Information processing device, information processing method, and program

Abstract

The information processing device includes: an acquisition unit that acquires observation data including an explanatory variable and a target variable at a predetermined time; a summary unit that generates, for each piece of summary information of the observation data, a predetermined number or less of summary data indicating the observation data; and a learning unit that generates a learning completion model on the basis of learning data including the observation data acquired in the learning target section and the summary data.

Description

Technical Field

[0001] This disclosure relates to an information processing apparatus, information processing method, and program. Background Technology

[0002] Techniques for using fully learned models for prediction already exist. To prevent the fully learned models from deteriorating over time, techniques have been proposed to periodically update them using the most recent data.

[0003] For example, Patent Document 1 describes a predictive model relearning device that, when it detects performance degradation of the predictive model caused by changes in data characteristics over time, uses the relearning data used in the relearning of the predictive model to perform relearning of the predictive model.

[0004] <Prior art documents>

[0005] <Non-Patent Literature>

[0006] Patent Document 1: (Japanese) Patent No. 7176656 Summary of the Invention

[0007] <Problem to be solved by this invention>

[0008] However, existing technologies suffer from a problem where prediction accuracy decreases as the numerical variance of the predicted object changes over time. For example, a learned model that was trained when the predicted object was in a stable state may become a constant expression or exhibit parameter divergence. In such cases, when the predicted object transitions to a changing state, significant prediction errors may occur.

[0009] In view of the problems mentioned above, one aspect of this disclosure aims to improve the prediction accuracy of a learned model.

[0010] <Methods for solving problems>

[0011] This disclosure includes the structure shown below.

[0012] [1] An information processing device, comprising:

[0013] The data acquisition department is responsible for acquiring observational data, including explanatory and target variables, within a specified timeframe.

[0014] The summarization unit generates a specified number of summary data representing the observation data according to each summary information of the observation data; and

[0015] The learning department generates a fully learned model based on the learning data, which includes the observation data and the summary data obtained in the learning target range.

[0016] [2] The information processing device according to [1] above further includes:

[0017] The prediction unit predicts the target variable by inputting the explanatory variables obtained in the prediction target interval into the learned model.

[0018] [3] According to the information processing device described in [2] above, wherein,

[0019] The prediction target interval is a time interval that includes the current time.

[0020] The learning target interval is the time interval immediately preceding the prediction target interval.

[0021] [4] According to the information processing device described in [3] above, wherein,

[0022] The summarizing unit generates the summarizing data, which includes the observation data obtained before the learning target interval.

[0023] [5] According to the information processing device described above [4], wherein,

[0024] The summarizing unit generates summary data representing the observation data re-acquired according to a predetermined number of each summary information.

[0025] [6] The information processing apparatus according to any one of [1] to [5] above, wherein,

[0026] The summarizing unit generates the summary information by rounding the observed data.

[0027] [7] The information processing apparatus according to any one of [1] to [6] above, wherein,

[0028] In the observed data, the numerical variance of the explanatory variable or the target variable changes over time.

[0029] [8] The information processing apparatus according to any one of [1] to [7] above, wherein,

[0030] The explanatory variables include the water level of the dam.

[0031] The target variable includes the power generation of the hydroelectric power station using the dam.

[0032] [9] The information processing apparatus according to any one of [1] to [7] above, wherein,

[0033] The explanatory variables include parameters representing the working conditions of the factory.

[0034] The target variable includes parameters representing the state of the factory.

[0035]

[10] An information processing method, executed by a computer:

[0036] The steps to obtain observational data containing explanatory and target variables within a specified time;

[0037] The step of generating a predetermined number of summary data representing the observation data according to each summary information of the observation data; and

[0038] The step of generating a fully learned model based on learning data that includes the observation data and the summary data obtained in the learning target range.

[0039]

[11] A program for causing a computer to execute:

[0040] The steps to obtain observational data containing explanatory and target variables within a specified time;

[0041] The step of generating a predetermined number of summary data representing the observation data according to each summary information of the observation data; and

[0042] The step of generating a fully learned model based on learning data that includes the observation data and the summary data obtained in the learning target range.

[0043] <The Effects of the Invention>

[0044] According to one aspect of the present invention, the prediction accuracy of a learned model can be improved. Attached Figure Description

[0045] Figure 1 This is a graph representing an example of observed data.

[0046] Figure 2 This is a diagram illustrating an example of the relationship between observed data and a predictive model.

[0047] Figure 3 This is a block diagram representing an example of the overall structure of a prediction system.

[0048] Figure 4 This is a block diagram illustrating an example of a computer's hardware structure.

[0049] Figure 5 This is a block diagram illustrating an example of the functional structure of an information processing device.

[0050] Figure 6 This is a graph representing an example of observed data.

[0051] Figure 7 This is a diagram illustrating an example of how summary data is updated.

[0052] Figure 8 This is a flowchart representing an example of learning processing.

[0053] Figure 9 This is a diagram used to illustrate an example of learning processing.

[0054] Figure 10 This is another diagram used to illustrate learning processing.

[0055] Figure 11 This is a flowchart illustrating an example of predictive processing.

[0056] Figure 12 This is a graph representing an example of prediction accuracy. Detailed Implementation

[0057] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. Furthermore, in this specification and the drawings, structural elements having substantially the same functional structure are labeled with the same reference numerals to omit redundant descriptions.

[0058] [Implementation Method]

[0059] One embodiment of this disclosure is a prediction system for predicting a given target variable based on a learned prediction model. The prediction system in this embodiment has the function of periodically updating the prediction model using learning data that has been updated through appropriate selection and filtering, thereby preventing the prediction model from deteriorating over time.

[0060] In this embodiment, the prediction system uses time-series observation data observed at specified times. The observation data is conceived as having a variable numerical variance over time. Specifically, the observation data includes periods with small numerical variance (hereinafter also referred to as "stable periods") and periods with large numerical variance (hereinafter also referred to as "variable periods"). The timing of transitions from stable periods to variable periods or from variable periods to stable periods in the observation data can be periodic or irregular.

[0061] Figure 1 This is a graph representing an example of observed data. In Figure 1 In this example, the numerical variance of the observed data changes over time, representing the shift in the water level of the dam and the shift in the amount of electricity generated by the hydroelectric power plant utilizing the dam.

[0062] like Figure 1As shown, water level x and power generation y are data that alternate between periods of large fluctuations and periods of stable operation with small variances. The relationship between water level x and power generation y, with water level x as the explanatory variable and power generation y as the target variable, can be modeled as y = ax + b. The model parameters a and b vary with external factors. External factors could include, for example, changes in flow velocity due to weather or the accumulation of dirt in the waterway.

[0063] In addition, Figure 1 In the example shown, the lengths of the variable period and the stable period are different, and the time variation of the numerical variance is not periodic. However, over long spans, the time variation of the numerical variance can be periodic. For example, the time variation of the numerical variance may not be periodic when measured in months, but it may be periodic when measured in years.

[0064] When using a learned model for prediction, high-precision predictions are possible if the variance of the learning data is greater than that of the observed data. Conversely, prediction accuracy decreases if the variance of the learning data is less than that of the observed data. When using recent observations to periodically update the prediction model, if the observations transition from a stable period to a variable period, the prediction model can be learned using the stable period's observations, while predictions are made using the variable period's observations; therefore, the prediction error will increase.

[0065] Figure 2 This is a graph illustrating an example of the relationship between observed data and a predictive model. Figure 2 In the diagram, dotted lines represent the observed shifts in the target variable. Solid lines represent the observed data used as training data, and dashed lines represent the predicted values ​​of the target variable.

[0066] like Figure 2 As shown, when the numerical variance of the observed data changes over time, the parameters of the prediction model vary depending on the time period used as the training data. Figure 2 In the example shown, model A, learned using observational data from the variable period, has a value of a=2.09 and b=50.45, while model B, learned using observational data from the stable period, has a value of a=3.88 and b=49.51. The value of parameter a varies considerably. Figure 2 The figure indicates that when a model B, which was learned from observations during a stable period, is used to make predictions using observations during a variable period, a large prediction error will occur.

[0067] In this embodiment, the goal is to improve the prediction accuracy of the fully trained model. Specifically, in this embodiment, the fully trained model is updated with the goal of ensuring that prediction accuracy is not reduced even with observational data whose numerical variance changes over time. Furthermore, according to this embodiment, it is possible to accurately predict the target variable using observational data whose numerical variance changes over time.

[0068] <Overall Structure>

[0069] Regarding the overall structure of the prediction system in this embodiment, refer to... Figure 3 Please provide an explanation. Figure 3 This is a block diagram representing an example of the overall structure of a prediction system.

[0070] like Figure 3 As shown, the prediction system 1000 includes an information processing device 10 and a terminal device 20. The information processing device 10 and the terminal device 20 are connected via a communication network N such as a LAN (Local Area Network) or the Internet in a manner that enables data communication.

[0071] Information processing device 10 is an information processing device such as a personal computer, workstation, or server that predicts a specified target variable. Information processing device 10 acquires observation data, including explanatory variables and target variables, within a specified time period, and generates a completed learning prediction model based on the observation data observed during a specified learning period. Information processing device 10 predicts the target variable by inputting the explanatory variables observed during the specified prediction period into the completed learning prediction model, and sends the prediction results to terminal device 20.

[0072] The prediction period can be a time interval including the current time. The learning period can be a time interval immediately preceding the prediction period. A time interval refers to the range of the time axis divided by any point in time. An example of a time interval is the update interval of the prediction model. The update interval of the prediction model can be arbitrarily set according to the characteristics of the prediction object. The update interval of the prediction model can be constant or non-constant, and can be the same as or different from the aforementioned time intervals. The following explanation assumes that the time interval is a predefined time period and that observation data is obtained at specified time intervals.

[0073] Terminal device 20 is an information processing terminal such as a personal computer, smartphone, or tablet terminal operated by a user of the prediction system 1000. Terminal device 20 displays prediction results received from information processing device 10 to the user. The prediction results include predicted values ​​of the target variable for a specified prediction period. Terminal device 20 can also request prediction results from information processing device 10 based on user actions.

[0074] in addition, Figure 3 The overall structure of the prediction system 1000 shown is only one example, and various other system structures can be adopted depending on the application or purpose. For example, more than one of the information processing device 10 and the terminal device 20 may be included in the prediction system 1000. For example, the information processing device 10 can be implemented by multiple computers or as a cloud computing service. Figure 3 The information processing device 10 and terminal device 20 shown are examples of such devices.

[0075] <Hardware Structure>

[0076] Regarding the hardware structure of the prediction system 1000 in this embodiment, refer to... Figure 4 Please provide an explanation.

[0077] <<Computer>>

[0078] The information processing device 10 and terminal device 20 of this embodiment are implemented, for example, by a computer. Figure 4 This is a block diagram illustrating an example of the hardware structure of the computer 500 according to this embodiment.

[0079] like Figure 4 As shown, computer 500 includes a CPU (Central Processing Unit) 501, ROM (Read Only Memory) 502, RAM (Random Access Memory) 503, HDD (Hard Disk Drive) 504, input device 505, display device 506, communication I / F (Interface) 507, and external I / F 508. CPU 501, ROM 502, and RAM 503 constitute what is called a computer. The various hardware components of computer 500 are interconnected via bus 509. Here, input device 505 and display device 506 can be connected to external I / F 508.

[0080] CPU 501 is a computing device that reads programs or data from storage devices such as ROM 502 or HDD 504 into RAM 503 and performs processing, thereby realizing the overall control and functions of computer 500.

[0081] ROM 502 is an example of a non-volatile semiconductor memory (storage device) that can retain programs and data even when the power is turned off. ROM 502 functions as the main storage device, storing various programs and data required by the CPU 501 to execute various programs installed in HDD 504. Specifically, ROM 502 stores data such as boot programs for BIOS (Basic Input / Output System), EFI (Extensible Firmware Interface), OS (Operating System) settings, and network settings that are executed when the computer 500 starts up.

[0082] RAM503 is an example of a volatile semiconductor memory (storage device) whose program and data are erased when power is cut off. RAM503 is, for example, DRAM (Dynamic Random Access Memory), SRAM (Static Random Access Memory), etc. RAM503 provides a working area when various programs installed in HDD504 are executed by CPU501.

[0083] HDD504 is an example of a non-volatile storage device that stores programs and data. The programs and data stored in HDD504 include the OS, the basic software used to control the computer 500 as a whole, and application programs that provide various functions on the OS. Alternatively, computer 500 can use a storage device that uses flash memory as the storage medium (e.g., SSD: Solid State Drive, etc.) instead of HDD504.

[0084] Input device 505 is a touch panel, operation key or button, keyboard or mouse, microphone, etc., for users to input various signals, such as sound data.

[0085] The display device 506 consists of a display such as a liquid crystal or organic EL (Electro-Luminescence) display showing images, and a speaker that outputs sound data such as sound.

[0086] The Communication I / F507 is an interface for connecting to a communication network and for data communication between the computer 500 and the network.

[0087] External I / F508 is the interface for external devices. External devices include drive units 510, etc.

[0088] The drive unit 510 is a device for setting the recording medium 511. Here, the recording medium 511 includes media that record information optically, electrically, or magnetically, such as CD-ROMs, floppy disks, and magneto-optical disks. Alternatively, the recording medium 511 may also include semiconductor memories that record information electrically, such as ROMs and flash memory. Therefore, the computer 500 can read from and / or write to the recording medium 511 via an external I / F 508.

[0089] Additionally, various programs installed in the HDD504 can be installed, for example, by placing the distributed recording medium 511 in the drive unit 510 connected to the external I / F 508, and having the drive unit 510 read the various programs recorded in the recording medium 511. Alternatively, various programs installed in the HDD504 can also be downloaded from the communication network N or a network other than the communication network N via the communication I / F 507.

[0090] <Functional Structure>

[0091] Regarding the functional structure of the prediction system in this embodiment, refer to... Figure 5 Please provide an explanation. Figure 5 This is a block diagram illustrating an example of the functional structure of an information processing device.

[0092] like Figure 5 As shown, the information processing device 10 includes an acquisition unit 101, an observation data storage unit 102, a summarization unit 103, a summary data storage unit 104, a generation unit 105, a learning unit 106, a model storage unit 107, a prediction unit 108, and a result output unit 109.

[0093] The acquisition unit 101, the summarization unit 103, the generation unit 105, the learning unit 106, the prediction unit 108, and the result output unit 109 are executed by the CPU 501 from... Figure 4 The process of expanding the HDD504 onto RAM503 is shown in the diagram.

[0094] The observation data storage unit 102, the summary data storage unit 104, and the model storage unit 107 are connected via... Figure 4 The HDD504 implementation shown.

[0095] The acquisition unit 101 acquires observational data, including explanatory and target variables, at prescribed time intervals. The time interval for acquiring the observational data may vary depending on the object being predicted. For example, in acquiring the water level of a dam and the power generation of a hydroelectric power plant using that dam, observational data may be acquired at 24-hour intervals. The observational data may include multiple data points observed at different times. Hereinafter, each data point included in the observational data will also be referred to as a "record".

[0096] Figure 6 This is a graph representing an example of observed data. For example... Figure 6 As shown, the observation data includes time t, water level x, and power generation y as data items. Each row containing time t represents one record.

[0097] Time *t* represents the time at which each observation data was acquired. Time *t* can be the elapsed time since a specified reference time. Time *t* can be the date and time in a specified time zone. Water level *x* represents the water level of a specified dam. Water level *x* is an example of an explanatory variable. Power generation *y* represents the power generation of a hydroelectric power station using a specified dam. Power generation *y* is an example of a target variable. It can contain multiple explanatory variables or target variables.

[0098] The observation data storage unit 102 stores the observation data acquired by the acquisition unit 101. The observation data storage unit 102 can store the observation data in a time series. The observation data storage unit 102 can also store only the observation data observed during the learning period, or it can store the observation data observed during the learning period and the observation data observed in the period immediately preceding the learning period, and it can store all the observation data within the storage capacity of the HDD504 or the like.

[0099] The summarization unit 103 generates summary data based on the observation data obtained from the observation data storage unit 102. More specifically, the summarization unit 103 generates summary data based on observation data observed during the learning period or observation data observed in a period immediately preceding the learning period. Furthermore, if summary data has already been created, existing summary data stored in the summary data storage unit 104 is read and updated. Summary data represents one or more observation data points, with each summary information representing a summary of the observation data. Specifically, summary data is data that associates summary information with one or more observation data points that become that summary information.

[0100] Each summary piece of observation included in the summary data can be the most recently acquired (in other words, the latest observation at time t). Each summary piece of observation included in the summary data can also be a specified number of observations starting from the beginning of the new data (in other words, starting from the beginning of the descending order of time t).

[0101] Summary information can be information summarizing part or all of the explanatory variables, or information summarizing both the explanatory and target variables. Summarizing, for example, can involve rounding. Specifically, summarizing can involve discarding a specified number of decimal places (e.g., the second decimal place), or rounding up or down.

[0102] Summarizing data, as an example, can be achieved using associative arrays in computer programming. When summarizing data using associative arrays, the summary information can be used as the key, and the observed data as the value.

[0103] Figure 7 This is a diagram illustrating an example of updating summary data. Figure 7 This indicates how the summary data is updated starting from a state where the summary data is 0, based on the observation data obtained at each time t. Figure 7 In the update of the summary data shown, the prescribed number of elements to be retained for each summary information is 1.

[0104] like Figure 7 As shown, suppose the observation data x=0.03 and y=0.09 were obtained at time t=1. The data is rounded down to the second decimal place, resulting in the summary information (0.0, 0.0). Since there is no element with the key (0.0, 0.0) in the summary data, an element with the key (0.0, 0.0) and values ​​x=0.03 and y=0.09 is added.

[0105] Next, assume that observation data x=0.24 and y=0.11 were obtained at time t=2. The data is rounded down to the second decimal place, resulting in the summary information (0.2, 0.1). Since there is no element with the key (0.2, 0.1) in the summary data, an element with the key (0.2, 0.1) and values ​​x=0.24 and y=0.11 is added.

[0106] Next, assume that observation data x=0.24 and y=0.21 were obtained at time t=3. Rounding to the second decimal place, the summarized information becomes (0.2, 0.2). In the summarized data, there is no element with the key (0.2, 0.2). Therefore, an element with the key (0.2, 0.2) and values ​​x=0.24 and y=0.21 is added.

[0107] Then, assume that the observation data x=0.09 and y=0.07 were obtained at time t=4. The values ​​are rounded down to the second decimal place, resulting in a summary of (0.0, 0.0). Since there is only one element with the key (0.0, 0.0) in the summary data, the value of the element with the key (0.0, 0.0) is rewritten as x=0.09 and y=0.07.

[0108] The amount of summary information included in the summary data can be limited or unlimited. For example, in the case where the summary data contains a predetermined upper limit of summary information, the summary unit 103 can delete the summary information representing the oldest observation data when adding new summary information. This allows the amount of summary information included in the summary data to be limited to below the upper limit. Furthermore, for example, the summary unit 103 can periodically delete summary information representing observation data older than a predetermined time. This allows the summary data to be limited to not containing observation data older than the predetermined time.

[0109] The summary data storage unit 104 stores the summary data generated by the summary unit 103.

[0110] The generation unit 105 generates learning data for learning a prediction model based on observation data read from the observation data storage unit 102 and summary data read from the summary data storage unit 104. Alternatively, the generation unit 105 may generate learning data by associating observation data observed during the learning process with the observation data shown in the summary data.

[0111] The learning unit 106 generates a learned prediction model based on the learning data generated by the generation unit 105. The prediction model is a machine learning model that takes specified explanatory variables as input and outputs predicted values ​​for specified target variables. For example, the prediction model can be a regression model. However, the prediction model is not limited to regression models and can be any type of machine learning model.

[0112] The model storage unit 107 stores the learned prediction model generated by the learning unit 106. When an existing prediction model is stored, the model storage unit 107 can overwrite it with a new prediction model, or it can store a new prediction model based on a backup of the existing prediction model.

[0113] The prediction unit 108 predicts the specified target variable based on the prediction model read from the model storage unit 107. The prediction unit 108 can predict the target variable during the prediction period by inputting the explanatory variables observed during the prediction period into the prediction model.

[0114] The result output unit 109 outputs the prediction results from the prediction unit 108. The prediction results include the predicted values ​​of the target variable during the prediction period. The result output unit 109 can send a result screen containing the prediction results to the terminal device 20. The result output unit 109 can also output the prediction results to the display device 506.

[0115] <Processing Steps>

[0116] Regarding the information processing method performed by the prediction system 1000 in this embodiment, please refer to... Figure 8 and Figure 9This will be explained. Information processing methods include learning processing and prediction processing. Learning processing is the process of generating a learned prediction model. Prediction processing is the process of using the learned prediction model to predict the target variable.

[0117] Learning Processing

[0118] Figure 8 This is a flowchart illustrating an example of a learning process. The learning process is executed repeatedly at specified time intervals.

[0119] In step S1, the acquisition unit 101 of the information processing device 10 acquires observation data including explanatory variables and target variables. Then, the acquisition unit 101 stores the acquired observation data in the observation data storage unit 102.

[0120] In step S2, the summarization unit 103 of the information processing device 10 reads observation data observed during the learning period or in a period immediately preceding the learning period from the observation data storage unit 102. Next, the summarization unit 103 generates summary information based on the read observation data.

[0121] In step S3, the summarization unit 103 of the information processing device 10 reads summary data from the summary data storage unit 104. Here, it is assumed that the summary data storage unit 104 stores summary data containing one or more elements. Then, the summarization unit 103 updates the summary data based on the summary information generated in step S2. If the summary data does not contain a predetermined number of elements with the summary information generated in step S2 as the key, the summarization unit 103 appends elements with the summary information as the key and observation data as the value to the summary data. On the other hand, if the summary data contains a predetermined number of elements with the summary information generated in step S2 as the key, the summarization unit 103 uses observation data to overwrite the values ​​of the elements with the summary information as the key. If the value of an element is set to the most recently acquired (in other words, the latest time t) observation data, the overwriting is performed by replacing existing observation data. If the value of an element is set to less than a predetermined number of observation data starting from the new data (in other words, starting from the beginning of the descending order of time t), the overwriting is performed by deleting the oldest data in the existing observation data and appending new observation data. The summary unit 103 stores the updated summary data in the summary data storage unit 104.

[0122] In step S4, the generation unit 105 of the information processing apparatus 10 reads the observation data observed during the learning period from the observation data storage unit 102. Additionally, the generation unit 105 reads summary data from the summary data storage unit 104. Then, the generation unit 105 obtains the observation data shown in the summary data. Next, the generation unit 105 generates learning data, which includes the observation data observed during the learning period and the observation data shown in the summary data. Then, the generation unit 105 sends the generated learning data to the learning unit 106.

[0123] In step S5, the learning unit 106 of the information processing device 10 receives learning data from the generation unit 105. Then, the learning unit 106 generates a learned prediction model based on the received learning data. Finally, the learning unit 106 stores the generated prediction model in the model storage unit 107.

[0124] Figure 9 This is a diagram used to illustrate an example of learning processing. Here, at time t... i This updates the prediction model. In this case, the learning object is from time t. i-1 By time t i The time period. Time t i-1 From time t i Tracing back to a specified time period. In this example, the associative array M, as an example of summary data, is based on time t. i-2 The observation data D obtained so far i-2 D i-3 ... has been updated.

[0125] At time t i The latest observation data obtained from Department 101 D i Next, the acquisition unit 101 will obtain the most recent observation data D. i It is stored in the observation data storage unit 102.

[0126] The summarization unit 103 obtains the associative array M from the summarization data storage unit 104. Next, the summarization unit 103 obtains the observation data D from the observation data storage unit 102. i-1 And based on the obtained observation data D i-1 Update the associative array M. Specifically, summary section 103 updates the observed data D. i-1 Each record d contained j Repeat the following process. Here, j is the record index, which is an integer greater than or equal to 1.

[0127] First, the summary unit 103 analyzes record d. j Perform a summary and obtain the summary information h. j Next, the summary unit 103 will use the key as the summary information h.j The value is record d. j Elements are appended to or overwritten in the associative array M. When the specified number of elements to be maintained is set to the most recently acquired (in other words, time t is the latest) element, this is achieved by using record d. j Overwrite by replacing existing records. When the specified number of elements to be retained is multiple, the oldest record d that was deleted from the existing records is appended starting from the new record (in other words, starting from the beginning of the descending order at time t). j To cover it.

[0128] Generation unit 105 collects the records shown by each element of the associative array M and generates data D. M The generation unit 105 connects to the most recent observation data D. i and data D M Therefore, learning data D is generated.

[0129] Learning unit 106 uses learning data D to fit a prediction model. This generates a prediction model learned from the most recent observation data D. i The observed data is generated after discarding the most recent data. The learning unit 106 updates the existing prediction model with the generated prediction model.

[0130] Figure 10 This is a diagram used to illustrate another example of learning processing. In this example, the associative array M, as an example of summary data, is based on time t. i-1 The observation data D obtained so far i-1 D i-2 ... has been updated.

[0131] At time t i The latest observation data obtained from Department 101 D i Next, the acquisition unit 101 will obtain the most recent observation data D. i It is stored in the observation data storage unit 102.

[0132] The summarization unit 103 retrieves the associative array M from the summarization data storage unit 104. Next, the summarization unit 103 retrieves the most recent observation data D from the observation data storage unit 102. i And based on the obtained observation data D i Update the correlation array M. Specifically, the summary section 103 updates the observed data D. i Each record d contained j Repeat the following process.

[0133] First, the summary unit 103 analyzes record d. j To obtain summary information h by performing a summary jNext, the summary unit 103 will use the key h as the summary information. j Record d with values j The elements are appended to or overwritten in the associative array M. If the specified number of elements maintained is the most recently acquired (in other words, the latest time t) 1, then by using record d... j The overwrite is performed by replacing existing records. When the specified number of elements to be retained is multiple, the oldest existing record, d, is appended and deleted starting from the new record (in other words, from the beginning of the descending order at time t). j To perform the coverage.

[0134] Generation unit 105 collects the records shown by each element of the associative array M and generates data D. M The learning unit 106 uses data DM as training data to fit the prediction model. This generates a prediction model based on selected observations after discarding the most recent observation data Di. The learning unit 106 then uses the generated prediction model to update the existing prediction model.

[0135] <<Predictive Processing>>

[0136] Figure 11 This is a flowchart illustrating an example of predictive processing. Predictive processing can be executed repeatedly at specified time intervals, or it can be executed at any time according to the user's requirements.

[0137] In step S11, the acquisition unit 101 of the information processing device 10 acquires the explanatory variables during the prediction period. Then, the acquisition unit 101 sends the acquired explanatory variables to the prediction unit 108.

[0138] In step S12, the prediction unit 108 of the information processing device 10 receives explanatory variables from the acquisition unit 101. Next, the prediction unit 108 reads the learned prediction model from the model storage unit 107. Then, the prediction unit 108 inputs the received explanatory variables into the read prediction model.

[0139] The prediction model predicts the target variable based on the input explanatory variables and outputs its predicted value. The prediction unit 108 obtains the predicted value of the target variable output by the prediction model. Then, the prediction unit 108 sends the obtained predicted value of the target variable to the result output unit 109.

[0140] In step S13, the result output unit 109 of the information processing device 10 receives the predicted value of the target variable from the prediction unit 108. Then, the result output unit 109 sends the prediction result to the terminal device 20. The prediction result contains the predicted value of the target variable. The prediction result may also contain other information related to the output of the prediction model. Other information, for example, may include statistical information about the predicted value or the reliability of the predicted value.

[0141] Terminal device 20 receives the prediction results from information processing device 10. Then, terminal device 20 displays the received prediction results to the user. Terminal device 20 may also display a result screen with the prediction results embedded on display device 506.

[0142] <Prediction Accuracy>

[0143] Figure 12 This is a graph representing an example of prediction accuracy. Figure 12 (A) is a graph showing the relationship between the target variable and the predicted value when no model update is performed. Figure 12 (B) is a graph showing the relationship between the target variable and the predicted value when updating the model using only the most recent data. Figure 12 (C) is a graph showing the relationship between the target variable and the predicted value when the model is updated using the most recent and summary data. Figure 12 (C) is equivalent to the prediction result of the prediction system 1000 in this embodiment.

[0144] Figure 12 In this context, the root mean squared error (RMSE) is used as an example to represent prediction accuracy. A smaller RMSE value indicates higher prediction accuracy, while a larger RMSE value indicates lower prediction accuracy.

[0145] like Figure 12 As shown, the mean squared error is 0.61 without model updates. However, when only the most recent data is used for model updates, the mean squared error increases to 0.22. Figure 2 As shown, when using a model learned from observations during a stable period to make predictions using observations during a variable period, prediction errors were identified. Conversely, when the model was updated using the most recent and summary data, the prediction errors were not identified, and the mean squared error increased to 0.18. Figure 12 This indicates that the prediction system 1000 according to this embodiment can improve prediction accuracy.

[0146] <Effects of the Example>

[0147] In the information processing apparatus 10 of this embodiment, summary data representing the observation data of a predetermined number or less is generated according to each summary information of the observation data, and a fully learned model is generated based on the learning data and the summary data included in the learning target interval. Since the fully learned model is a novel learning method using learning data with increased numerical variance, even for observation data whose numerical variance changes over time, the decrease in prediction accuracy can be suppressed. In one aspect, according to this embodiment, the prediction accuracy of the fully learned model can be improved.

[0148] The information processing device 10 can predict the target variable by inputting the explanatory variables obtained from the prediction target interval into the learning model. According to this embodiment, the target variable can be predicted with high accuracy.

[0149] The learning target interval can be a time period that includes the current time. The learning target interval can also be a time period immediately preceding the prediction target interval. According to this embodiment, it is possible to predict the target variable within a time period including the current time with high accuracy based on the most recent observation data.

[0150] The information processing device 10 can generate summary data containing observation data acquired up to the learning target interval. The information processing device 10 can generate summary data representing up to a predetermined number of newly acquired observation data for each summary piece of information. According to this embodiment, the numerical variance of the learning data can be improved with fewer observation data.

[0151] The information processing device 10 can generate summary information by rounding the observation data. According to this embodiment, observation data can be summarized through simple calculations.

[0152] In the observational data, the numerical variance of the explanatory or target variable can change over time. According to this implementation method, even observational data with numerical variance changing over time can improve the prediction accuracy of the learned model.

[0153] The explanatory variables may include the water level of the dam, and the target variable may include the power generation of the hydroelectric power plant using the dam. According to this embodiment, the power generation of the hydroelectric power plant can be predicted with high accuracy.

[0154] The explanatory variables may also include parameters representing the factory's operating conditions, and the target variable may include parameters representing the factory's state. According to this embodiment, factory anomalies can be detected with high precision.

[0155] [Application Example]

[0156] The above embodiments illustrate an example of a prediction system performing the task of predicting the power generation of a hydroelectric power station utilizing a dam. As another example, the prediction system described in these embodiments can be applied to the task of monitoring the status of a plant.

[0157] Techniques for monitoring plant status include using regression parameters between state variables as indicators of plant status. By periodically updating the regression formula using the most recent data, the most up-to-date regression parameters can always be utilized.

[0158] For example, the relationship between a factory's raw material input level *x* and its pressure *y* can be modeled as *y = ax + b*, with *x* as the explanatory variable and *y* as the target variable. The model parameters *a* and *b* vary depending on external factors. These external factors could include, for example, ambient air temperature or dust accumulation in pipes. In factory condition monitoring, regression parameters *a* and *b* are used as criteria for determining when pipes are clean.

[0159] When state variables are in a stable state, and the numerical variance of the most recent data used for regression is scarce, the regression equation may become a constant expression, and the regression parameters may diverge. However, if a predictive system is implemented that can calculate the regression parameters by pre-keeping the latest state variables in the summary data for each summary of state variable information and adding the summary data to the most recent data, the regression parameters can be obtained.

[0160] The prediction system in the application example can suppress the decrease in numerical variance of the data used in the regression and obtain the regression parameters with high accuracy. In another aspect, based on the application example, it can detect factory anomalies with high accuracy.

[0161] Other applications will not be detailed here, but the prediction system can also be used to predict meteorological or natural disasters such as earthquakes and volcanic eruptions.

[0162] [Replenish]

[0163] The functions of the embodiments described above can be implemented by one or more processing circuits. Here, "processing circuit" in this specification includes devices such as processors programmed to perform functions via software, such as CPUs (Central Processing Units) or GPUs (Graphics Processing Units) installed by electronic circuits, ASICs (Application Specific Integrated Circuits), DSPs (Digital Signal Processors), FPGAs (Field Programmable Gate Arrays), or conventional circuit modules designed to perform the functions described above.

[0164] The embodiments of this disclosure have been described in detail above. These embodiments are illustrative in all respects and not limiting. Various modifications and variations of the embodiments are possible without departing from the scope and spirit of the appended claims. The items described in the above embodiments can be employed in other structures or combined within the scope of non-conflict.

[0165] This application claims priority to Japanese Patent Application No. 2023-184702, filed with the Japanese Patent Office on October 27, 2023, and the entire contents of that Japanese patent application are incorporated herein by reference.

[0166] Symbol Explanation

[0167] 10. Information processing device

[0168] 20 Terminal devices

[0169] 101 Obtained Department

[0170] 102 Observation Data Storage Department

[0171] 103 Summary Department

[0172] 104 Summary Data Storage Department

[0173] Production Department 105

[0174] 106 Study Department

[0175] 107 Model Storage Department

[0176] 108 Forecasting Department

[0177] 109 Result Output Department

[0178] 1000 Prediction System

Claims

1. An information processing apparatus, comprising: The data acquisition department is responsible for acquiring observational data, including explanatory and target variables, within a specified timeframe. The summarization unit generates a specified number of summary data representing the observation data according to each summary information of the observation data; and The learning department generates a fully learned model based on the learning data, which includes the observation data and the summary data obtained in the learning target range.

2. The information processing apparatus according to claim 1, further comprising: The prediction unit predicts the target variable by inputting the explanatory variables obtained in the prediction target interval into the learned model.

3. The information processing apparatus according to claim 2, wherein, The prediction target interval is a time interval that includes the current time. The learning target interval is the time interval immediately preceding the prediction target interval.

4. The information processing apparatus according to claim 3, wherein, The summarizing unit generates the summarizing data, which includes the observation data obtained before the learning target interval.

5. The information processing apparatus according to claim 4, wherein, The summarizing unit generates summary data representing the observation data re-acquired according to a predetermined number of each summary information.

6. The information processing apparatus according to any one of claims 1 to 5, wherein, The summarizing unit generates the summary information by rounding the observed data.

7. The information processing apparatus according to any one of claims 1 to 6, wherein, In the observed data, the numerical variance of the explanatory variable or the target variable changes over time.

8. The information processing apparatus according to any one of claims 1 to 7, wherein, The explanatory variables include the water level of the dam. The target variable includes the power generation of the hydroelectric power station using the dam.

9. The information processing apparatus according to any one of claims 1 to 7, wherein, The explanatory variables include parameters representing the working conditions of the factory. The target variable includes parameters representing the state of the factory.

10. An information processing method, executed by a computer: The steps to obtain observational data containing explanatory and target variables within a specified time; The step of generating a predetermined number of summary data representing the observation data according to each summary information of the observation data; and The step of generating a fully learned model based on learning data that includes the observation data and the summary data obtained in the learning target range.

11. A program for causing a computer to perform: The steps to obtain observational data containing explanatory and target variables within a specified time; The step of generating a predetermined number of summary data representing the observation data according to each summary information of the observation data; and The step of generating a fully learned model based on learning data that includes the observation data and the summary data obtained in the learning target range.

Images