JP2972939B2 - Ultrasonic object discriminating apparatus, ultrasonic object discriminating method, and method for forming neural network in ultrasonic object discriminating apparatus - Google Patents

Description

The present invention relates to an ultrasonic object discriminating apparatus for discriminating an object using an ultrasonic sensor, an ultrasonic object discriminating method, and a neural network. The present invention relates to a method for forming a neural network in an ultrasonic object discriminating apparatus.

[Conventional technology]

2. Description of the Related Art Conventionally, as an ultrasonic object discriminating apparatus, there is an apparatus that discriminates whether an object is metal or rock by giving a spectrum of a sonar reflected sound to a neural network.

[Problems to be solved by the invention]

However, according to such a conventional ultrasonic object discriminating apparatus, it is difficult to generate data on the spectrum of the sonar reflected sound, and there has been a problem that the configuration is complicated.

[Means for solving the problem]

The present invention has been proposed to solve such a problem, and a first invention (the invention according to claim 1) is an ultrasonic wave receiving apparatus which applies ultrasonic waves to an object and receives reflected ultrasonic waves from the object. An ultrasonic object discriminator comprising: an ultrasonic sensor; conversion means for converting the output of the ultrasonic sensor into distance information from the ultrasonic sensor to the object; and sensitivity information indicating the intensity of reflected ultrasonic waves; and a neural network. Using the device, distance information and sensitivity information obtained from the conversion means are given to the input layer of the neural network, and each unit of the output layer of the neural network is defined as a unit representing each object to be determined. The type of the target object is determined on the basis of the output.

A second invention (an invention according to claim 2) is directed to an ultrasonic sensor for applying ultrasonic waves to a target and receiving reflected ultrasonic waves from the target, and outputting an output of the ultrasonic sensor from the ultrasonic sensor to the target. Conversion means for converting the distance information and sensitivity information representing the intensity of the reflected ultrasonic wave; and object discrimination means (neural network) for discriminating the type of the object based on the distance information and sensitivity information obtained by the conversion means. To
This is provided in an ultrasonic object discriminating apparatus.

A third invention (an invention according to a third invention) is a method for forming a neural network in the ultrasonic object discriminating apparatus according to the first invention or the second invention, wherein a conversion means for each object measured as a sample is used. Distance information and sensitivity information are used as teacher information, and the teacher information is provided to the input layer of the neural network, and an error between an output of the output layer of the neural network obtained by this and an ideal output corresponding to each object is used. A weighting coefficient between each layer unit is changed according to a learning rule to form a neural network that finally eliminates an error from the ideal output.

[Action]

Therefore, according to the present invention, the type of the target object can be determined via the neural network based on the distance information and the sensitivity information for which data can be easily generated.

〔Example〕

 Hereinafter, the present invention will be described in detail.

FIG. 1A is a schematic configuration diagram showing a sensor unit in the ultrasonic object discriminating apparatus according to the present invention.

The sensor unit 10 applies an ultrasonic wave to the object X,
An ultrasonic sensor 1 that receives reflected ultrasonic waves from the ultrasonic sensor 1 and a receiving sensitivity (hereinafter, referred to as a distance) from the ultrasonic sensor 1 to the object X and an intensity of the reflected ultrasonic waves.
(To simply call sensitivity). By using the sensor unit 10, the distance from the ultrasonic sensor 1 to the object X and the sensitivity indicating the intensity of the reflected ultrasonic wave from the object X can be measured.

A system using the sensor unit 10 was actually constructed, and distance-sensitivity characteristics of various objects were measured.

FIG. 2 shows the distance-sensitivity characteristics of each object actually measured. In the actual measurement, the object X is placed as flat as possible in an area of about 60 cm square, and the sensitivity is measured while changing the inclination of the ultrasonic sensor 1 at a predetermined distance from above, and the sensitivity is maximized. At that time, the sensitivity was taken as the sensitivity at that distance. Here, the sensitivity is obtained by applying a certain conversion to the intensity voltage of the reflected ultrasonic wave, and is about 10 dB at about 350 (the reference voltage of the reflected ultrasonic wave is about 50 mmV), 20 dB at 500, and
It is almost proportional at the rate of 0.1dB / digit. For example, 570
Then it shows about 27dB. FIG. 2 also shows that the characteristics differ depending on the object. Fig. 3 shows sand with different particle sizes.
It is a distance-sensitivity characteristic of soil. FIG. 4 shows distance-sensitivity characteristics of cloth files having different surface roughness. As shown in FIG. 3 and FIG. 4, it can be seen that the distance-sensitivity characteristic changes when the particle size and the surface roughness are different even for the same material.

Therefore, if these characteristics are measured in advance for each object, for example, when a certain object is placed in the dark, the ultrasonic sensor 1 is held over the object. By measuring the distance and sensitivity by using the distance-sensitivity characteristics measured in advance, the type of the object, that is, the type of the object (the type of the object, the roughness of the surface of the object, etc.) Can be determined.

As described above, by using the sensor unit 10 shown in FIG. 1A, the type of the object can be determined based on the distance information and the sensitivity information that can easily generate data. This makes it possible to simplify the configuration. Further, if the sensor unit 10 is used, the type of the object X can be easily determined even if the distance between the ultrasonic sensor 1 and the object X changes.

In the case where the characteristics are clearly divided depending on the object, such as the distance-sensitivity characteristics of sand and soil having different particle diameters shown in FIG. 3, the discrimination range is set on the graph of the distance-sensitivity characteristics. It is possible to determine the object based on whether or not it is within the range. However, in the case of discriminating an object such as a floor, an acrylic board, or a curtain shown in FIG. It is necessary to set a determination range, and setting of the determination range becomes very difficult.

Therefore, in the ultrasonic object discriminating apparatus according to the present invention,
As shown in FIG. 1 (b), by combining the sensor unit 10 and the neural network 3, even in the case of an object whose characteristic lines are close to each other, an artificially higher order discrimination range can be obtained. Is not set, and the type can be accurately determined.

The present apparatus 11 includes an ultrasonic sensor 1 that applies ultrasonic waves to an object X and receives reflected ultrasonic waves from the object X, and outputs an output of the ultrasonic sensor 1 to a distance from the ultrasonic sensor 1 to the object X and The apparatus includes a conversion device 2 for converting the intensity of the reflected ultrasonic wave into sensitivity, and a neural network 3. The neural network 3 includes an input layer 3-1 to which distance information and sensitivity information output from the conversion device 2 are input, an intermediate layer 3-2 to which an output of the input layer 3-1 is input, and an intermediate layer 3-2. Has an output layer 3-3 to which the output of the above is input.
Each of the layers 3-1 to 3-2 is a set of units U, and the units of the adjacent layers are connected without any loss. Through this connection, information is transmitted from the unit U of the preceding layer to the unit U of the following layer. At this time, each connection has a unique weighting factor (determined by a back propagation learning rule described later), and the value obtained by multiplying the output value of the preceding unit by the above weighting factor is used as the input value to the subsequent unit. Also, the sum of input values from all units in the preceding layer is calculated, and a value obtained by performing a non-linear transformation defined as an output function is used as an output value of the subsequent unit.

In the present embodiment, the number of units in the input layer 3-1 is two, and distance information and sensitivity information, which are output values of the conversion device 2, are given to each unit U in the input layer 3-1. The number of units of the intermediate layer 3-2 is set to an arbitrary number.
The output layer 3-3 has a unit U corresponding to each object to be determined. In the present embodiment, seven units U are provided so that seven objects can be distinguished. The number of objects to be determined can be freely set by changing the number of units in the output layer 3-3. In the output layer 3-3, the positive output from each unit U is
It indicates that the unit U has recognized the representative object. If a value of “1” is approximately output from a certain unit U in the output layer 3-3, it means that it is recognized that “the measurement target is a target represented by the unit U”. I do.

The object discrimination by the neural network 3 is roughly divided into two processes. The first is the learning process. Here, distance information and sensitivity information in the conversion device 2 for each measured object as a sample are used as teacher information, and the teacher information is given to the input layer 3-1 of the neural network 3, and the neural network 3 obtained by this is obtained. Of the output layer 3-3 and the ideal output (the unit U representing the measured object)
Output of "1", output of other units U is "0")
The neural network 3 is improved using the error of
Form the correct neural network 3. The improvement of the neural network 3 is performed by a method called a back propagation learning rule (back propagation). That is,
The weight coefficient between each layer unit is changed by back propagation, and finally the conversion device 2 for each object
When the distance information and the sensitivity information are given, only the unit U representing the correct object outputs "1", that is, the neural network 3 is formed such that there is no error from the ideal output. In the second process, using the neural network 3 completed by the above learning, the distance information and the sensitivity information of the unknown object in the conversion device 2 are given to the input layer 3-1. The object is determined from the output.

Therefore, if the distance-sensitivity characteristics unique to each object as shown in FIGS. 2 to 4 are given to the neural network 3 and learned by back propagation, the neural network 3 can accurately determine the type of the object. After the network 3 is formed and the learning is completed, the distance information and the sensitivity information for the unknown object with respect to the unknown object are given to the neural network 3 to immediately determine exactly what the object is. Can be.

In the above-described embodiment, the intermediate layer 3-2 is set to 1
Although a layer is used, a multilayer structure may be used. Further, instead of the back propagation, another learning rule may be used.

〔The invention's effect〕

As is apparent from the above description, according to the present invention, based on distance information and sensitivity information that can be easily generated, based on a neural network,
The type of the object (the type of the object, the roughness of the surface of the object, and the like) can be accurately determined, and the configuration is simplified as compared with the related art.

Further, according to the present invention, it is possible to easily determine various objects even when the distance between the ultrasonic sensor and the object changes.

[Brief description of the drawings]

FIG. 1A is a schematic configuration diagram showing a sensor unit in the ultrasonic object discriminating apparatus according to the present invention, and FIG. 1B is an embodiment of the ultrasonic object discriminating apparatus according to the present invention. FIG. 2 is a diagram showing a distance-sensitivity characteristic of each object,
FIG. 3 is a diagram showing the distance-sensitivity characteristics of sand and soil having different particle diameters, and FIG. 4 is a diagram showing the distance-sensitivity characteristics of cloth files having different surface roughness. 1 ... Ultrasonic sensor, 2 ... Converter, 3 ... Neural network, 3-1 ... Input layer, 3-2 ... Intermediate layer, 3-3 ... Output layer, U ... Unit, X ... … Object, 10… Sensor, 11… Ultrasonic type object discriminating device.

Continuation of the front page (72) Inventor Kenichi Hanari 1-27-1, Kichijoji Minamicho, Musashino City, Tokyo Inside NTA Advanced Technology Co., Ltd. (56) References JP-A-2-39286 (JP, A (58) Field surveyed (Int. Cl. ⁶ , DB name) G01V 1/00 G01S 15/04 G01S 15/88

Claims (3)

(57) [Claims] An ultrasonic sensor for applying an ultrasonic wave to an object and receiving reflected ultrasonic waves from the object, and outputs an output of the ultrasonic sensor to distance information from the ultrasonic sensor to the object and to the object. A conversion unit that converts the intensity of the reflected ultrasonic waves into sensitivity information; an input layer; an intermediate layer having an output of the input layer as an input; and an output layer having an output of the intermediate layer as an input. Is a set of, the units of adjacent layers are completely connected, and each connection has a unique weighting factor, the output value of the previous layer unit multiplied by the weighting factor to the subsequent layer unit And a neural network that calculates the sum of the input values from all the units in the previous layer and performs a non-linear transformation defined as an output function to obtain the output value of the unit in the subsequent layer. Super sound A distance information and a sensitivity information obtained by the conversion means are given to an input layer of the neural network using a formula object discriminating device, and a unit representing each object to be discriminated from each unit of the output layer of the neural network. And determining the type of the object based on the output from the unit. 2. An ultrasonic sensor for irradiating an ultrasonic wave to an object and receiving reflected ultrasonic waves from the object, and outputs an output of the ultrasonic sensor to distance information from the ultrasonic sensor to the object and the ultrasonic sensor. A conversion unit that converts the intensity of the reflected ultrasonic wave into sensitivity information; and an object determination unit that determines the type of the object based on the distance information and the sensitivity information obtained by the conversion unit. Has an input layer receiving the distance information and the sensitivity information, an intermediate layer receiving an output of the input layer, and an output layer receiving an output of the intermediate layer.
Each layer is a set of units, and the units of adjacent layers are completely connected to each other, and each connection has a unique weighting factor, and a value obtained by multiplying the output value of the preceding layer unit by the weighting factor is As the input value to the rear layer unit, calculate the sum of the input values from all the units in the previous layer, the value obtained by performing a non-linear transformation defined as an output function as the output value of the rear layer unit, the output An ultrasonic object discriminating apparatus, which is a neural network in which each unit in a layer is a unit representing each object to be discriminated. 3. A method for forming a neural network in an ultrasonic object discriminating apparatus according to claim 1, wherein distance information and sensitivity information in a conversion means for each object measured as a sample are used as teacher information. And giving the teacher information to the input layer of the neural network, and using the error between the output of the output layer of the neural network and the ideal output corresponding to each of the objects, the weight between the respective layer units according to a predetermined learning rule. Change the coefficient,
Eventually, a neural network that eliminates the error from the ideal output is formed. When the distance information and the sensitivity information of the unknown object by the conversion unit are given to the input layer of the neural network, A method of forming a neural network in an ultrasonic object discriminating apparatus, wherein the type of the unknown object can be accurately discriminated from the output of an output layer.