JP7657141B2 - Prediction system, prediction method, and program - Google Patents

Description

This disclosure relates to a prediction system, a prediction method, and a program. More specifically, this disclosure relates to a prediction system, a prediction method, and a program for predicting the lifespan of a volcanic ash filter.

When a volcano erupts, volcanic ash will affect buildings and equipment within the area where it falls. As one measure against volcanic ash, nuclear plants install volcanic ash filters in the path of volcanic ash to prevent it from entering equipment. In order to limit the impact on a plant when an eruption actually occurs, it is desirable to clean or replace filters before they reach their performance limits. To do this, it is necessary to know in advance how long it will take for the filter's performance to reach its limits.

As a related technique, Patent Document 1 discloses a diagnostic method for predicting the lifespan of a filter installed in a general air conditioning system based on the filter differential pressure, ambient temperature, humidity, and the period of actual use of the filter, and notifying the user when to replace the filter. The diagnostic method of Patent Document 1 collects online information on the filter after use has started, such as the differential pressure and ambient temperature, and uses these parameters to correct the period of use of the filter to predict the filter lifespan; however, it does not predict the lifespan of the filter before use begins.

JP 2017-217616 A

There is a need for technology that can predict the lifespan of volcanic ash filters in advance.

The present disclosure provides a prediction system, a prediction method, and a program that can solve the above problems.

The prediction system of the present disclosure comprises an acquisition unit that acquires specific evaluation conditions for a volcanic ash filter to be evaluated; a prediction model that, when the evaluation conditions are input, outputs a predicted value for the lifespan of an unused volcanic ash filter related to the evaluation conditions; a prediction unit that predicts the lifespan of the volcanic ash filter to be evaluated based on the evaluation conditions acquired by the acquisition unit; and a prediction model construction unit that constructs the prediction model by learning through machine learning a relationship between the evaluation conditions and the time until the pressure loss of the unused volcanic ash filter reaches a threshold value based on learning data that uses the evaluation conditions as explanatory variables and the time measured in a filter test that measures the time until the pressure loss of the unused volcanic ash filter reaches a threshold value based on the evaluation conditions as a response variable , wherein the evaluation conditions include the particle size distribution pattern of the volcanic ash captured by the volcanic ash filter, the mesh coarseness of the volcanic ash filter, and the threshold value for the pressure loss of the volcanic ash filter, and the prediction model predicts the time until the pressure loss of the unused volcanic ash filter reaches the threshold value.

The prediction method disclosed herein is executed by a prediction system, and uses machine learning to learn the relationship between predetermined evaluation conditions and the time until the pressure loss of an unused volcanic ash filter reaches a threshold based on learning data using predetermined evaluation conditions as explanatory variables and the time measured in a filter test that measures the time until the pressure loss of an unused volcanic ash filter reaches a threshold based on the evaluation conditions as a response variable, to construct a prediction model, obtain the evaluation conditions for the volcanic ash filter to be evaluated , and predict the life of the volcanic ash filter to be evaluated by inputting the obtained evaluation conditions into the prediction model, where the evaluation conditions include the particle size distribution pattern of volcanic ash captured by the volcanic ash filter, the mesh coarseness of the volcanic ash filter, and the threshold for the pressure loss of the volcanic ash filter, and the prediction model predicts the time until the pressure loss of an unused volcanic ash filter reaches the threshold.

The program of the present disclosure causes a computer to use machine learning to learn a relationship between predetermined evaluation conditions and the time until the pressure loss of an unused volcanic ash filter reaches a threshold based on learning data having predetermined evaluation conditions as explanatory variables and the time measured in a filter test measuring the time until the pressure loss of an unused volcanic ash filter reaches a threshold based on the evaluation conditions as a target variable, to construct a prediction model, obtain the evaluation conditions for the volcanic ash filter to be evaluated , and predict the life of the volcanic ash filter to be evaluated by inputting the obtained evaluation conditions into the prediction model, which outputs a predicted value of the life of the volcanic ash filter , wherein the evaluation conditions include the particle size distribution pattern of volcanic ash captured by the volcanic ash filter, the mesh coarseness of the volcanic ash filter, and the threshold value for the pressure loss of the volcanic ash filter, and the prediction model executes a process of predicting the time until the pressure loss of an unused volcanic ash filter reaches the threshold .

The above-mentioned prediction system, prediction method, and program make it possible to predict the lifespan of a filter before it is first used.

FIG. 1 is a block diagram illustrating an example of a prediction system according to an embodiment. FIG. 4 is a diagram illustrating an example of learning data according to the embodiment. FIG. 1 is a diagram illustrating an example of a main part of a prediction system according to an embodiment. FIG. 11 is a diagram showing an example of a parameter search result according to the embodiment. FIG. 13 is a diagram illustrating an example of a prediction model construction log according to the embodiment. FIG. 1 is a first diagram showing an example of a scatter diagram of predicted values and actual measured values according to an embodiment. FIG. 2 is a second diagram showing an example of a scatter diagram of predicted values and actual measured values according to the embodiment. FIG. 13 is a diagram illustrating an example of a graph showing a transition of a prediction error according to the embodiment. 10 is a flowchart illustrating an example of an operation of the prediction system according to the embodiment. FIG. 2 is a diagram illustrating an example of a hardware configuration of a prediction system according to an embodiment.

Hereinafter, a prediction system according to the present disclosure will be described with reference to FIGS.
<Embodiment>
(composition)
FIG. 1 is a block diagram showing an example of a prediction system according to an embodiment. The prediction system 10 is a system for predicting the lifespan of a volcanic ash filter. The lifespan of a volcanic ash filter is when the pressure loss of the filter reaches a threshold value. Hereinafter, the lifespan of a volcanic ash filter is also referred to as the time to replace the filter. The prediction system 10 includes an input receiving unit 11, a control unit 12, a memory unit 13, and a display unit 14.

The input reception unit 11 is configured using input devices such as a keyboard, mouse, touch panel, buttons, etc. The input reception unit 11 acquires various information input using the input device and outputs the information to the control unit 12. For example, the input reception unit 11 acquires information such as evaluation conditions used for life prediction, such as the particle size distribution pattern of volcanic ash and the mesh coarseness of the volcanic ash filter, and test conditions and test results of filter tests previously conducted on volcanic ash filters. The test conditions and test results of past filter tests become learning data for constructing a prediction model.

The control unit 12 controls the process of predicting the lifespan of the volcanic ash filter. In the learning phase, the control unit 12 constructs a prediction model to be used for lifespan prediction, and in the prediction phase, the control unit 12 predicts the lifespan of the volcanic ash filter using the constructed prediction model. The control unit 12 includes a prediction model construction unit 121, a prediction unit 122, and a judgment unit 123. The prediction model construction unit 121 constructs a prediction model that outputs the lifespan of the volcanic ash filter when evaluation conditions such as the particle size distribution pattern of volcanic ash are input, using test results of past filter tests as learning data. The prediction unit 122 predicts the lifespan of the volcanic ash filter to be evaluated from a prediction model used for unused lifespan prediction regarding evaluation conditions based on a certain amount of accumulated learning data, and information such as evaluation conditions used for lifespan prediction, such as the particle size distribution pattern of the volcanic ash and the coarseness of the mesh of the volcanic ash filter, and test conditions of filter tests previously conducted on volcanic ash filters. Furthermore, the prediction model constructed by the prediction model construction unit 121 is used to predict the lifespan of the volcanic ash filter being evaluated. When learning data is accumulated by carrying out a filter test, the determination unit 123 determines whether to reconstruct (or construct a new) prediction model in order to reflect the newly accumulated learning data in the prediction model. Furthermore, the control unit 12 outputs the prediction results by the prediction unit 122 to a display device or electronic file.

The storage unit 13 stores various information necessary for predicting the service life of the volcanic ash filter. For example, the storage unit 13 stores learning data and a constructed prediction model.
The display unit 14 is configured using a display device such as a liquid crystal display.

An example of learning data is shown in Figure 2. In the testing of volcanic ash filters, the test data shown in Figure 2 is recorded. The test data includes the "item number", which is the test identification number, "customer information" such as the customer name and the nuclear plant site name, the "test date" when the volcanic ash filter test was conducted, the "test time", test implementation information such as the "temperature" and "humidity" at the time the test was conducted, "grain size 1", "grain size 2", ..., "grain size N" indicating the grain size (grain size, particle size) of the volcanic ash used in the test, "concentration" indicating the weight per unit volume of the volcanic ash used in the test, and the configuration of the "upstream labyrinth" on the upstream side of the volcanic ash filter used in the test. The test conditions include the presence or absence of a "downstream labyrinth" downstream, the coarseness of the "mesh" of the volcanic ash filter, the "allowable pressure loss" which is a threshold value set for the pressure loss, which is the pressure upstream of the filter minus the pressure downstream of the filter when volcanic ash is fed into the volcanic ash filter at the rated air volume, and the "allowable pressure loss arrival time" which indicates the time it takes for the pressure loss of an unused volcanic ash filter to reach the "allowable pressure loss" when volcanic ash determined by the test conditions is supplied to the volcanic ash filter at a specified flow rate. The "allowable pressure loss arrival time" is the test result. The test conditions "grain size 1", "grain size 2", ... "grain size N" are associated in advance with the grain size of volcanic ash, and the percentage of ash of that grain size contained in the volcanic ash is set for each item of "grain size 1", "grain size 2", ... "grain size N" (grain size distribution pattern). For example, in the example of FIG. 2, in the test of item number 1, volcanic ash was used having a particle size distribution pattern in which ash of particle size "1" was 70%, ash of particle size "2" was 20%, and ash of particle size "N" was 10%. The test conditions and test results of the test data illustrated in FIG. 2 are used as learning data when constructing a prediction model, and the test conditions are explanatory variables and the test results are target variables. Note that the test conditions illustrated in FIG. 2 are only an example. For example, the test conditions (explanatory variables) may include the flow rate of the volcanic ash. When predicting the life of the volcanic ash filter, the values of each item of the test conditions are input to the prediction model, and the allowable pressure loss arrival time output by the prediction model is obtained as the predicted value of the volcanic ash filter life. The test data collected when the test was previously performed at the nuclear power plant is acquired by the input receiving unit 11 and stored in the memory unit 13. The test data may be edited, for example, using general spreadsheet software and stored in the memory unit 13 in a data format managed by the spreadsheet software. By using spreadsheet software, for example, when adding explanatory variables, the explanatory variables can be added by adding columns, and when adding training data, the training data can be added by adding test data in the row direction.

(Predictive model building function)
Fig. 3 is a diagram showing an example of the main parts of the prediction system of Fig. 1. The prediction model construction function and the life prediction function of the volcanic ash filter possessed by the prediction system 10 will be described with reference to Fig. 3. The prediction model construction unit 121 includes a search condition setting unit 1211, a search unit 1212, a search result display unit 1213, a construction setting unit 1214, a construction unit 1215, a construction result display unit 1216, a prediction error confirmation unit 1217, and a prediction error display unit 1218. The storage unit 13 also stores past test data 131 described in Fig. 2.

The predictive model construction unit 121 has the following functions: when constructing a predictive model, it searches for appropriate settings for the machine learning parameters used in constructing the predictive model, and presents the search results to the user; it constructs a predictive model using a combination of setting values for the parameters selected by the user from the search results; and it presents the prediction accuracy (prediction error) of the constructed predictive model to the user.

The search condition setting unit 1211 sets the search range of the parameter setting value. For example, the prediction model construction unit 121 uses a random forest to construct a prediction model, and performs oversampling by, for example, SMOTE (Synthetic Minority Oversampling Technique) as a preprocessing. In this case, the search condition setting unit 1211 sets the search range for each parameter, for example, the number of decision trees as a parameter of the random forest, the number of samples of k nearest neighbor data as a parameter of the SMOTE, the sampling threshold, and the outlier removal threshold as a parameter for the prediction result. Here, the number of decision trees, the number of samples of k nearest neighbor data, and the sampling threshold are parameters required for each algorithm. The outlier removal threshold is used for the purpose of first performing regression learning by random forest using learning data, and then finding the difference between the predicted value of the created learned model and the actual value of the learning data, and removing the prediction result when the difference is equal to or greater than the "outlier removal threshold". For example, the user sets a value in the range of 20 to 50 for the number of decision trees, a value in the range of 2 to 5 for the number of samples of k-nearest neighbor data, a value in the range of 0 to 60 for the sampling threshold, and a value in the range of 0 to 180 for the outlier removal threshold. Note that the parameters used to construct the prediction model given here are merely examples. For example, in addition to the number of decision trees, the depth of the decision trees and the number of leaves for a random forest may be included in the parameters to be searched. Hereinafter, the parameters of the number of decision trees, the number of samples of k-nearest neighbor data, the sampling threshold, and the outlier removal threshold may be referred to as parameters 1 to 4, and the values set for each parameter may be referred to as settings 1 to 4. For example, the search condition setting unit 1211 sets {15, 25, 35, 45} as the search range for setting value 1, {2, 3, 4} as the search range for setting value 2, {none, 30, 60} as the search range for setting value 3, and {none, 120, 180} as the search range for setting value 4 according to the user's settings.

The search unit 1212 searches for appropriate setting values 1 to 4 for the search range set by the search condition setting unit 1211. Specifically, the search unit 1212 creates combinations for the setting values 1 to 4 set by the search condition setting unit 1211. In the above example, 4 x 3 x 3 x 3 combinations of setting values are created. The search unit 1212 performs preprocessing and machine learning for each combination using the test conditions and test results of the past test data 131 as learning data, and evaluates the prediction accuracy when a prediction model is constructed with the combination of the setting values 1 to 4. For example, in the case of a combination of setting value 1 = 25, setting value 2 = 2, setting value 3 = 30, and setting value 4 = 120, the search unit 1212 sets the number of samples of k-nearest neighbor data to 2 and the sampling threshold to 30, increases the learning data (oversampling) by SMOTE, removes outliers from the learning data after oversampling, and uses the data to perform regression analysis by random forest with the number of decision trees set to 25, and calculates the prediction error of each parameter of the learned model. Similarly, the search unit 1212 calculates the prediction error of each parameter for each combination.

The search result display unit 1213 displays the search results by the search unit 1212. For example, the search result display unit 1213 displays a predetermined number of combinations of setting values 1 to 4 with small prediction errors from the search results by the search unit 1212, arranged in order of smallest prediction errors. An example of the display of the search results is shown in FIG. 4. In the example of FIG. 4, combinations of setting values 1 to 4 are displayed in order of best prediction accuracy. The combination of setting values 1 to 4 with the highest prediction accuracy is "25" for setting value 1, "2" for setting value 2, "30" for setting value 3, and "120" for setting value 4, which indicates that the prediction error was "26" when learning was performed with these values set for parameters 1 to 4. The lower and upper limits of the search results are the lower and upper limits of the "allowable pressure loss arrival time" (objective variable) included in the learning data used for the search. This value indicates the range of application of the prediction model. For example, if the lower limit is 30 minutes and the upper limit is 500 minutes, the filter life prediction result using this learning data will be a value in the range of 30 to 500 minutes.

The construction setting unit 1214 accepts the setting values 1 to 4 used to construct the prediction model. For example, the user refers to the search results displayed by the search result display unit 1213 and determines the combination of setting values 1 to 4 to be used to construct the prediction model. For example, the user may set the same as Search No. 1, which had the smallest prediction error among the search results illustrated in FIG. 4. Alternatively, the user may adopt a combination of setting values 1 to 4 (not shown) in which the range of the lower limit and upper limit is wide (the range of the prediction result is wide) even if the value of the prediction error is somewhat large. The construction setting unit 1214 sets the upper limit and lower limit of the parameters and the learning data used to construct the prediction model according to the user's settings. For example, the construction setting unit 1214 sets "25", "2", "30", and "120" to the setting values 1 to 4, respectively. It is generally considered that a more accurate prediction model can be constructed if the number of learning data is large. Therefore, if you want to build a prediction model with high prediction accuracy even if the prediction range is narrow, you can adjust the input data to narrow down the range where a lot of training data is obtained.

The construction unit 1215 constructs a prediction model (without constraints) 133 and/or a prediction model (with constraints) 134 using the past test data 131 as learning data, using the setting values 1 to 4 set by the construction setting unit 1214. Here, the prediction model (without constraints) indicates a case where the model is constructed using all data without checking the value of the allowable pressure loss arrival time, and the prediction model (with constraints) indicates a case where the model is constructed by checking the value and narrowing down the data to be input when constructing the model (for example, 600 (minutes) or less). For example, when the construction setting unit 1214 performs the above setting, the construction unit 1215 sets the number of samples of k nearest neighbor data to 2 and the sampling threshold to 30, increases the learning data by SMOTE, removes data that is an outlier from the increased data, and uses the data to perform a regression analysis of random forest by setting the number of decision trees to 25, creates a prediction model, and calculates the prediction error. The construction unit 1215 stores the constructed prediction model (without constraints) 133 in the storage unit 13. The construction unit 1215 also records a log (prediction model construction log 132) at the time of construction of the prediction model (without constraint) 133 in the storage unit 13. FIG. 5 shows an example of data recorded as the prediction model construction log 132. As shown in the figure, the prediction model construction log 132 includes the following items: No., model name, setting value 1, setting value 2, setting value 3, setting value 4, lower limit, upper limit, and prediction error. No. is set to the identification number of the constructed prediction model. The model name and date are set to the name of the constructed prediction model and the date on which the prediction model was constructed, respectively. The other items are the same as the items of the same names in FIG. 4. The construction unit 1215 also records the item number (item number of the test data in FIG. 2) of the learning data used to construct the prediction model in the storage unit 13 in association with, for example, the identification number of the prediction model. The construction unit 1215 also checks the value of the "allowable pressure loss arrival time" to narrow down the data to be input when constructing the model (for example, 600 (minutes) or less), constructs a prediction model (with constraints) 134 (first time) with the same parameter settings, and calculates the prediction error of the constructed prediction model (with constraints) 134. In the case of the above setting example, the prediction model (with constraints) 134 is constructed using only learning data with the "allowable pressure loss arrival time" in the range of 33 to 395, and the same values are set for the setting values 1 to 4, and the prediction model construction log 132 is also recorded for the prediction model (with constraints) 134. FIG. 5 shows an example of a log for the prediction model (with constraints) 134. The construction unit 1215 may construct both the prediction model (without constraints) 133 and the prediction model (with constraints) 134, or may construct only one of them.

Furthermore, the construction of a prediction model can be repeated. If the same values are set for setting values 1 to 4 to construct a prediction model from the second time onwards, in the second construction, the construction unit 1215 uses the added data as input and performs oversampling, outlier removal, and machine learning in the same way as the first construction to construct a prediction model (unconstrained) 133 (second time) and a prediction model (constrained) 134 (second time), and records a prediction model construction log 132 for each. Thereafter, prediction models can be constructed in the same way. (When there is no need to distinguish between with and without constraints on the range of the training data and the number of times of training, it will simply be described as a "prediction model.")

The construction result display unit 1216 displays the construction result of the prediction model on the display unit 14 based on the prediction model construction log 132 recorded by the construction unit 1215. For example, the construction result display unit 1216 displays the log information shown in FIG. 5, which is recorded in the construction of the first to fifth prediction models, on the display unit 14. The construction result display unit 1216 also creates a scatter plot of the predicted values and actual values by the prediction model, and displays the created scatter plot on the display unit 14. FIG. 6A shows an example of a scatter plot for the prediction model (without constraints) 133 (nth time), and FIG. 6B shows an example of a scatter plot for the prediction model (with constraints) 134 (nth time). The vertical axis of the graphs in FIG. 6A and FIG. 6B shows the actual measured value of the "allowable pressure loss arrival time" included in the learning data, and the horizontal axis shows the predicted value of the "allowable pressure loss arrival time" predicted by the explanatory variables included in the learning data and the prediction model. Each point corresponds to the actual measured value and the predicted value obtained from one learning data. The more points there are near L1, the higher the performance of the prediction model. (L2 shows an example of a predicted value output by the prediction model when evaluation conditions are input to the prediction model constructed in the prediction phase; L2 is not displayed in the learning phase.)

The prediction error confirmation unit 1217 reads and acquires the history of prediction errors of the prediction model recorded in the prediction model construction log 132. For example, the prediction error confirmation unit 1217 reads and acquires the first to fifth prediction errors (58, 34, 38, 30, 28) from the prediction model construction log 132.

The prediction error display unit 1218 creates a graph showing the progress of the prediction error according to the number of learning iterations from the prediction error history information acquired by the prediction error confirmation unit 1217, and displays this graph. Figure 7 shows an example of a graph showing the progress of the prediction error. In this way, even if the setting values 1 to 4 are the same, the prediction accuracy can be improved by increasing the amount of learning data each time the construction of a prediction model is repeated.

(Lifespan prediction function)
Next, the life prediction function of the volcanic ash filter will be described. The prediction unit 122 includes an evaluation condition setting unit 1221, a prediction model setting unit 1222, a filter replacement time prediction unit 1223, and a prediction error display unit 1224. For example, the storage unit 13 stores N prediction models (without constraints) 133 and/or N prediction models (with constraints) 134 created in N constructions, and a prediction model construction log 132 recorded when the N prediction models were constructed.

The evaluation condition setting unit 1221 records the evaluation conditions input by the user in the storage unit 13. As a result, evaluation condition data 135 is stored in the storage unit 13. The evaluation conditions are explanatory variables exemplified in FIG. 2. For example, when creating a test plan for a volcanic ash filter, the user inputs the test conditions for the test to be performed as evaluation conditions. Or, not limited to filter testing, when a volcano near a nuclear plant erupts and a volcanic ash filter is about to be installed, the user predicts the particle size distribution pattern of the volcanic ash, and inputs the evaluation conditions. The evaluation condition setting unit 1221 saves the input evaluation conditions in the storage unit 13.

The prediction model setting unit 1222 displays a list of prediction models constructed by the prediction model construction unit 121 on the display unit 14. For example, the prediction model setting unit 1222 displays log information as exemplified in FIG. 5 and a graph showing the progress of prediction error as exemplified in FIG. 7. Then, the prediction model setting unit 1222 outputs the identification number (model No.) of the prediction model selected by the user from the list to the filter replacement time prediction unit 1223.

The filter replacement time prediction unit 1223 predicts the filter replacement time of the volcanic ash filter to be evaluated. The filter replacement time prediction unit 1223 inputs the evaluation conditions set by the evaluation condition setting unit 1221 into the prediction model identified by the prediction model setting unit 1222 (the prediction model selected by the user from the list), obtains the allowable pressure loss arrival time output by the prediction model, and sets this value as the filter replacement time (filter life).

The prediction error display unit 1224 displays the filter replacement time predicted by the filter replacement time prediction unit 1223 on the display unit 14. For example, the prediction error display unit 1224 displays the predicted value of the filter replacement time (L2 in FIG. 6A, L2 in FIG. 6B) obtained by inputting the evaluation conditions on a scatter plot of the actual measured values and predicted values of the prediction model used for the prediction shown in FIG. 6A and FIG. 6B by superimposing it on the scatter plot.

(Operation)
Next, a flow of the process of constructing a prediction model and the process of predicting the filter replacement time in this embodiment will be described.
FIG. 8 is a flowchart showing an example of the operation of the prediction system.
It is assumed that the situation is one in which a life prediction of a volcanic ash filter is to be performed. First, the control unit 12 determines whether or not to construct a prediction model (step S1). For example, if a prediction model has not been constructed, if a certain amount of learning data has been accumulated, if the number of explanatory variables has increased, if the number of test data performed in a different type of plant has increased, etc., the user determines to construct a prediction model and instructs the prediction system 10 to construct a prediction model. The input reception unit 11 notifies the control unit 12 that construction of a prediction model has been instructed. In this case, the control unit 12 determines to construct a prediction model. In addition, if a prediction model has already been constructed and there has been no addition of learning data or change in explanatory variables after construction of the prediction model, or no addition of test data performed in a different type of plant, the user instructs the prediction system 10 to execute a life prediction of the volcanic ash filter. The input reception unit 11 notifies the control unit 12 that a life prediction of the volcanic ash filter has been instructed. In this case, the control unit 12 determines not to construct a prediction model.

When it is determined that a predictive model is to be constructed (step S1; Yes), the user uses spreadsheet software to edit the past test data (step S2). For example, the user checks whether there are any additions to explanatory variables or learning data, and adds explanatory variables or learning data. The input reception unit 11 acquires the data entered by the user and makes changes or additions to the past test data 131.

After completing the editing of the test data (learning data), the user then sets the search range for the parameter setting values to be used in machine learning and preprocessing (step S3). For example, the user sets {15, 25, 35, 45} for the number of decision trees in the random forest. The user also sets search ranges for other parameters. The input receiving unit 11 acquires the search range set by the user and outputs it to the search condition setting unit 1211. The search condition setting unit 1211 records the search range for the setting values set by the user in the memory unit 13.

Next, the user instructs the prediction system 10 to search for setting values. The input reception unit 11 receives the user's instruction and notifies the search unit 1212 that a search for setting values has been instructed. The search unit 1212 searches for setting values (step S4). The search unit 1212 searches for a combination of setting values that results in a good prediction error from within the search range of setting values set for each parameter. Next, the search result display unit 1213 displays a list of the search results (step S5). For example, the search result display unit 1213 displays the search results exemplified in FIG. 4 on the display unit 14.

The user then refers to the list of search results and determines the setting value to be set for each parameter (step S6). For example, the user checks the setting values 1 to 4 determined for parameters 1 to 4 and the value of the allowable pressure loss arrival time, and inputs the settings to the prediction system 10 to narrow down the data to be input when constructing the model. The input reception unit 11 outputs the setting values 1 to 4 input by the user to the construction setting unit 1214. The construction setting unit 1214 sets setting value 1 to parameter 1, setting value 2 to parameter 2, setting value 3 to parameter 3, and setting value 4 to parameter 4.

Next, the user instructs the prediction system 10 to build a prediction model. The input reception unit 11 receives the user's instruction and notifies the construction unit 1215 that construction of a prediction model has been instructed. Based on this instruction, the construction unit 1215 constructs a prediction model (step S7). Based on the setting values 1 to 4, the construction unit 1215 performs preprocessing (oversampling) and regression analysis using random forest to construct a prediction model (unconstrained) 133. The construction unit 1215 also constructs a prediction model (with constraints) 134 using data that has been oversampled and has had outliers removed. The construction unit 1215 also calculates the prediction error between the prediction model (unconstrained) 133 and the prediction model (with constraints) 134. The construction unit 1215 stores the constructed prediction model in the memory unit 13. The construction unit 1215 records a prediction model construction log 132 for the prediction model (unconstrained) 133 and the prediction model (with constraints) 134. In addition, the construction unit 1215 records the item numbers of the learning data used to construct the prediction model in the storage unit 13 in association with the identification number of the prediction model.

Next, the construction result display unit 1216 displays a scatter plot of the predicted values and the actual measured values (step S8). The construction result display unit 1216 displays the scatter plots (excluding L2) shown in Figs. 6A and 6B. The user checks the prediction errors of the prediction model (without constraints) 133 and the prediction model (with constraints) 134. If the prediction accuracy is not sufficient, the user can repeat the processes from step S6 onwards. For example, as described above, a prediction model may be constructed by setting the same setting values 1 to 4 five consecutive times until the prediction error falls below a predetermined value.

Next, the prediction error confirmation unit 1217 obtains the history of prediction errors from the prediction model construction log 132. Next, the prediction error display unit 1218 displays log information based on the prediction model construction log 132 and the trend in prediction errors (step S9). For example, the prediction error display unit 1218 displays, on the display unit 14, the log information of the prediction model construction illustrated in FIG. 5 and a graph showing the trend in prediction errors illustrated in FIG. 7. The user can confirm that the prediction error has become smaller and that a prediction model with high prediction accuracy has been constructed. Once a prediction model has been constructed, learning is completed and a life prediction for the volcanic ash filter can be performed.

If it is determined in step S1 that a prediction model will not be constructed (step S1; No), the prediction system 10 predicts the service life of the volcanic ash filter. First, the user sets the evaluation conditions using spreadsheet software (step S10). For example, the user sets the test conditions for a performance test of the filter as the evaluation conditions. The input reception unit 11 acquires the evaluation conditions entered by the user and stores the evaluation condition data 135 in the memory unit 13.

Next, the user instructs the prediction system 10 to predict the lifespan of the volcanic ash filter under the evaluation conditions. The input reception unit 11 receives the user's instruction and notifies the prediction model setting unit 1222 that a filter lifespan prediction has been instructed. The prediction model setting unit 1222 then displays a list of prediction models on the display unit 14 (step S11). For example, the prediction model setting unit 1222 displays a list of constructed prediction models (log information) as exemplified in FIG. 5.

Next, the user selects a prediction model to be used for prediction (step S12). The input reception unit 11 outputs the model number of the prediction model selected by the user to the prediction model setting unit 1222. The prediction model setting unit 1222 outputs the model number of the prediction model selected by the user to the filter replacement time prediction unit 1223. Next, the filter replacement time prediction unit 1223 predicts the replacement time of the filter (step S13). The filter replacement time prediction unit 1223 inputs the evaluation conditions set in step S10 into the prediction model of the model number output from the prediction model setting unit 1222, and predicts the replacement time of the volcanic ash filter. Next, the prediction error display unit 1224 displays the prediction result (step S14). For example, the prediction error display unit 1224 displays a graph in which the currently predicted "allowable pressure loss arrival time" is superimposed (L2) on the scatter diagram of the actual measurement value and the predicted value as exemplified in FIG. 6A and FIG. 6B. The user can grasp the prediction accuracy along with the predicted value. For example, when the prediction result shown in FIG. 6B is obtained, that is, when P2 is obtained as the predicted value, the actual measured values and predicted values near P2 are concentrated near L1, so the prediction accuracy can be expected to be high.

(effect)
As described above, according to this embodiment, the life of a volcanic ash filter can be predicted before it is used. As a result, for example, the filter replacement time can be predicted in advance at the stage of planning a test of a volcanic ash filter, and the test results for various test condition patterns can be predicted in advance, which can be useful for planning a filter test. In addition, not only in a filter test, but also in a situation where a volcano actually erupts, the filter life can be predicted and understood in advance, so that the volcanic ash filter can be replaced at an appropriate time, and the impact on the plant can be reduced. In addition, according to this embodiment, by simply setting the test data of a past filter test and the search range of parameters related to machine learning, it is possible to search for parameter setting values that can maintain good prediction accuracy. Since a prediction model can be constructed using the searched setting values, a prediction model with high prediction accuracy can be constructed. In addition, the accuracy of the prediction model can be confirmed by referring to the scatter diagrams exemplified in FIG. 6A and FIG. 6B. In addition, by limiting the range of the learning data by the upper limit value and the lower limit value, a prediction model according to the prediction range can be constructed. In addition, an appropriate prediction model can be selected from prediction models created in the past to predict the filter life, and it is possible to check whether the learning of the prediction error of the constructed prediction model is progressing. For example, by referring to the graph showing the transition of prediction error shown in FIG. 7, if the prediction error is reduced by repeating the construction of the prediction model, the last constructed prediction model can be selected. Also, for example, when selecting a prediction model by referring to the log information shown in FIG. 5, by referring to the upper and lower limit values of the learning data, multiple prediction models can be selected from prediction models constructed with learning data having a wide range of upper and lower limit values and prediction models constructed with learning data having a narrow range of upper and lower limit values, and the filter life can be predicted, thereby obtaining various prediction results for an unknown volcanic ash filter. Also, when learning data is added, the prediction model can be reconstructed by adjusting parameters 1 to 4, so that a model with a smaller prediction error can be constructed.

Example 1
In the above embodiment, the prediction model to be used for prediction is selected from the list of constructed learning models to predict the life of the volcanic ash filter (steps S11 to S13), but it is also possible to suggest which prediction model to use at the stage of setting the evaluation conditions (step S10). For example, assume that the design policy of the volcanic ash filter is planned to be "flow velocity is XX or less, mesh is XX or more, and no labyrinth is upstream or downstream." In this case, when testing this volcanic ash filter, it is preferable to predict the life using a prediction model that has learned a lot of learning data that meets these conditions. In such a case, it is difficult for the user to manually check the learning data used to construct each prediction model and identify a prediction model that has been constructed using learning data that contains a lot of data that meets the above conditions. Therefore, the prediction model setting unit 1222 displays a list of prediction models according to the conditions set by the user. The user sets the priority of each condition. For example, the user sets "flow velocity is XX or less" as priority 1, sets "mesh XX or more" as priority 2, and sets "no labyrinth" as priority 3. Then, in step S10, the evaluation condition setting unit 1221 records the priorities 1 to 3 set together with the evaluation conditions in the storage unit 13. Next, in step S12, when selecting and displaying a list of constructed prediction models, the prediction model setting unit 1222 selects a prediction model constructed using a learning data group containing a large amount of learning data satisfying the condition for which a high priority has been set. For example, the prediction model setting unit 1222 investigates the learning data used to construct the prediction model based on the item number of the learning data recorded in the storage unit 13 when constructing the prediction model, and calculates the number of learning data that satisfy the highest priority condition of "flow velocity is XX or less" and what percentage of the learning data that satisfies the condition is included in the learning data group used to construct each prediction model. The prediction model setting unit 1222 extracts a prediction model constructed using a learning data group that contains a large amount of learning data that satisfies the condition of priority 1. For example, the prediction model setting unit 1222 selects a predetermined number of learning data in descending order of the number of learning data that satisfy the condition of priority 1. Alternatively, the prediction model setting unit 1222 selects a predetermined number of learning data in descending order of the proportion of learning data that satisfies the condition of priority 1. Similarly, the prediction model setting unit 1222 extracts a prediction model constructed by learning a group of learning data that contains a large amount of learning data that satisfies the condition of "mesh XX or more" that is assigned the second highest priority. The same is true for "no labyrinth" that is assigned the third highest priority. Then, the prediction model setting unit 1222 displays a predetermined number of prediction models constructed using a large amount of the extracted learning data that satisfies priority 1 in descending order of the number and/or proportion of the data, and similarly displays a predetermined number of prediction models extracted for priority 2 and priority 3. The prediction model setting unit 1222 displays, in a list, the number and proportion of data that satisfies the conditions contained in the learning data used to construct the prediction model together with the extracted prediction model. This allows the user to utilize a prediction model constructed using a large amount of learning data that matches the test conditions of the filter test to be performed to predict the life of the volcanic ash filter to be evaluated.

Example 2
In the above embodiment, in step S1, the user determines whether or not to construct (reconstruct) a prediction model, but this determination may be automatically performed by the determination unit 123. The user sets priorities for the conditions that trigger model construction. For example, the user sets "the number of explanatory variables has been changed" as priority 1, "the test data has increased by more than a predetermined number" as priority 2, "the number of types of nuclear power plants that have not been learned has increased by more than a predetermined number" as priority 3, and "the test data for the rainy season that have not been learned has increased by more than a predetermined number" as priority 4. Then, the determination unit 123 acquires these setting information through the input reception unit 11, and sets the priorities of the conditions that trigger model construction in the storage unit 13. In step S1, the determination unit 123 determines whether or not to construct a model, and displays the determination result on the display unit 14. For example, the determination unit 123 determines whether or not it is necessary to construct a model according to the priorities. First, for priority 1, the determination unit 123 compares the test conditions, i.e., explanatory variables, of the latest test data stored in the storage unit 13 with the explanatory variables of the learning data used to build the already-constructed prediction model to check whether the number of explanatory variables has increased. If the number of explanatory variables has increased, the determination unit 123 determines to build a prediction model. If there is no change in the number of explanatory variables, the determination unit 123 determines whether to build a prediction model based on priority 2, which has the next highest priority. For example, the determination unit 123 subtracts the number of learning data used to build the already-constructed prediction model from the number of all learning data including the latest test data stored in the storage unit 13, and if the difference is equal to or greater than a predetermined number, determines to build a prediction model. Regarding priority 3, the determination unit 123 compares the customer information of the test information of the test data corresponding to the item number of the learning data not used in constructing the prediction model stored in the storage unit 13 with the customer information of the test data corresponding to the item number of the learning data used in constructing the prediction model, and determines that a prediction model will be constructed when the number of unused test data having customer information (e.g., the site name of a nuclear power plant) not used in constructing the prediction model is equal to or greater than a predetermined number. This is because it is considered that if the site name of the nuclear power plant is different, the nearby volcano is different and the particle size distribution pattern of the volcanic ash changes, so that it is necessary to reconstruct the prediction model using learning data with different properties. Next, regarding priority 4, the determination unit 123 compares the customer information of the test information of the test data corresponding to the item number of the learning data not used in constructing the prediction model stored in the storage unit 13 with the customer information of the test data corresponding to the item number of the learning data used in constructing the prediction model, and determines that a prediction model will be constructed when the number of unused test data whose test date corresponds to the rainy season is equal to or greater than a predetermined number. This eliminates the need for the user to check the contents of the test data and determine whether or not to build (reconstruct) a prediction model. When the determination unit 123 determines that a prediction model should be built, it does not automatically build the prediction model. For example, if the determination unit 123 determines that a prediction model should be built based on priority 1, it may display a message such as "Test data with a changed number of explanatory variables has been accumulated. Please consider building a prediction model" on the display unit 14 to prompt the user to build a prediction model. The determination of whether or not to build a prediction model in the second embodiment is not limited to step S1 of the flowchart in FIG. 8 (i.e., the timing when the life expectancy of the volcanic ash filter is about to be predicted), but may be automatically performed at a predetermined cycle (such as once a month).

FIG. 9 is a diagram illustrating an example of a hardware configuration of the prediction system according to the embodiment.
The computer 900 includes a CPU 901, a main memory device 902, an auxiliary memory device 903, an input/output interface 904, and a communication interface 905. The prediction system 10 described above is implemented in the computer 900. The above-described functions are stored in the auxiliary memory device 903 in the form of a program. The CPU 901 reads the program from the auxiliary memory device 903, loads it in the main memory device 902, and executes the above-described processing in accordance with the program. The CPU 901 also secures a storage area in the main memory device 902 in accordance with the program. The CPU 901 also secures a storage area in the auxiliary memory device 903 for storing data being processed in accordance with the program.

A program for realizing all or part of the functions of the prediction system 10 may be recorded on a computer-readable recording medium, and the program recorded on the recording medium may be read into a computer system and executed to perform processing by each functional unit. The term "computer system" here includes hardware such as an OS and peripheral devices. In addition, if a WWW system is used, the term "computer system" also includes a homepage providing environment (or display environment). In addition, the term "computer-readable recording medium" refers to portable media such as CDs, DVDs, and USBs, and storage devices such as hard disks built into a computer system. In addition, if the program is distributed to a computer 900 via a communication line, the computer 900 that receives the program may expand the program into the main storage device 902 and execute the above processing. In addition, the above program may be for realizing part of the functions described above, and may further be capable of realizing the functions described above in combination with a program already recorded in the computer system.

As described above, several embodiments of the present disclosure have been described, but all of these embodiments are presented as examples and are not intended to limit the scope of the invention. These embodiments can be implemented in various other forms, and various omissions, substitutions, and modifications can be made without departing from the gist of the invention. These embodiments and their modifications are included in the scope of the invention and its equivalents as described in the claims, as well as in the scope and gist of the invention.

<Additional Notes>
The prediction system, prediction method, and program described in each embodiment can be understood, for example, as follows.

(1) The prediction system of the first aspect includes an acquisition unit (input receiving unit 11) that acquires specific evaluation conditions related to the volcanic ash filter to be evaluated, a prediction model that outputs a predicted value of the life of an unused volcanic ash filter related to the evaluation conditions when the evaluation conditions are input, and a prediction unit 122 that predicts the life of the volcanic ash filter to be evaluated based on the evaluation conditions acquired by the acquisition unit.
This makes it possible to predict the life of the volcanic ash filter before it is first used.

(2) A prediction system according to a second aspect is the prediction system of (1), wherein the evaluation conditions include a particle size distribution pattern of the volcanic ash captured by the volcanic ash filter, a mesh coarseness of the volcanic ash filter, and a pressure loss threshold of the volcanic ash filter, and the prediction model predicts the time until the pressure loss of an unused volcanic ash filter reaches the threshold.
This makes it possible to predict the life of the volcanic ash filter before it is first used.

(3) A prediction system according to a third aspect is the prediction system of (2), further comprising a prediction model construction unit that constructs the prediction model by learning the relationship between the evaluation conditions and the time until the pressure loss reaches the threshold value through machine learning based on learning data that uses the evaluation conditions as explanatory variables and the time measured in a filter test that measures the time until the pressure loss of an unused volcanic ash filter reaches the threshold value under the evaluation conditions as the objective variable.
This makes it possible to develop a predictive model that can predict the lifespan of a volcanic ash filter before it is put into use.

(4) A prediction system according to a fourth aspect is the prediction system of (3), wherein the prediction model construction unit searches a predetermined search range for setting values for parameters to be used in the machine learning, identifies a predetermined number of setting values in order of the smallest prediction error of the prediction model, displays the identified setting values together with the prediction error associated with the setting values, and sets the parameter to a setting value selected from the identified setting values to construct the prediction model.
This makes it possible to grasp the setting values for the parameters that will provide high prediction accuracy. The setting values that will provide high prediction accuracy are used as a reference to determine the setting values that will actually be used to build a prediction model, so a prediction model with high prediction accuracy can be built.

(5) A prediction system according to a fifth aspect is a prediction system according to any one of (3) to (4), in which, when a filter test is conducted to measure the service life of the volcanic ash filter, the learning data is generated based on the results of the filter test, and the prediction system further includes a determination unit that determines whether or not to execute construction of the prediction model based on any one of the following conditions: a change in the number of evaluation items included in the evaluation conditions in the learning data, an increase in the number of learning data, an increase in the number of plants that have conducted the filter test related to the learning data, and a time when the test related to the increased learning data was conducted.
This makes it possible to automatically determine when to build a predictive model as training data accumulates.

(6) A sixth aspect of the prediction system is the prediction system of (5), in which the determination unit compares, for the condition, the latest learning data and the learning data not used in constructing the prediction model with the learning data used in constructing the prediction model, in order of decreasing priority, and makes the determination based on the result of this comparison.
This makes it possible to automatically determine whether or not a predictive model needs to be built.

(7) A prediction system according to a seventh aspect is a prediction system according to any one of (1) to (6), wherein the prediction unit performs machine learning on learning data having the evaluation conditions as explanatory variables and the lifespan measured in a filter test measuring the lifespan of an unused volcanic ash filter under the evaluation conditions as a target variable, displays a list of a plurality of prediction models constructed using various different machine learning parameters and prediction errors of the prediction models, and predicts the lifespan of the volcanic ash filter using a prediction model selected from the list.
This makes it possible to select a prediction model to be used for predicting the service life of the volcanic ash filter with reference to the prediction accuracy.

(8) A prediction system according to an eighth aspect is the prediction system of (7), wherein, when displaying the list, the prediction unit selects a predetermined number of the learning data used to construct the prediction model in order of the number of the learning data whose evaluation conditions satisfy a predetermined condition, and displays the selected number in the list.
This makes it easier to select a prediction model that is suitable for the evaluation conditions related to the prediction.

(9) In the prediction method according to the ninth aspect, a prediction system obtains predetermined evaluation conditions for the volcanic ash filter to be evaluated, and predicts the lifespan of the volcanic ash filter to be evaluated by inputting the obtained evaluation conditions into a prediction model that outputs a predicted value of the lifespan of the volcanic ash filter related to the evaluation conditions when the evaluation conditions are input.

(10) The program according to the tenth aspect causes the computer 900 to acquire specific evaluation conditions for the volcanic ash filter to be evaluated, and execute a process of predicting the lifespan of the volcanic ash filter to be evaluated by inputting the acquired evaluation conditions into a prediction model that outputs a predicted value of the lifespan of the volcanic ash filter related to the evaluation conditions when the evaluation conditions are input.

10 Prediction system 11 (Acquisition unit) Input reception unit 12 Control unit 121 Prediction model construction unit 1211 Search condition setting unit 1212 Search unit 1213 Search result display unit 1214 Construction setting unit 1215 Construction unit 1216 Construction result display unit 1217 Prediction error confirmation unit 1218 Prediction error display unit 122 Prediction unit 1221 Evaluation condition setting unit 1222 Prediction model setting unit 1223 Filter replacement time prediction unit 1224 Prediction error display unit 123 Determination unit 13 Memory unit 14 Display unit 900 Computer 901 CPU
902: Main memory device 903: Auxiliary memory device 904: Input/output interface 905: Communication interface

Claims (8)

an acquisition unit that acquires predetermined evaluation conditions for the volcanic ash filter to be evaluated;
a prediction model that, when the evaluation conditions are input, outputs a predicted value of the life of an unused volcanic ash filter related to the evaluation conditions; and a prediction unit that predicts the life of the volcanic ash filter to be evaluated based on the evaluation conditions acquired by the acquisition unit;
a prediction model construction unit that constructs the prediction model by learning a relationship between the evaluation conditions and the time until the pressure loss of the unused volcanic ash filter reaches the threshold value by machine learning based on learning data that uses the evaluation conditions as explanatory variables and the time measured in a filter test that measures the time until the pressure loss of the unused volcanic ash filter reaches the threshold value based on the evaluation conditions as a response variable;
Equipped with
The evaluation conditions include a particle size distribution pattern of the volcanic ash captured by the volcanic ash filter, a mesh coarseness of the volcanic ash filter, and the threshold value of the pressure loss of the volcanic ash filter, and the prediction model predicts the time until the pressure loss of the unused volcanic ash filter reaches the threshold value.
Prediction system. The prediction model construction unit is
searching a predetermined search range for setting values for the parameters used in the machine learning, and identifying a predetermined number of the setting values in order of decreasing the prediction error of the prediction model;
Displaying the identified setpoints together with the prediction errors associated with the setpoints;
constructing the prediction model by setting the parameter to the setting value selected from the identified setting values;
The prediction system of claim 1 . When a filter test is conducted to measure the service life of the volcanic ash filter, the learning data is generated based on the results of the filter test;
a determination unit that determines whether or not to execute construction of the prediction model based on any one of the following conditions: a change in the number of evaluation items included in the evaluation conditions in the learning data, an increase in the learning data, an increase in the number of plants in which the filter test related to the learning data has been performed, and a time when the filter test related to the increased learning data has been performed;
The prediction system according to claim 1 or 2, further comprising: The determination unit is
comparing, in descending order of priority, the latest learning data or the learning data not used in constructing the prediction model with the learning data used in constructing the prediction model for the condition, and making the determination based on the result of this comparison;
The prediction system of claim 3 . The prediction unit is
When machine learning is performed on learning data using the evaluation conditions as explanatory variables and the lifetime measured in a filter test for measuring the lifetime of an unused volcanic ash filter based on the evaluation conditions as a response variable, a list of a plurality of the prediction models constructed using various different machine learning parameters and the prediction errors of the prediction models is displayed, and the lifetime of the volcanic ash filter is predicted using a prediction model selected from the list.
A prediction system according to any one of claims 1 to 4. The prediction unit is
When displaying the list, a predetermined number of the learning data used in constructing the prediction model are selected in descending order of the number of the learning data whose evaluation conditions satisfy a predetermined condition, and are displayed in the list.
The prediction system of claim 5 . The prediction system is
A prediction model is constructed by learning a relationship between a predetermined evaluation condition and the time until the pressure loss of an unused volcanic ash filter reaches a threshold value by machine learning based on learning data having the predetermined evaluation condition as an explanatory variable and the time measured in a filter test in which the time until the pressure loss of an unused volcanic ash filter reaches a threshold value based on the evaluation condition as a response variable,
Obtaining the evaluation conditions for the volcanic ash filter to be evaluated;
The obtained evaluation conditions are input into the prediction model to predict the service life of the volcanic ash filter to be evaluated;
The evaluation conditions include a particle size distribution pattern of the volcanic ash captured by the volcanic ash filter, a mesh coarseness of the volcanic ash filter, and the threshold value of the pressure loss of the volcanic ash filter, and the prediction model predicts the time until the pressure loss of the unused volcanic ash filter reaches the threshold value.
Forecasting methods. On the computer,
A prediction model is constructed by learning a relationship between a predetermined evaluation condition and the time until the pressure loss of an unused volcanic ash filter reaches a threshold value by machine learning based on learning data having the predetermined evaluation condition as an explanatory variable and the time measured in a filter test in which the time until the pressure loss of an unused volcanic ash filter reaches a threshold value based on the evaluation condition as a response variable,
Obtaining the evaluation conditions for the volcanic ash filter to be evaluated;
a process of predicting a service life of the volcanic ash filter to be evaluated by inputting the acquired evaluation conditions into the prediction model, the evaluation conditions including a particle size distribution pattern of the volcanic ash captured by the volcanic ash filter, a mesh coarseness of the volcanic ash filter, and the threshold value of the pressure loss of the volcanic ash filter, and the prediction model predicts a time until the pressure loss of the unused volcanic ash filter reaches the threshold value;
A program that executes the following.

Images