JPH08106564A - Metal piece discriminating device - Google Patents

Abstract

PURPOSE: To shorten the time required for calculating feature amount by making the classification of a metal piece be recognized from the waveform data of the metal piece of an unknown classification by using a neural network where learning is completed. CONSTITUTION: The calculation that a metal piece 1 is discriminated from sensor waveforms 40A and 40B by using sensors 4A and 4B which are capable of detecting the distance up to the metal piece 1 and the projecting and recessing parts of the metal piece 1 is performed by using a neural network 12. The structure of this neural network 12 determines the range to be used for a discrimination of the sensor waveform data for which a sampling is performed, performs an N-division for the range and imparts the range to an input layer. A prescribed waveform signal is stored as data and the neural network 12 is made to learn this waveform signal data. When the data that the sensor signal waveform data obtained for the unknown metal piece 1 to be discriminated is worked is inputted in the neural network 12 where the learning is completed, a discrimination is performed because only the output showing the kind of the metal piece 1 shows a large value.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an apparatus for identifying the shape of a metal piece using a coil sensor, such as a coin sorting apparatus used in a vending machine. In the drawings, the same reference numerals indicate the same or corresponding parts.

[0002]

2. Description of the Related Art As a prior art closest to the present invention, there is Japanese Patent Application No. 6-113276 "Metal Fragment Identification Device" which is a prior application of the present applicant. Next, the contents of this prior application will be briefly described. FIG. 7 shows a configuration example of a 4-sensor type coin sorting device in the above-mentioned prior application. Note that FIG. 1A is a configuration diagram of the entire apparatus, and FIG. 1B is a cross-sectional view of a sensor portion of the coin passage.

In the figure, reference numeral 1 is a coin (generally referred to as a metal piece in the following), 2 is a coin passage, and 3 is a coin passing direction. Next, 4 (4A to 4D) is a sensor also called a coil sensor provided on the wall surfaces 2a and 2b facing the coin surface of the coin passage 2, and a coil is put in a so-called pot-shaped ferrite core having a smaller diameter than the coin 1. Is configured as one. The sensors 4A to 4D generate an alternating magnetic field. This coin sorting device extracts characteristics of the coin 1, for example, material, size, thickness, surface pattern and the like. For this reason, the sensor is arranged for each of these features and is connected to a separate detection circuit 5 (5A-5D) respectively. In general, these sensors 4 are arranged on one side or both sides of the metal piece (coin) 1 and have different connection methods (in-phase or anti-phase) and detection circuit methods.

In the example of FIG. 7, the sensors 4A and 4B are on the coin passage wall surfaces 2a and 2b, respectively (that is, on both sides of the coin 1).
The sensors 4C and 4D are arranged to face each other, and the sensors 4C and 4D are arranged on the wall surface 2a (that is, on one side of the coin 1). 6 (6
A to 6D) are A / D converters for sampling the waveforms detected by the detection circuits 5A to 5D, and 7 is the A / D converter.
The waveform data sampled by the / D converters 6A to 6D is taken in, and the feature amount representing the unevenness of the coin is calculated,
It is a CPU that outputs the identification result of the coin 1. Reference numeral 8 is a memory for storing a database and the like for this calculation of the CPU 7.

By the way, in the device of the above-mentioned prior application, as a sensor capable of detecting the distance between the metal pieces and the unevenness of the metal pieces so that the characteristics of the surface of the metal pieces can be detected, as described above, a pot-shaped ferrite having a smaller diameter than a coin A sensor with a coil in the core is used. However, the output waveform of such a sensor is detected differently even with the same waveform of the metal piece, depending on the passage route of the sensor on the surface of the metal piece and the variation in the size of the gap between the metal piece and the sensor. In addition, the metal pieces are easily affected by wear and deformation. In order to prevent such a problem, in FIG. 6, a weighting calculation means 7B is provided together with a feature amount calculation means 7A as a main means for sharing the function in the CPU 7, and a weighting function 8A is stored in the memory 8. Has been done.

The procedure for identifying a metal piece using the apparatus shown in FIG. 6 is as follows. Here, the item (2) and subsequent items are the operation procedure of the CPU 7 when actually identifying an unknown metal piece. (1) A plurality of characteristic quantities including the influence of variations in the passage routes of various kinds of known metal pieces are calculated in advance from the sensor output. Particularly, the gaps a ₁ , a ₂ , a between the metal pieces and the sensor are calculated.
_3, each · · · a _m, the size of each feature quantity, obtains a relationship between the detected frequency and the feature amount is variable based on this measurement data, the feature amount as shown in FIG. 8 the frequency of occurrence A weighting function 8A represented by a weight as a function of is created and stored in the memory 8 as described above. The weighting function 8A exists for each of the metal piece type, the feature amount type, the number of sensors, and the gap type. It should be noted that each feature amount is configured so as not to be easily influenced by the variation factor of the passing speed of the metal piece. Further, the frequency distribution of the weights of the weighting function 8A is widened as shown in FIG. 9 so that the influence of wear and deformation of the metal pieces can be absorbed. Also,
The sampling signal level at each gap is also stored in the memory 8 in advance.
To memorize it. This is used for waveform normalization.

(2) The above characteristic amount is calculated from the sensor waveform when the metal piece to be identified is passed through the metal piece passage. (3) Extract the information of the gap from a part of this feature amount,
Regarded as closest a _n to the estimated gap of the metal pieces from a plurality of gaps prepared in advance gaps of the metal pieces in this case.

(4) The signal level of the actually measured sensor waveform is
Since it is slightly different from the sampling signal level in the case of the gap a _n obtained in the item (3), the actually measured waveform is amplified or attenuated and normalized so that the signal level of the actually measured waveform of the sensor matches the signal level. (5) Each feature amount is extracted from the above-mentioned normalized waveform, and a weighting function of the gap a _n is calculated for each feature amount and a known type of metal piece (for example, 50 yen) corresponding to each feature amount. The weighting is performed for each feature amount (that is, the weight is obtained) as if the metal piece actually took the gap a _n .

(6) The sum of the weights for each feature quantity weighted as in the items (5) is obtained for each type of known metal piece, and when the sum of the weights exceeds the threshold level, the identification target The metal piece of is identified as the same kind as the metal piece of the type corresponding to the weighting function. (7) Further, when there are a plurality of types of metal pieces whose total sum of weights exceeds a certain threshold level as in the item (6), the weighted metal piece of the case where the total sum of weights is maximum is It is assumed to be the same type as the type.

[0010]

As described above, in the device of the prior application, each feature amount is obtained from the sensor waveform normalized so as to match the signal level of the gap a _n prepared closest to the estimated gap of the metal piece. Is extracted, and metal pieces of a known type (for example, 50 pieces) corresponding to the respective characteristic quantities and the respective characteristic quantities are extracted.
Known for a circle), the weighting of each feature quantity as really or metal pieces took gap a _n with a weighting function of the gap a _n, the sum of the weights of each feature quantity making this weighting Each of the metal pieces is obtained, and when the sum of the weights exceeds the threshold level, the metal piece to be identified is identified as the same kind as the metal piece of the type corresponding to the weighting function.

However, since such identification processing is large in quantity and complicated, it is possible to perform high-speed calculation using a high-speed CPU or DSP (digital signal processor). For example, a coin identification device of a vending machine, etc. There is a problem that mass production equipment is difficult to use for actual commercialization because it leads to cost increase. Therefore, the present invention is a general-purpose and inexpensive CP for identifying a metal piece using the surface shape feature amount.
An object is to provide a metal piece identifying device that can be realized by U or the like.

[0012]

In order to solve the above-mentioned problems, a metal piece identification device according to a first aspect of the present invention is a metal piece passage (2) through which a flat plate-like metal piece (1 or the like) whose type is to be identified passes. Etc.) and the wall surfaces (2a, 2b, etc.) of the passage so as to face the front and back surfaces of the metal piece with the passage interposed therebetween, depending on the distance to the unevenness of the surface of the passing metal piece. Two sensors (4A, 4B, etc.) whose output changes by
Circuits (A / D converters 6A, 6) for sampling the signal waveforms (40A, 40B, etc.) obtained from the above-mentioned sensors based on the passage of the metal piece at fixed time intervals, respectively.
B)) and a data sequence based on this sampling of the signal waveform, or a data sequence obtained by thinning out the data sequence at a predetermined interval for the two sensors (1)
2A, etc.) (hereinafter referred to as the first data string) in the input layer (1
21) to output signals (metal piece type signals 12B, etc.) indicating respective predetermined types of metal pieces, from an output layer (123, etc.), which indicates the type of the relevant metal piece to be identified. And a learned neural network (12, etc.) for outputting

A metal piece identifying device according to a second aspect is the metal piece identifying device according to the first aspect, wherein the signal waveform region of the sensor relating to the first data string input to the input layer of the neural network is: It is limited to a region (between, etc.) corresponding to a flat plate surface including minute irregularities representing the characteristics from the front end to the rear end of the metal piece when passing through the metal piece passage, and the maximum of the signal waveform of this area. Means for processing the output data of the sampling circuit such that the value and the minimum value become the normalized maximum value (1.0) and minimum value (0.0) of the first data string, respectively (data Processing part 11 etc.)
Shall be provided.

Further, a metal piece identifying device according to a third aspect is the metal piece identifying device according to the first or second aspect, which corresponds to a difference between signal waveforms of the two sensors from output data of the sampling circuit. Is obtained, and the first data sequence and the second data sequence are combined and used as an input to the input layer of the neural network.

Further, a metal piece identifying device according to a fourth aspect is the metal piece identifying device according to the first or second aspect, which corresponds to a sum of signal waveforms of the two sensors from output data of the sampling circuit. Is obtained, and the first data sequence and the third data sequence are combined and used as the input to the input layer of the neural network.

Further, a metal piece identifying device according to a fifth aspect is the metal piece identifying device according to the third aspect, further comprising the third data corresponding to the sum of the signal waveforms of the two sensors from the output data of the sampling circuit. A data string is obtained, and the first data string, the second data string, and the third data string are combined and used as an input to the input layer of the neural network.

[0017]

Operation: A sensor capable of detecting the distance to the metal piece and the unevenness of the metal piece is used, and the calculation for identifying the metal piece from the sensor waveform is performed using the neural network. The structure of this neural network determines a range to be used for identification among the sampled sensor waveform data, divides the range into N, and gives the divided range to the input layer. The number of output layers is the number of types of metal pieces to be identified (double the number when identifying the front and back sides). The optimum value for the middle layer is obtained by trial and error. And in order to let the neural network learn, prepare multiple pieces of known metal pieces to be identified, those with different wear conditions, those with scratches and those with no scratches, and actually pass many passages, The waveform signal is stored as data and the neural network is made to learn this waveform signal data.

When the data of the sensor signal waveform obtained for the unknown metal piece to be identified or the data obtained by processing this data is input to the learned neural network, the metal piece in the output of the neural network is input. The identification is performed by showing a large value only in the output indicating the type.

[0019]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG. 1 is a block diagram of the essential parts of a coin sorting apparatus as an embodiment of the present invention, and this figure corresponds to FIG. The upper part of FIG. 1 is a cross-sectional view of the metal piece (coin) passage 2 as viewed from above, and the lower part of FIG. 1 is a functional block diagram of the metal piece identifying device. The sensors according to the present invention are two sensors 4A and 4B that face each other with the coin passage 2 in between. Note that the sensor 4
A is also referred to as sensor # 1 and sensor 4B is also referred to as sensor # 2. As described above, this sensor is a magnetic coil sensor which uses a pot-shaped core smaller than the diameter of the metal piece and whose output changes depending on the unevenness of the surface of the metal piece. Sensor 4A, 4B
Are respectively connected to the detection circuits 5A and 5B, and the detected waveforms are sampled by the A / D converters 6A and 6B, respectively, and are taken into the CPU 7 via the data processing unit 11 or directly. The CPU 7 incorporates the neural network 12 which has been learned in advance and is processed.
Alternatively, direct input data is input to the neural network 12.

FIG. 2 shows the input of the neural network 12.
It is explanatory drawing of an output signal. In the figure, the control circuit 10
Is the A / D converter 6, data processing unit 11, CP of FIG.
U7, and CPU7 includes memory. The input signal 12A to the neural network 12 is A / D
Sensor (# 1) 4 consisting of sampling data of converter 6A or data obtained by thinning this data at predetermined intervals
The output waveform 40B of the sensor (# 2) 4B is composed of a data string obtained by time-sequentially dividing the output waveform 40A of A into N and sampling data of the A / D converter 6B or data obtained by thinning this data at predetermined intervals. This is a combination of a data string that is serially divided into N. In other words, the number of input points is 2
N.

Further, the metal piece type signal 12B as an output signal from the neural network 12 respectively identifies a predetermined plurality of kinds of metal pieces (further, if necessary, which side of the wall faces the front and back sides). It consists of a signal.
In this example, a neural network having a structure in which the numbers of neurons in the input layer, the intermediate layer, and the output layer described in FIG. 4 are 22, 22 and 11, respectively is shown. The number of neurons in the input layer corresponds to the number of input points of the input signal 12A, and the number of neurons in the output layer corresponds to the metal piece type signal 12B.
Corresponding to the number of types.

In FIG. 2, the data processing unit 11 causes
The case where only the waveforms between the peaks and valleys of the peak portions of the sensor output waveforms 40A and 40B are extracted and normalized waveform data is input to the neural network 12. The neural network 12 that has completed learning calculates the output using the above-mentioned waveform data string that has been input. The output signal 12B indicates the type of metal piece, and only one specific output becomes 1 and other outputs become 0. (Actually, it does not become 1,0 but a value close to 1 and a value close to 0.) The structure and features of the neural network 12 will be described below in the order of the invention.

1) Regarding the invention relating to claim 1 (referred to as the first invention): FIG. 3 is an explanatory diagram of a sensor detection waveform when the metal piece 1 passes. In this example, the sensor 4B is shown for the sake of convenience. In FIG. 1B, the position in the front-rear direction when the metal piece 1 passes over the metal piece passage 2 in the direction of arrow 3 and in front of the sensor 4B is shown. The same figure (A) shows the detected waveform output by the sensor 4B at this time. Then, in FIG.
The positions (1) to (4) on the waveform in FIG.

In FIG. 3 (B), the vertical distance between the metal piece 1 and the sensor 4B seems to change, but it is difficult to see the change in position due to the overlapping of the metal pieces 1. To prevent this, the vertical position is intentionally shifted, and the actual vertical distance is almost constant.
As shown in FIG. 3, when the metal piece 1 passes in front of the sensor 4B, the detection waveform 40B becomes a waveform close to the unevenness due to the unevenness of the surface of the metal piece 1. The waveform 40B is sampled by the A / D converter 6B at a cycle t and stored in the memory as a digital value at each sampling time.

This detection waveform 40B is obtained by the metal piece 1 and the sensor 4.
It changes depending on the positional relationship with B. The uneven waveform is detected by starting to increase from the position, reaching the maximum at the position, and changing to the distance of the metal surface up to. Position ~ is the opposite of. In the first invention, the neural network 12
Waveform sampling data to be input to
As shown in (A), the value is from the position where the increase of the waveform starts to the position where the decrease of the waveform ends.

When the number of input points is to be reduced, the sampling time is thinned out (for example, every two bits are input) as necessary. Then, considering the case where the sensor passing speed of the metal piece 1 varies, the position-to-position is equally divided into N pieces, and this waveform data string is further divided into the sensors 4A, 4
Neural network 12 is the combination of B
The input is 12A.

Next, regarding the neural network 12
explain. The neural network 12 is shown in FIG.
Generally, as shown, the input layer 121, the middle layer 122, the output
The layers 123 are configured in multiple stages. Each layer 121-1
23 represents a human brain cell called a neuron 124
All the neurons 124 between each layer are connected.
Have been combined. The degree of connection between the neurons is shown in FIG.
As shown in _i(I = 1, 2, 3, ...)
A parameter that represents the strength of the coupling called (the resistance in an electrical circuit).
Equivalent to).

The neuron 124 itself produces an output.
With a threshold for
When focusing on the input layer 124, the front layer (in this case, the input layer 12
The output of each neuron in 1) is X₁, X₂,···Also,
As the degree of connection between each neuron in the previous layer and the neuron of interest
The weight of the output X₁, X₂,, W respectively corresponding to
₁, W₂, ..., input to the target neuron
The total sum of forces (sum of products operation) Y is expressed by the following equation (1).
If the threshold value of the neuron is θ,
The output Z is, for example, a sigmoid function like the following equation (2).
Is represented by.

[0029]

[Number 1] _{Y = Σ (W i · X} i) ··· (1) Z = 1 / (1 + EXP (-Y + θ)) ··· (2) output Z of the sigmoid function (2) is, of input If the sum Y is sufficiently larger than the threshold value θ, it is 1.0, and if it is sufficiently small, it is 0.
When both 0, Y and θ are equal, it becomes 0.5. Therefore, when the total sum Y of the inputs exceeds the threshold value θ, the output Z has a significant value.

Therefore, when the input 12A in the input layer 121 is input, the neural network 12 performs the calculation based on the weight and the threshold value, and finally the output layer 123.
To output a certain value 12B. The learning of the neural network 12 means the data to be learned by the input layer 121 (hereinafter referred to as learning data, in the case of the present embodiment, N waveform data strings 12A converted from the A / D data of a known metal piece). ) Is input, and the output 12B of the output layer 123 at that time has a desired value (hereinafter, referred to as teacher data, each output 12B of the output layer 123 in the present embodiment indicates the type of the metal piece. The weight of each neuron and the threshold value of each neuron are adjusted so that only the desired type of output is 1 and the others are 0) 12C.
For this learning, a large number of sets of learning data 12A and teacher data 12C are prepared, and the difference 1 between the output 12B of the neural network 12 and the teacher data 12C is 1 for all data sets.
Learning is completed when 2D becomes a certain value or less.

Here, the learning data 12A and the teacher data 1
2C is obtained by actually passing a piece of metal 1 of a known type to be identified through the portion of the sensor 4A, 4B of the metal piece passage 2. Since the metal piece 1 rotates when passing through the sensor unit, a large amount of data is randomly used for learning. Further, in order to remove the influence of the difference in the wear state of the metal pieces and the presence or absence of scratches, many pieces of the same type of metal that can be actually used in various wear and scratch states are prepared and data is obtained.

By using the neural network learned as described above, the metal piece is identified from the sensor waveform. 2) Regarding the invention related to claim 2 (referred to as a second invention): In the second invention, the waveform data of the positions up to and are not used in the sensor waveform 40B of FIG. To do. Then, as shown in FIG. 5, the data processing unit 11 of FIG. 1 finds the minimum value and the maximum value in the waveform data of the position to, which is the maximum value (1.0) and the minimum value of the input 12A of the neural network 12. The value is normalized so as to be a value (0.0) and input to the neural network 12.

In this way, by enlarging the sensor waveform of the uneven portion on the surface of the metal piece, it becomes possible to emphasize and identify the features of the unevenness. 3) Regarding the invention relating to claims 3, 4 and 5, FIG. 6 shows the positional relationship of the metal piece 1 passing on the metal piece passage 2. Here, the width of the metal piece passage 2 is D, and the average thickness of the surface of the metal piece 1 is d. The signal detected by the sensor 4 is used within a range in which it linearly changes with respect to the distance between the sensor 4 and the metal piece 1.

Further, the position in the passage direction of the metal piece (left-right direction in the figure) with respect to the sensors 4A, 4B arranged on both sides of the metal piece passage 2 is x, and the metal piece passage wall surfaces 2a, 2b to the metal piece 1 are located. The values of the detection signal waveforms 40A and 40B when the distance in the thickness direction of the metal piece (vertical direction in the figure) is y are respectively Va.
(X, y), Vb (x, y), and the thickness of the metal piece 1 is a function d (x) of x, although the metal surface changes depending on the shape of the surface unevenness in the x direction.

Assuming now that the metal piece 1 passes through the center of the passage 2, the values Va and Vb of the detection signal waveforms of the sensors 4A and 4B at the position x are expressed by the following equation (3).

[0036]

## EQU00002 ## Va (x, y) = Vb (x, y) = Va ((D-d (x)) / 2) (3) Further, the center of the metal piece 1 is from the center of the passage 2. Assuming that the signals have passed by Lc, the values Va and Vb of the detection signal waveforms of the respective sensors 4A and 4B are expressed by the following equations (4) and (5).

[0037]

## EQU3 ## Va (x, y) = Va ((D-d (x)) / 2-Lc) (4) Vb (x, y) = Va ((D-d (x)) / 2 + Lc) (5) Since the sensors 4A and 4B are used in a linear range with respect to the distance y, equations (4) and (5) are expressed by the following equations (6) and (7), respectively. Can be rewritten as

[0038]

## EQU00004 ## Va (x, y) = Va ((D-d (x)) / 2) -Va (Lc) (6) Vb (x, y) = Va ((D-d (x )) / 2) + Va (Lc) (7) Therefore, the following equations (8) and (9) are obtained by taking the sum and difference of equation (6) and equation (7).

[0039]

## EQU5 ## | Va + Vb | = 2Va ((D-d (x)) / 2) ... (8) | Va-Vb | = 2Va (Lc) ... (9) Here, the difference (9 ) Is a function only for the positional deviation Lc of the metal piece from the center of the passage in the thickness direction, and the equation (8) of the sum is a function only for the thickness d because D is known.

Therefore, by adding the difference between the waveform signals of the sensors 4A and 4B to the signal waveform input to the neural network 12 in the first invention or the second invention to learn, the position where the metal piece passes is the thickness. Even if there is a variation in the direction, the neural network can determine and correct the positional deviation Lc. In order to perform this learning, in addition to the above learning, the positional deviation Lc of the metal piece to be identified from the center of the passage is variously changed.

The signal waveform input to the neural network 12 in the first invention or the second invention is further added to the sensor 4.
By adding and learning the sum of the waveform signals of A and 4B, it becomes easy or possible to determine the thickness of the metal piece by the neural network. Particularly, in the case of the second invention, the thickness obtained by normalizing the sensor waveform The problem that cannot be identified can be solved. In order to perform this learning, in addition to the above-described learning, particularly various variations in the thickness of the metal piece to be identified are selected.

In addition, the signal waveform input to the neural network according to the first or second invention is added to the sensor 4A,
By additionally learning the difference and the sum of the waveform signals of 4B, it is possible to further enhance the discrimination ability of the neural network with respect to the metal piece.

[0043]

According to the present invention, the sensors capable of detecting the unevenness of the surface of the metal piece are opposed to each other across the metal piece passage, and the waveform data and the type of the metal piece when the metal piece is passed through the passage are classified. Since the neural network is made to learn the relationship in advance using a known type of metal piece, and the learned neural network is used to identify the type of the metal piece from the waveform data of the unknown type of metal piece, the sensor The enormous amount of time required to calculate the feature amount from the waveform data can be significantly reduced. Further, since the neural network is made to learn the difference and the sum of the sensor outputs facing each other, it is possible to identify the variation in the passing position of the metal piece and the difference in the thickness.

[Brief description of drawings]

FIG. 1 is a diagram showing a main configuration of a metal piece identifying device as an embodiment of the present invention.

FIG. 2 is an explanatory diagram of input / output signals of the neural network.

FIG. 3 is an explanatory view of the relationship between the position of the metal piece and the sensor detection waveform.

FIG. 4 is an explanatory diagram of the structure of the neural network.

FIG. 5 is an explanatory diagram of processing of a sensor detection waveform similarly.

FIG. 6 is an explanatory diagram of variables for analytical expression of sensor output.

FIG. 7 is a diagram showing a configuration example of a main part of a conventional metal piece identifying device.

8 is a diagram showing an example of the weighting function of FIG.

FIG. 9 is a diagram showing another example of the weighting function.

[Explanation of symbols]

1 metal piece (coin) 2 metal piece passage (coin passage) 2a, 2b wall surface 3 passing direction 4 (4A, 4B) sensor 4A sensor # 1 4B sensor # 2 5 (5A, 5B) detection circuit 6 (6A, 6B) A / D converter 7 CPU 10 Control circuit 11 Data processing unit 12 Neural network 12A Neural network input signal 12B Neural network output signal (metal piece type signal) 40 (40A, 40B) Sensor output waveform 121 Input layer 122 Intermediate layer 123 Output layer 124 neuron

Claims (5)

[Claims] 1. A metal piece passage through which a flat plate-like metal piece whose type should be identified passes, and a metal piece passage disposed on the wall surface of the passage so as to face the front and back surfaces of the metal piece with the passage interposed therebetween. Two sensors whose output changes according to the distance to the unevenness of the surface of the metal piece inside, and a circuit for sampling the signal waveforms obtained from the respective sensors based on the passage of the metal piece at fixed intervals. And a data string based on this sampling of the signal waveform, or the data string thinned out at a predetermined interval.
A data string (hereinafter, referred to as a first data string) formed by combining two sensors is input to the input layer, and the output layer capable of outputting a signal indicating each of a plurality of predetermined types of metal pieces is used to identify the metal to be identified. A learned piece of neural network that outputs a signal indicating the type of piece, the piece identification device. 2. The metal piece identifying device according to claim 1, wherein the signal waveform area of the sensor relating to the first data string input to the input layer of the neural network is passed through the metal piece passage. The area of the metal piece is limited to the area corresponding to the flat plate surface including minute irregularities representing the characteristics from the front edge to the rear edge, and the maximum and minimum values of the signal waveform in this area are normalized respectively in the first data string. A metal piece identifying device comprising means for processing the output data of the sampling circuit so as to obtain the maximum value and the minimum value. 3. The metal piece identifying device according to claim 1, wherein a second data string corresponding to a difference between signal waveforms of the two sensors is obtained from the output data of the sampling circuit, and the first data string is obtained. A piece of metal identifying device characterized in that a data string and a second data string are combined and used as an input to an input layer of the neural network. 4. The metal piece identifying device according to claim 1, wherein a third data string corresponding to a sum of signal waveforms of the two sensors is obtained from output data of the sampling circuit, and the first data string is obtained. A metal piece identifying device characterized in that a data string and a third data string are combined and used as an input to an input layer of the neural network. 5. The metal piece identifying device according to claim 3, wherein a third data string corresponding to a sum of signal waveforms of the two sensors is obtained from output data of the sampling circuit, and the first data is obtained. A metal piece identifying device, characterized in that a row, a second data row and a third data row are combined and used as an input to an input layer of the neural network.