 56 to which a database (DB) 54 is connected is connected to the network 52, and information can be exchanged with the DB 54. FIG. The computer 30 may also serve as the data server 56 and the DB 54 .

The DB 54 stores in advance data of a 3D (three-dimensional) model of a vehicle for which vehicle body performance is to be predicted, data related to vehicle body performance prediction, and the like. The storage of information in the DB 54 may be data storage via the computer 30 or the data server 56 , or data storage via another device connected to the network 52 . The 3D model data stored in the DB 54 may be CAD (Computer Aided Design) vehicle data, but may also be data whose geometry has been adjusted in advance so as to facilitate arithmetic processing.

In this embodiment, it is assumed that the DB 54 connected to the data server 56 stores 3D model data of the vehicle subject to vehicle body performance prediction, observed values of vehicle body vibration, data related to vehicle body performance prediction, and the like. As will be described, the information in the DB 54 may be stored in an external storage device such as the HDD 46 built into the computer 30 or an external hard disk.

A vehicle body performance prediction program for predicting vehicle body performance is installed in the HDD 46 of the computer 30 . In the present embodiment, the CPU 32 executes a vehicle body performance prediction program to predict vehicle body performance. Also, the CPU 32 causes the display 40 to display the prediction result. There are several methods for installing the vehicle body performance prediction program of the present embodiment in the computer 30. For example, the vehicle body performance prediction program is stored in a CD-ROM, DVD, or the like, , and the vehicle performance prediction program is installed in the HDD 46 by executing the setup program for the CPU 32 . Alternatively, the vehicle body performance prediction program may be installed in HDD 46 by communicating with other information processing equipment connected to computer 30 via a public telephone line or network 52 .

FIG. 2 is a schematic diagram showing an example of a neural network system 100 constructed by a vehicle body performance prediction program. The neural network system 100 according to the present embodiment includes an input unit 110 for inputting vehicle parameters, which are observed values (variables) related to vehicle body performance prediction, and an intermediate calculation parameter, which is a predicted value of the input vehicle parameters. It includes a simulator section 130, an intermediate calculation parameter output section 150 that outputs the intermediate calculation parameters calculated by the simulator section, and a calculation section 170 that performs calculation for vehicle body performance prediction.

The input unit 110 inputs the vehicle parameters and intermediate calculation parameters based on the vehicle parameters to the vehicle parameter unit 112 to which the vehicle parameters are input, and inputs the intermediate calculation parameters based on the vehicle parameters to the simulator unit 130 . It includes series data input units 114A, 114B, 114C, and 114D for calculating differences from calculation parameters.

Simulator section 130 includes a coupling layer 132 which is a neural network, and LSTM (Long Short-Term Memory) sections 134A, 134B, 134C, and 134D. The coupling layer 132 and a coupling layer 174 of the computing unit 170, which will be described later, are hierarchical neural networks, and more specifically, multilayer perceptrons are used. The LSTM units 134A, 134B, 134C, and 134D and the LSTM units 172A, 172B, 172C, and 172D of the arithmetic unit 170, which will be described later, are recursive neural networks.

Intermediate calculation parameter output section 150 includes sequence data output sections 152A, 152B, 152C, and 152D.

The computing unit 170 includes LSTM units 172A, 172B, 172C, and 172D, a coupling layer 174 that is a neural network, a responding unit 176 that outputs the computation result of the coupling layer 174, and an MSE (Mean Squared Error) section 178 .

As shown in FIG. 2, neural network system 100 according to the present embodiment outputs (series data) ₁ calculated by simulator section 130 based on vehicle parameters and (series data) ₀ to operation section 170. Together, it is fed back to the input unit 110 . In this embodiment, for example, a series of steps from the sequence data input unit 114A to the joint layer 132, the LSTM unit 134A, the sequence data output unit 152A, and the LSTM unit 172A predict the output at the time T with the sequence do.

In the input unit 110, the fed back (series data) ₁ , the intermediate calculation parameter one step before (series data) ₀ , and the difference δ ₁ (not shown) between the (series data) ₁ and the (series data) ₀ ) and newly observed vehicle parameters are input to the simulator unit 130, new (series data) ₂ (not shown) calculated by the simulator unit 130 are output to the calculation unit 170, and the input unit 110 provide feedback.

After that, in the input unit 110, the fed back (series data) _t , the intermediate calculation parameter one step before (series data) _t-1 , and (series data) _t and (series data) _t-1 The difference δ _t and the newly observed vehicle parameters are input to the simulator unit 130, and the new (series data) _t+1 calculated by the simulator unit 130 is output to the calculation unit 170 and fed back to the input unit 110. This is repeated until the (sequence data) _T before input to the LSTM section 172D of the calculation section 170 is calculated.

Further, neural network system 100 according to the present embodiment includes skip connection 116 for inputting vehicle parameters, which are observed values, to coupling layer 174 of calculation unit 170, and series data input units 114A, 114B, 114C, and 114D. Skip connections 118A, 118B, 118C, and 118D input two intermediate calculation parameters input to 130 and their differences to LSTM units 172A, 172B, 172C, and 172D of the operation unit 170, respectively.

Since some of the input information may be missing in the process of calculating the intermediate calculation parameters in the simulator unit 130, in this embodiment, the skip connections 116, 118A, 118B, 118C and 118D are used.

The function of each part of the neural network system 100 according to this embodiment will be described below.
In the input unit 110, the vehicle parameters, which are observed values, are p = (p ₁ , p ₂ , ..., p _N ), and the steering input angle is s = {s ₁ , s ₂ , ..., s _T }. x, which is a combination of a vehicle parameter and a steering input angle of the vehicle, is defined as the following equation (1). In equation (1), N is the number of dimensions of the vehicle parameter, t is the number of time steps related to the sequence, and T is the length of the sequence, which is the maximum number of time steps. Such x _t is input to the simulator section 130 together with intermediate calculation parameters.

The simulator unit 130 calculates intermediate calculation parameters, which are estimated values of the input values in time series. First, in the simulator unit 130 , the feature amount is extracted from the input values in the coupling layer 132 . The LSTM units 134A, 134B, 134C, and 134D calculate intermediate calculation parameters, which are estimated values for the next stage in time series, based on the feature amount extracted by the coupling layer 132 . The intermediate calculation parameters calculated by the LSTM sections 134A, 134B, and 134C are output to the series data output sections 152A, 152B, and 152C, and are also output to the LSTM sections 134B, 134C, and 134D of the next series. The LSTM units 134B, 134C, and 134D of the next series calculate new intermediate calculation parameters based on the feature values extracted by the joint layer 132 and the intermediate calculation parameters input from the LSTM units 134A, 134B, and 134C of the previous series. do.

In this embodiment, the neural network of the coupling layer 132 is NN(x), and the neural network of the LSTM units 134A, 134B, 134C, and 134D is LSTM(x). If the hidden layer of the coupling layer 132 is N layers, W is the weight matrix, and f(x) is the activation function of the nonlinear function, NN(x) is expressed by the following equation (2).

Also, the output of each series of simulator section 130 is represented by the following equation (3). The left side of equation (3) is an estimate of the intermediate computational parameter.

In the present embodiment, as described above, the input unit 110 sends the intermediate calculation parameter fed back from the intermediate calculation parameter output unit 150 and the intermediate calculation parameter one step before holding the differentiation of the estimated parameter to the simulator unit 130. are entering in However, since the initial value input by the series data input unit 114A in FIG. 2 is one type of (series data) ₀ , there is no intermediate calculation parameter one step before. In this embodiment, two (sequence data) ₀ are used to set the output of the simulator section 130 as shown in the following equation (4).

In the calculation unit 170, the intermediate calculation parameters obtained for each series are input to the LSTM units 172A, 172B, 172C, and 172D, respectively, the feature amount necessary for the response is extracted in the joint layer 174, and the response is calculated from the extracted feature amount. It is calculated and output by the response unit 176 . The output of the calculation unit 170 is represented by the following equation (5). The left hand side of equation (5) is the estimate of the response.

As described above, in this embodiment, the skip connections 116, 118A, 118B, 118C, and 118D are added to improve accuracy. The skip connection 116 of this embodiment is provided to re-add information that has been lost by passing through the layers.

In this embodiment, the simulator unit 130 for calculating the intermediate calculation parameters and the calculation unit 170 for extracting the response are configured in two stages. information may be lost. In this embodiment, skip connections 116 compensate for missing information. Skip connection 116 combines the output value from LSTM section 172</b>D of calculation section 170 with the vehicle parameter input from vehicle parameter section 112 and inputs the result to coupling layer 174 . By the skip connection 116, the output of the calculation unit 170 is represented by the following equation (6). The left hand side of equation (6) is the estimate of the response.

The MSE unit 178 of the computing unit 170 defines a loss function L and performs error determination by hybrid loss that simultaneously optimizes the loss (error) of the intermediate calculation parameter and the loss (error) of the response.

The loss function L is defined by Equation (7) below. In the following equation (7), y is the true value of the response, z is the true value of the intermediate calculation parameter, and α and β are coefficients and arbitrary constants.

In this embodiment, in order to imitate the sequential calculation of the simulator, the predicted values of the intermediate calculation parameters of the preceding series are input. However, if a predicted value with a large error is input during learning of the neural network, then learning to remove the error is required. Therefore, in the early stage of learning, correct values are input instead of inputting predicted values. By inputting the correct answer, the neural network model is optimized to a solution with less error for each series. However, although it is true that a sequence in which the correct answer is input provides a solution with a small error, it is not always possible to obtain a solution with a small error for the entire system.

To obtain a low-error solution for the entire system requires minimizing the overall loss when inputting the predicted values. In this embodiment, the correct answer is input at the beginning of learning, the correct answer and the predicted value are randomly input at the middle of learning, and only the predicted value is input at the end of learning. gradually guide the system to a solution with less error. Formula (3) of the simulator unit 130 when the correct answer is treated as an input becomes the following formula (8). x _t-1 in equation (8) is the true value of the variable x, and z _t-1 is the true value of the intermediate calculation parameter.

In scheduled sampling, the input is changed from a true value to an estimated value by switching the arithmetic expression of the simulator unit 130 from expression (8) to expression (3) during learning, and an intermediate value is obtained using expression (3) during testing. Calculate calculation parameters.

FIG. 3 is a schematic diagram showing an example of calculated results of an NCAP fishhook test for testing vehicle roll stability. In the test of FIG. 3, when the vehicle speed is constant, the maximum lateral acceleration of the vehicle at the time of turning with steering operation in the left and right directions was predicted. As training data, DOE (Design of experiments) was used to generate 752 designs, and as an example, CarMaker (registered trademark), which is a vehicle driving simulator, was used.

The accuracy of the prediction model was evaluated using Root Mean Square Error (RMSE). RMSE is represented by the following formula. D in the following formula is the number of data.

FIG. 4 shows an example of the effects of the vehicle body performance prediction method according to this embodiment. Compared to the conventional method of predicting the maximum lateral acceleration from the vehicle parameters, this embodiment has the effect of reducing the prediction error (RMSE) by 76%.

As described above, according to the present embodiment, a true value is input at the beginning of learning, a true value and an estimated value are input at random during the middle of learning, and an estimated value is input at the end of learning. can be learned efficiently.

Further, according to the present embodiment, in the post-learning neural network system 100, an intermediate calculation parameter, which is an estimated value related to the behavior of the vehicle body calculated based on the vehicle parameter, which is a variable related to the behavior of the vehicle body, is fed back to the input. By repeating the process of calculating new intermediate calculation parameters by using

Further, according to the present embodiment, by outputting a response related to the behavior of the vehicle based on the calculated intermediate calculation parameter and vehicle parameter, information missing when calculating the intermediate calculation parameter can be replaced by the vehicle parameter. Compensation can be used to obtain a more accurate response.

Further, according to the present embodiment, by simultaneously optimizing the loss of each calculated intermediate calculation parameter and the loss of the response, it is possible to improve the accuracy of the response, which is the final output, and improve the vehicle performance. can be provided.

10 arithmetic processing unit 30 computer 38 input/output port 40 display 42 mouse 44 keyboard 50 disk drive 52 network 56 data server 100 neural network system 110 input section 112 vehicle parameter section 114A, 114B, 114C, 114D series data input section 116, 118A, 118B, 118C, 118D Skip connection 130 Simulator unit 132 Coupling layer 134A, 134B, 134C, 134D LSTM unit 150 Intermediate calculation parameter output unit 152A, 152B, 152C, 152D Sequence data output unit 170 Operation unit 172A, 172B, 172C, 172D LSTM Section 174 Coupling Layer 176 Response Section 178 MSE Section

Claims (1)

a step of inputting a true value at the beginning of learning, randomly inputting a true value and an estimated value at the middle of learning, and inputting an estimated value at the end of learning to train the system;
a step of calculating an estimated value related to the behavior of the vehicle body based on variables related to the behavior of the vehicle body in the system after learning;
A step of feeding back the calculated estimated value to the input and calculating the estimated value of the next stage in time series;
repeating the step of calculating the estimated value of the next stage in the time series a predetermined number of times;
a step of outputting a response related to the behavior of the vehicle based on the estimated values and the variables calculated in chronological order;
simultaneously optimizing the estimated loss and the response loss calculated in the time series;
A vehicle body performance prediction method including.

Images