Analytical equipment, analytical method, and analytical program

Abstract

To improve the accuracy of soft sensors. [Solution] The analysis device 10 has a creation unit 151 and an update unit 153. The creation unit 151 creates training data by combining a value based on a first value associated with a first time, a second value associated with a second time included in a period containing the first time, and a third value which is a value associated with the second value and is one of several types of values ​​associated with multiple times included in the period. The update unit 153 updates a model representing the relationship between the second value and the third value using the training data created by the creation unit 151.

Description

[Technical Field]

[0001] This invention relates to an analytical device, an analytical method, and an analytical program. [Background technology]

[0002] For example, factories are equipped with numerous physical sensors (hereinafter referred to as physical sensors), and operations are carried out using data collected by these physical sensors. On the other hand, there are cases where measurements cannot be taken using physical sensors due to reasons such as the high cost of physical sensors, their susceptibility to failure, lack of space to install physical sensors, or the unavailability of physical sensors corresponding to the physical quantity to be measured.

[0003] In such cases, it is conceivable to obtain measurement values ​​through manual analysis, but manual analysis has several problems: it is time-consuming, analytical equipment such as gas chromatography is expensive, and there are safety risks involved in collecting samples for analysis (for example, workers inhaling toxic gases).

[0004] In contrast, soft sensors are conventionally known. Soft sensors are a method that uses software calculations to obtain values ​​that have not actually been measured, based on measurements from physical sensors. [Prior art documents] [Non-patent literature]

[0005] [Non-Patent Document 1] [online], Analysis 2017 / 3 Gas Chromatography, [Retrieved January 23, 2025], Internet<URL:https: / / www.jsac.or.jp / ~gc / GCforum_takeda / pdf / bunseki_2017_2.pdf> [Non-Patent Document 2] [online], construction, operation, and maintenance of soft sensors, [searched January 23, 2025], Internet<URL:https: / / www.jstage.jst.go.jp / article / sicejl / 47 / 11 / 47_921 / _pdf> [Overview of the Initiative] [Problems that the invention aims to solve]

[0006] However, conventional soft sensors have the problem of low accuracy.

[0007] Here, the soft sensor calculates the predicted sensor value based on the predicted sensor value. One challenge in operating a soft sensor is distribution shift. Distribution shift refers to a change in the distribution of the predicted sensor value and the predicted sensor value, or the relationship between the predicted sensor value and the predicted sensor value, between the time the soft sensor is developed and the time it is put into operation. Distribution shift reduces the prediction accuracy of the soft sensor.

[0008] Furthermore, particularly with regard to soft sensors using machine learning models (hereinafter simply referred to as "models"), the Just-In-Time (JIT) method can be used to eliminate the effects of distribution shifts. The JIT method is a technique in machine learning that retrains a model using data obtained from past history under conditions similar to those at the time of prediction.

[0009] On the other hand, conventional JIT methods have a problem in that the time required for sample collection and analysis to prepare training data for retraining results in a time lag between the acquisition time of the source sensor value and the acquisition time of the target sensor value in the training data, leading to a decrease in accuracy. In addition, with conventional JIT methods, if the analysis work to obtain training data is performed periodically, depending on the frequency of that analysis work, sufficient training data may not be obtained. [Means for solving the problem]

[0010] To solve the above-described problems and achieve the objective, an analysis device comprises: a creation unit that creates learning data combining a second value associated with a second time included in a period containing a first time, the second value being a value based on a first value associated with the first time, and a third value that is a plurality of types of values associated with a plurality of times included in the period and is a value associated with the second value; and an update unit that updates a model representing the relationship between the second value and the third value using the learning data created by the creation unit. [Effects of the Invention]

[0011] According to the present invention, the accuracy of the soft sensor can be improved. [Brief Description of the Drawings]

[0012] [Figure 1] Figure 1 is a diagram illustrating a configuration example of the analysis device according to the first embodiment. [Figure 2] Figure 2 is a diagram for explaining the outline of a learning process. [Figure 3] Figure 3 is a diagram for explaining a method of adding an objective variable. [Figure 4] Figure 4 is a diagram for explaining augmentation. [Figure 5] Figure 5 is a flowchart illustrating the flow of the learning process. [Figure 6] Figure 6 is a flowchart illustrating the flow of an inference process and a relearning process. [Figure 7] Figure 7 is a diagram illustrating an example of a computer that executes an analysis program. [Modes for Carrying Out the Invention]

[0013] In the first embodiment, a soft sensor is implemented using a model. In addition, learning of the model is performed by a method that improves the conventional JIT method.

[0014] The model outputs a target variable based on the input explanatory variables. For example, the explanatory variables are sensor values ​​such as temperature, pressure, and flow rate, measured by physical sensors and acquired at least during a certain period prior to the acquisition time of the target variable. Alternatively, the target variable may be a sensor value from a soft sensor. The sensor value from the soft sensor may be a sensor value measurable by a physical sensor, or it may be an index determined by calculation.

[0015] Embodiments of the analytical apparatus, analytical method, and analytical program according to the present invention will be described in detail below with reference to the drawings. However, the present invention is not limited to the embodiments described below.

[0016] [First Embodiment] [Configuration of the first embodiment] The analysis device 10 will be described in detail using Figure 1. Figure 1 is a diagram showing an example of the configuration of the analysis device according to the first embodiment. As shown in Figure 1, the analysis device 10 has a communication unit 11, an input unit 12, an output unit 13, a storage unit 14, and a control unit 15.

[0017] The communication unit 11 performs data communication with other devices via the network. For example, the communication unit 11 is a NIC (Network Interface Card).

[0018] The input unit 12 is an interface connected to input devices such as a keyboard and mouse. The output unit 13 is an interface connected to output devices such as a display and speakers.

[0019] The storage unit 14 is a storage device such as an HDD (Hard Disk Drive), SSD (Solid State Drive), or optical disc. Alternatively, the storage unit 14 may be a rewritable semiconductor memory such as RAM (Random Access Memory), flash memory, or NVSRAM (Non-Volatile Static Random Access Memory).

[0020] The memory unit 14 stores the OS (Operating System) and various programs executed by the analysis device 10. The memory unit 14 also stores model information 141 and sample data 142.

[0021] Model information 141 is information such as parameters for building the model. For example, if the model is a neural network, model information 141 is the weights and biases of each layer. Alternatively, if the model is a linear model, the model information is the coefficients and constant terms for each input. Sample data 142 is the explanatory and dependent variables used for training.

[0022] The control unit 15 controls the entire analysis device 10. The control unit 15 is, for example, an electronic circuit such as a CPU (Central Processing Unit), MPU (Micro Processing Unit), or GPU (Graphics Processing Unit), or an integrated circuit such as an ASIC (Application Specific Integrated Circuit) or FPGA (Field Programmable Gate Array).

[0023] Furthermore, the control unit 15 has internal memory for storing programs and control data that define various processing procedures, and executes each process using the internal memory. In addition, the control unit 15 functions as various processing units when various programs are run. For example, the control unit 15 has a creation unit 151, a calculation unit 152, an update unit 153, and a correction unit 154.

[0024] The creation unit 151 creates training data. The calculation unit 152 inputs explanatory variables into the model and calculates the target variable. The update unit 153 updates the model parameters. For example, the update unit 153 updates the parameters using the training data so that the predicted value of the target variable output by the model approaches the actual value of the target variable. The correction unit 154 adjusts the target variable output by the model.

[0025] The learning process by the analysis device 10 will be described in detail using Fig. 2. Fig. 2 is a diagram for explaining the outline of the learning process.

[0026] As shown in Fig. 2, the explanatory variables are time-series data. In the example of Fig. 2, at time t s The objective variable is acquired in the period from -w / 2 to time t s +w / 2. For example, the objective variable is the data at any time t in the period from time t s -w / 2 to time t s +w / 2. At this time, t u s +w / 2 < t0, that is, if w < 2(t0 - t s ) is not satisfied, the objective variable is not acquired because it is an unacquired value. For the objective variable acquired at time t u , a plurality of types of data collected at times from t u -N-M to t u -N are acquired as explanatory variables. Each pair of the acquired objective variable and the corresponding explanatory variable is associated with the time of each objective variable as a sample and added to the sample data 142. For example, when t u = t s - w / 2, the explanatory variables are a plurality of types of data 201 collected in the period from time t s - w / 2 - N - M to time t s - w / 2 - N. Also, for example, when t u = t s + w / 2, the explanatory variables are a plurality of types of data 202 collected in the period from time t s + w / 2 - N - M to time t s + w / 2 - N.

[0027] The creation unit 151 creates learning data by adding explanatory variables corresponding to the objective variables. Fig. 3 is a diagram for explaining a method of adding the objective variables. (a) of Fig. 3 shows the sample data 142 before the objective variables are added. When the window width M is 3 steps and the prediction width N is 2, x1, x2, and x3 acquired from time "8:20" to time "8:40" are explanatory variables, and y1 acquired at time "9:00" is the objective variable.​

[0028] Here, let's assume that based on data acquired at time "9:00", the dependent variable "100" is obtained as the analysis result at time "10:00". The acquired data may also include the independent variables.

[0029] The creation unit 151 receives data indicating that the time the underlying sample was acquired was "9:00" and that the target variable obtained as a result of the analysis is "100". The creation unit 151 adds the target variable shown in the received data to the target variable of the sample data 142 corresponding to the time shown in the received data. In the example shown in Figure 3(b), the creation unit 151 adds "100" to the target variable corresponding to the time "9:00".

[0030] Furthermore, the creation unit 151 adds the same target variable as the reference time sample to each of the samples corresponding to multiple times included in the period based on the time indicated in the received data (hereinafter referred to as the reference time). In other words, the same target variable is set for multiple times included in the period and for the reference time.

[0031] In Figure 2, w represents the width of the period (hereinafter referred to as the augmentation width). The creation unit 151 generates data at time t s -w / 2 to time t s The time included in the period up to +w / 2, time t s This process involves adding an objective variable identical to the original objective variable, along with corresponding explanatory variables for each objective variable. This expansion of the training data by adding objective and explanatory variables is called augmentation.

[0032] In this way, the creation unit 151 creates training data by combining a target variable associated with each of several time points included in the period from a time point w / 2 before the reference time to a time point w / 2 after the reference time, and explanatory variables associated with each of the target variables.

[0033] Figure 4 is a diagram illustrating augmentation. Here, the sample acquisition time t s The time is "9:00". The augmentation width w is "60 minutes". As shown in Figure 4, the creation unit 151 adds "100" as the target variable for the times "8:30", "8:40", "8:50", "9:10", "9:20", and "9:30" included in the period from time "8:30" to time "9:30". Furthermore, it obtains the explanatory variables corresponding to each target variable and adds them to the sample data 142 along with the target variable. For example, if the window width M=3 and the prediction width N=2, the creation unit 151 adds the data for x1, x2, and x3 for "8:50", "9:00", and "9:10", along with the set of the target variable for "9:30", to the sample data 142. Similarly, for example, the creation unit 151 may add the data for x1, x2, and x3 for "8:40", "8:50", and "9:00", along with the set of the target variable for "9:20", to the sample data 142. Similarly, for example, the creation unit 151 may add the data for x1, x2, and x3 of "8:30", "8:40", and "8:50", along with the target variable set of "9:10", to the sample data 142.

[0034] Thus, the creation unit 151 creates training data by combining a first value associated with a first time, a second value associated with a second time included in the period containing the first time, and a third value, which is a value associated with the second value and is of multiple types associated with multiple times included in the period. "9:00" is an example of the first time. Also, "100," which is the value of y1 (dependent variable), is an example of the first value. Also, "9:30," which is included in the period from time "8:30" to time "9:30," is an example of the second time, and the value of y1 (dependent variable) at the second time is an example of the second value. Furthermore, the values ​​of x1, x2, and x3 (explanatory variables) for the period from time "8:50" to time "9:10," which are used to predict the second value, are examples of the third value.

[0035] Furthermore, the creation unit 151 creates training data by combining a second value, which is a value obtained by replicating the value acquired as an analysis value of a physical sensor, and a third value, which is the sensor value of the physical sensor over a certain period of time.

[0036] The explanatory variables may be sensor values ​​from physical sensors. The dependent variable may be a value calculated by a model, an existing soft sensor, or manually, based on the sensor values ​​and analysis results of the physical sensors.

[0037] The calculation unit 152 calculates the target variable by inputting explanatory variables into the model. The model is, for example, a linear model. During training, the calculation unit 152 adds the set of target and explanatory variables shown in Figure 4 to the sample data. For example, pairs of x1, x2, x3 for the period from time "8:50" to time "9:10" and the target variable at time "9:30" are added to the sample data, and similarly, the target variable from time "8:30" to time "9:20" is added to the sample data along with its corresponding explanatory variables. The calculation unit 152 inputs the explanatory variables of the training data into the model.

[0038] The update unit 153 updates the model representing the relationship between the target variable and the explanatory variables using the training data created by the creation unit 151. Specifically, the update unit 153 updates the parameters using the set of target and explanatory variables added to the sample data as training data so that the predicted value of the target variable output by the model approaches the actual value of the target variable. Note that this training process may also be called retraining, as it involves further training on an already trained model.

[0039] In this case, the retrained model may not necessarily have sufficient accuracy. Therefore, the correction unit 154 corrects the model's output in the inference process performed after retraining, based on the difference between the output value obtained by inputting the explanatory variables of the training data into the retrained model and the target variable of the training data.

[0040] Returning to Figure 2, line 301 represents the true value of the dependent variable. Line 302 represents the measured value of the dependent variable.

[0041] The measured value is the target variable obtained from the analysis of the acquired sample using an analytical instrument such as gas chromatography. (Time t) s Based on the samples obtained, the measured values ​​are obtained as analysis results at time t0. Furthermore, within the range shown in Figure 2, it is assumed that no measured values ​​are obtained at other times. Therefore, at time t s The measured value does not change during the period from [time] to time t0.

[0042] Furthermore, the true value here is the value obtained by applying the measured value obtained as an analysis result to the time when the sample on which the measured value was based was acquired. In other words, the true value is the value obtained at time t s These are the measured values ​​assuming that the samples obtained were analyzed immediately (at time 0) and the measured values ​​were determined.

[0043] Line 303 shows the value of the target variable output by the model before retraining. Line 304 shows the value of the target variable output by the model after retraining.

[0044] As shown in Figure 2, time t s After this point, the output values ​​of the model before retraining, shown on line 303, deviate significantly from the true values. On the other hand, at time t s After this, the output value of the retrained model shown on line 304 approaches the true value. In this way, the accuracy of the model improves through retraining.

[0045] On the other hand, there are cases where the accuracy of the model does not improve sufficiently through retraining. For example, if the retrained model is affected by time t s The output value when the explanatory variable is input, and time t s There is a difference of d between the target variable (true value) and the model after retraining. sIf the output value when the explanatory variable is input is "90", then the target variable in the training data is "100", so d is "10". Also, d is called the calibration value.

[0046] Therefore, the correction unit 154 corrects the output value by adding a calibration value d to the output value when the explanatory variables (explanatory variables other than the training data) of the samples from time t0 onward are input to the model.

[0047] In the examples explained so far, the target variable based on a sample taken at time "9:00" was obtained at time "10:00". On the other hand, the time it takes to obtain the target variable can be even shorter. For example, if it takes 20 minutes to obtain the target variable, the target variable based on a sample taken at time "9:40" will be obtained at time "10:00". In this case, since there is no sample at time "10:10", which is 30 minutes after time "9:40", the combination of the target variable and the corresponding explanatory variable for the period from time "9:10" to time "10:00" will be added as training data.

[0048] [Processing according to the first embodiment] Figure 5 is a flowchart showing the learning process flow. As shown in Figure 5, first, the analyzer 10 acquires the sample acquisition time and the analysis value based on the sample (step S101).

[0049] Next, the analyzer 10 sets the analysis value as the target variable for each time point included in the augmentation width period, which includes the time of sample acquisition (step S102). Then, the analyzer 10 acquires the explanatory variables for each target variable (step S103). Furthermore, the analyzer 10 adds the set of the target variable and corresponding explanatory variables for each time point to the sample data (step S104). After that, the analyzer 10 retrains the model (step S105) and then calculates the calibration value based on the analysis value used for training and the output value of the model (step S106).

[0050] Figure 6 is a flowchart showing the flow of the inference and retraining processes. As shown in Figure 6, first, the analysis device 10 acquires explanatory variables (step S201), and then acquires past sample data from the sample data that are close in time to the explanatory variables (step S202).

[0051] The analyzer 10 performs model training using sample data (step S203), then inputs explanatory variables into the model and calculates the model's output (step S204). Finally, the analyzer 10 corrects the model's output values ​​based on the acquired samples using calibration values ​​(step S205).

[0052] [Effects of the first embodiment] As explained above, the creation unit 151 creates training data by combining a first value associated with a first time, a second value associated with a second time included in the period containing the first time, and a third value associated with the second value, which is a value of multiple types associated with multiple times included in the period. The update unit 153 updates the model representing the relationship between the second value and the third value using the training data created by the creation unit 151. For example, the creation unit 151 creates training data by combining a second value, which is an analysis value of a physical sensor, and a third value, which is a sensor value of the physical sensor.

[0053] In this way, the analysis device 10 can expand the training data based on the values ​​at the first time point. This allows the model's accuracy to be further improved through retraining, even when the frequency of sample acquisition is limited. As a result, the analysis device 10 can improve the accuracy of the soft sensor using the model.

[0054] Alternatively, even if the analysis of the sample takes a long time, the analyzer 10 can retrain using explanatory variables corresponding to the target variable, thereby improving the accuracy of the soft sensor.

[0055] Furthermore, the creation unit 151 creates training data by combining a third value, which is associated with each of several time points included in the period from a time point one hour before the first time point to a time point one hour after the first time point, with the second value. This allows the analysis device 10 to reflect the distribution of data at time points around the time the sample was acquired in the training data.

[0056] Furthermore, the correction unit 154 corrects the value obtained by inputting a value other than the third value into the model, based on the difference between the value obtained by inputting a third value into the model and the second value. As a result, even if the accuracy of the model does not improve sufficiently through retraining alone, the analysis device 10 can obtain a highly accurate output value through correction.

[0057] [System configuration, etc.] Furthermore, the components of each illustrated device are functionally conceptual and do not necessarily need to be physically configured as shown. In other words, the specific forms of distribution and integration of each device are not limited to those shown, and all or part of them can be functionally or physically distributed or integrated in any unit according to various loads and usage conditions. Moreover, each processing function performed by each device can be implemented, in whole or in any part, by a CPU (Central Processing Unit) and a program that is analyzed and executed by the CPU, or by hardware using wired logic. Note that the program may be executed not only by the CPU but also by other processors such as a GPU.

[0058] Furthermore, among the processes described in this embodiment, all or part of the processes described as being performed automatically can be performed manually, or all or part of the processes described as being performed manually can be performed automatically by known methods. In addition, the processing procedures, control procedures, specific names, and information including various data and parameters shown in the above document and drawings can be arbitrarily changed unless otherwise specified.

[0059] [program] In one embodiment, the analysis device 10 can be implemented by installing an analysis program that performs the above-mentioned learning process as packaged software or online software on a desired computer. For example, by having the above-mentioned analysis program run on an information processing device, the information processing device can function as the analysis device 10. The information processing device referred to here includes desktop or notebook personal computers. In addition, the category of information processing device also includes tablet terminals, smartphones, mobile phone and PHS (Personal Handyphone System) mobile communication terminals, and slate terminals such as PDAs (Personal Digital Assistants).

[0060] Furthermore, the analysis device 10 can also be implemented as a server that provides the above-mentioned learning process to a terminal device used by a user, with the user's terminal device acting as the client. For example, the server can be implemented as a server device that provides a learning service, taking the specification of request points as input and outputting a trained model. In this case, the server may be implemented as a web server, or it may be implemented as a cloud service that provides the above-mentioned learning process through outsourcing.

[0061] Figure 7 shows an example of a computer running an analysis program. Computer 1000 has, for example, memory 1010 and a CPU 1020. Computer 1000 also has a hard disk drive interface 1030, a disk drive interface 1040, a serial port interface 1050, a video adapter 1060, and a network interface 1070. These components are connected by a bus 1080.

[0062] Memory 1010 includes ROM (Read Only Memory) 1011 and RAM (Random Access Memory) 1012. ROM 1011 stores, for example, a boot program such as BIOS (Basic Input Output System). The hard disk drive interface 1030 is connected to the hard disk drive 1090. The disk drive interface 1040 is connected to the disk drive 1100. For example, a removable storage medium such as a magnetic disk or optical disk is inserted into the disk drive 1100. The serial port interface 1050 is connected to, for example, a mouse 1110 and a keyboard 1120. The video adapter 1060 is connected to, for example, a display 1130.

[0063] The hard disk drive 1090 stores, for example, the OS 1091, application programs 1092, program modules 1093, and program data 1094. That is, the programs defining each process of the analytical device 10 are implemented as program modules 1093 containing code executable by a computer. The program modules 1093 are stored, for example, on the hard disk drive 1090. For example, a program module 1093 for performing processes similar to those in the analytical device 10 is stored on the hard disk drive 1090. Note that the hard disk drive 1090 may be replaced by an SSD (Solid State Drive).

[0064] Furthermore, the configuration data used in the processing of the above-described embodiment is stored as program data 1094 in, for example, memory 1010 or hard disk drive 1090. The CPU 1020 then reads the program module 1093 and program data 1094 stored in memory 1010 or hard disk drive 1090 into RAM 1012 as needed and executes the processing of the above-described embodiment.

[0065] Furthermore, the program module 1093 and program data 1094 are not limited to being stored in the hard disk drive 1090; for example, they may be stored in a removable storage medium and read by the CPU 1020 via a disk drive 1100 or the like. Alternatively, the program module 1093 and program data 1094 may be stored in another computer connected via a network (LAN (Local Area Network), WAN (Wide Area Network), etc.). The program module 1093 and program data 1094 may then be read by the CPU 1020 from the other computer via a network interface 1070. [Explanation of Symbols]

[0066] 10 Analyzer 11 Communications Department 12 Input section 13 Output section 14 Storage section 15 Control Unit 141 Model Information 142 sample data 151 Creation Department 152 Calculation section 153 Update Department 154 Correction section

Claims

1. A creation unit that creates learning data by combining a value based on a first value associated with a first time, a second value associated with a second time included in a period including the first time, and a third value associated with the second value, which is a plurality of types of values ​​associated with a plurality of times included in the period. An update unit updates a model representing the relationship between the second value and the third value using the training data created by the creation unit, An analytical apparatus characterized by having the following features.

2. The creation unit creates training data by combining the second value, which is the analysis value of the physical sensor, and the third value, which is the sensor value of the physical sensor. The analytical apparatus according to feature 1.

3. The creation unit creates training data by combining the third value, which is associated with each of the multiple time points included in the period from a time point one hour before the first time point to a time point one hour after the first time point, with the second value. The analytical apparatus according to feature 1.

4. A correction unit corrects the value obtained by inputting a value other than the second value into the model, based on the difference between the value obtained by inputting the third value into the model and the second value. The analytical apparatus according to claim 1, further comprising the features.

5. An analytical method performed by an analytical device, A creation step to create learning data by combining a value based on a first value associated with a first time, a second value associated with a second time included in a period containing the first time, and a third value associated with the second value, which is a plurality of types of values ​​associated with a plurality of times included in the period; An update step, which updates the model representing the relationship between the second value and the third value using the training data created in the above creation step, An analytical method characterized by including [a certain component].

6. A creation step to create training data by combining a value based on a first value associated with a first time, a second value associated with a second time included in a period containing the first time, and a third value associated with the second value, which is a plurality of types of values ​​associated with a plurality of times included in the period; An update step, which updates the model representing the relationship between the second value and the third value using the training data created in the creation step, An analysis program characterized by having a computer execute it.

Images