US-12626037-B2 - Vehicle response predicting device, training device, method and recording medium on which a program is recorded

US12626037B2US 12626037 B2US12626037 B2US 12626037B2US-12626037-B2

Abstract

A vehicle response predicting device comprising: a memory; and a processor coupled to the memory, wherein the processor is structured so as to, by using a convolutional neural network model that has been trained in advance, and that is for predicting response of a vehicle, and whose input is input data expressing a time series of an input to the vehicle and a characteristic of the vehicle, predict response of a vehicle that is an object of prediction, from input data expressing a time series of an input to the vehicle that is the object of prediction and a characteristic of the vehicle that is the object of prediction.

Inventors

Kohei SHINTANI
Yutaka Sasaki
Makoto Miwa
Kohei Makino

Assignees

Kohei SHINTANI
Yutaka Sasaki
Makoto Miwa
Kohei Makino

Dates

Publication Date: 20260512
Application Date: 20210816
Priority Date: 20200817

Claims (7)

1 . A vehicle response predicting device comprising: a memory; and a processor coupled to the memory, wherein the processor is configured to: acquire a 3D model of a vehicle; acquire a characteristic of the vehicle based on the 3D model, the characteristic comprising a roll angle of the vehicle; prepare input data that expresses a time series of preset input to the vehicle and the acquired characteristic of the vehicle, the input data comprising a matrix, the matrix comprising a plurality of vectors, each vector of the plurality of vectors comprising a steering angle at a particular time step and the acquired characteristic of the vehicle; and by using a convolutional neural network model that has been trained in advance, and that is for predicting response of a vehicle, and whose input is input data expressing a time series of an input to the vehicle and a characteristic of the vehicle, predict a response of the vehicle, wherein a filter of a convolutional layer includes a plurality of filters having different intervals, in a time series direction of values convolved by the filters, wherein output of at least one convolutional layer of the convolutional neural network model is results obtained by applying an activation function to a sum of convolution results, which are based on output of a layer that is one before, and the output of the layer that is one before.
2 . A training device comprising: a memory; and a processor coupled to the memory, wherein the processor is structured so as to acquire data of simulation results determined by computer aided engineering that relate to vehicle response prediction; prepare, based on the simulation results, training data that expresses input data, which expresses a time series of an input to a vehicle and a characteristic of the vehicle, the characteristic comprising a roll angle of the vehicle, and response of the vehicle, from response of a vehicle that is obtained as results of simulation, the input data comprising a matrix, the matrix comprising a plurality of vectors, each vector of the plurality of vectors comprising a steering angle at a particular time step and the characteristic of the vehicle, and train a convolutional neural network model for predicting response of the vehicle by using the input data as input, based on the training data, wherein output of at least one convolutional layer of the convolutional neural network model is results obtained by applying an activation function to a sum of convolution results, which are based on output of a layer that is one before, and the output of the layer that is one before.
3 . The training device of claim 2 , wherein the processor prepares training data that expresses input data, which expresses a time series of an input to a vehicle and a characteristic of the vehicle, a time series of a state of the vehicle, and response of the vehicle, from a time series of a state of a vehicle and response of the vehicle that are obtained as results of simulation based on a time series of input to the vehicle and a characteristic of the vehicle, and trains the convolutional neural network model such that intermediate output of the convolutional neural network model corresponds to the time series of the state of the vehicle of the training data, and final output of the convolutional neural network model corresponds to the response of the vehicle of the training data.
4 . A vehicle response predicting method comprising predicting processing that, acquires a 3D model of a vehicle; acquires a characteristic of the vehicle based on the 3D model, the characteristic comprising a roll angle of the vehicle; prepares input data that expresses a time series of preset input to the vehicle and the acquired characteristic of the vehicle, the input data comprising a matrix, the matrix comprising a plurality of vectors, each vector of the plurality of vectors comprising a steering angle at a particular time step and the acquired characteristic of the vehicle; and by using a convolutional neural network model that has been trained in advance, and that is for predicting response of a vehicle, and whose input is input data expressing a time series of an input to the vehicle and a characteristic of the vehicle, predicts a response of the vehicle, wherein a filter of a convolutional layer includes a plurality of filters having different intervals, in a time series direction, of values convolved by the filters, wherein output of at least one convolutional layer of the convolutional neural network model is results obtained by applying an activation function to a sum of convolution results, which are based on output of a layer that is one before, and the output of the layer that is one before.
5 . A training method comprising: acquiring data of simulation results determined by computer aided engineering that relate to vehicle response prediction; preparing, based on the simulation results, processing that prepares training data that expresses input data expressing a time series of an input to a vehicle and a characteristic of the vehicle, the characteristic comprising a roll angle of the vehicle, and response of the vehicle, from response of a vehicle that is obtained as results of simulation, the input data comprising a matrix, the matrix comprising a plurality of vectors, each vector of the plurality of vectors comprising a steering angle at a particular time step and the characteristic of the vehicle, and training processing that trains a convolutional neural network model for predicting response of the vehicle by using the input data as input, based on the training data, wherein output of at least one convolutional layer of the convolutional neural network model is results obtained by applying an activation function to a sum of convolution results, which are based on output of a layer that is one before, and the output of the layer that is one before.
6 . A non-transitory recording medium on which is recorded a vehicle response predicting program executable by a computer to perform processing of: acquiring a 3D model of a vehicle; acquiring a characteristic of the vehicle based on the 3D model, the characteristic comprising a roll angle of the vehicle; preparing input data that expresses a time series of preset input to the vehicle and the acquired characteristic of the vehicle, the input data comprising a matrix, the matrix comprising a plurality of vectors, each vector of the plurality of vectors comprising a steering angle at a particular time step and the acquired characteristic of the vehicle; and by using a convolutional neural network model that has been trained in advance, and that is for predicting response of a vehicle, and whose input is input data expressing a time series of an input to the vehicle and a characteristic of the vehicle, predicting a response of the vehicle, wherein a filter of a convolutional layer includes a plurality of filters having different intervals, in a time series direction, of values convolved by the filters, wherein output of at least one convolutional layer of the convolutional neural network model is results obtained by applying an activation function to a sum of convolution results, which are based on output of a layer that is one before, and the output of the layer that is one before.
7 . A non-transitory recording medium on which is recorded a training program executable by a computer to perform processing of: acquiring data of simulation results determined by computer aided engineering that relate to vehicle response prediction; preparing, based on the simulation results, training data that expresses input data expressing a time series of an input to a vehicle and a characteristic of the vehicle, the characteristic comprising a roll angle of the vehicle, and response of the vehicle, from response of a vehicle that is obtained as results of simulation, the input data comprising a matrix, the matrix comprising a plurality of vectors, each vector of the plurality of vectors comprising a steering angle at a particular time step and the characteristic of the vehicle; and training a convolutional neural network model for predicting response of the vehicle by using the input data as input, based on the training data, wherein output of at least one convolutional layer of the convolutional neural network model is results obtained by applying an activation function to a sum of convolution results, which are based on output of a layer that is one before, and the output of the layer that is one before.

Description

CROSS-REFERENCE TO RELATED APPLICATION This application is based on and claims priority under 35 USC 119 from Japanese Patent Application No. 2020-137639 filed on Aug. 17, 2020, the disclosure of which is incorporated by reference herein. BACKGROUND Technical Field The present disclosure relates to a vehicle response predicting device, a training device, a method, and a recording medium on which a program is recorded. Related Art In vehicle development, the behavior of a vehicle at the time of traveling is inferred from data of the vehicle body by using CAE (Computer Aided Engineering) that is a simulation by a computer, and research into predicting performances of vehicles is advancing on the basis of such inferences. The approach of using machine learning as a substitute model for CAE is known. For example, a technique using random forests is disclosed in L'ubor Ladick'y, SoHyeon Jeong, Barbara Solenthaler, Marc Pollefeys, and Markus Gross. Data-Driven Fluid Simulations using Regression Forests. ACM Trans. Graph., Vol. 34, No. 6, pp. 199:1-199:9, October 2015. However, the approach of the above-referenced L'ubor Ladick'y, SoHyeon Jeong, Barbara Solenthaler, Marc Pollefeys, and Markus Gross. Data-Driven Fluid Simulations using Regression Forests. ACM Trans. Graph., Vol. 34, No. 6, pp. 199:1-199:9, October 2015 does not use a model that takes the entire time series into consideration, and is difficult to apply to an object that requires consideration of a continuous time series such as CAE analysis of vehicle driving performance. SUMMARY An object of the present disclosure is to provide a vehicle response predicting device, a training device, a method and a program, which can accurately predict response of a vehicle with respect to the time series of an input to the vehicle. A first aspect is a vehicle response predicting device including a predicting section that, by using a convolutional neural network model that has been trained in advance, and is for predicting response of a vehicle, and whose input is input data expressing the time series of an input to the vehicle and a characteristic of the vehicle, predicts response of a vehicle that is an object of prediction, from input data expressing the time series of an input to the vehicle that is the object of prediction and a characteristic of the vehicle that is the object of prediction. Here, the characteristic of the vehicle is a parameter relating to a physical structure of the vehicle. In accordance with the vehicle response predicting device of the first aspect, the predicting section predicts response of a vehicle that is an object of prediction by using a convolutional neural network model that is for predicting response of a vehicle and whose input is input data expressing the time series of an input to the vehicle and a characteristic of the vehicle. At this time, at the convolutional neural network model, convolution processing is carried out on the input data that expresses the time series of an input to the vehicle and a characteristic of the vehicle. Due thereto, response of the vehicle with respect to the time series of an input to the vehicle can be predicted accurately. Note that, in a vehicle response predicting device of a second aspect, in the vehicle response predicting device of the first aspect, output of at least one convolutional layer of the convolutional neural network model is results obtained by applying an activation function to a sum of convolution results, which are based on output of a layer that is one before, and the output of the layer that is one before. In accordance with the vehicle response predicting device of the second aspect, the convolutional neural network model can be trained efficiently. Further, in a vehicle response predicting device of a third aspect, in the vehicle response predicting device of the first or second aspect, the input data is a matrix in which vectors, which express an input to the vehicle and a characteristic of the vehicle, are lined up in time series order. In accordance with the vehicle response predicting device of the third aspect, response of the vehicle can be predicted by the convolution processing at the convolutional neural network model, while taking the time series of an input to the vehicle into consideration. Further, in a vehicle response predicting device of a fourth aspect, in the vehicle response predicting device of any of the first through third aspects, a filter of the convolutional layer includes plural filters having different intervals, in a time series direction, of values convolved by the filters. In accordance with the vehicle response predicting device of the fourth aspect, convolution processing can be carried out at different time periods at a convolutional neural network model. A fifth aspect is a training device including: a preparing section that prepares training data that expresses input data, which expresses a time series of an input to a vehicle and a cha