JP2024021532A - Concentration measuring device, abnormality detection method, and method for training learning model used in abnormality detection method - Google Patents

Abstract

To provide a concentration measuring device, an abnormality detection method, and a method for training a learning model that is used in the abnormality detection method, with which it is possible to detect the abnormality of digital waveform data of multiplexed light.SOLUTION: The present invention comprises: light sources 2, 3 that emit light in difference wavelengths; a multiplexer 5 that multiplexes the emitted light; a measurement cell 6 for accommodating a fluid G to be measured which the multiplexed light passes through; a transmitted light detector 7 that detects transmitted light having passed through the measurement cell 6; a signal converter 9 that converts the detected analog detection signal into a digital signal; a concentration computing unit 13 that is configured to frequency analyze the digital signal and calculate a concentration; a data acquisition unit 15 that is configured to acquire a detection data group in a prescribed time range from digital waveform data from the signal converter 9; and a determination unit 16 that is configured to input the detection data group of a detection object to an auto encoder 16a having been trained by normal digital waveform data and determine whether or not the restoration error of the restored detection data group exceeds a maximum value.SELECTED DRAWING: Figure 2

Description

The present invention is a concentration measuring device for measuring gas concentration using absorbance, and in particular, a concentration measuring device equipped with an abnormality detection function, an abnormality detection method during concentration measurement, and a learning model used in the abnormality detection method. Concerning learning methods.

Conventionally, there has been known a concentration measuring device that is incorporated into a gas supply line that supplies raw material gas such as an organic metal (MO) gas to semiconductor manufacturing equipment and measures the concentration of gas flowing through the gas supply line (see Patent Document 1, etc.). ).

This type of concentration measuring device measures absorbance by inputting light of a predetermined wavelength from a light source into a measurement cell and using a transmitted light detector to detect the light that is absorbed by the gas as it passes through the measurement cell. The concentration is determined from the absorbance.

In semiconductor manufacturing equipment and the like, multiple types of gases are supplied, and the absorption ratio with respect to the wavelength of light differs depending on the gas type. Therefore, absorbance can be measured using light of a plurality of different wavelengths (ultraviolet light). A plurality of lights of different wavelengths are combined by a multiplexer and then input into a measurement cell. The combined light incident on the measurement cell is absorbed by the gas within the measurement cell, and then received by one transmitted light detector. The detection signal of the transmitted light detector is amplified, A/D converted, and then subjected to frequency analysis using fast Fourier transform and converted into an amplitude spectrum of each frequency component.The amplitude value of the amplitude spectrum is output as the intensity of transmitted light. be done. The amplitude of the amplitude spectrum at the wavelength where there is absorption decreases.

Based on the change in the amplitude of the amplitude spectrum, the concentration C is calculated using the following formula (1) for determining the absorbance A _λ based on the Beer-Lambert law.

A _λ = log ₁₀ (I ₀ /I) = αLC (1)
Here, I ₀ is the initial intensity of the incident light that enters the measurement cell, I is the intensity of the transmitted light that has passed through the measurement cell, α is the molar extinction coefficient (m ² /mol), and L is the optical path length of the measurement cell ( m), C is the concentration (mol/m ³ ). The molar extinction coefficient α is a coefficient determined by the substance.

FIG. 1 is a functional block diagram showing a conventional concentration measuring device. This concentration measuring device 1A includes a plurality of light sources 2 and 3, a light emitting circuit 4 for causing each of the light sources 2 and 3 to emit light of a different wavelength, and a combination of light emitted by each of the light sources 2 and 3. A multiplexer 5 that generates a wave, a measurement cell 6 into which the light multiplexed by the multiplexer 5 enters, a transmitted light detector 7 that detects the light that has passed through the measurement cell 6, and a transmitted light detector 7. a light receiving circuit 8 for amplifying and/or adjusting the offset of the detection signal of the transmitted light detected by the light receiving circuit 8; a signal converter 9 for converting the analog signal amplified and offset-adjusted by the light receiving circuit 8 into a digital signal; It includes a processing device 10A that collects digital signals converted into digital signals by the converter 9 and performs arithmetic processing. Although there are two light sources in the illustrated example, three or more light sources may be provided as necessary.

The light sources 2 and 3 are composed of light emitting elements such as light emitting diodes, and the light emitting circuit 4 operates based on a command signal from the processing device 10A in order to flow drive currents of different frequencies to each of the plurality of light sources 2 and 3. , has the function of adjusting and controlling the current value 4a, frequency 4b, and current offset 4c of the drive current, respectively.

The measurement cell 6 includes a gas inlet 6a and a gas outlet 6b through which the gas to be measured flows in and out, a light entrance window 6c, and a light exit window 6d.

The multiplexer 5 includes a half mirror as in FIG. 6 of Patent Document 1, for example, multiplexes light incident from two directions, and outputs the multiplexed light in two directions. One of the two directions of light output from the multiplexer 5 enters the measurement cell 6, and the other light does not enter the measurement cell 6 and is detected as a reference light by the non-incident light detector 11. . A splitter that branches the multiplexed light as a reference light may be provided separately from the multiplexer (FIG. 1 of Patent Document 1). The detection signal of the non-incident light detector 11 is used as reference light to correct the detection signal of the transmitted light detector 7 (Patent Document 1). The correction may be performed by arithmetic processing based on equation (2) below.

I _cor = I _cell × (I _{ref, 0} /I _ref ) ... (2)
In the above equation (2), I _cor is the light intensity of the transmitted light after correction, I _cell is the intensity of the transmitted light detected by the transmitted light detector at the time of density measurement, and I _ref,0 is the intensity of the transmitted light detected by the transmitted light detector. The initial intensity of the detected non-incident light, I _ref , is the intensity at the time of density measurement of the non-incident light detected by the non-incident light detector.

The light receiving circuit 8 includes a fixed amplification section 8a, an offset adjustment section 8b, and a variable amplification section 8c. They perform fixed amplification, offset adjustment, and variable amplification of the transmitted light and reference light detection signals, respectively. The offset adjustment amount (ie, offset voltage) by the offset adjustment section 8b and the amplification factor of the variable amplification section 8c are determined based on a command signal from the processing device 10A.

The signal converter 9 converts the analog detection signals of the transmitted light and reference light adjusted by the light receiving circuit 8 into digital signals.

The processing device 10A includes a data collection unit 10a that collects digital signals from the signal converter 9, an analysis unit 10b that performs frequency analysis using fast Fourier transform, and an amplitude value of the amplitude spectrum obtained by the frequency analysis as an intensity output 12. It is provided with a spectrum output section 10c for outputting. The processing device 10A further includes a concentration calculation unit 13, and the concentration calculation unit 13 uses the intensity output 12 to calculate the concentration of the gas to be measured G based on the Lambert-Beer equation described above. The processing device 10A uses an MPU (Micro Processor Unit). Reference numeral 14 is a clock generation circuit.

In the concentration measuring device described above, in order to increase the measurement sensitivity, the target value of the intensity output is set so that the output of the amplitude spectrum, that is, the intensity output, is as large as possible. At this time, the target value of the intensity output is set within a range that does not exceed the rating of the A/D converter constituting the signal converter 9. Then, in the light receiving circuit 8, while checking the intensity output so that the intensity output reaches the set target value, the parameters (amplification factor, offset voltage, ) had been adjusted.

International Publication No. WO2017/029791

However, in the above-mentioned conventional concentration measuring device, the input signal of the A/D converter constituting the signal converter is within the rated range and is within the range below the maximum digital value of the digitally converted digital waveform. Even with this, digital waveform saturation could occur irregularly. That is, even if the analog waveform data of the combined light is within the rated range of the A/D converter, the digital waveform may saturate randomly. In such a case, even if an attempt was made to detect saturation, it was unclear at which digital value saturation would occur, depending on the parameters used to adjust the current value of the light emitting circuit, etc. Furthermore, even if the analog waveform data of the combined light is within the rated range of the A/D converter, if saturation occurs randomly, the digital value output by the A/D converter will fluctuate, for example, from 0 to 65535 digits. , it was not possible to set the reference value to be detected as a fixed value.

The main object of the present invention is to provide a concentration measuring device capable of detecting the above abnormality, an abnormality detection method, and a learning method for a learning model used in the abnormality detection method.

In order to achieve the above object, an abnormality detection method according to one aspect of the present invention is an abnormality detection method that detects an abnormality in a combined light of a plurality of different frequencies used for concentration measurement, The process of calculating the maximum value of the restoration error of the autoencoder, which is a trained learning model adjusted to restore the digital value of the digital waveform data, as a threshold, and the digital value of the digital waveform data of the combined light to be detected. inputting a sensed data group consisting of the above into the learned autoencoder, and determining whether a restoration error of the digital value of the sensed data group output from the autoencoder exceeds the threshold; including.

Further, a method for learning a learning model according to one aspect of the present invention is a method for learning a learning model used in an abnormality detection method for detecting an abnormality in combined light of a plurality of lights of different frequencies used for concentration measurement, the method comprising: A learning data group is generated by acquiring a first data group in a first predetermined time range from a single normal digital waveform data of the combined light, and acquiring a plurality of second data groups from the first data group. and a learning step of inputting the learning data group to an autoencoder and adjusting the weights of the neural network so that the input value and the output value match.

The plurality of second data groups are time series data of a second predetermined time range shorter than the first predetermined time range, and may include a portion where data overlaps.

Further, the concentration measuring device according to one aspect of the present invention includes a plurality of light sources that each emit light of a different wavelength, and a multiplexer that combines the light of a plurality of different wavelengths emitted by the plurality of light sources. a measurement cell for containing a fluid to be measured through which the light multiplexed by the multiplexer passes; a transmitted light detector for detecting the transmitted light that has passed through the measurement cell; and detection by the transmitted light detector. a signal converter that converts the detected analog detection signal into a digital signal; a concentration calculation unit configured to calculate a concentration using an amplitude spectrum obtained by frequency analysis of the digital signal; A data acquisition unit configured to acquire a data group in a predetermined time range from the digitally converted digital waveform data, and an autoencoder that has learned the normal digital waveform data to detect the data acquired from the digital waveform data to be detected. and a determination unit configured to input a data group and determine whether a restoration error of the restored sensed data group exceeds a maximum value of the restoration error of the normal digital waveform data.

According to the present invention, the above saturation, which has been difficult to detect conventionally, can be detected by abnormality detection using the autoencoder having the above configuration.

FIG. 1 is a functional block diagram showing a conventional concentration measuring device. FIG. 1 is a functional block diagram showing a first embodiment of a concentration measuring device according to the present invention. It is a graph which shows an example of the digital waveform extracted in a predetermined time range from the normal digital waveform data of the multiplexed light output from the signal converter. 5 is a graph appended with an explanation for extracting learning data from the digital waveform data of FIG. 4. FIG. This is an example of a set of digital values obtained from the waveform data of FIG. 4 at sampling numbers 1 to 32. FIG. 2 is an explanatory diagram illustrating an example of a neural network of an autoencoder that is a component of the present invention. 7 is a graph in which the input waveform and output waveform of the learning data input to the trained autoencoder of FIG. 6 are displayed in an overlapping manner. This is an example of a table showing input values, output values, absolute errors, and average absolute errors of a set of learning data. It is a graph which shows an example of the digital waveform extracted in a predetermined time range from the abnormal digital waveform data of the multiplexed light output from the signal converter. 10 is an enlarged view of a waveform graph of restored data outputted by inputting a data group obtained from the digital waveform data of FIG. 9 to a learned autoencoder. FIG. This is an example of a set of data representing the input value, output value, absolute error, and average absolute error of the learned autoencoder obtained from the abnormal digital waveform data of FIG. 9. It is a flowchart which shows the algorithm which determines abnormality from the digital data of multiplexed light. FIG. 2 is a functional block diagram showing a second embodiment of the concentration measuring device according to the present invention.

Embodiments of the concentration measuring device according to the present invention will be described below with reference to FIGS. 2 to 13. Note that the same or similar components as in the prior art are given the same reference numerals.

Referring to FIG. 2, the concentration measuring device 1 according to the first embodiment of the present invention includes a plurality of light sources 2 and 3, each of which emits light of a different wavelength, and a plurality of different light sources emitted by the plurality of light sources 2 and 3. A multiplexer 5 that multiplexes light of different wavelengths, a measurement cell 6 for containing the gas to be measured G through which the light multiplexed by the multiplexer 5 passes, and a transmitted light that has passed through the measurement cell 6 is detected. a transmitted light detector 7, a light receiving circuit 8 for adjusting the amplification and/or offset of the transmitted light signal of the transmitted light detected by the transmitted light detector 7, and an analog signal 8s adjusted by the light receiving circuit 8. It includes a signal converter 9 that converts into a digital signal 9s, and a processing device 10 that performs arithmetic processing on the digital signal converted by the signal converter 9.

Although light emitting diodes are used as the light sources 2 and 3 in the illustrated example, other light emitting elements such as laser diodes can also be used. Based on the command signal from the processing device 10, the light emitting circuit 4 adjusts and controls the current value 4a, frequency 4b, and current offset 4c of the drive current. Although the wavelengths of the light sources 2 and 3 are in the ultraviolet region in the illustrated example, it is also possible to use light in wavelength regions other than the ultraviolet region.

The multiplexer 5 includes a half mirror 5a, multiplexes light incident from two directions, and outputs the multiplexed light in two directions. One of the two directions of light outputted from the multiplexer 5 enters the measurement cell 6, and the other light does not enter the measurement cell 6 and is detected by the non-incident photodetector 11.

Although photodiodes are used in the illustrated example as the transmitted light detector 7 and the non-incident light detector 11, other optical sensors such as phototransistors may also be used.

Offset adjustment in the light emitting circuit 4 and the light receiving circuit 8 can be performed by adjusting the offset voltage of the operational amplifier. The variable amplification section 8c in the light receiving circuit 8 can adjust the amplification factor by adjusting the variable resistor provided in the operational amplifier.

Similar to FIG. 1, the processing device 10 includes a data collection unit 10a that collects digital signals from the signal converter 9, an analysis unit 10b that performs frequency analysis using fast Fourier transform, and an amplitude value of an amplitude spectrum obtained by frequency analysis. It includes a spectrum output section 10c that outputs an intensity output 12, and a concentration calculation section 13 that calculates the concentration of the gas G to be measured using an amplitude spectrum obtained by frequency-analyzing the digital signal 9s. The density calculation in the density calculation unit 13 and the density correction calculation using the reference light are well known from the above-mentioned Patent Document 1, etc., so a detailed explanation will be omitted.

The processing device 10 is configured by an MPU (Micro Processor Unit). The data collection unit 10a is constituted by a storage device such as a buffer memory, and can record digital data of the combined light sampled from the signal converter 9. The analysis section 10b, the spectrum output section 10c, and the concentration calculation section 13 can be constructed by programs stored in a storage device built in the processing device 10 or externally connected. The processing device 10 performs fast Fourier transform on the digital signal collected from the signal converter 9 according to a program recorded in a built-in storage device, outputs the intensity of the amplitude value, and calculates the density. The recording device that recorded the program may be externally connected to the processing device 10. The processing device 10 is not limited to an MPU, and may be another computer system such as a microcomputer or an MCU (Micro Controller Unit).

The processing device 10 further includes a data acquisition unit 15 that acquires a data group in a predetermined time range from the digital waveform data converted into digital data by the signal converter 9, and an autoencoder 16a that has been trained using the normal digital waveform data. Input the sensed data group acquired from the digital waveform data to be detected, and check whether the restoration error of the restored sensed data group exceeds the maximum value of the restoration error of the learned normal digital waveform data. A determination unit 16 that makes a determination is provided.

The data acquisition section 15 is constituted by a storage device such as a buffer memory, and can record a group of detection data of digital waveform data of the combined light extracted and acquired within a predetermined time range from the data acquisition section 10a. A determination program that executes the arithmetic processing of the determination unit 16 including the autoencoder 16a is stored in a storage device that is built in or externally connected to the processing device 10, and is executed by the processing device 10.

Next, an example of the process in the data acquisition unit 15 and determination unit 16 will be described.

FIG. 3 is a graph showing an example of a digital waveform extracted over a predetermined time range from normal digital waveform data of the combined light output from the signal converter 9. Specifically, it is a graph of digital waveform data extracted in time series from the data collection unit 10a. The horizontal axis in FIG. 3 is the sampling number. The vertical axis in FIG. 3 is the digital value (digits) of amplitude. The sampling numbers of the digital waveform data in FIG. 3 are 1 to 1024.

Referring to FIG. 4, by dividing the 1024 sampling data digital waveform data into 32 consecutive sampling data in time series and sliding them one point at a time in the time axis direction (in the direction in which the number of samples increases), 993 (=1024-31) data groups are obtained. That is, if 1024 pieces of sampling data are expressed in order as d(0) to d(1023), d(0) to d(31) are one set, and d(1) to d(32) are the next set. , d(2) to d(33) to the next set, d(3) to d(34) to the next set, and so on, 933 sets of data, with 32 consecutive pieces of data as one set. Create a group. The number of sampling data can be set as appropriate.

FIG. 5 is an example of one set of data of digital values obtained at sampling numbers 1 to 32. These 993 sets of data consisting of 32 points of time series data are input to the autoencoder 16a, which is a machine learning model, as a learning data group, and the autoencoder 16a is automatically encoded so that the output value of the autoencoder 16a matches the input value. The encoder 16a is made to learn by deep learning.

FIG. 6 is an explanatory diagram illustrating the neural network of the autoencoder 16a. In the autoencoder (self-encoding), there are 32 nodes in the input layer, and one set of 32 digital values (input x) is inputted to each node. 993 sets of data are input as learning data to the input layer of the autoencoder 16a. The input digital values of the 993 data group (input x) are compressed, decompressed, and output to the output layer. There are also 32 nodes in the output layer. The autoencoder 16a compares the input x and the output y to find a restoration error so that they match, and adjusts the weights w of the two neural networks of the encoder and decoder by error backpropagation. The meaning of the weights in the learning model shown in FIG. 6 is known, and a known method can be used to determine the weights (for example, see Japanese Patent Application Laid-Open No. 2021-9441).

FIG. 7 is a graph in which the input waveform and output waveform of learning data input to the learned autoencoder 16a are displayed in a superimposed manner. The output waveform restored by the learned autoencoder 16a is almost the same as the input waveform, but as shown in the partially enlarged view of FIG. 7, the restoration is not complete and there are slight errors. This error occurs because noise is removed during restoration by the autoencoder 16a.

FIG. 8 is an example of a table showing input values, output values, absolute errors, and average absolute errors of one set of learning data (data numbers 441 to 472) out of 993 sets of learning data. The mean absolute error (MAE) is a function used as the evaluation function or loss function of the output layer, and calculates the absolute value of the difference (=error) between the predicted value (output value) and the correct value for each data, This is a function that outputs the value (=average value) obtained by dividing the sum by the number of data.

In the present invention, since the correct value = input value, the input value to the autoencoder 16a (in(i):i=1 to 32) and the output value output from the autoencoder 16a based on the input value (out( i):i=1 to 32) (in(i)-out(i)) is the "absolute error" (Δ(i)). The value obtained by dividing the total of these 32 absolute errors by the number of data in one set, 32, is displayed in FIG. 8 as the average absolute error.

This average absolute error is calculated using all the data of the 993 sets described above, the average absolute error that takes the maximum value among the calculated average absolute errors of the 993 sets is calculated, and the maximum value of the average absolute error is set as the threshold for abnormality detection. Set to . For example, if the average absolute error of 495.61 shown in FIG. 8 is the maximum value among the 993 sets of average absolute errors, 495.61 is set as the threshold value.

FIG. 9 is a graph showing an example of a digital waveform extracted over a predetermined time range from the abnormal digital waveform data of the combined light output from the signal converter 9. The vertical and horizontal axes in FIG. 9 are the same as in FIG. 3. There are abnormalities in the area surrounded by six horizontally long ellipses at the bottom of FIG. 9 and the area surrounded by the vertically long ellipse at the right end.

The digital waveform data in FIG. 9 is digital waveform data with a sampling number of 1024 points, similar to the above, and is divided into sampling points of 32 points and slid one point at a time in the time axis direction (in the direction in which the sampling number increases). As a result, 993 data groups are obtained as examples of detected data groups.

FIG. 10 shows an enlarged view of the waveform graph of the restored data outputted by inputting the 993 data groups obtained from the digital waveform data of FIG. 9 to the learned autoencoder 16a. In FIG. 10, the input waveform and the output waveform are displayed in an overlapping manner. FIG. 10 shows that the abnormal location shown in FIG. 9 cannot be accurately restored, and the restoration error is large.

FIG. 11 is an example of a set of data representing the input value, output value, absolute error, and average absolute error of the learned autoencoder obtained from the abnormal digital waveform data of FIG. 9. In FIG. 11, the average absolute error is shown as 600.25. For example, if the threshold is set to 495.61, the average absolute error of 600.25 for the set of data shown in FIG. 11 exceeds the threshold of 495.61, and therefore is determined to be abnormal.

In this way, the data acquisition unit 15 acquires a detection data group of digital waveform data of the combined light (S1), and the determination unit 16 inputs the detection data group to the learned autoencoder 16a (S2), Calculate the output value restoration error (average absolute error) (S3), determine whether the output value restoration error (average absolute error) exceeds a threshold (S4), and calculate the data portion exceeding the threshold. can be determined to be abnormal, and data portions that do not exceed the threshold can be determined to be normal.

FIG. 12 is a flowchart showing the algorithm of steps S1 to S4. The processing device 10 executes the processes of steps S1 to S4 according to a program stored in a storage device that is built into the processing device 10 or externally connected. If the restoration error exceeds a threshold value in step S4, the program may be configured to issue a warning such as an alarm indicating an abnormality.

If the determination unit 16 determines that there is an abnormality, the current value 4a of the drive current in the light emitting circuit 4, the frequency 4b, the current offset 4c, the offset voltage by the offset adjustment unit 8b in the light receiving circuit 8, and the amplification factor of the variable amplification unit 8c. One or more parameters are readjusted.

In the above embodiment, the average absolute error is used as the restoration error for determining the threshold value and the restoration error of the detected data group to be compared with the threshold value, but the invention is not limited to this. (MSE), relative absolute error, or relative squared error can also be used.

The abnormal digital waveform in FIG. 9 is a digital waveform in which the input signal of the A/D converter constituting the signal converter 9 is within the rated range (0 to 3.3 V in the illustrated example) and is a digitally converted digital waveform. Even within the range below the maximum digital value (0 to 65535 digits in 16 bits), it may occur irregularly. That is, even if the analog waveform data of the combined light is within the rated range of the A/D converter, saturation may occur randomly. In such a case, even if you try to detect saturation, depending on the parameters that adjust the current value 4a of the light emitting circuit 4, etc., it is difficult to know at what digital value the saturation will occur, and if the digital value is 0 to 65535 digits. It was not possible to set the reference value for detection as a fixed value because the value fluctuates. Such saturation can be detected by anomaly detection using an autoencoder.

As described above, since the learning data group can be obtained from a single piece of normal digital waveform data, there is no need to collect a large amount of learning data.

As in the above example, the learning data group is obtained by acquiring the first data group from a single normal digital waveform data in the first predetermined time range (the number of samplings is 1024 in the above example), and It is obtained by acquiring the second data group. Similarly, the detection data group is obtained by acquiring a first data group from the digital waveform data in a first predetermined time range (the number of samplings is 1024 in the above example), and acquiring a plurality of second data groups from the first data group. It will be done. The data acquisition unit 15 acquires a time-series first data group and a second data group from the data collection unit 10a.

The plurality of second data groups are time series data of a second predetermined time range (in the above example, the number of samples is 32) shorter than the first predetermined time range, and include portions where data overlaps. For example, in the sample data sets “d(0) to d(31)” and “d(1) to d(32)” described above, d(1) to d(31) overlap.

In the embodiment described above, the plurality of second data groups include overlapping data portions, but they may also be data groups that do not overlap.

FIG. 13 shows a functional block diagram of a concentration measuring device according to a second embodiment of the present invention. The concentration measuring device 1 of the second embodiment is different from the first embodiment in that it includes a branching device 17 that branches the combined light as a reference light separately from the multiplexer 5, and other points. The configuration is the same as that of the first embodiment, so detailed explanation will be omitted.

The present invention is not limited to the above-described embodiments, and various modifications can be made without departing from the spirit of the present invention. For example, in the above embodiment, two light sources are provided, but a configuration may also be adopted in which three or more light sources are provided, and each light source emits light of a different wavelength. Alternatively, a configuration may be adopted in which no correction processing using the reference light is performed.

1 Concentration measuring device 2, 3 Light source 4 Light emitting circuit 5 Multiplexer 6 Measuring cell 7 Transmitted light detector 8 Light receiving circuit 9 Signal converter 10 Processing device 11 Non-incident light detector 15 Data acquisition section 16 Judgment section 16a Auto encoder

Claims (4)

An abnormality detection method for detecting an abnormality in combined light of a plurality of different frequencies of light used for concentration measurement,
calculating as a threshold the maximum value of the restoration error of the autoencoder, which is a trained learning model adjusted to restore the digital value of the normal digital waveform data of the combined light;
A detection data group consisting of digital values of digital waveform data of combined light to be detected is input to the learned autoencoder, and the restoration error of the digital value of the detection data group output from the autoencoder is a step of determining whether a threshold is exceeded;
The above abnormality detection method. A learning method for a learning model used in an anomaly detection method for detecting an anomaly in a combined light of a plurality of different frequencies used for concentration measurement, the method comprising:
A learning data group is obtained by acquiring a first data group in a first predetermined time range from the single normal digital waveform data of the combined light, and acquiring a plurality of second data groups from the first data group. The process of generating;
a learning step of inputting the learning data group to an autoencoder and adjusting the weights of the neural network so that the input value and the output value match;
The said learning method including. 3. The learning method according to claim 2, wherein the plurality of second data groups are time series data of a second predetermined time range shorter than the first predetermined time range, and include portions where data overlaps. multiple light sources each emitting light of a different wavelength,
a multiplexer that multiplexes the light of a plurality of different wavelengths emitted by the plurality of light sources;
a measurement cell for containing a fluid to be measured through which the light multiplexed by the multiplexer passes;
a transmitted light detector for detecting transmitted light passing through the measurement cell;
a signal converter that converts the analog detection signal detected by the transmitted light detector into a digital signal;
a concentration calculation unit configured to calculate concentration using an amplitude spectrum obtained by frequency analysis of the digital signal;
a data acquisition unit configured to acquire a data group in a predetermined time range from digital waveform data digitally converted by the signal converter;
The detection data group acquired from the digital waveform data to be detected is input to the autoencoder that has learned the normal digital waveform data, and the restoration error of the restored detection data group is equal to the restoration error of the normal digital waveform data. a determination unit configured to determine whether the maximum value is exceeded;
A concentration measuring device comprising: