Information processing device, control system, index value calculation method, and index value calculation program

Abstract

To perform appropriate processing related to update of a machine learning model.SOLUTION: An information processor (2) comprises an outlier determination unit (204) that determines whether or not input data inputted to a machine learning model is outlier using a threshold for determining whether or not training data included in a training data set is outlier and an index value calculation unit (205) that calculates an index value indicating the suitability of the machine learning model for the input data based on a determination result.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The present invention relates to an information processing device or the like that performs processing related to the exchange and relearning of machine learning models. [Background technology]

[0002] The use of machine learning models is progressing in various fields. For example, Patent Document 1 below describes a waste incineration plant facility that incinerates waste and generates electricity using the heat generated during incineration, where a machine-learned neural network model is used to predict the amount of steam generated after a predetermined time. The same document also describes generating a new machine learning model using newly acquired training data, calculating an evaluation value for the prediction accuracy of the machine learning model, and using the new machine learning model if the evaluation value is higher than that of the previous machine learning model. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Application Laid-Open No. 2005-249349 Summary of the Invention [Problem to be solved by the invention]

[0004] However, in the conventional technology described above, it is difficult to properly evaluate the prediction accuracy of the machine learning model. This is because, when the technology of Patent Document 1 is applied, if it is predicted that the amount of steam generated will decrease after a predetermined time, intervention control is performed to prevent the decrease in the amount of steam. In other words, if the amount of steam generated does not decrease after a prediction that the amount of steam generated will decrease, it is impossible to distinguish whether the prediction was wrong or whether the prediction was correct but the decrease was prevented by intervention control, making it difficult to properly evaluate the prediction accuracy.

[0005] As described above, when it is difficult to verify the accuracy of inference results from a machine learning model, there is a problem that it becomes difficult to appropriately perform various processes related to updating the machine learning model, such as determining the need to replace the machine learning model or the need for re-learning, or selecting data to be used for re-learning. This problem is not limited to predictive models used in plants, but is a common problem that arises when using machine learning models that make inferences about phenomena for which it is difficult to verify the accuracy of inference results.

[0006] One aspect of the present invention aims to realize an information processing device, etc. that enables appropriate processing related to updating a machine learning model even when it is difficult to verify the accuracy of the inference results produced by the machine learning model. [Means for solving the problem]

[0007] In order to solve the above problem, an information processing device according to one embodiment of the present invention includes an outlier determination unit that determines whether each of a plurality of input data input to a machine learning model is an outlier using a threshold for determining whether training data included in a training dataset used to generate the machine learning model is an outlier, and an index value calculation unit that calculates an index value indicating the suitability of the machine learning model for the input data based on the determination result of the outlier determination unit.

[0008] In addition, in order to solve the above-mentioned problem, an information processing device according to another aspect of the present invention includes an outlier determination unit that determines whether each of multiple input data input to a machine learning model is an outlier using a threshold for determining whether training data included in a training dataset used to generate the machine learning model is an outlier, and an alarm unit that prompts re-learning of the machine learning model depending on the determination result of the outlier determination unit.

[0009] In addition, in order to solve the above-mentioned problem, an information processing device according to yet another aspect of the present invention includes an outlier determination unit that determines whether each of a plurality of input data input to a machine learning model is an outlier using a threshold for determining whether training data included in a training dataset used to generate the machine learning model is an outlier, and a learning data extraction unit that extracts data to be used for re-learning the machine learning model from the plurality of input data based on the determination result of the outlier determination unit.

[0010] In addition, in order to solve the above-mentioned problem, an information processing device according to yet another aspect of the present invention includes an outlier determination unit that determines whether each of multiple input data input to the machine learning model is an outlier using a threshold for determining whether training data included in a training dataset used to generate the machine learning model is an outlier, and a learning unit that re-trains the machine learning model at a timing determined based on the determination result of the outlier determination unit.

[0011] In addition, in order to solve the above-mentioned problem, an index value calculation method according to one embodiment of the present invention is an index value calculation method executed by at least one information processing device, and includes an outlier determination step of determining whether or not each of multiple input data input to the machine learning model is an outlier using a threshold for determining whether or not training data included in a training dataset used to generate the machine learning model is an outlier, and an index value calculation step of calculating an index value indicating the suitability of the machine learning model for the input data based on the determination result in the outlier determination step. [Effects of the Invention]

[0012] According to one aspect of the present invention, it is possible to appropriately perform processing related to updating a machine learning model. [Brief explanation of the drawings]

[0013] [Figure 1]1 is a diagram showing an overview of a control system according to an embodiment of the present invention. [Figure 2] A block diagram showing an example of the main configuration of an information processing device (a device that calculates an index value indicating the suitability of an inference model) according to one embodiment of the present invention. [Figure 3] 1 is a block diagram showing an example of a configuration of a main part of an information processing device (a device that searches for optimal values ​​of parameters) according to one embodiment of the present invention. [Figure 4] 10A and 10B are diagrams illustrating a method for determining a threshold value and a method for determining whether input data is an outlier. [Figure 5] FIG. 1 is a diagram showing the transition of calculated index values ​​and recurrence rates in an incineration power plant that generates electricity by utilizing heat generated by incinerating waste. [Figure 6] FIG. 10 is a diagram showing an example of a display screen of a search result. [Figure 7] 10 is a flowchart showing an example of processing performed by the information processing device (a device that calculates an index value indicating the suitability of an inference model). [Figure 8] 6 is a flowchart showing an example of processing executed by the information processing device (device that searches for optimal values ​​of parameters). DETAILED DESCRIPTION OF THE INVENTION

[0014] [System Overview] 1 is a diagram showing an overview of a control system 7 according to one embodiment of the present invention. As shown in the figure, the control system 7 includes a learning device 1, an information processing device 2, an information processing device 3, a control device 4, and a control target 5. The control system 7 is a system in which the control device 4 controls the operation of the control target 5, and the control of the control target 5 is performed based on the results of inference using an inference model.

[0015] The controlled object 5 is an object controlled by the control device 4 and includes a device 51. A measuring device 52 is attached to the controlled object 5. The measuring device 52 measures data related to the device 51, and may be adapted to suit the device 51 and the data to be measured. For example, if it is desired to measure the internal or external temperature of the device 51, a temperature sensor may be used as the measuring device 52.

[0016] The controlled object may include one or more of each of the equipment 51 and the measuring device 52. For example, if the controlled object 5 is a plant, the equipment 51 is an equipment within the plant. In this case, the measuring device 52 may be attached to the equipment 51, or may be attached to another location within the plant, or may be installed outside the plant.

[0017] The learning device 1 generates an inference model used to determine the content of control for the control target 5. The inference model can also be called a machine learning model because it is generated by machine learning using training data. The learning device 1 also re-learns the inference model and updates the inference model. Note that any machine learning algorithm can be used, and an appropriate algorithm can be applied depending on the training data used, the inference content, etc.

[0018] The inference model may be any model that performs inference on matters that serve as guidelines for determining the content of control for the control target 5. For example, the inference model may be one that predicts whether the future operating state of the control target 5 will be normal or not. Also, for example, the inference model may be one that predicts future measurement values ​​of the measuring device 52, or one that predicts the optimal content of control for the control target 5.

[0019] The information processing device 2 calculates an index value indicating the suitability of the inference model for input data input to the inference model generated by the learning device 1 when making an inference using the inference model. As will be described in detail later, when calculating the index value, the information processing device 2 determines, for each of the multiple input data input to the inference model, whether the input data is an outlier, using a threshold value for determining whether training data included in the training dataset used to generate the inference model is an outlier. The information processing device 2 then calculates the index value based on the determination result. This makes it possible to accurately determine the suitability of the inference model for the input data. The index value is used to determine when to cause the learning device 1 to update the inference model.

[0020] The information processing device 3 searches for optimal values ​​of parameters related to the control of the controlled object 5. The information processing device 3 may also search for optimal values ​​of parameters related to inference by an inference model instead of or in addition to the parameters related to the control of the controlled object 5. The parameters detected by this search are provided to the control device 4, and the control device 4 applies these parameters to control the controlled object 5.

[0021] As will be described in detail later, the information processing device 3 calculates a predictive distribution of a function indicating the relationship between the above-mentioned parameters related to at least one of the inference by the inference model and the control based on the result of the inference, and the operating state of the control target 5 after the control, and searches for candidates for the optimal value of the parameter based on the calculated predictive distribution. This makes it possible to appropriately set the parameter without being influenced by the skill of the operator.

[0022] The control device 4 performs inference using the above-mentioned inference model, and determines the control content for the control target 5 based on the result of the inference. Specifically, the control device 4 includes a data acquisition unit 41, an inference unit 42, and a control content determination unit 43, and the above-mentioned functions are realized by these units.

[0023] The data acquisition unit 41 acquires data necessary for inference. For example, the data acquisition unit 41 may acquire measurement data measured by a measurement device 52. The data acquired by the data acquisition unit 41 is used as input data for the inference model either directly or after undergoing predetermined data processing. Then, the inference unit 42 inputs the input data into the inference model and obtains output data.

[0024] The control content determination unit 43 determines the control content according to the result of inference by the inference unit 42, i.e., the output data described above. For example, if the inference model predicts whether the future operating state of the controlled object 5 will be normal or not, the control content determination unit 43 may not change the control content from the previous content when output data indicating that the operating state is likely to be normal is output. On the other hand, when output data indicating that the operating state is likely to be abnormal is output, the control content determination unit 43 may determine the control content for normalizing the operating state. The control content is determined based on parameters detected by the information processing device 3.

[0025] When the learning device 1 updates the inference model, the data acquisition unit 41 acquires the updated inference model, the inference unit 42 obtains output data by inputting input data into the updated inference model, and the control content determination unit 43 determines the control content for the control object 5 based on the output data.

[0026] As described above, the control system 7 includes an information processing device 2 that determines whether input data input to an inference model is an outlier using a threshold for determining whether training data is an outlier, and calculates an index value indicating the suitability of the inference model for the input data based on the determination result, a learning device 1 that updates the inference model at a timing determined based on the calculated index value, and a control device 4 that determines the control content for the control target 5 based on output data obtained by inputting the input data to the updated inference model. When the inference model is updated in this manner, the information processing device 2 acquires the input data input to the updated inference model and calculates an index value indicating the suitability of the inference model for the input data.

[0027] In this way, the control system 7 can repeatedly re-learn and update the inference model and calculate an index value that indicates the compatibility between the inference model and input data. This makes it possible to re-learn at an appropriate time and maintain the appropriateness of the control content for the controlled object 5.

[0028] As described above, the control system 7 includes an information processing device 3 that calculates a predictive distribution of a function that indicates the relationship between a certain parameter and the operating state of the controlled object 5 during a period in which the parameter is applied, and searches for a candidate optimal value for the parameter based on the calculated predictive distribution, and a control device 4 that applies the detected candidate optimal value to control the controlled object 5. The certain parameter is a parameter related to at least one of the control of the controlled object 5 and inference using an inference model.

[0029] The information processing device 3 then acquires result data indicating the operating state during the period in which the detected candidate is applied, updates the predictive distribution based on the result data, and searches for new candidates for the optimal value of the parameter based on the updated predictive distribution. This allows the control system 7 to bring the parameter closer to the optimal value while controlling the controlled object 5.

[0030] [Configuration of information processing device 2] 2 is a block diagram showing an example of the main configuration of the information processing device 2. As shown in the figure, the information processing device 2 includes a control unit 20 that controls all of the units of the information processing device 2, and a storage unit 21 that stores various data used by the information processing device 2. The information processing device 2 also includes a communication unit 22 that enables the information processing device 2 to communicate with other devices, an input unit 23 that accepts input of various data to the information processing device 2, and an output unit 24 that enables the information processing device 2 to output various data.

[0031] The control unit 20 also includes a data acquisition unit 201, an average distance calculation unit 202, a threshold determination unit 203, an outlier determination unit 204, an index value calculation unit 205, a relearning necessity determination unit 206, a recall calculation unit 207, and a learning data extraction unit 208.

[0032] The data acquisition unit 201 acquires various data used in the information processing device 2. For example, the data acquisition unit 201 acquires a training dataset including multiple pieces of training data used to generate an inference model, and input data input to the inference model for inference. For example, the data acquisition unit 201 may acquire a training dataset from the learning device 1 and input data from the control device 4.

[0033] The average distance calculation unit 202 calculates the average distance between data. Specifically, the average distance calculation unit 202 performs a process for each training data item to calculate the average value of the distances between one piece of training data included in the training data set acquired by the data acquisition unit 201 and each of a predetermined number of training data items that are closest to the training data item. Furthermore, the average distance calculation unit 202 performs a process for each input data item to calculate the average value of the distances between one piece of input data item acquired by the data acquisition unit 201 and each of a predetermined number of other input data items that are closest to the input data item.

[0034] The threshold determination unit 203 determines a threshold for determining whether training data included in the training data set acquired by the data acquisition unit 201 is an outlier. More specifically, the threshold determination unit 203 calculates a deviation indicating the degree to which each training data item included in the training data set deviates from other training data items, and determines a deviation of a predetermined rank among the calculated deviations as the threshold. The deviation may be, for example, an average value calculated by the average distance calculation unit 202 for each training data item.

[0035] The outlier determination unit 204 determines whether or not each of the plurality of input data acquired by the data acquisition unit 201 is an outlier. This determination is made using the threshold determined by the threshold determination unit 203. That is, the outlier determination unit 204 determines whether or not the input data is an outlier based on the threshold determined by the threshold determination unit 203.

[0036] The index value calculation unit 205 calculates an index value that indicates the suitability of the inference model for the input data based on the determination result of the outlier determination unit 204. The index value will be explained later in the section [Example of Index Value].

[0037] The relearning necessity determination unit 206 determines whether or not the inference model needs to be re-learned based on the index value calculated by the index value calculation unit 205. When the relearning necessity determination unit 206 determines that the inference model needs to be re-learned, it instructs the learning device 1 to re-learn the inference model. In other words, the timing when the relearning necessity determination unit 206 determines that relearning is necessary is the timing when relearning should be performed, and it can be said that the relearning necessity determination unit 206 determines the timing of relearning.

[0038] The recall calculation unit 207 calculates the recall for each inference performed by inputting multiple pieces of input data into the inference model. The recall is an index indicating the proportion of cases where the true value is predicted as a positive case among positive cases. For example, in the case of an inference model that predicts whether a state is normal or abnormal, the recall can be calculated using the formula (number of cases correctly determined as abnormal) / {(number of cases correctly determined as abnormal)+(number of cases where an abnormality was erroneously determined as normal)}. In other words, the recall calculation unit 207 calculates the recall for the determination by dividing the number of cases correctly determined as abnormal by the sum of the number of cases correctly determined as abnormal and the number of cases where an abnormality was erroneously determined as normal.

[0039] The learning data extraction unit 208 extracts input data to be used for relearning the inference model from the multiple input data acquired by the data acquisition unit 201, based on the index values ​​calculated by the index value calculation unit 205. As described above, the index values ​​calculated by the index value calculation unit 205 are calculated based on the determination result of the outlier determination unit 204. Therefore, it can be said that the learning data extraction unit 208 extracts input data to be used for relearning based on the determination result of the outlier determination unit 204.

[0040] Furthermore, the learning data extraction unit 208 may extract input data to be used for relearning, taking into consideration the recall calculated by the recall calculation unit 207. For example, the extracted input data may be transmitted to the learning device 1 when the relearning necessity determination unit 206 determines that relearning of the inference model is necessary.

[0041] As described above, the information processing device 2 includes an outlier determination unit 204 that determines whether each of multiple input data input to the inference model is an outlier using a threshold value for determining whether the training data included in the training dataset used to generate the inference model is an outlier, and an index value calculation unit 205 that calculates an index value indicating the suitability of the inference model for the input data based on the determination result of the outlier determination unit 204.

[0042] According to the above configuration, whether or not input data input to the inference model is an outlier is determined using a threshold for determining whether or not training data is an outlier. In this process, input data that would be determined to be an outlier if it were included in the training data set, i.e., input data that deviates from the learned range, is determined to be an outlier.

[0043] When the number or proportion of input data that deviates from the learned range increases, it can be said that the suitability of the inference model for the input data is decreasing. In addition, at this time, it can be said that the need for re-learning or replacement of the inference model is increasing. Therefore, the above configuration that calculates the index value based on the above judgment result makes it possible to calculate a reasonable index value.

[0044] Furthermore, this index value can be calculated without using the inference results of the inference model. Therefore, with the above configuration, it is possible to calculate a valid index value even when it is difficult to verify the accuracy of the inference result, that is, to accurately determine the suitability of the inference model for the input data used in the inference. Furthermore, by using this index value, it is also possible to appropriately perform various processes related to updating the inference model.

[0045] As described above, the information processing device 2 is equipped with an outlier determination unit 204 that determines whether or not each of the multiple input data input to the inference model is an outlier using a threshold value for determining whether or not the training data included in the training dataset used to generate the inference model is an outlier, and a learning data extraction unit 208 that extracts data to be used for re-learning the inference model from the multiple input data input to the inference model based on the determination result of the outlier determination unit 204.

[0046] As described above, when the number or proportion of input data that deviates from the learned range increases, it can be said that the suitability of the inference model for the input data is decreasing. Furthermore, at this time, it can be said that the need for re-learning or replacement of the inference model is increasing. Therefore, with the above configuration that extracts input data to be used for re-learning the inference model based on the determination result of the outlier determination unit 204, it is possible to extract appropriate data for re-learning.

[0047] The learning data extraction unit 208 may extract input data using the index value calculated by the index value calculation unit 205, or may extract input data using the determination result of the outlier determination unit 204 without using the index value. In other words, it is not essential to provide the index value calculation unit 205 when extracting input data to be used for relearning.

[0048] Furthermore, with the above configuration, it is not necessary to use the inference results of the inference model when extracting input data to be used for re-learning. Therefore, with the above configuration, even when it is difficult to verify the accuracy of the inference results, it is possible to appropriately perform the process of extracting data for re-learning, which is a process related to updating the inference model.

[0049] [Configuration of information processing device 3] 3 is a block diagram showing an example of the configuration of the main parts of the information processing device 3. As shown in the figure, the information processing device 3, like the information processing device 2, includes a control unit 30, a storage unit 31, a communication unit 32, an input unit 33, and an output unit 34. The control unit 30 of the information processing device 3 includes a data acquisition unit 301, an evaluation value calculation unit 302, a predictive distribution calculation unit 303, a search unit 304, an optimization control unit 305, and a display control unit 306.

[0050] The data acquiring unit 301 acquires various data used in the information processing device 3. Specifically, the data acquiring unit 301 acquires result data indicating the operating state of the control target 5 during a period in which control was performed using parameters detected by the information processing device 3 (more precisely, the searching unit 304). For example, the data acquiring unit 301 may acquire measurement data of the measuring device 52 during that period as the result data.

[0051] The evaluation value calculation unit 302 calculates an evaluation value that evaluates the operating state of the control target 5 using the result data acquired by the data acquisition unit 301. The evaluation value may be any value that represents the quality of the operating state of the control target 5. It can also be said that this evaluation value represents the quality of the applied parameters.

[0052] For example, in the control system 7, when an inference model that predicts a value indicating the future operating state of the controlled object 5 is used to perform control to maintain the operating state of the controlled object 5 in a normal state, the evaluation value calculation unit 302 may use the abnormal time rate as the evaluation value. The abnormal time rate is the ratio of the time during which the operating state of the controlled object 5 is abnormal to the operating time of the controlled object 5. Whether the operating state of the controlled object 5 is abnormal or normal can be determined based on the result data acquired by the data acquisition unit 301.

[0053] The predictive distribution calculation unit 303 calculates the predictive distribution of a function indicating the relationship between the above-mentioned parameters and the operating state of the controlled object 5 during the period in which the parameters are applied. The operating state of the controlled object 5 is represented by an evaluation value calculated by the evaluation value calculation unit 302. For example, when the evaluation value calculation unit 302 calculates the abnormal time rate, the predictive distribution calculation unit 303 uses the abnormal time rate as information indicating the operating state of the controlled object 5. In this case, it becomes possible to set the parameters to values ​​that can reduce the abnormal time rate. Note that instead of having the evaluation value calculation unit 302 calculate the evaluation value, the evaluation value may be input via the input unit 33 or the like, and the evaluation value calculation unit 302 may be omitted.

[0054] The search unit 304 searches for candidates for the optimal values ​​of the parameters based on the predictive distribution calculated by the predictive distribution calculation unit 303. Note that the method for calculating the predictive distribution and the method for searching for the candidates will be described later in the section [Details of the method for calculating the predictive distribution and the method for searching for the optimal value candidates].

[0055] The optimization control unit 305 controls the parameter optimization by the predictive distribution calculation unit 303 and the search unit 304. For example, when a predetermined condition is satisfied, the optimization control unit 305 controls the predictive distribution calculation unit 303 and the search unit 304 to terminate the optimization.

[0056] The display control unit 306 causes a display device to display and output the candidates detected by the search unit 304. The display device may be a device included in the information processing device 3. For example, if the output unit 34 is a display device, the display control unit 306 may cause the output unit 34 to display and output the candidates detected by the search unit 304. Furthermore, the display control unit 306 may cause a display device connected to the information processing device 3 to display and output, or may cause a display device connected to another device via another device that can communicate with the information processing device 3 via the communication unit 32 to display and output.

[0057] As described above, the information processing device 3 includes a predictive distribution calculation unit 303 that calculates the predictive distribution of a function that indicates the relationship between the above-mentioned parameter and the operating state of the control object 5 during the period in which the parameter is applied, and a search unit 304 that searches for candidates for the optimal value of the parameter based on the calculated predictive distribution.

[0058] According to the above configuration, it is possible to detect appropriate parameters even when it is difficult to formulate the relationship between the parameters and the operating state. Therefore, according to the above configuration, it is possible to appropriately set parameters without being influenced by the skill of individuals in the control system 7 that controls a predetermined control object according to the result of inference by the inference model.

[0059] Furthermore, the predictive distribution calculation unit 303 updates the predictive distribution based on the candidates detected by the search unit 304 and the operating state of the control target 5 during the period in which the candidates are applied. Then, the search unit 304 searches for new candidates for the optimal values ​​of the parameters based on the updated predictive distribution.

[0060] With this configuration, a new candidate for the optimum value is searched for based on the operating state when the previously detected candidate for the optimum value is applied, making it possible to detect a candidate with a higher degree of validity. Furthermore, even if the state of the controlled object changes due to seasonal fluctuations or aging, for example, it is possible to detect a candidate that is appropriate for the changed state. While this configuration may involve some trial and error, the number of such trials is kept to a minimum.

[0061] [Method of determining threshold and method of determining whether or not a value is an outlier] 4 is a diagram illustrating a method for determining a threshold and a method for determining whether input data is an outlier. In this section, a method for determining a threshold by the threshold determination unit 203 and a method for determining whether input data is an outlier by the outlier determination unit 204 will be described based on FIG.

[0062] The threshold determination unit 203 determines a threshold using a training data set used in the machine learning of the inference model. In determining the threshold, the average distance calculation unit 202 performs a process of calculating, for each training data, the average value of the distances between one piece of training data included in the training data set and each of a predetermined number of training data sets that are closest to the training data. As described above, the average value calculated in this manner indicates the degree to which each training data set deviates from other training data sets, and can also be called the degree of deviation.

[0063] The left side of FIG. 4 shows plots in the feature space of each piece of training data included in the training dataset. For example, assume that the above-mentioned predetermined number is 5. In this case, for data D1, which is one piece of training data, the average distance calculation unit 202 calculates the distance between data D1 and each of the five pieces of training data that are closest to data D1 in the feature space, and calculates the average value of these distances. Similarly, for data D2, the average distance calculation unit 202 calculates the average value of the distance between data D2 and each of the five pieces of training data that are closest to data D2. The average distance calculation unit 202 performs this process for each piece of training data included in the training dataset. This process is similar to the process performed in the k-nearest neighbor method, which is a data classification method.

[0064] Next, the threshold determination unit 203 sorts the average values ​​calculated as described above in ascending order, finds the average value that ranks in a predetermined order, and determines the average value as the threshold. The training data with a larger calculated average value is more deviated from the other training data and is more likely to be an outlier from the perspective of the entire training data set. For example, since the average value calculated for data D2 shown in Figure 4 is larger than the average value calculated for data D1, data D2 is more likely to be an outlier than data D1.

[0065] The predetermined ranking may be set as appropriate, and there are no particular limitations on the value to which it is set. For example, the predetermined ranking may be set so that a predetermined percentage of training data in the training data set is ranked higher than the predetermined ranking. To give a specific example, if 96% of the training data is to be ranked higher than the predetermined ranking, the predetermined ranking is (total number of training data) x 0.96. In this case, 4% of the training data will be outliers.

[0066] On the other hand, plots in the feature space of each piece of input data input to the inference model are shown on the right side of Fig. 4. The following describes how the outlier determination unit 204 determines whether or not a value is an outlier.

[0067] The outlier determination unit 204 uses the threshold determined as described above to determine whether or not input data input to the inference model is an outlier. The determination of whether or not an input data is an outlier also begins with the calculation of an average value by the average distance calculation unit 202. More specifically, the average distance calculation unit 202 performs a process for each input data item to calculate the average value of the distances between one piece of input data and each of a predetermined number of input data items that are closest to the input data item. The average value calculated by the average distance calculation unit 202 indicates the degree to which each piece of input data deviates from the other input data items, and may also be called the degree of deviation.

[0068] 4, the predetermined number is 5, as in the example on the left side of the figure. In this case, for data d1, which is one of the input data, the average distance calculation unit 202 calculates the distance between data d1 and each of the five input data that are closest to data d1 in the feature space, and calculates the average value of these distances.

[0069] Then, the outlier determination unit 204 compares the calculated average value with the threshold determined by the threshold determination unit 203, and determines whether or not the data d1 is an outlier based on the magnitude relationship between them. For example, as shown in the figure, the outlier determination unit 204 may determine that the data d1 is an outlier when the average value calculated for the data d1 is equal to or greater than the threshold. In the same manner, the outlier determination unit 204 determines whether or not each of the other input data is an outlier.

[0070] The method of detecting outliers is not limited to the above example. For example, outliers may be detected by LOF (Local Outlier Factor). In this case, the threshold value determination unit 203 determines a threshold value for local density, and the outlier determination unit 204 determines input data whose local density is equal to or less than the threshold value as an outlier. Alternatively, outliers may be detected by Hotelling's theory, for example. In this case, the threshold value determination unit 203 determines a threshold value for the degree of anomaly, and the outlier determination unit 204 determines input data whose degree of anomaly is equal to or greater than the threshold value as an outlier.

[0071] As described above, the threshold determination unit 203 may calculate a deviation indicating the degree to which each piece of training data included in the training dataset deviates from other pieces of training data, and determine a deviation of a predetermined rank from the calculated deviations as the threshold.The outlier determination unit 204 may then determine whether each piece of input data is an outlier by comparing the deviation calculated for each piece of input data included in the plurality of input data, indicating the degree to which the input data deviates from other pieces of input data, with the threshold determined by the threshold determination unit 203.

[0072] If the number of input data items determined to be outliers increases, it can be said that the characteristics of the input data used for inference have changed to become different from the characteristics of the training data. Therefore, with the above configuration, it is possible to accurately determine the suitability of the inference model for the input data.

[0073] As with the method for detecting outliers, the method for calculating the deviation is not particularly limited. For example, as described above, the deviation may be the average distance from a predetermined number of nearby data. In this case, the distance may be, for example, the Euclidean distance, or another distance such as the Manhattan distance or the Mahalanobis distance. Furthermore, for example, the deviation may be the degree of anomaly in Hotelling's theory, or the local density used in LOF.

[0074] [Example of index value] The index value calculated by the index value calculation unit 205 may be any value that indicates the suitability of the inference model for the input data. For example, the index value calculation unit 205 may use the total number of input data that the outlier determination unit 204 has determined to be outliers over a predetermined period as the index value. For example, when the control target 5 is operated every day, it is assumed that each of the multiple input data used in inference during a day is determined to be an outlier. In this case, the index value calculation unit 205 may calculate the total number of input data determined to be outliers as the index value that indicates the suitability of the inference model at the end of that day.

[0075] Furthermore, for example, the index value calculation unit 205 may calculate the index value using the number of pieces of input data determined to be outliers by the outlier determination unit 204. For example, the index value calculation unit 205 may calculate, as the index value, the proportion of input data determined to be outliers by the outlier determination unit 204 among the input data measured by the measuring device 52 during a predetermined period.

[0076] Furthermore, if the inference model is a model that predicts whether the future state of a specified object (e.g., a controlled object 5) will be normal or abnormal based on input data about the object, the index value calculation unit 205 may calculate as an index value the number of input data that were determined to be outliers and for which the state of the object was abnormal after prediction using the input data, or a value calculated using the number of input data.

[0077] Input data determined to be outliers can be classified into those for which the target state was abnormal after prediction using the input data and those for which the target state was normal. Of these, those for which the target state was abnormal after prediction can be said to contain characteristics corresponding to an abnormal state.

[0078] Generally, a normal state does not change, but an abnormal state varies, so when there are many outliers in input data that contain characteristics corresponding to an abnormal state, it is possible that the abnormal state has changed since learning, and re-learning is highly necessary. Therefore, with the above configuration, it is possible to calculate an appropriate index value that takes into account the changing nature of the abnormal state.

[0079] [Determining whether or not re-learning is necessary, taking recall into account] As described above, the inference model may be a model that predicts whether a future state of a given object will be normal or abnormal based on input data about the object. In this case, it is preferable to evaluate the accuracy of the inference model using recall, in order to reduce the possibility of overlooking an abnormal state, i.e., determining an abnormal state as normal when it is actually abnormal.

[0080] It can be difficult to verify the accuracy of the inference results of the above-mentioned inference model, but by using the recall rate together with the index value calculated by the index value calculation unit 205 rather than using it alone, effective re-learning becomes possible.

[0081] This will be explained with reference to Fig. 5. Fig. 5 is a diagram showing the transition of the index value calculated by the index value calculation unit 205 and the recall rate in a waste incineration power plant that generates power by using heat generated by incinerating waste. That is, in this example, the control object 5 is a waste incineration power plant. The index value is the total number of input data determined to be outliers per day. The figure also shows the transition of the set value of the steam flow rate in the waste incineration power plant. The horizontal axis of the figure represents time (days), and the values ​​on the vertical axis are normalized to values ​​between 0 and 1.

[0082] In Figure 5, during the period indicated by circle C1, the recall rate remained at a high level. Meanwhile, the index value, which had been low until recently, suddenly increased and then remained at a high level. Since the setting value changed to a large value immediately before the sudden increase in the index value, it is believed that the sudden increase in the index value was caused by a change in the setting value, which changed the characteristics of the input data.

[0083] On the other hand, in the period indicated by circle C2, the index value increased and the recall rate decreased, even though there was no significant change in the setting value. During this period, factors other than the change in the setting value are thought to have caused an increase in the amount of input data judged to be outliers, and the recall rate to decrease. This period of increased index value and decreased recall rate can be said to be a period in which the inference model's ability to classify normal and abnormal conditions is low.

[0084] In this way, when an inference model is used that predicts whether the future state of a specified object will be normal or abnormal based on input data about the object, the reproducibility calculation unit 207 may calculate the reproducibility of a prediction made by inputting multiple input data into the inference model, i.e., the number of cases that were correctly determined to be abnormal divided by the sum of the number of cases that were correctly determined to be abnormal and the number of cases that were incorrectly determined to be abnormal as normal.

[0085] A period in which recall is declining can be considered a period in which the accuracy of the inference model's predictions of whether something is normal or abnormal is declining. Therefore, recall can be used as an indicator of the level of the inference model's ability to classify normal and abnormal data. However, the number of cases correctly identified as abnormal and the number of cases incorrectly identified as normal, which are used to calculate recall, are affected by intervention control based on the prediction results of the inference model and operator actions, so recall is not an absolute evaluation indicator. On the other hand, by using the above-mentioned indicator value, it is possible to identify periods in which the input data used tends to be unlearned.

[0086] For this reason, the learning data extraction unit 208 may extract input data to be used for re-learning the inference model from among the multiple input data, based on the recall calculated by the recall calculation unit 207 and the index value calculated by the index value calculation unit 205. This makes it possible to estimate a period in particular that requires learning (a period in which the input data has a tendency to be unlearned and has become an anomaly that should be avoided), identify a data period that requires efficient learning, and extract input data from that period as data for re-learning. For example, in the example of Figure 5, effective re-learning can be performed using input data from the period indicated by circle C2 among the acquired input data.

[0087] [Details of how to calculate the predictive distribution and how to search for optimal value candidates] The following describes in detail the method for calculating the predictive distribution (including updating) by the predictive distribution calculation unit 303 and the method for searching for optimal value candidates by the search unit 304. Note that the following description relates to the method for calculating the predictive distribution and the search method in Bayesian optimization. However, it is also possible to apply methods other than those described below as long as they use the predictive distribution to search for optimal value candidates.

[0088] If there are N parameters to be optimized, the control parameters are

[0089]

number

[0090] is expressed as The evaluation value for this is

[0091]

number

[0092] It is expressed as:

[0093] The predictive distribution calculation unit 303 calculates the predictive distribution of a function indicating the relationship between the parameters to be optimized and the operating state of the controlled object after control, based on the result data acquired by the data acquisition unit 301. This function will be referred to as the evaluation function f(θ) hereinafter. When new result data (for example, result data indicating the result of control when a candidate detected by the search unit 304 is applied) is acquired, the predictive distribution calculation unit 303 updates the predictive distribution so that the result data is reflected.

[0094] The relationship between the parameters and the operating state is calculated using Gaussian noise ε n Using ~N(0,β)

[0095]

number

[0096] Assuming this, the following distribution is obtained as the predictive distribution of the evaluation function using a Gaussian process.

[0097]

number

[0098]

number

[0099]

number

[0100] where k * =k(θ,θ), and K Θ is [K Θ ] i,j =k(θ i ,θ j ) is the Gram matrix obtained by

[0101]

number

[0102] k Θ,* is [k Θ,* ] i =k(θ i ,θ), and k(·,·) is the kernel function. Here, the parameter of the kernel function is α k Let's say.

[0103] The mean function μ(θ) indicates the average value of the evaluation function predicted from the result data (and the evaluation value calculated by the evaluation value calculation unit 302 from the result data). The variance function σ(θ) is the variance of the evaluation function predicted from the result data (and the evaluation value calculated by the evaluation value calculation unit 302 from the result data). σ(θ) indicates the uncertainty of the prediction, and its value tends to be large in areas where result data is insufficient. When σ is large, it can be said that the prediction is uncertain, and that there is a lack of result data necessary to increase the certainty of the prediction. As is clear from formula (3), the kernel function included in the variance function σ(θ) and the parameter α of the kernel function k affects the calculation of the predictive distribution. When calculating the predictive distribution, the parameter α k The optimization method is not particularly limited, and for example, various optimization methods that are generally used in Bayesian optimization can be applied.

[0104] The search unit 304 searches for optimal control parameter candidates to determine the optimal control parameters. Specifically, the search unit 304 searches for parameters that maximize the acquisition function a(θ) using the mean function μ(θ) and the variance function σ(θ), as shown in equations (4) and (5) below. The parameters detected in this search become optimal parameter candidates. This search is based on the UCB (Upper Confidence Bound) strategy.

[0105] In formula (5), κ is a parameter for adjusting search and utilization. When the reception unit 104 receives a selection of whether to prioritize the search around a value where the feedback control result is known or the range where the feedback control result is unknown, the search unit 304 sets κ to a value according to the received selection and searches for candidates.

[0106] Of course, it is also possible to search for new parameters by other methods. For example, optimal parameter candidates may be searched for using the PI (Probability of Improvement) strategy or the EI (Expected Improvement) strategy. In addition, optimal parameter candidates may be searched for by applying strategies such as PTR (Probability in Target Range) or MI (Mutual Information) and using an acquisition function appropriate for each strategy.

[0107] When determining the optimal parameters as those that minimize the value of the evaluation function (for example, when using the abnormal time rate as information indicating the operating status), it is sufficient to search for parameters that minimize the acquisition function a(θ).

[0108]

number

[0109]

number

[0110] The candidate optimal values ​​for the parameters detected by the search unit 304 may be presented to the operator of the information processing device 3 by the display control unit 306 displaying them on a display device, etc. Then, the control device 4 starts control of the control target 5 to which the candidate is applied, and measurement data measured by the measuring device 52 during the period in which the control is performed is acquired and accumulated in the control device 4. The applied parameters are then associated with the accumulated measurement data, and the data is input to the information processing device 3 as new result data.

[0111] Then, the predictive distribution calculation unit 303 updates the predictive distribution based on the new result data, and the search unit 304 searches for optimal values ​​of the parameters based on an evaluation function configured based on the predictive distribution updated by the predictive distribution calculation unit 303. In this way, by repeating the updating of the predictive distribution and the search for optimal values, it becomes possible to detect optimal parameters.

[0112] [Display screen example (result display)] The display control unit 306 may present the search results to the operator by displaying a display screen such as that shown in Fig. 6. Fig. 6 is a diagram showing an example of a display screen showing the search results. More specifically, Fig. 6 is a diagram showing, on a parallel coordinate system, the history of candidates detected for each of elements a to d included in the parameters to be searched.

[0113] In this way, the parameters to be searched may include multiple elements, thereby allowing multiple elements to be optimized in parallel. For example, if the above-mentioned incineration power plant is the controlled object 5, the above elements a to d may be various control parameters related to automatic combustion control. Examples of control parameters related to automatic combustion control include parameters related to adjustment of the combustion rate, adjustment of the dry air supply flow rate, adjustment of the combustion air supply flow rate, adjustment of the combustion air temperature, and adjustment of the height of the waste layer to be incinerated.

[0114] The parameters to be searched may also include, for example, parameters indicating the content and amount of control of the control target 5 according to the result of inference by the inference model. This makes it possible to appropriately set the content and amount of control according to the result of inference by the inference model.

[0115] Furthermore, the parameters to be searched for may include inference parameters used for inference by the inference model, in addition to the above parameters indicating the content and amount of control of the control object 5. In this case, the search unit 304 searches for both candidates for the optimal value of the content and amount of control of the control object 5 and candidates for the optimal value of the inference parameters, based on the predictive distribution calculated by the predictive distribution calculation unit 303.

[0116] In addition, if the inference model is a neural network model, the inference parameters may be hyperparameters such as weight values ​​obtained by learning. Also, if the inference model ensembles the inference results of multiple machine learning models to output a final inference result, the inference parameters may be ensemble ratios (weights for each inference result).

[0117] The operating state of the control object 5 is affected by the results of inference by the inference model and intervention control based on those results. Therefore, if the optimal values ​​of the control content and control amount for the control object 5 and the optimal values ​​of the inference parameters are calculated separately, the parameter combination may not be optimal overall. In this regard, with the above configuration, both the optimal values ​​of the control content and control amount for the control object 5 and the optimal values ​​of the inference parameters are searched for, making it possible to detect the optimal parameter combination overall.

[0118] Furthermore, when an inference model is used to predict whether the future state of the controlled object 5 will be normal or abnormal, the output value of the inference model indicates the probability of normality and / or the probability of abnormality. Therefore, in order for the control content determination unit 43 in the control device 4 to determine the control content depending on whether the state is normal or abnormal, it is necessary to determine whether the state is normal or abnormal using a threshold value for the above probability.

[0119] Since the value at which the threshold value is set is important for determining appropriate control content according to the situation, the elements a to d may include a threshold value for determining whether the future state of the controlled object 5 is normal or abnormal. This makes it possible to appropriately determine whether it is normal or abnormal and to perform appropriate control according to the determination result. Note that the number of elements included in the parameters to be optimized is arbitrary, and the content of each element is not limited to the above example.

[0120] Returning now to the explanation of FIG. 6, the parallel coordinates shown in FIG. 6 have their vertical axis representing the value of each element included in the parameters. On the parallel coordinates, a graph is shown in which broken lines connect the elements included in the parameters detected in one search. In other words, the intersections of the broken lines and the vertical axis represent the values ​​of each element. Such a graph is called a parallel coordinate plot. The parallel coordinate plot shown in FIG. 6 shows the transition of the values ​​of the candidates detected in each search by the search unit 304, and can be called transition information.

[0121] In this way, the display control unit 306 may display transition information indicating the transition of the candidate values ​​detected in each search by the search unit 304. This allows the operator to easily recognize how the candidate optimal values ​​for the parameters have transitioned as the search is repeated, and to confirm whether the optimization is proceeding normally.

[0122] In particular, when the parameter to be optimized includes multiple elements, the display control unit 306 may display a parallel coordinate plot, i.e., a graph showing the value of each candidate detected in each search for each element, on parallel coordinates. This allows the operator to easily recognize how each candidate detected for each element included in the parameter has changed as the search is repeated, and to confirm whether the optimization is proceeding normally.

[0123] The parallel coordinate plot is merely one example of a display format for the transition information, and the display format for the transition information is arbitrary. For example, the display control unit 306 may display, as the transition information, a parallel set graph, a contour plot, a correlogram, or the like, which shows the transition of the values ​​of the candidates detected in each search by the search unit 304.

[0124] 6, the line segments constituting the line graph include solid lines, dashed lines, and dashed-dotted lines. The patterns of these line segments correspond to the operating state of the control target 5 after the control target 5 is controlled by applying the values ​​of the elements on the line segments. Specifically, the line segments connecting elements when the abnormal time rate is less than the first threshold are solid lines, the line segments connecting elements when the abnormal time rate is equal to or greater than the first threshold and less than the second threshold are dashed lines, and the line segments connecting elements when the abnormal time rate is equal to or greater than the second threshold are dashed lines.

[0125] In this way, the display control unit 306 may change the display mode of the value of each candidate included in the transition information depending on the operating state of the control target 5 after the control of the control target 5 is performed by applying the detected candidate. This allows the user to easily recognize whether each candidate is good or bad. Note that the display mode depending on the operating state may be any mode and is not limited to the example shown in the figure. For example, the display control unit 306 may display the value of each candidate in a color depending on the operating state. In this case, it is sufficient to determine in advance the correspondence between the evaluation value (e.g., abnormal time rate) that evaluates the operating state and the display color. This allows the display control unit 306 to display the value of each candidate in a display color depending on the evaluation value calculated by the evaluation value calculation unit 302.

[0126] [Flow of processing executed by information processing device 2] The flow of the process (index value calculation method) executed by the information processing device 2 will be described with reference to Fig. 7. Fig. 7 is a flowchart showing an example of the process executed by the information processing device 2.

[0127] In S11, the data acquisition unit 201 acquires a training dataset. The method for acquiring the training dataset is arbitrary. For example, the data acquisition unit 201 may acquire the training dataset input via the input unit 23, or may acquire the training dataset from the learning device 1 via the communication unit 22.

[0128] In S12, the average distance calculation unit 202 calculates, for each training data item included in the training data set acquired in S11, the average value of the distance between the training data item and a predetermined number of nearby training data items. Note that the predetermined number of nearby training data items are the predetermined number of other training data items that are closest in distance to the training data item.

[0129] In S13, the threshold determination unit 203 determines the average value of a predetermined rank among the multiple average values ​​calculated in S12 as the threshold for determining whether the input data is an outlier. Note that the processes of S11 to S13 only need to be performed by the time the process of S16 is performed at the latest, and do not necessarily have to be performed immediately before S14.

[0130] In S14, the data acquisition unit 201 acquires input data used for inference by the inference model. The method of acquiring the input data is arbitrary. For example, the data acquisition unit 201 may acquire input data input via the input unit 23. Furthermore, for example, the data acquisition unit 201 may acquire input data from the control device 4 via the communication unit 22, or may acquire measurement data from the measurement device 52 to generate input data.

[0131] The data acquisition unit 201 may acquire input data for a predetermined period of time in one lump sum, or may acquire the input data in real time, that is, at the time the input data is used or will be used for inference by the inference model.

[0132] In S15, the average distance calculation unit 202 calculates the average value of the distances between the input data acquired in S14 and a predetermined number of input data in the vicinity of the input data. If one input data is acquired in S14, the average distance calculation unit 202 calculates the average value of the distances between the input data and multiple input data acquired previously. On the other hand, if multiple input data are acquired in S14, the average distance calculation unit 202 performs the process of calculating the average value of the distances between the input data and a predetermined number of input data in the vicinity of the input data for each of the multiple acquired input data.

[0133] In S16 (outlier determination step), the outlier determination unit 204 determines whether the input data acquired in S14 is an outlier based on the average value calculated in S15. More specifically, the outlier determination unit 204 compares the average value calculated in S15 with the threshold value determined in S13, and determines whether the input data is an outlier based on the comparison result. Note that if multiple pieces of input data are acquired in S14, the outlier determination unit 204 determines for each piece of input data whether the input data is an outlier.

[0134] In S17 (index value calculation step), the index value calculation unit 205 calculates an index value indicating the suitability of the inference model for the input data based on the determination result of S 16. For example, the index value calculation unit 205 may calculate, as the index value, the number of input data determined to be outliers in S16 among the input data corresponding to the measurement data measured by the measurement device 52 during a predetermined period.

[0135] In S18, the relearning necessity determination unit 206 determines whether relearning is necessary based on the index value calculated in S17. For example, the relearning necessity determination unit 206 may determine that relearning is necessary (YES in S18) if the index value calculated in S17 is equal to or greater than a predetermined threshold, and may determine that relearning is not necessary (NO in S18) if the index value is less than the threshold. If NO is determined in S18, the processing in FIG. 7 ends. Note that when input data is acquired in real time, if NO is determined in S18, the processing may return to S14 and new input data may be acquired.

[0136] If the answer is YES in S18, the process proceeds to S19, where the re-learning necessity determination unit 206 sends the input data acquired in S14 to the learning device 1 to re-learn the inference model, and the processing in Figure 7 is then terminated.

[0137] The recall calculation unit 207 may calculate the recall of the inference performed using the input data acquired in S14. The learning data extraction unit 208 may then extract input data to be used for relearning the inference model from the multiple input data acquired in S14, based on the index value calculated in S17 and the recall calculated by the recall calculation unit 207. The learning data extraction unit 208 may then transmit the input data to be used for relearning to the learning device 1 when the determination in S18 is YES.

[0138] [Structure to encourage relearning] Furthermore, in S19, the relearning necessity determination unit 206 may perform processing to prompt the operator to relearn the inference model, instead of controlling the learning device 1 to perform relearning. In this case, the relearning necessity determination unit 206 functions as a notification unit that prompts the operator to relearn the inference model in accordance with the determination result of the outlier determination unit 204. Note that the notification method and the notification target are not particularly limited. For example, the notification unit 210 may notify by outputting information indicating that relearning is necessary. The output destination is not particularly limited. For example, the notification unit 210 may output the information to the output unit 24, the learning device 1, or another device such as a terminal device carried by the operator.

[0139] Furthermore, when processing to prompt relearning is performed, the index value calculation unit 205 may be omitted, and the relearning necessity determination unit 206 may determine whether relearning is necessary using the determination result of the outlier determination unit 204. For example, the relearning necessity determination unit 206 may count the number of input data determined to be outliers in S16, and determine that relearning is necessary if the counted number is equal to or greater than a predetermined threshold.

[0140] In this way, the information processing device 2 may be configured to include an outlier determination unit 204 that determines whether or not each of a plurality of input data input to the inference model is an outlier using a threshold for determining whether or not the training data included in the training dataset used to generate the inference model is an outlier, and a re-learning necessity determination unit (notification unit) 206 that prompts re-learning of the inference model in accordance with the determination result of the outlier determination unit 204. This makes it possible to appropriately perform processing that prompts re-learning, which is processing related to updating the inference model, even when it is difficult to verify the accuracy of the inference result.

[0141] [Configuration for performing relearning in information processing device 2] The information processing device 2 may also include a learning unit that performs re-learning of the inference model. In this case, the learning device 1 can be omitted from the control system 7. In this case, if it is determined in S18 of Fig. 7 that re-learning is necessary, re-learning is performed by the learning unit included in the information processing device 2 in S19.

[0142] In this case, too, as explained above in [Configuration for Promoting Re-learning], the index value calculation unit 205 may be omitted. When the index value calculation unit 205 is omitted, the timing of relearning is determined using the determination result of the outlier determination unit 204. For example, the relearning necessity determination unit 206 may determine whether relearning is necessary using the determination result of the outlier determination unit 204, in which case the time when it is determined that relearning is necessary becomes the timing of relearning. Also, the relearning necessity determination unit 206 may be omitted, and the learning unit may determine whether or not to perform relearning, in which case the learning unit will determine the timing of relearning based on the determination result of the outlier determination unit 204.

[0143] In this way, the information processing device 2 may be configured to include an outlier determination unit 204 that determines whether or not each of multiple input data input to the inference model is an outlier using a threshold value for determining whether or not the training data included in the training dataset used to generate the inference model is an outlier, and a learning unit that re-learns the inference model at a timing determined based on the determination result of the outlier determination unit 204.

[0144] According to the above configuration, whether or not input data input to the inference model is an outlier is determined using a threshold for determining whether or not training data is an outlier. In this process, input data that would be determined to be an outlier if it were included in the training data set, i.e., input data that deviates from the learned range, is determined to be an outlier.

[0145] When the number or proportion of input data that deviates from the learned range increases, it can be said that the suitability of the inference model for the input data is decreasing. At this time, it can be said that the need for re-learning or replacement of the inference model is increasing. Therefore, with the above configuration in which the inference model is re-learned at a timing determined based on the judgment result of the outlier judgment unit 204, it is possible to perform re-learning at an appropriate timing.

[0146] Furthermore, with the above configuration, it is not necessary to use the inference results from the inference model. Therefore, with the above configuration, even when it is difficult to verify the accuracy of the inference results, it is possible to appropriately perform re-learning, which is a process for updating the inference model.

[0147] [Flow of processing executed by information processing device 3] The flow of the process (search method) executed by the information processing device 3 will be described with reference to Fig. 8. Fig. 8 is a flowchart showing an example of the process executed by the information processing device 3. Note that the timing for executing the process of Fig. 8 is not particularly limited. For example, the process may be executed at predetermined intervals, or may be executed at a timing when it is determined that the operating status of the control target 5 has deteriorated or is showing a tendency to deteriorate based on information such as an abnormal time rate that indicates the operating status of the control target 5.

[0148] In S21, the data acquisition unit 301 acquires the initial values ​​of the parameters to be optimized and the result data corresponding to the initial values. For example, the data acquisition unit 301 may acquire the initial values ​​and the result data input via the input unit 33. The result data corresponding to the initial values ​​of the parameters indicates the operating state of the control target 5 during the period in which the control of the control target 5 was performed using the initial values, and may be, for example, a measurement value measured by the measuring device 52 during that period.

[0149] In addition, in S21, the data acquisition unit 301 may acquire, in addition to the initial value and the resulting data, an upper limit value of the parameter and the resulting data when the upper limit value is applied, and a lower limit value of the parameter and the resulting data when the lower limit value is applied.

[0150] The method for determining the initial values ​​of the parameters is not particularly limited. For example, the initial values ​​of the parameters may be determined by a descriptive statistical method or the like. In this case, for example, a frequency distribution table showing the distribution of parameter values ​​before and after the occurrence of an abnormality may be created, and the parameter values ​​that are considered to be optimal may be identified based on the table, and the identified values ​​may be used as the initial values. Once the initial values ​​are determined, the parameter values ​​may be set to the initial values, and the control target 5 may be controlled for a predetermined period, and result data indicating the operating state during that period may be obtained.

[0151] Furthermore, the method for determining the upper and lower limit values ​​of the parameters is not particularly limited. For example, they may be determined by a descriptive statistical method or the like, as with the initial value, or the initial value may be used as a reference and values ​​deviating from the initial value by a predetermined range may be set as the upper and lower limit values. In the latter case, for example, the predetermined range may be set to 10%, and the upper limit value may be set by adding 10% of the initial value to the initial value, and the lower limit value may be set by subtracting 10% of the initial value from the initial value. Result data for each of the upper and lower limit values ​​can be obtained in the same manner as for the initial value.

[0152] In S22, the evaluation value calculation unit 302 uses the result data acquired in S21 to calculate an evaluation value that evaluates the operating state of the control target 5 during the period in which the control target 5 was controlled by applying the initial value. As the evaluation value, for example, the abnormal time rate described above may be applied. Note that the definition of "abnormal" when calculating the abnormal time rate may be the same as that used when learning the inference model, or a broader definition may be applied.

[0153] In S23 (step of calculating the predictive distribution), the predictive distribution calculation unit 303 calculates the predictive distribution of a function indicating the relationship between the parameters to be optimized and the operating state of the controlled object 5, based on the initial values ​​acquired in S21 and the evaluation values ​​calculated in S22. Note that the display control unit 306 may present the calculated predictive distribution to the operator by, for example, outputting it to the output unit 34.

[0154] In S24 (a step of searching for candidates), the search unit 304 searches for candidates for the optimal values ​​of the parameters based on the predictive distribution calculated in S23. In addition, the display control unit 306 displays the detected candidates to the operator by, for example, outputting the candidates to the output unit 34.

[0155] Thereafter, the operator checks whether there is any problem in applying the presented candidate, and if it is determined that there is no problem, inputs the candidate into the control device 4 and starts control of the control target 5 applying the candidate. Then, the operator inputs measurement data etc. measured by the measuring device 52 during the period in which control applying the candidate was performed, together with the values ​​of the applied parameters (values ​​of the presented candidate) into the information processing device 3 as result data indicating the operating state of the control target 5. These processes may be performed automatically without the intervention of the operator.

[0156] If the operator determines that there is a problem with applying the candidate presented as a result of the search in S24, the processing from S25 onward is skipped. In this case, the search unit 304 may search for a candidate different from the candidate previously detected. Also, a normal range for the parameter may be determined in advance. In this case, regardless of the operator's judgment, if the candidate detected in S24 is outside the normal range, the processing from S25 onward is skipped.

[0157] In S25, the data acquisition unit 301 acquires the applied parameter values ​​and result data input as described above. Note that the data acquisition unit 301 may acquire the applied parameter values ​​and result data from the control device 4. Then, in S26, the evaluation value calculation unit 302 calculates an evaluation value using the result data acquired in S25.

[0158] In S27, the optimization control unit 305 determines whether or not to end the optimization based on the evaluation value calculated in S26. For example, the optimization control unit 305 may determine to end the optimization (YES in S27) if the evaluation value is equal to or greater than a predetermined threshold, and may determine to continue the optimization (NO in S27) if the evaluation value is less than the threshold. If the determination in S27 is YES, the processing in FIG. 8 ends, and if the determination in S27 is NO, the processing returns to S23. In S23, to which a transition is made from S27, the predictive distribution is updated using the parameter values ​​and result data acquired in S25.

[0159] [Modification] The execution entity of each process described in each of the above-described embodiments is arbitrary and is not limited to the above-described examples. In other words, the functions of the learning device 1, information processing devices 2 and 3, and control device 4 can be realized by multiple information processing devices (which can also be called processors) that can communicate with each other. For example, each process described in the flowcharts of Figures 7 and 8 can be shared among multiple information processing devices. In other words, the execution entity of the control method in each of the above-described embodiments may be one information processing device or multiple information processing devices.

[0160] Furthermore, as described above, the functions of learning device 1 may be provided in information processing device 2, and these may be integrated into a single device, or some or all of the functions of learning device 1, information processing device 2, and information processing device 3 may be provided in control device 4. In this way, it is possible to appropriately change the devices that make up control system 7.

[0161] [Software implementation example] The functions of the learning device 1, the information processing devices 2 and 3, and the control device 4 (hereinafter referred to as "devices") can be realized by a program (index value calculation program / search program) that causes a computer to function as the device, and that causes a computer to function as each control block of the device (particularly each part included in the control unit 20 and the control unit 30).

[0162] In this case, the device includes a computer having at least one control device (e.g., a processor) and at least one storage device (e.g., a memory) as hardware for executing the program. The control device and storage device execute the program, thereby realizing the functions described in each of the above embodiments.

[0163] The program may be non-transitory and may be recorded on one or more computer-readable recording media. The recording media may or may not be included in the device. In the latter case, the program may be supplied to the device via any wired or wireless transmission medium.

[0164] Furthermore, some or all of the functions of the control blocks can be realized by logic circuits. For example, an integrated circuit in which a logic circuit that functions as each of the control blocks is formed is also included in the scope of the present invention. In addition, the functions of the control blocks can also be realized by, for example, a quantum computer.

[0165] 〔summary〕 An information processing device according to aspect 1 of the present invention includes an outlier determination unit that determines whether each of a plurality of input data input to a machine learning model is an outlier using a threshold for determining whether training data included in a training dataset used to generate the machine learning model is an outlier, and an index value calculation unit that calculates an index value indicating the suitability of the machine learning model for the input data based on the determination result of the outlier determination unit.

[0166] The information processing device according to aspect 2 of the present invention is, in aspect 1, provided with a threshold determination unit that calculates, for each training data included in the training dataset, a deviation indicating the degree to which the training data deviates from other training data, and determines a deviation of a predetermined rank from the calculated deviations as the threshold, and the outlier determination unit calculates, for each input data included in the plurality of input data, a deviation indicating the degree to which the input data deviates from other input data, and determines whether each input data is an outlier by comparing the calculated deviation with the threshold.

[0167] In an information processing device according to aspect 3 of the present invention, in aspect 1 or 2, the machine learning model is a model that predicts whether a future state of a specified object will be normal or abnormal based on the input data related to the object, and the index value calculation unit calculates, as the index value, the number of input data items that were determined to be outliers and for which the state of the object was abnormal after prediction using the input data, or a value calculated using the number of input data items.

[0168] The information processing device according to aspect 4 of the present invention is, in any of aspects 1 to 3, a model in which the machine learning model predicts whether a future state of a predetermined object will be normal or abnormal based on input data related to the object, and includes a recall calculation unit that calculates a recall in the prediction by dividing the number of cases in which an abnormality was correctly determined in a prediction made by inputting a plurality of pieces of input data into the machine learning model by the sum of the number of cases in which an abnormality was correctly determined and the number of cases in which an abnormality was incorrectly determined as normal, and a learning data extraction unit that extracts data to be used for re-training the machine learning model from the plurality of pieces of input data based on the index value and the recall.

[0169] An information processing device according to aspect 5 of the present invention includes an outlier determination unit that determines whether each of a plurality of input data input to a machine learning model is an outlier using a threshold for determining whether training data included in a training dataset used to generate the machine learning model is an outlier, and an alarm unit that prompts re-learning of the machine learning model depending on the determination result of the outlier determination unit.

[0170] An information processing device according to aspect 6 of the present invention includes an outlier determination unit that determines whether each of a plurality of input data sets input to a machine learning model is an outlier using a threshold value for determining whether training data included in a training dataset used to generate the machine learning model is an outlier, and a learning data extraction unit that extracts data to be used for re-learning the machine learning model from among the plurality of input data sets based on the determination result of the outlier determination unit.

[0171] An information processing device according to aspect 7 of the present invention includes an outlier determination unit that determines whether each of a plurality of input data input to a machine learning model is an outlier using a threshold for determining whether training data included in a training dataset used to generate the machine learning model is an outlier, and a learning unit that re-trains the machine learning model at a timing determined based on the determination result of the outlier determination unit.

[0172] A control system according to aspect 8 of the present invention includes the information processing device according to aspect 1, a learning device that updates the machine learning model at a timing determined based on the index value, and a control device that determines control content for a specified control object based on output data obtained by inputting input data into the updated machine learning model, wherein the information processing device acquires the input data input into the updated machine learning model and calculates an index value indicating the suitability of the machine learning model for the input data.

[0173] An index value calculation method according to aspect 9 of the present invention is an index value calculation method executed by at least one information processing device, and includes an outlier determination step of determining whether each of a plurality of input data input to the machine learning model is an outlier using a threshold for determining whether training data included in a training dataset used to generate the machine learning model is an outlier, and an index value calculation step of calculating an index value indicating the suitability of the machine learning model for the input data based on the determination result in the outlier determination step.

[0174] An index value calculation program according to aspect 10 of the present invention is an index value calculation program for causing a computer to function as the information processing device described in aspect 1, and causes the computer to function as the outlier determination unit and the index value calculation unit.

[0175] The present invention is not limited to the above-described embodiments, and various modifications are possible within the scope of the claims. Embodiments obtained by appropriately combining the technical means disclosed in different embodiments are also included in the technical scope of the present invention. [Explanation of symbols]

[0176] 1 Learning device 2. Information processing equipment 203 Threshold determination unit 204 Outlier detection section 205 Index value calculation unit 207 Recall rate calculation part 208 Learning Data Extraction Unit 4. Control device 7. Control System

Claims

1. an outlier determination unit that determines whether each of a plurality of input data input to the machine learning model is an outlier, using a threshold value for determining whether training data included in a training dataset used to generate the machine learning model is an outlier; an index value calculation unit that calculates an index value indicating the suitability of the machine learning model for the input data based on the determination result of the outlier determination unit, a threshold value determination unit that calculates a deviation degree indicating the degree to which each training data included in the training data set deviates from other training data, and determines a deviation degree of a predetermined rank among the calculated deviation degrees as the threshold value; The outlier determination unit determines whether each input data is an outlier by comparing a deviation degree calculated for each input data included in the plurality of input data, which indicates the degree to which the input data deviates from other input data, with the threshold value.

2. An outlier determination unit that determines whether each of a plurality of input data input to a machine learning model is an outlier, using a threshold value for determining whether training data included in a training dataset used to generate the machine learning model is an outlier; an index value calculation unit that calculates an index value indicating the suitability of the machine learning model for the input data based on the determination result of the outlier determination unit, the machine learning model is a model that predicts whether a future state of a predetermined object will be normal or abnormal based on the input data regarding the object; The index value calculation unit calculates, as the index value, the number of input data items that were determined to be outliers and for which the target state was abnormal after prediction using the input data, or a value calculated using the number of input data items.

3. An outlier determination unit that determines whether or not each of a plurality of input data input to the machine learning model is an outlier, using a threshold value for determining whether or not training data included in a training dataset used to generate the machine learning model is an outlier; 、 an index value calculation unit that calculates an index value indicating the suitability of the machine learning model for the input data based on the determination result of the outlier determination unit, the machine learning model is a model that predicts whether a future state of a predetermined object will be normal or abnormal based on the input data regarding the object; a recall calculation unit that calculates a recall in the prediction by dividing the number of cases that were correctly determined to be abnormal in a prediction made by inputting a plurality of pieces of input data into the machine learning model by the sum of the number of cases that were correctly determined to be abnormal and the number of cases that were erroneously determined to be abnormal as normal; and a learning data extraction unit that extracts data to be used for re-learning the machine learning model from among the plurality of input data based on the index value and the recall rate.

4. an outlier determination unit that determines whether each of a plurality of input data input to the machine learning model is an outlier, using a threshold value for determining whether training data included in a training dataset used to generate the machine learning model is an outlier; a notification unit that prompts the machine learning model to re-learn in response to a determination result of the outlier determination unit; a threshold value determination unit that calculates a deviation degree indicating the degree to which each training data included in the training data set deviates from other training data, and determines a deviation degree of a predetermined rank among the calculated deviation degrees as the threshold value; The outlier determination unit determines whether each input data is an outlier by comparing a deviation degree calculated for each input data included in the plurality of input data, which indicates the degree to which the input data deviates from other input data, with the threshold value.

5. an outlier determination unit that determines whether each of a plurality of input data input to the machine learning model is an outlier, using a threshold value for determining whether training data included in a training dataset used to generate the machine learning model is an outlier; and a learning data extraction unit that extracts data to be used for re-learning the machine learning model from the plurality of input data based on the determination result of the outlier determination unit.

6. an outlier determination unit that determines whether each of a plurality of input data input to the machine learning model is an outlier, using a threshold value for determining whether training data included in a training dataset used to generate the machine learning model is an outlier; a learning unit that re-learns the machine learning model at a timing determined based on the determination result of the outlier determination unit; a threshold value determination unit that calculates a deviation degree indicating the degree to which each training data included in the training data set deviates from other training data, and determines a deviation degree of a predetermined rank among the calculated deviation degrees as the threshold value; The outlier determination unit determines whether each input data is an outlier by comparing a deviation degree calculated for each input data included in the plurality of input data, which indicates the degree to which the input data deviates from other input data, with the threshold value.

7. The information processing device according to claim 1 ; a learning device that updates the machine learning model at a timing determined based on the index value; a control device that determines control content for a predetermined control target based on output data obtained by inputting input data into the updated machine learning model, The information processing device acquires the input data input to the updated machine learning model, and calculates an index value indicating the suitability of the machine learning model for the input data. Your system.

8. An index value calculation method executed by at least one information processing device, an outlier determination step of determining, for each of a plurality of input data input to the machine learning model, whether or not the input data is an outlier, using a threshold for determining whether or not training data included in a training dataset used to generate the machine learning model is an outlier; an index value calculation step of calculating an index value indicating the suitability of the machine learning model for the input data based on the determination result in the outlier determination step; a threshold determination step of calculating a deviation indicating the degree to which each training data included in the training data set deviates from other training data, and determining a deviation of a predetermined rank among the calculated deviations as the threshold, In the outlier determination step, a deviation degree calculated for each input data included in the plurality of input data, which indicates the degree to which the input data deviates from other input data, is compared with the threshold value to determine whether each input data is an outlier.

9. 2. An index value calculation program for causing a computer to function as the information processing device according to claim 1, the index value calculation program causing a computer to function as the outlier determination unit, the index value calculation unit, and the threshold value determination unit.

Images