JPH04148828A - Spectroscopic inspecting apparatus - Google Patents

Abstract

PURPOSE:To make it possible to set the optimum inspecting reference automatically, to omit the man-hours for statistical analysis and to realize the shortening of inspecting time by performing judging operation based on the reference value by using a neural network. CONSTITUTION:The light which is transmitted through or reflected from a specimen 4 undergoes spectroscopic action in a diffraction grating 5. The spectrum data in one or a plurality of wavelength regions required for inspection are extracted out of the spectroscopic data by a window circuit 13. In the window circuit 13, the spectroscopic data are checked with the data stored in a window table 15, and whether the data are in the wavelength region required for the inspection or not is judged. When the data are located in the region, the data are added to the data in a memory 14 and stored. The integrated value obtained in this way is inputted into a neural network 16 which is learned with many inspecting samples beforehand. The spectroscopic data are judged with the neural network 16, and outputted as the result of the inspection.

Description

DETAILED DESCRIPTION OF THE INVENTION [Industrial Application Field 1] The present invention relates to a spectroscopic inspection device for inspecting a specimen based on the results of spectrum analysis of the specimen.

[Prior art] Conventional spectroscopic inspection equipment determines the spectral distribution of a specimen, measures the amount of fluorescence or absorbed light in one or more specific wavelength regions in the spectral distribution, and uses it as the inspection target. The content of atoms and molecules to be tested in the sample is determined to be within the standard range by comparing it with the standard value of the amount of fluorescence or absorption light in the above specific wavelength range, which is determined in advance from the spectral distribution of the sample specimen containing atoms and molecules. It is determined whether or not it is within the range.

Here, the reference values for the amount of fluorescence and the amount of absorbed light are determined by humans based on the results of statistical processing on a large number of sample specimens. Furthermore, when testing multiple types of atoms or molecules on the same specimen, the determination of whether or not the test object is within a reference range is performed separately for each test object.

[Problems to be Solved by the Invention] As mentioned above, in conventional spectroscopic inspection devices, the reference value for inspection was determined by humans using statistical methods, etc., so it was not always the best value. In order to determine the reference value, it was necessary to process a large number of sample data, which was time-consuming. On the other hand, when testing multiple types of atoms or molecules on the same sample,
The judgment time required depends on the number of types of atoms and molecules to be tested.

The present invention has been made to address the above-mentioned circumstances, and its purpose is to eliminate the need for time-consuming determination of reference values, and to
It is an object of the present invention to provide a spectroscopic inspection device capable of simultaneously performing determinations on the same specimen even when there are many types of atoms and molecules to be examined.

[Means for Solving the Problems] The outline of the present invention will be explained with reference to FIG. 1. As shown in the figure, in the present invention, spectral data in one or more wavelength regions necessary for inspection is extracted from the spectral data of the specimen, the spectral intensity in each wavelength region is integrated, and the integrated value is is input into a neural network that has been trained in advance using a large number of test samples, and this neural network judges the spectral data and outputs it as a test result.

[Operation] According to the present invention, since the judgment operation based on the reference value is performed by using a neural network, all that is required is to present sample data and give a desired judgment result for the sample data using the learning function of the neural network. Now, the neural network will automatically have the ability to make judgments, and there will be no need for humans to perform the process of determining reference values as in the past, which will greatly reduce the amount of effort required, and the reference values will also be closer to the optimal values. . Furthermore, due to the parallel nature of the neural network, it is possible to judge multiple types of atoms and molecules at the same time, reducing processing time.

[Example] Hereinafter, an example of the spectroscopic inspection apparatus according to the present invention will be described with reference to FIGS. 2 to 4. FIG. 2 is a block diagram showing the configuration of the embodiment. Light from a light source 1 passes through a slit 2 and a half mirror 3 and illuminates a specimen 4. The light transmitted or reflected by the specimen 4 is separated by the diffraction grating 5 and received by the line sensor 6. The line sensor 6 is composed of a large number of light receiving elements arranged in a straight line, and arranged so that the direction in which the light receiving elements are arranged and the direction of the grid of the diffraction grating 5 are perpendicular to each other. As a result, each light receiving element receives only spectral lights of different wavelengths.

On the other hand, a part of the illumination light is reflected by the half mirror 3 and is received by the optical sensor 7. The output of the optical sensor 7 is A/
After being converted into a digital quantity by the D converter 8, the real number converter 9
is converted into a floating point number and stored in the memory 10 as normalization data.

Data from the line sensor 6 is read out for each light receiving element.

First, the data of the first light receiving element is read out, A,
After being converted into a digital quantity by the /D converter 11, it is sent to the normalization calculator 12. The normalization calculator 12 converts the integer data sent from the A/D converter 11 into a floating point number, divides it by the data stored in the memory lO,
The division result is sent to the window circuit 13. That is, the normalization calculator 12 normalizes the spectral data according to the intensity of illumination light.

The window circuit 13 refers to the information stored in the window table 15 in advance based on the order of the data sent from the normalization calculator 12, and determines whether the data sent from the normalization calculator 12 is necessary for inspection. It is determined whether the wavelength range is within the wavelength range or the test wavelength range, and if it is, the data stored in the address corresponding to the wavelength range in the memory 14 is sent from the normalization calculator 12. Add the data that has been collected. Furthermore, if the data sent from the normalization calculator 12 does not belong to any of the wavelength ranges required for inspection, nothing is done. Note that in the initial state, the contents of all addresses in the memory 14 are set to 0.

When the series of processes described above are completed for the data of the first light receiving element of the line sensor 6, the same process is repeated for the second light receiving element. When the data of the second light receiving element is processed in the same manner as the data of the first light receiving element, the same processing as that of the data of the first and second light receiving elements is repeated for the data of the third light receiving element. .

In this way, when the above processing is completed for the data of all the light receiving elements of the line sensor 6, the memory 4
At each address, a quantity corresponding to the integral value of the spectral intensity in the corresponding wavelength region is stored.

Next, among the data in the memory 4, data at addresses corresponding to the test wavelength range are sent one after another to the neural network 16. The neural network of this example has the first
As shown in the conceptual diagram in the figure, it has a three-layer hierarchical structure, with no connections between neuron units belonging to the same layer, and only connections between neuron units in adjacent layers.

Starting from the input side, each layer is called an input layer, a middle layer, and an output layer. Each two-neelon unit performs the process according to the following equation (1).

Here, xi is the output of the i-th neuron unit that belongs to the layer one layer before (on the input layer side) the layer to which the neuron unit of interest belongs, and wi is the output of the i-th neuron unit that belongs to the layer that is one layer before the neuron unit of interest. This is a weighting coefficient (degree of synaptic connection) representing the strength of connection with the 7-1-ron unit. n is the number of all neuron units that belong to the previous layer to which the neuron unit of interest belongs, θ is the directivity between the inputs of the neuron unit of interest, y is the output, and f is shown in the following equation (2). It is a sigmoid function.

f (x) = 1/ (1+e-X) =-(
2) FIG. 3 shows the configuration of the neural network 16, and FIG. 4 shows the configuration of each neuron unit. In this example, the number of neuron units belonging to the intermediate layer is 4,
Although the number of neuron units belonging to the output layer is two, this number changes depending on the object to be inspected and the purpose of the inspection, and is not limited to this. The number of output units is determined by the number of types of atoms or molecules to be inspected, and one output unit is responsible for inspecting one type of atom or molecule. The sigmoid function in equation (2) has a function value f when X changes from -ω to +ω.
(x) changes from O to 1, so in the learning described later, if the content of the atom or molecule to be tested in the sample is within the standard range, the output unit responsible for that atom or molecule is 1. , and if it is outside the reference range, output O. Furthermore, in this example,
The memory 14 corresponds to the input layer. That is, the data corresponding to the integral value of the spectral intensity in the test wavelength region stored in the memory 14 corresponds to the output of the neuron unit belonging to the input layer.

Data for each inspection wavelength region is transferred from the memory 14 to the buffer 18.
When the data is simultaneously sent to the neuron units 20 and 21 through the neuron units 23, each neuron unit 20 to 23 receives the data previously stored in the weight memory 30 by the arithmetic unit 26, as shown in FIG. The product of the weight data of the connection with the first neuron unit of the input layer and the data corresponding to the output of the first neuron unit of the input layer sent from the memory 14 is calculated, and the product is stored in the register 27. is memorized. Next, the product of the second data in the memory 14 and the second data in the weight memory is calculated, and the sum of this product and the data already stored in the register 27 (first product) is calculated. The sum is stored in register 27 again. Thereafter, the above processing is repeated for the data for each test wavelength region supplied one after another from the memory 14, and when the processing for the last data in the memory 14 is completed, the register 27 contains Σ of equation (1).
wi-xi will be stored. Here, 1=1 n is the number of data stored in the memory 14 and corresponds to the number of input neuron units.

Next, from the contents of the register 27, the value corresponding to O in equation (1) stored in advance in the threshold value memory 29 is subtracted, and y in equation (1) is determined from the correspondence between the function value f (x) and the function value f (x) at that time, and is stored in the output buffer 28 in the unit as an output of each two-knee-long unit 20 to 23. The above processing is performed simultaneously and in parallel for each neuron unit 20-23.

The same processing as the neuron units 20 to 23 is performed on the neuron units 24 and 25 as well. However, in the case of the neuron units 24 and 25, unlike the case of the neuron units 20 to 23, data in the output buffer 28 of the neuron units 20 to 23 is sequentially read as input data instead of data in the memory 14. That is, first, the data in the output buffer 28 of the two neuron units 20 is read, and each of the neuron units 24 and 25 stores the data in the weight memory 30 in advance.
The product with the weight of the connection with the neuron unit 2o stored in the register 2 is calculated by the arithmetic unit 26, and the product is stored in the register 2.
7 is stored. Next, regarding the data in the output buffer 28 of the neuron unit 21, the neuron unit 2
In each of steps 4 and 25, the product with the weight of the connection with the neuron unit 21 stored in advance in the weight memory 30 is calculated by the calculator 26, and the product is added to the contents of the register 27 (the above product). Thereafter, the sum is stored in register 27.

By repeating the above processing for the neuron units 23.24, the neuron units 24.25
Note that 4, which is the number of Σ, is the number of intermediate neuron units in this embodiment.

Specifically, the input threshold O of each neuron unit 24 and 25 stored in advance in the threshold value memory 29 is found from the contents of the register 27, and X in equation (2) stored in advance in the function table 31 and the function at that time are determined. From the correspondence of the values f (x), y in equation (1) is determined and stored in the output buffer 28 as the output of each of the neuron units 24 and 25.

Output buffer 28 of this neuron unit 24.25
The postcomputer 17 reads out the values stored in the postcomputer 17 and presents the results to the user.

If the output value of the neuron units 24 and 25 is close to 1, it is determined that the content of atoms or molecules assigned to each output neuron unit is within the standard value range, and if it is close to 0, it is outside the standard value range. It was determined that The controller 19 has neuron units 20 to 25
Generates a synchronization signal to operate simultaneously and in parallel.

As described above, the neuron units in the intermediate layer and the output layer each operate in parallel, allowing multiple atoms or molecules to be tested simultaneously, thereby shortening the determination time.

Next, neuron units 20 belonging to the intermediate layer and the output layer
How to determine the weight wi and the input threshold value O in equation (1) in ~25 will be explained. In this embodiment, Ramelhart (R
The pack propagation learning algorithm, also known as the membrane delta rule, was devised by D. E. Ramelhart and J. L. McClelland et al.

Written by PDP Research Group, translated and supervised by Toshi Seri, rPDP
Model, Chapter 8; Sangyo Tosho, 1989), and the obtained results are stored in the weight memory 30 and threshold memory 29 of each neuron unit 20-25 prior to the spectroscopic examination of the specimen.

The weight wi and input threshold O of the neuron unit are determined as follows. First, for a sample specimen for which the test result is known, the spectroscopic inspection apparatus shown in FIG. 2 is operated to store in the memory 14 the integral value of the spectral intensity of the wavelength region to be tested. host computer 17
reads this data, and at the same time, the correct judgment result for the sample specimen is associated with the unit of the output layer of the neural network 16 and expressed as a vector as training data, and is used as training data along with the data from the memory 14''2. 1 in the host computer 17. Next, another sample is registered as learning data 2 using the same procedure. The above procedure is performed for a large number of sample specimens to obtain a set of learning data.

m pieces of learning data obtained in this way are (Xl.

DI), −(Xm, Dm). here,
X1 (i=1 to m) represents a plurality of data obtained in memory 4 for one sample specimen as a vector. The number of elements of this vector corresponds to the number of test wavelength regions in FIG. 1, and is the number of neuron units in the input layer of the neural network 16 (here, 4). Di (i=l−m) is the correct test result for the neuron unit 24 of the output layer of the neural network 16.
．． This is the vector pattern that we want to obtain in the 25 output buffers, and the number of elements is the number of neuron units in the output layer (here, 2).

In learning using the pack propagation algorithm, the weight wi of each neuron unit and the input threshold O are changed so as to minimize the error shown in the following equation. Here, Y 1 (i=l
~m) uses the learning data Xi as the neural network 16
Neural network 16 obtained when inputting
This is the output of If the learning data of m silk is input once at a time and the weight Wi is changed so that the error E is reduced the most, the amount of change in the weight between the neuron unit j of the output layer and the neuron unit i of the intermediate layer △wji is given by Δwji, and the amount of change in weight between neuron unit i in the hidden layer and neuron unit l-h in the input layer is Δwih
is given by △ wih ・w, 1i(1-y+p)yt" 2 ・15)h. Here, ε is a positive constant, MJ+y, and P are the output layers when the learning data Xp is input to the neural network, respectively. is the output of neuron unit j in the middle layer and neuron unit l-i in the intermediate layer. yhp is the h-th element of If we consider that −1 is always the weight of the input terminal, it can be determined by equations (4) and (5).

Repeat equations (4) and (5) for each connection weight, and end the process when the error E in equation (3) becomes sufficiently small. Then, calculate the weight and input threshold value at that time. is stored in the weight memory 30 and threshold memory 29 of each neuron unit.

This completes the learning.

[Effects of the Invention] As explained above, according to the present invention, the learning function of the neural network allows the neural network to automatically have classification ability by simply providing sample data and a desired classification result for the sample data. By using this neural network, optimal inspection standards can be automatically set, eliminating the need for statistical analysis, etc.
It is possible to provide a spectroscopic inspection device that can shorten inspection time due to the parallel processing performance of the neural network. Note that the present invention is not limited to the above-described embodiments, and can be implemented with various modifications.

[Brief explanation of drawings]

FIG. 1 is a diagram schematically showing a spectroscopic inspection device according to the present invention, FIG. 2 is a block diagram of an embodiment of the spectroscopic inspection device according to the present invention, and FIG. 3 is a block diagram showing details of the neural network of the embodiment. FIG. 4 is a block diagram showing the neuron unit of the embodiment. ]...Light source, 3...Half mirror 15...Diffraction grating, 6...Line sensor, 7...Light sensor, 8
, 1]... A/D converter, 9... Real number converter, 12
...Normalization arithmetic unit, 13...Window circuit,
14...Memory, 16...Neural network, 17...Host computer. Applicant's agent Patent attorney Jun Tsuboi

Claims (1)

[Claims] means for obtaining spectral data for each wavelength of light transmitted or reflected by the specimen; means for integrating the spectral data for each wavelength region to be inspected; and means for obtaining inspection results according to the integrated value of the spectral data for each wavelength region. and a neural network means which is trained to output , and into which the output of the integrating means is input.