JPH08137509A - Control method using map prepared by utilizing neural network - Google Patents

Abstract

PURPOSE: To provide a control method using a map prepared by utilizing a neural network capable of letting a user clearly recognize a processing by the neural network and the processing result and stably outputting an appropriate processing result no matter what kind of input is supplied to the neural network. CONSTITUTION: All the sets of six values normalized corresponding to a car speed for which 0.0-1.0 is equally divided into five and the six values for which division into four frequency ranges, for instance, is performed, road noise is averaged for the respective frequency ranges, then normalization is performed corresponding frequencies and 0.0-1.0 is equally divided into five, are inputted to the neural network and learned. By using the learned network, road surface conditions corresponding to the input values of all the sets are discriminated and the discriminated results are turned to the map corresponding to the input value of the set and stored in a memory. Then, by using the map, the road surface state is discriminated based on the values normalized corresponding to the car speed detected by a car speed detection means and normalized corresponding the road noise detected by a road noise detection means.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a control method using a map created by using a neural network.

[0002]

2. Description of the Related Art Conventionally, as a control method using a neural network, there has been proposed a so-called road surface condition detecting device for detecting a condition of a road surface on which a vehicle is traveling based on road noise. As such a road surface condition detecting device, the applicant of the present application detects the vehicle speed and the road noise generated from the wheels as a running sound as parameters indicating the running road surface condition in Japanese Patent Application No. 5-340052, and detects the road surface condition. We proposed a road surface condition detection device that determines the road surface condition using a neural network from the pattern of each frequency component of road noise.

[0003]

However, in the above-mentioned conventional control method using a neural network, particularly in the road surface state detecting apparatus, the processing after detecting the road noise is entirely left to the neural network. It cannot be said that unexpected processing results are not output depending on the values input to the network.
For example, if the frequency component of the input road noise deviates from the learning data range of the neural network due to noise or the like, there arises a problem that an appropriate processing result cannot be obtained. In addition, the user could not know whether the processing result by the neural network was appropriate.

The present invention has been made in view of the above problems, and makes it possible for a user to clearly understand the processing by a neural network and the processing result thereof, and to perform proper processing regardless of any input to the neural network. An object of the present invention is to provide a control method using a map created by using a neural network capable of stably outputting a result.

[0005]

In order to achieve the above object, the present invention selects a parameter to be given to a neural network as teacher data, learns by giving the selected parameter to the neural network, and the learning is performed. The output for the value of the parameter is obtained using the generated neural network, the correspondence between the value of the parameter and the output value from the neural network corresponding to the value is mapped and stored in the storage means, and the parameter is stored. When a parameter is detected by the parameter detecting means for detecting, a corresponding output value is read from the stored map according to the parameter value, and the controlled object is controlled based on the output value.

[0006]

According to the structure of the present invention, when a parameter is detected by the parameter detecting means, the corresponding output value is read from the map stored in the storing means according to the parameter value, and the output value is read. The controlled object is controlled based on this.

[0007]

Embodiments of the present invention will now be described in detail with reference to the drawings.

FIG. 1 is a block diagram showing a schematic configuration of a road surface state detecting device when a control method using a map created by using a neural network according to an embodiment of the present invention is applied to the device. is there.

In the figure, reference numeral 1 is a road noise detecting means for detecting road noise, which is constituted by a microphone in this embodiment. On the output side of the road noise detecting means 1, an amplifier 2 for amplifying the road noise detected by the road noise detecting means 1 and a filter 3 for analyzing the frequency of the road noise amplified by the amplifier 2, As will be described later in detail, the determination unit 4 for determining the road surface state is connected to one input side of the determination unit 4 using a map created by a neural network, and the other input side of the determination unit 4 detects the speed of the vehicle. The output side of the vehicle speed detecting means 5 is connected. The determination unit 4 determines the road surface state according to these two input signals, and outputs the determination results to an alarm device and a motion control device (neither is shown).

FIG. 2 is a diagram for explaining the arrangement positions of the road noise detecting means 1 and the vehicle speed detecting means 5.

In FIG. 1, it is assumed that the vehicle C is a front engine vehicle, and the road noise detecting means 1 which is a microphone has stones or stones on at least one of the inside of the wheel house of the left and right rear wheels, which is less affected by engine noise. The vehicle speed detecting means 5 is arranged so as not to be directly hit by water, and the vehicle speed detecting means 5 is arranged at a predetermined position in at least one of the left and right sides in the rear wheel. Here, the vehicle speed detecting means 5 generates an electric pulse signal according to the wheel speed.

FIG. 3 is an enlarged view of the main part of the region r in FIG. 2, and is a diagram for explaining the structure in the vicinity of the microphone 1 in detail.

In FIG. 1, (a) shows a microphone 1
Is a diagram showing a case where the wheel house 11 is mounted at a predetermined position on the inner side of the vehicle interior. For example, when the entire wheel house 11 is made of steel, the vicinity of the mounting portion of the microphone 11 to which road noise is transmitted is shown. The wheel house portion is cut out, and another member 11 ₁ such as a rubber grommet is embedded in this portion.
In this way, the high frequency component of the road noise, that is, the frequency component unnecessary for determining the road surface condition is removed.
It is possible to accurately detect the road surface condition. Further, when a resin inner fender (not shown) is provided outside the vehicle interior of the wheel house 11, in order to easily detect road noise from the microphone 1, the inner fender is provided near the mounting portion of the microphone 1. Make sure to cut out the part. Further, when the wheel house 11 is coated for rust prevention, for the same reason as the above-mentioned inner fender, if a portion near the microphone 11 is cut out and another member is embedded in the cutout portion, road noise will be reduced. Transparency will change.

FIG. 3B shows the diaphragm 1 of the microphone 1.
It is a figure which shows the case where it is directly attached to the iron plate of the _1- wheel house 11, and when doing so, the diaphragm ₁₁ becomes the tire 1
Since it is in the immediate vicinity of 2, it becomes easy to detect road noise, and it is also possible to directly detect water splashing and the like.

FIG. 4 is a diagram for explaining a method of making a neural network learn. As shown in FIG. 4, road noise from wheels is sampled by the road noise detecting means 1 and sampled. After frequency analysis of the data, it is input to the input layer (not shown) of the neural network together with the vehicle speed data. Then, each element (weight) of the network is determined so that the current road surface state matches the road surface type based on the final output (probability) from the output layer (not shown).

FIG. 5 is a flow chart showing a procedure of a subroutine of the above-mentioned learning processing (hereinafter referred to as "learning mode") and processing for actually applying the learned network (hereinafter referred to as "execution mode"). This subroutine is executed by the CPU (not shown) built in the determination unit 4 of FIG.

First, in step S1, the CPU, memory (not shown), etc. are initialized (initialized),
In step S2, parameter data such as the type of neural network (for example, the number of layers, the number of cells in each layer, etc.), the weight initially set, and the information input to the input layer are input.

Next, in step S3, it is determined whether the operator has selected the "learning mode" or the "execution mode". When the "learning mode" is selected, the process proceeds to step S4, Step S5 described later
~ The upper limit number of times α for repeating the process of step S10 is compared with the current number of times of repetition N, and when the current number of times of repetition N is smaller than the number of times α, the process proceeds to step S5. In step S5, if the number of times N is "0", the above step S5 is performed.
When the number N is other than "0", the input data processing for inputting the weight changed in step S8, which will be described later, is performed.

In the following step S6, the step S5 is executed.
The output data (the road surface type) output from each cell of the output layer based on the weight input in step S2 and the information (vehicle speed and sound pressure value of a specific frequency) input in the input layer input as parameter data in step S2. TYPEn, n =
Probability of 1, 2, ...) is calculated. Further, in step S7, an error check is performed by comparing the calculated certainty (value in the range of 0 to 1) of each road surface type TYPEn with the current road surface condition, and checking the difference between the two, and in step S8, The error ε detected in step S7 is compared with a predetermined value ε ₀ , and when the error ε is smaller than the predetermined value ε ₀ , the process proceeds to step S11 described below, while when the error ε is the predetermined value ε ₀ or more, the next step S9 is performed. Proceed to.

In step S9, the weight is changed according to the result of this error check, and in step S10, the number of repetitions N is incremented by 1, and then the process returns to step S4.

On the other hand, if it is determined in step S4 that the current number of repetitions N is greater than or equal to the number α, or if the error ε is smaller than the predetermined value ε _{0 in} the determination in step S8 as described above, step S11. Then, the process proceeds to step S8, the network is saved, that is, the coupling matrix (weight matrix) based on the weight determined in step S8 is stored, and the present subroutine ends.

On the other hand, when the "execution mode" is selected in the determination in step S3, the process proceeds to step S11,
The network saved in step S10 is read out, and the weight is set in step S12 based on the network read out in step S11. In the following step S13, data of each cell in the output layer is calculated based on the information input to the input layer input in step S2 and the weight set in step S12, and in step S14, each road surface state TYPEn is calculated. After outputting the certainty of, the present subroutine processing is ended.

FIG. 6 shows a graph for determining the parameter value relating to the vehicle speed among the parameters input to each cell in the input layer in step S2 of FIG. 5, where the horizontal axis represents the vehicle speed (km / h). ), And the vertical axis represents the input parameter value.

As shown in FIG. 6, in this embodiment, the vehicle speeds are [0,30), [30,60), [60,90),
[90, 120), [120, 150), [150,
∞) is divided into 6 sections, and parameter values are 0.0, 0.2, 0.4, 0.6, 0.8 and 1.0, respectively.
Is given the value of. Needless to say, the number of divisions and the parameter values given need not be limited to this.

FIG. 7 shows a frequency analysis of the road noise detected when traveling at 60 km / h for each road surface type (dry road, bad road, wet road, and snowy road). Similarly, is used to determine the frequency parameter value input to the neural network. In FIG. 7, the horizontal axis represents frequency (Hz) and the vertical axis represents sound pressure level (dB) of road noise.

As shown in FIG. 7, the frequency domain is divided so that the characteristics of each road surface type can be discriminated.
That is, in this embodiment, the frequencies are (63, 200), (2
50,630), (800,2.0k), (2.5k,
It is divided into 4 areas (4.0k) (referred to as areas I to IV, respectively). In areas I and II, bad roads and other road surface types can be distinguished, and in area III, snow-covered roads and other road surface types can be distinguished. In the area IV, the wet road (WET) can be distinguished from other road surface types. In this example, (0, 63),
(200,250), (630,800), (2.0
This is because the frequencies in the regions (k, 2.5k) and (4.0k, ∞) are not used, but these frequencies are determined to be sound components other than road noise. For example, (2
00,250), (630,800), (2.0k,
Sounds belonging to the area of 2.5k) are considered to be those of the engine sound, and (0, 63), (4.0k, ∞)
Sounds belonging to the area of are considered to be frequency components that are ineffective for detecting the road surface condition, that is, noise, and by removing these components, the characteristics of each road surface can be better expressed. As a result, learning to the neural network can be effectively performed, and the road surface condition can be detected better.

Further, the frequency domain is divided as described above, and each time the road noise is detected from the microphone 1, the average sound pressure level for each frequency domain is calculated. For example, the average value of the road noise signal S detected on the rough road in the area II is 98.3 dB, and if 110 dB is taken as the maximum sound pressure level, 98.3 / 110≈0.8.
9, and this value is input as frequency data in the area II to a predetermined cell in the input layer of the neural network. Thus, by averaging the sound pressure level of the road noise for each of the divided frequency regions, noise other than the road noise can be removed, and the road surface condition can be better detected. The area of road noise may be calculated for each frequency region without being limited to this.

FIG. 8 is a diagram showing a map in which the input / output relations with respect to the neural network are mapped and stored in the memory. The input parameters include the vehicle speed described in FIG. 6 and the four frequency regions described in FIG. Sound pressure level is used.

As described with reference to FIG. 6, the vehicle speed is 0,0.
Six values of 2, 0.4, 0.6, 0.8, and 1.0 are taken, and as described in FIG. 7, the sound pressure level is also the same.
Since it is configured to take six values of 0, 0.2, 0.4, 0.6, 0.8, 1.0, the map is 6 ⁶
It has a set of input parameters. Then, after the input parameter values of each set are actually applied to the neural network, that is, after the neural network is trained by the learning mode of FIG. 5, the road surface state is discriminated by the execution mode, and the discrimination result is inputted. It is mapped corresponding to a set of parameters and stored in memory.

Using the map created in this way, the discriminating unit 4 in FIG. 1 determines the road surface condition from the road noise detected by the microphone 1 and subjected to frequency analysis and the vehicle speed detected by the vehicle speed detecting means 5. Determine.

As described above, according to this embodiment,
Since the road surface condition judged by the neural network is mapped in correspondence with the input parameters, the processing by the neural network and the processing result can be clearly understood by the user, and any input to the neural network is given. It is possible to stably output an appropriate processing result. Further, since the road surface condition is determined by the map, it is not necessary to perform the calculation by the neural network at the time of control, and the control processing time can be shortened.

In the present embodiment, the average value or area calculated for each frequency region is normalized by the maximum sound pressure level and input to the neural network. However, the present invention is not limited to this, and the calculated average value or area is not limited to this. May be coded and then input to the neural network.

FIG. 9 is a diagram showing an example of a correspondence table for coding the vehicle speed and the sound pressure level. In this embodiment, the vehicle speed in the range of 0 to 120 km / h is divided every 8 km / h. , The sound pressure level in the range of 60 to 120 dB is 4 dB.
It is divided for each and a 4-bit code is assigned to each. For example, at a vehicle speed of 40 km / h, “0101
The code "B" is assigned, and the code "1010B" is assigned to the sound pressure level of 100 dB.
"B" indicates that the value immediately before it is a binary number. When coded in this way, the difference between the parameter values of the same type input to the neural network becomes clear, and the convergence in the learning mode of FIG. 5 becomes faster (that is, the error ε is a predetermined value with a small number of iterations N). ε
_Less than ₀ ).

In this embodiment, the example in which the control method of the present invention is applied to the detection of the road surface condition has been described, but the present invention is not limited to this and may be applied to the control of other controlled objects.

[0035]

As described above, according to the present invention,
A parameter to be given to the neural network as teacher data is selected, learning is performed by giving the selected parameter to the neural network, an output corresponding to the value of the parameter is obtained using the learned neural network, and When a parameter is detected by the parameter detecting means for detecting the parameter by mapping the correspondence between the value and the output value from the neural network corresponding to the value and storing the result in the storage means, the parameter is detected according to the parameter value. Since the corresponding output value is read from the stored map and the controlled object is controlled based on the output value, it is possible to let the user clearly understand the processing by the neural network and the processing result, and Force proper processing result be given to can be stably output, an effect which is possible to a shortened further control the processing time.

[Brief description of drawings]

FIG. 1 is a block diagram showing a schematic configuration of a road surface state detection device when a control method using a map created by using a neural network according to an embodiment of the present invention is applied to the device.

FIG. 2 is a diagram illustrating the positions where the road noise detecting means and the vehicle speed detecting means in FIG. 1 are arranged.

FIG. 3 is an enlarged view of a main part of a region r in FIG. 2 and is a diagram for explaining in detail a structure near a microphone.

FIG. 4 is a diagram for explaining a method for making a neural network learn.

FIG. 5 is a flowchart showing a procedure of a subroutine of learning mode processing and execution mode processing.

FIG. 6 is a diagram showing a graph for determining the value of a parameter related to vehicle speed among parameters input to each cell of the input layer of the neural network in step S2 of FIG.

FIG. 7 is a diagram showing frequency analysis results of road noise detected when traveling at 60 km / h for each road surface type.

FIG. 8 is a diagram showing a map in which an input / output relationship for a neural network is mapped and stored in a memory.

FIG. 9 is a diagram showing an example of a correspondence table for encoding vehicle speed and sound pressure level.

[Explanation of symbols]

 1 Road Noise Detecting Means (Parameter Detecting Means) 4 Judging Section (Memory Means) 5 Vehicle Speed Detecting Means (Parameter Detecting Means)

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Atsushi Itagaki 1-4-1 Chuo, Wako-shi, Saitama Inside the Honda R & D Co., Ltd. (72) Inventor Hideki Kubota 4-1-1 Chuo, Wako, Saitama Stock Company Honda Technical Research Institute

Claims (1)

[Claims] 1. A parameter to be given to a neural network as teacher data is selected, learning is performed by giving the selected parameter to the neural network, and an output corresponding to the value of the parameter is obtained by using the learned neural network. Obtained, the correspondence between the value of the parameter and the output value from the neural network corresponding to the value is mapped and stored in the storage means, and the parameter is detected by the parameter detection means for detecting the parameter, A control method using a map created by using a neural network, wherein a corresponding output value is read from the stored map according to a parameter value, and a control target is controlled based on the output value.