JPH04149605A - Vehicle control method - Google Patents

Abstract

PURPOSE:To control the movement of a vehicle in a simple way by providing a neural net for simulation which simulates the vehicle movement, and a main neural net which performs the integrated control of the vehicle based on the result obtained by learning the movement of the vehicle through those neural nets. CONSTITUTION:A neural net B is provided for simulation of the movement of a vehicle together with a neural net A which gives the integrated control to the vehicle. Then a teacher signal is calculated from the signal that is specified by the net B, and the weight of the net A is corrected based on the teacher signal. This corrected net A is attached to an actual vehicle and the integrated control is given to the vehicle according to the steering wheel angle, the accelerator opening amount, and the brake application force. Thus it is possible to omit the construction of an integrated algorithm, and the construction of a system is simplied.

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention relates to a vehicle control method for controlling vehicle motion by integrally controlling front and rear wheel steering angles, engine output, etc.

(Prior art and selection points therebetween) Conventionally, there have been known technologies that integrally control steering control, engine control, drive system, etc. and,
In this case, the system #I was constructed using an integrated algorithm for each individual control.

In the conventional control method as described above, it is necessary to construct an integrated algorithm, but this task is said to become more difficult in proportion to the square of the number of control elements.

Therefore, there is a problem in that as the number of control elements increases, an enormous number of man-hours are required to develop algorithms.

An object of the present invention is to provide a control method that allows vehicle motion to be controlled in a simple manner.

(H. Means for Solving the Problem) This invention includes a simulation neural network that simulates vehicle motion and a main neural network that integrally controls the vehicle, and based on signals specified by the simulation neural network, The feature is that a teacher signal for the neural network is calculated, and the weights of the main neural network are modified based on this teacher signal.

(Action of the present invention) The present invention allows a simulation neural network to learn vehicle motion. Then, the simulation neural network that has been trained is connected to the main neural network, a teacher signal for the main neural network is specified, and the weights of the main neural network are adjusted based on this teacher signal.

(Effects of the present invention) According to the control method of the present invention, a neural network is used to learn vehicle motion, and the learning result is used to integrally control the vehicle. This eliminates the need to construct a system, making system construction that much easier.

(Embodiment of the present invention) In FIG. 2, B is a simulation neural network, and this simulation neural network B is
The front wheel steering angle α1, the rear wheel steering angle α1, the throttle opening degree δ0, the fuel injection amount Qa, the ignition timing t1, the cylinder pressure PA and the brake pressure Pa are human power, and the vehicle motion results are yaw angular velocity φ, roll angular velocity φ, and pitch. Angular velocity δ, longitudinal acceleration X,
Angular velocity θ, longitudinal acceleration X, lateral acceleration y, vertical acceleration i, and vehicle speed V are used as teacher signals.

The output value as described above is compared with the teacher signal, and the weights of the neural network are modified so that the difference is below an acceptable level. There are backpropagation methods etc. as methods for modifying the weights. Any method may be used for this purpose.

The simulation neural network B, which has completed its learning as described above, is connected to the main neural network A as shown in FIG. 3, and the main neural network A is caused to learn desired integrated control of vehicle motion.

In this case, the ideal vehicle motion result is given as the teacher signal, but this ideal motion result is, for example, a steady yaw determined by the steering wheel angle αH and the vehicle speed V with respect to the yaw angular velocity φ. It is possible to match the angle. The method for finding it is as follows.

Eight V Q, ... (1) φ“1+Av-.

n Steering gear ratio A: stability factor However, the specification of this teacher signal is not limited to the above. In addition, any suitable teaching signal may be used.

Then, the weights of the main neural network A are modified so that they are below the allowable level obtained from the neural network. At this time, since the simulation neural network B has already been completed in the first step to simulate the motion of the target vehicle, the problem is how to make the neural network A learn integrated control.

In order to make the main neural network A learn the above-mentioned integrated control, the teacher signals α1, α1.9t, Qa
, Ql, QA, and Qa need to be given only △, so they must be replaced with the teacher signal A on the main neural network A side, , public 1, sum t, Qa, ↑1, PA, PI3. The replacement is performed as follows using the calculation system C shown in FIG. For example, the input set of simulation net B (α;, α;
, δ: ...), an output of 1 is obtained, and the teacher signal at that time is φ, then it is replaced with the teacher signal of the main neural network A as follows.

First, the difference ΔΔφ−φ−φ (2) between the calculated value and the teacher signal is determined. At this time, the simulation news 1 net B has been completed, so it is easy to obtain the partial differential value b↓/α, = Cory], (3) aαf α minus αf. As a result, the teacher signal is set as Δα,=αt+Δαf−◆・・−(5) as a correction.

Other teacher signals are calculated in the same way.

Once each teacher signal, Q, , A, , Q, etc., is determined in this way, the main neural network A Correct the weight of.

Once the main neural network A with the weights modified in this way is completed, this neural network A is installed on an actual vehicle as shown in Figure 1, and integrated according to the steering wheel hook, accelerator opening, and brake pedal force. control.

As is clear from the above, the simulation neural network B is constructed for the purpose of training the main neural network A, and is not used after the training is completed.

[Brief explanation of the drawing]

The drawings show an embodiment of the present invention. Fig. 1741 is a block diagram of the main neural network installed in a vehicle, Fig. 2 is a block diagram of the simulation neural network, and Fig. 183 shows the main neural network and the simulation system. FIG. 2 is a block diagram showing a state of connection with a neural network for use in the computer. A-Main neural network, B-Simulation neural network.

Claims (1)

[Claims] It is equipped with a simulation neural network that simulates vehicle motion and a main neural network that provides integrated control of the vehicle.A teacher signal for the main neural network is calculated based on the signal specified by the simulation neural network, and a teacher signal is calculated based on this teacher signal. A vehicle control method characterized by modifying weights of a main neural network.