Information processing apparatus, learning method, and learning program

Abstract

To allow for continuously performing appropriate updating of an inference model.SOLUTION: An information processing apparatus (1) comprises: a data acquisition unit (101) which successively acquires successively generated data for machine learning of an inference model (111); and a learning unit (107) which updates the inference model (111) on the basis of acquired data if the acquired data satisfies a prescribed condition for determining normalcy of the data.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to an information processing device or the like that updates an inference model. [Background technology]

[0002] In recent years, inference models constructed by machine learning have been put to practical use in various fields. For example, Patent Document 1 below discloses an information processing device that uses a predictive learning model constructed by machine learning to predict the concentration of acid gases in exhaust gas emitted from a waste incinerator. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Application Laid-Open No. 2022-13163 Summary of the Invention [Problem to be solved by the invention]

[0004] When predicting the concentration of acid gases in the exhaust gas emitted from a waste incinerator, it is desirable to periodically update the predictive learning model to obtain prediction results that adapt to such fluctuations, since the amount of waste being incinerated varies from season to season. Furthermore, prediction accuracy can decline due to factors such as the deterioration of incinerators and other equipment over time, so updating the predictive learning model is also important for these reasons.

[0005] Generally, an inference model is updated using a large amount of training data generated using data measured over a predetermined period of time. However, when this type of update is applied, the inference model cannot be updated until the required amount of training data has been accumulated. It is also possible to generate training data based on measurement data measured sequentially without waiting for the training data to be accumulated, and to sequentially update the inference model using the generated training data. However, in this case, if the measurement data used is not suitable for learning, the inference accuracy of the inference model may actually decrease.

[0006] This is a problem that arises not only in the predictive learning model described in Patent Document 1, but also in any inference model that is constructed by machine learning and that infers any inference item. The present invention has been made in light of this problem, and its purpose is to provide an information processing device or the like that enables appropriate and continuous updating of an inference 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 a data acquisition unit that sequentially acquires data for machine learning of an inference model that is generated sequentially, and a learning unit that updates the inference model based on the data when the acquired data satisfies predetermined conditions for determining the normality of the data.

[0008] In addition, in order to solve the above-mentioned problems, a learning method according to one embodiment of the present invention is a learning method executed by one or more information processing devices, and includes a data acquisition step of sequentially acquiring data for machine learning of an inference model that is generated sequentially, and a learning step of updating the inference model based on the data when the acquired data satisfies predetermined conditions for determining the normality of the data. [Effects of the Invention]

[0009] One aspect of the present invention enables continuous and appropriate updates of inference models. [Brief explanation of the drawings]

[0010] [Figure 1] 1 is a block diagram showing an example of a main configuration of an information processing device according to an embodiment of the present invention; [Figure 2] 1 is a block diagram illustrating an example of the configuration of a control system according to an embodiment of the present invention. [Figure 3] 2 is a flowchart showing the flow of processing executed by the information processing device shown in FIG. [Figure 4] A block diagram showing an example of the main configuration of an information processing device that uses multiple inference models in combination. DETAILED DESCRIPTION OF THE INVENTION

[0011] [System Configuration] An overview of a control system 5 according to one embodiment of the present invention will be described with reference to Fig. 2. Fig. 2 is a block diagram showing an example configuration of the control system 5. The control system 5 is a system for automatically controlling the operation of a device 3 installed in a facility to which the control system 5 is applied. The control system 5 shown in Fig. 2 includes an information processing device 1, a control device 2, a device 3, and a sensor 4.

[0012] The facility to which the control system 5 is applied may be any facility that is equipped with the device 3 to be controlled, and the control system 5 can be applied to any facility. For example, the control system 5 can be applied to industrial product manufacturing facilities, vegetable, flower, crop, etc. production facilities, waste treatment facilities, power generation facilities, etc. The following mainly describes an example in which the control system 5 is applied to a waste treatment facility that incinerates waste and generates power by utilizing the exhaust heat generated by the incineration of the waste.

[0013] As described above, the device 3 is a device installed in a facility to which the control system 5 is applied. The control system 5 can also control multiple devices 3. That is, although one device 3 is shown in FIG. 2, the control system 5 may include multiple devices 3. Furthermore, the multiple devices 3 may all be the same type of device, or may be different types. When the control system 5 is applied to a waste treatment facility, the device 3 may be, for example, an exhaust gas treatment device that supplies a treatment agent to remove harmful components (e.g., hydrogen chloride) from exhaust gas generated by incineration of waste, or a supply device that supplies combustion air.

[0014] The sensor 4 is a device provided to monitor the status of at least one of the facility to which the control system 5 is applied and the equipment 3 provided in the facility, and can also be referred to as a measuring device or a detection device. Although one sensor 4 is shown in FIG. 2, the control system 5 may include multiple sensors 4. Furthermore, the multiple sensors 4 may all be the same type of sensor, or may be different types. When the control system 5 is applied to a waste treatment facility, the sensor 4 may be, for example, a sensor that measures the concentration of acid gases (e.g., hydrogen chloride, nitrogen oxides, sulfur oxides, etc.) in the exhaust gas generated by the incineration of waste, or a sensor that measures the amount of combustion air supplied.

[0015] The control device 2 is a device that controls the operation of the device 3. Although one control device 2 is shown in FIG. 2, the control system 5 may include multiple control devices 2. For example, if the control system 5 includes multiple devices 3, a control device 2 that controls each device 3 may be provided for each device 3. Note that the information processing device 1 may also control the device 3 without going through the control device 2, in which case the control device 2 is omitted.

[0016] The information processing device 1 controls the operation of the device 3 via the control device 2. In performing this operation control, the information processing device 1 performs inference using data collected in the facility to which the control system 5 is applied (for example, data measured by the sensor 4) and an inference model. Then, the information processing device 1 controls the operation of the device 3 based on the inference result.

[0017] Furthermore, the information processing device 1 updates the inference model using data collected at the facility. As will be described in detail later, the information processing device 1 sequentially acquires data for machine learning of the inference model, and updates the inference model based on the acquired data if the acquired data satisfies predetermined conditions for determining the normality of the data. Note that the inference and the update of the inference model may be performed by separate independent devices.

[0018] As described above, the control system 5 includes an information processing device 1 that sequentially acquires data for machine learning of an inference model and, if the acquired data satisfies predetermined conditions for determining the normality of the data, updates the inference model based on the data, and a control device 2 that controls the device 3 based on the results of inference using the updated inference model. Therefore, it is possible to continuously update the inference model appropriately, thereby continuously controlling the device 3 appropriately.

[0019] [Configuration of information processing device] A more detailed configuration of the information processing device 1 will be described with reference to Fig. 1. Fig. 1 is a block diagram showing an example of the configuration of the main parts of the information processing device 1. As shown in the figure, the information processing device 1 includes a control unit 10 that controls each unit of the information processing device 1, and a storage unit 11 that stores various data used by the information processing device 1. The information processing device 1 also includes a communication unit 12 that enables the information processing device 1 to communicate with other devices, an input unit 13 that accepts input to the information processing device 1, and an output unit 14 that enables the information processing device 1 to output various data.

[0020] The control unit 10 also includes a data acquisition unit 101, an inference unit 102, a device control unit 103, a judgment unit 104, a training data generation unit 105, a reference value update unit 106, a learning unit 107, an anomaly detection unit 108, and a reset unit 109. The memory unit 11 stores an inference model 111.

[0021] The data acquisition unit 101 acquires various data used by the information processing device 1. For example, the data acquisition unit 101 acquires data for inference used for inference by the inference model 111. The data for inference can also be referred to as explanatory variables of the inference model 111, or input data to be input to the inference model 111.

[0022] The data for inference may correspond to the inference content (which may also be referred to as the target variable) of the inference model 111. For example, the data for inference may be a measurement value measured by the sensor 4, or a set value set in the control device 2 or the device 3 (for example, a set value for the amount of combustion air supplied). These data may also be time-series data. The data for inference may also include both the measurement value and the set value. The data for inference may also be a feature value calculated using at least one of the measurement value and the set value. In this case, the calculation of the feature value may be performed by the data acquisition unit 101, or the data acquisition unit 101 may acquire a feature value calculated by another device.

[0023] Although details will be described later, the training data generation unit 105 uses the above-mentioned data for inference to generate training data for machine learning of the inference model 111. In other words, it can be said that the data acquisition unit 101 acquires data for machine learning of the inference model 111. Data for inference (which can also be referred to as data for machine learning) is generated sequentially during the operation period of the facility, and the data acquisition unit 101 sequentially acquires the sequentially generated data for inference.

[0024] The inference unit 102 performs inference using an inference model 111. The inference model 111 is an inference model generated by supervised machine learning. The inference content (objective variable) of the inference model 111 is related to the facility to which the control system 5 is applied, and is not particularly limited as long as it provides an inference result that can be used to determine whether or not to control the device 3 or to determine the control content of the device 3.

[0025] For example, the inference model 111 may be an inference model (which may also be called a prediction model) that outputs prediction results for a predetermined item in a facility to which the control system 5 is applied. As a specific example, the inference model 111 may be a prediction model that predicts the concentration of acid gases a predetermined time after a certain time from time-series measurement data measured by the sensor 4 up to that time in a facility to which the control system 5 is applied.

[0026] Furthermore, for example, the inference model 111 may be an inference model (which may also be called a classification model) that outputs a classification result for classifying a predetermined object in a facility to which the control system 5 is applied. As a specific example, the inference model 111 may be a classification model that classifies the operating status of the facility at a certain time into normal or abnormal based on data measured by the sensor 4 at that time in the facility to which the control system 5 is applied.

[0027] The device control unit 103 controls the device 3 based on the inference result of the inference unit 102. More specifically, the device control unit 103 determines whether or not control of the device 3 is necessary based on the inference result of the inference unit 102, and controls the operation of the device 3 via the control device 2 if it determines that control is necessary. The device control unit 103 may also directly control the device 3. The criteria for determining whether or not control is necessary, the device 3 to be controlled, and the content of the control to be performed by the device 3 may be determined in advance. For example, when an inference model 111 that predicts the concentration of acid gas after a predetermined time is used, the device control unit 103 may determine that control is necessary if the predicted value output by the inference model 111 exceeds a predetermined upper limit. In this case, the device control unit 103 performs control to reduce the concentration of the acid gas (e.g., control to increase the amount of treatment agent supplied by the exhaust gas treatment device). The device control unit 103 may also perform control by continuously varying the control amount according to the predicted value.

[0028] The determination unit 104 determines whether the data acquired by the data acquisition unit 101 satisfies a predetermined condition for determining the normality of the data. The predetermined condition will be explained later in the section "Regarding the predetermined condition."

[0029] The training data generation unit 105 generates training data for machine learning of the inference model 111. More specifically, the training data generation unit 105 generates training data by associating input data input to the inference model 111 to output an inference result with correct answer data to be output by the inference model 111. The correct answer data may be generated by the training data generation unit 105, or may be input by a user of the control system 5 via the communication unit 12 or the input unit 13.

[0030] The reference value update unit 106 updates the predetermined reference value used in the determination by the determination unit 104. Details of the reference value and the method of updating it will be explained later in the section "Regarding Predetermined Conditions."

[0031] The learning unit 107 performs machine learning of the inference model 111. More specifically, when data acquired by the data acquiring unit 101 satisfies a predetermined condition for determining the normality of the data, the learning unit 107 updates the inference model 111 based on the data. As described above, the determination unit 104 determines whether the predetermined condition is satisfied, and therefore the learning unit 107 updates the inference model 111 based on data that the determination unit 104 determines satisfies the predetermined condition. Also, as described above, the training data generation unit 105 generates training data from the data acquired by the data acquiring unit 101, and therefore the learning unit 107 updates the inference model 111 using this training data. Note that the learning unit 107 may not only update the inference model 111 but also generate the inference model 111.

[0032] As described above, in the information processing device 1, the training data generation unit 105 generates training data using data sequentially acquired by the data acquisition unit 101, and the learning unit 107 sequentially updates the inference model 111 using the generated training data. This makes it possible to start operation of a facility using the inference model 111 at a stage when there is not enough training data to generate the inference model 111, and to update the inference model 111 to one that is suitable for the facility during the operation period of the facility. Furthermore, when the condition of the facility changes due to aging or the like, or when the condition of the waste being treated at the facility changes due to seasonal changes or the like, the inference model 111 can be adapted to these changes.

[0033] The anomaly detection unit 108 detects anomalies in the facility to which the control system 5 is applied. The anomalies that are the target of detection by the anomaly detection unit 108 are anomalies in the facility, and are anomalies that affect the input data input to the inference model 111. The anomaly detection by the anomaly detection unit 108 is performed continuously while the facility is in operation. While detecting an anomaly, the anomaly detection unit 108 may notify the occurrence of an anomaly by outputting an anomaly signal, for example. The anomaly detection unit 108 may also record the above detection results in the memory unit 11, etc.

[0034] The reset unit 109 returns the inference model 111, which has been updated by the learning unit 107, to the state before the update. The reset unit 109 is not an essential component, but is useful when it is discovered after updating the inference model 111 that data unsuitable for learning was used in the update, or when the inference accuracy of the inference model 111 has decreased due to the update. Any method can be used to return the inference model 111 to the state before the update. For example, when the learning unit 107 updates the inference model 111, the value of each parameter in the inference model 111 before the update may be recorded together with the date and time of the update. This allows the reset unit 109 to return the inference model 111 to the state before the update using the recorded value of each parameter.

[0035] As described above, the information processing device 1 includes a data acquisition unit 101 that sequentially acquires data for machine learning of the inference model 111, which is generated sequentially, and a learning unit 107 that updates the inference model 111 based on the acquired data when the acquired data satisfies predetermined conditions for determining the normality of the data.

[0036] According to the above configuration, the inference model 111 is updated using data that satisfies predetermined conditions for determining normality from among the sequentially generated data for machine learning. This makes it possible to continuously update the inference model 111 appropriately.

[0037] As described above, the data acquired by the data acquisition unit 101 may be data for inference. Furthermore, the data acquired by the data acquisition unit 101 may be training data for machine learning. In this case, the learning unit 107 updates the inference model 111 using the acquired training data. In this case, the training data generation unit 105 is omitted.

[0038] Furthermore, the data acquisition unit 101 may acquire multiple pieces of data (each of the inference data used in multiple inferences performed in time series, or multiple pieces of training data generated based on them). In this case, the learning unit 107 may update the inference model 111 based on the multiple pieces of data if all of the multiple pieces of data satisfy a predetermined condition. However, if too much data is acquired at one time, the update interval of the inference model 111 may become too long. For this reason, it is preferable that the data acquisition unit 101 acquires data in such a quantity and frequency that it can be updated at the desired update interval.

[0039] [About inference models] As described above, various models can be applied as the inference model 111. For example, the inference model 111 may be one that outputs a predicted value of the amount of change in predetermined data at a predetermined facility based on predetermined data generated at the facility at a certain point in time. In this case, the inference model 111 that outputs the predicted value of the amount of change in the predetermined data, rather than the predicted value of the predetermined data, is the one that is subject to update. This makes it possible to prevent fraudulent learning from occurring.

[0040] For example, suppose that an inference model 111 that outputs a predicted value of measurement data three minutes later is updated based on measurement data measured at a certain time. In this case, if the measurement data at a certain time is output as a predicted value as is, the time-series changes in the predicted value will follow the time-series changes in the measurement data, and therefore, it appears that a valid predicted value is being output. This may result in improper learning that outputs a value close to the measurement data at a certain time, rather than proper learning that improves prediction accuracy.

[0041] In this regard, when updating the inference model 111 that outputs a predicted value of the change in specified data, no fraudulent learning is performed that outputs a value close to the specified data generated at a certain point in time, but rather appropriate learning is performed that improves the accuracy of the prediction.

[0042] [Regarding training data generation] As described above, the inference model 111 may output a predicted value for a specific facility based on data measured at the facility. In this case, the training data generation unit 105 can generate training data by associating the actual measured value measured at a specific time as correct answer data with the input data input to the inference model 111 to calculate a predicted value at the specific time.

[0043] In other words, the above configuration makes it possible to automate a series of tasks from generating training data to updating the inference model 111. Note that, if the inference model 111 is a classification model, for example, it is possible to semi-automate a series of tasks from generating training data to updating the inference model 111 by having the user input correct answer data.

[0044] [Regarding specified conditions] As described above, the determination unit 104 determines whether the data acquired by the data acquisition unit 101 satisfies a predetermined condition for determining the normality of the data. If it is determined that the predetermined condition is satisfied, the learning unit 107 updates the inference model 111 based on the data.

[0045] For example, the predetermined condition may include a condition that the value of the data acquired by the data acquisition unit 101 is within a predetermined normal range set based on the design value of the measurement device (e.g., the sensor 4) that measured the data. According to this configuration, if data outside the predetermined normal range set based on the design value is acquired, the inference model 111 is not updated. This makes it possible to avoid updating the inference model 111 using data that is likely to be unsuitable for learning. For example, at least one of the upper and lower design output limits of the measurement device can be set as the "design value of the measurement device." Note that, in addition to the normal range, an abnormal range may also be set. In this case, data within the normal range is used to update the inference model 111, while data within the abnormal range and data outside both the normal and abnormal ranges are not used to update the inference model 111.

[0046] As described above, the inference model 111 may output an inference result regarding a specific facility based on data acquired at the facility. The information processing device 1 also includes an anomaly detection unit 108 that detects anomalies in the facility. In this case, the specific condition may include a condition that the data acquired by the data acquisition unit 101 was acquired at the facility during a period in which the anomaly detection unit 108 did not detect any anomalies.

[0047] According to the above configuration, if data that is suspected to be an abnormal value is acquired at a facility during a period when an abnormality is occurring, the inference model 111 is not updated. This makes it possible to avoid updating the inference model 111 using data that is likely to be unsuitable for learning.

[0048] The predetermined condition may also include a condition that the value of the acquired data is not a statistically outlier. For example, the predetermined condition may include a condition that the acquired data is within a range from a value obtained by subtracting the standard deviation of a plurality of previously acquired data from the average value of the plurality of data to a value obtained by adding the standard deviation to the average value.

[0049] It should be noted that the "plurality of data acquired in the past" may include the latest data acquired by the data acquisition unit 101. In the following, the average value is referred to as μ and the standard deviation as σ. That is, this condition is expressed as "μ-σ≦(value of acquired data)≦μ+σ". This condition can also be expressed as the value of the acquired data being within the range of "μ±σ". It should be noted that the reference value for determining whether or not a value is a statistical outlier is not limited to this example, and may be determined appropriately depending on the properties of the data used, the inference content, the accuracy required for the inference result, etc.

[0050] According to the above configuration, when data is acquired that is statistically likely to deviate significantly from the average value μ of previously acquired data, the inference model 111 is not updated. This makes it possible to avoid updating the inference model 111 using data that is statistically likely to be unsuitable for learning.

[0051] Furthermore, with the above configuration, when data is acquired that is statistically small in deviation from the average value μ of previously acquired data, the inference model 111 is updated, which makes it possible to adapt the inference model 111 to changes in the trends of the acquired data.

[0052] The predetermined condition may also be changeable by the user of the control system 5 (which may also be referred to as the user of the information processing device 1). In this case, the predetermined condition may be that the value of the acquired data is within the range of (μ-Xσ) to (μ+Xσ). X is a positive number that can be set by the user. This allows the user to set the value of X to a desired value and adjust the criteria for updating the inference model 111. The change in the value of X may be received by the reference value update unit 106 via the communication unit 12 or the input unit 13.

[0053] When the above μ and σ are used as reference values ​​for determining whether to update the inference model 111, the reference value update unit 106 updates μ and σ. Specifically, the reference value update unit 106 updates μ and σ by adding or changing data used to calculate μ and σ and recalculating μ and σ.

[0054] For example, when calculating μ and σ using all previously acquired data, the reference value update unit 106 adds newly acquired data to all previously acquired data and calculates μ and σ using all of that data. Alternatively, the reference value update unit 106 may update μ and σ using data acquired within the most recent predetermined period, for example.

[0055] The conditions for updating μ and σ are not particularly limited. For example, the reference value update unit 106 may update μ and σ when a predetermined period of time has elapsed since the most recent update, or when a predetermined number of new data items have been acquired.

[0056] However, if μ and σ are updated using data that is not valid, it may not be possible to properly determine whether to update the inference model 111. For this reason, it is desirable that the reference value update unit 106 not use data that is statistically considered to have an extremely large deviation from past data to update μ and σ.

[0057] Therefore, for example, when the value of the acquired data is within the range from the value obtained by subtracting twice the standard deviation σ from the average value μ to the value obtained by adding twice the standard deviation σ to the average value μ, the reference value update unit 106 may update the average value μ and the standard deviation σ using the data. This condition is expressed as "μ - 2σ ≤ (value of data) ≤ μ + 2σ". Also, this condition can be expressed as the value of the data being included in the range of "μ ± 2σ".

[0058] According to the above configuration, when data is acquired such that the degree of deviation from the average value of the data acquired in the past is extremely large statistically, that data is not reflected in μ and σ (i.e., the reference value). This makes it possible to avoid updating the reference value to an inappropriate value.

[0059] Also, according to the above configuration, when data is acquired such that the degree of deviation from the average value of the data acquired in the past is not extremely large statistically, the reference value is updated using that data. This makes it possible to reflect changes in the trend of the acquired data in the reference value.

[0060] Also, the user of the control system 5 (which can also be referred to as the user of the information processing apparatus 1) may be able to change the condition for updating the reference value. In this case, the condition for updating the reference value may be that the value of the acquired data is within the range from (μ - Yσ) to (μ + Yσ). Here, Y is a positive number that can be set by the user and is larger than X described above. Thus, the user can set the value of Y to a desired value and adjust the criteria for updating the reference value. Also, since X < Y, it is possible to perform the update of the reference value and the update of the inference model 111 at appropriate timings. Note that the change in the value of Y may be received by the reference value update unit 106 via the communication unit 12 or the input unit 13.

[0061] 〔Flow of processing〕 The flow of processing executed by the information processing device 1 will be described with reference to Fig. 3. Fig. 3 is a flowchart showing the flow of processing executed by the information processing device. This flowchart includes each step of the learning method according to this embodiment.

[0062] In S1, the data acquisition unit 101 acquires data for inference. For example, the data acquisition unit 101 may acquire the latest data measured by the sensor 4 as the data for inference. The process of S1 is performed every time new data for inference is generated. In other words, in S1, data for inference that is generated sequentially is acquired sequentially.

[0063] In S2, the inference unit 102 executes inference using the data acquired in S1. More specifically, the inference unit 102 inputs the data acquired in S1 into the inference model 111 and acquires the output value as the inference result.

[0064] In S3, the device control unit 103 determines whether or not control of device 3 is necessary based on the inference result of S2. If the determination in S3 is YES, the process proceeds to S4. In S4, the device control unit 103 controls the operation of device 3. On the other hand, if the determination in S3 is NO, the process proceeds to S5.

[0065] In S5 (data acquisition step), the data acquisition unit 101 acquires data for machine learning of the inference model 111. The processing of S5 is performed in conjunction with S1. That is, in S5 as well, data for machine learning that is generated sequentially is acquired sequentially.

[0066] The data acquired in S5 is inference data acquired in the past. For example, assume that an inference model 111 is used to predict the value of measurement data three minutes later from measurement data of sensor 4 at a certain time. In this case, the data acquisition unit 101 may acquire the inference data used for the prediction three minutes earlier. Since the actual measurement value of sensor 4 has been obtained at that time, training data can be generated by associating the actual measurement value with the measurement value as ground truth data.

[0067] If the inference model 111 is a classification model or the like, the data acquired in S1 may be used as data for machine learning. In this case, S1 becomes a data acquisition step, and the processing of S5 is omitted. In this case, the correct answer data (e.g., the correct classification result) may be input by the user.

[0068] In S6, the determination unit 104 determines whether the data acquired in S5 is data from a period during which the anomaly detection unit 108 did not detect any anomalies in the facility. As described above, the anomaly detection unit 108 may continuously detect anomalies throughout the operation of the facility, and the detection results may be recorded in the storage unit 11 or the like. Furthermore, the data acquisition unit 101 may record the acquisition or measurement time of each piece of data acquired in S1. This enables the determination unit 104 to make the determination in S6. If the determination in S6 is NO, the process returns to S1, and if the determination in S6 is YES, the process proceeds to S7.

[0069] In S7, the determination unit 104 determines whether the value of the data acquired in S5 is within a predetermined normal range set based on the design value of the sensor 4 that measured the data. If the determination in S7 is NO, the process returns to S1, and if the determination in S7 is YES, the process proceeds to S8.

[0070] In S8, the judgment unit 104 judges whether the value of the data acquired in S5 is within the range of "μ±2σ". If the judgment in S8 is NO, the value of the data is obviously not within the range of "μ±σ", so the process returns to S1 without updating the inference model 111. On the other hand, if the judgment in S8 is YES, the process proceeds to S9.

[0071] In S9, the reference value update unit 106 updates μ and σ, which are reference values ​​for determining whether or not to update the inference model 111. For example, the reference value update unit 106 may generate a new data group by adding the data acquired in S5 to the data used to calculate μ and σ before the update, and then calculate μ and σ for that data group.

[0072] In S10, the determination unit 104 determines whether the value of the data acquired in S5 is within the range of "μ±σ" using μ and σ updated in S9. If the determination in S10 is NO, the process returns to S1, and if the determination in S10 is YES, the process proceeds to S11.

[0073] In S11, the training data generation unit 105 acquires correct answer data to be associated with the data acquired in S5. As described above, if the training data generation unit 105 can acquire an actual measurement value corresponding to the data acquired in S5, the training data generation unit 105 acquires the actual measurement value as correct answer data. Alternatively, the training data generation unit 105 may acquire correct answer data input by a user.

[0074] In S12, the training data generation unit 105 generates training data by associating the data acquired in S5 with the correct answer data acquired in S11. Then, in S13 (learning step), the learning unit 107 updates the inference model 111 using the training data generated in S12. After this, the process returns to S1. As described above, the training data generated in S12 is generated from the data acquired in S5, so it can be said that in S13, the inference model 111 is updated based on the data acquired in S5.

[0075] 3, if the judgment unit 104 judges YES in all of S6, S7, and S10, that is, if all three predetermined conditions are satisfied, the learning unit 107 updates the inference model 111. However, it is sufficient to use at least one predetermined condition. For example, the processes of S6 and S7 may be omitted, and the inference model 111 may be updated if the judgment in S10 is YES.

[0076] Furthermore, the training data generation unit 105 may generate training data by executing the processes of S11 and S12 after the process of S1 and before the processes of S6 and subsequent steps. In this case, the data acquisition unit 101 acquires the generated training data as data for machine learning. Thereafter, the processes of S6 and subsequent steps are performed on the training data, and if the determination in S10 is YES, the process of S13 is performed.

[0077] The learning method according to this embodiment includes a data acquisition step (S5) of sequentially acquiring data for machine learning that is generated sequentially, and a learning step (S13) of updating the inference model 111 based on the acquired data if the acquired data satisfies a predetermined condition for determining the normality of the data. This makes it possible to continuously update the inference model 111 appropriately.

[0078] [Combined use of multiple inference models] In the above explanation, an example of performing inference using one inference model 111 has been described, but it is also possible to use multiple inference models in combination. The use of multiple inference models in combination will be explained based on Figure 4. Figure 4 is a block diagram showing an example of the main configuration of an information processing device 1A that uses multiple inference models in combination. Note that components common to the information processing device 1 are given the same reference numbers and their explanations will not be repeated.

[0079] 4, the information processing device 1A includes a control unit 10A that controls the various units of the information processing device 1A, and a storage unit 11A that stores various data used by the information processing device 1A. The control unit 10A also includes a data acquisition unit 101, a training data generation unit 105, and a learning unit 107, as well as three inference units (inference units 102A to 102C), three determination units (determination units 104A to 104C), three inference units (reference value update units 106A to 106C), and an integration unit 110. The storage unit 11A also stores three inference models (inference models 111A to 111C).

[0080] The information processing device 1A can use two inference models, or can be expanded to a configuration using four or more inference models. Similarly to the information processing device 1, the information processing device 1A may include a device control unit 103, an abnormality detection unit 108, and a reset unit 109.

[0081] The inference models 111A to 111C are inference models that share common explanatory variables and objective variables. Therefore, the inference models 111A to 111C can be updated using the same training data. The inference models 111A to 111C before updating may be inference models generated using the same training data set, or may be inference models generated using different training data sets. Note that the explanatory variables of the inference models 111A to 111C may be different for each model.

[0082] As will be explained in detail below, the reference values ​​(specifically, the average value μ and standard deviation σ of the data) used to determine whether or not the inference models 111A to 111C need to be updated are different for each model. As a result, the data used to update the inference models 111A to 111C also differs for each model. As a result, the characteristics of the inference models 111A to 111C also differ as updates are repeated.

[0083] The inference units 102A to 102C perform inference using inference models 111A to 111C, respectively. The integration unit 110 integrates the inference results of the inference units 102A to 102C. The method of integrating the inference results is not particularly limited. For example, when the inference results of the inference units 102A to 102C are numerical values ​​(e.g., predicted values), the integration unit 110 may output a weighted average (which may also be referred to as a weighted average value) of the inference results of the inference units 102A to 102C as the integrated inference result. In this case, the weight may be a fixed value, or may be changed depending on the operating status of the facility, the data used for the inference, etc.

[0084] The determination units 104A to 104C determine whether the data acquired by the data acquisition unit 101 satisfies a predetermined condition for determining the normality of the data. The predetermined condition includes a condition that the acquired data is within the range of "μ±σ" (which can also be said as within the normal range). As described above, the reference values ​​μ and σ used by the determination units 104A to 104C to determine this condition are different from each other. In the following, the reference value used by the determination unit 104A is referred to as μ A , σ ASimilarly, the reference value used by the determination unit 104B is called μ B , σ B The reference value used by the determination unit 104C is called μ C , σ C It is called.

[0085] The judgment result of judgment unit 104A determines whether to update inference model 111A. Similarly, the judgment result of judgment unit 104B determines whether to update inference model 111B, and the judgment result of judgment unit 104C determines whether to update inference model 111C. Note that, below, an example is described in which the "predetermined condition" is the condition that the acquired data is within the range of "μ±σ", but the "predetermined condition" may also include other conditions such as those used by judgment unit 104.

[0086] The reference value update units 106A to 106C update the reference values ​​used in the determinations made by the determination units 104A to 104C, respectively. A and σ A The reference value update unit 106B updates the above μ B and σ B The reference value update unit 106C updates the above μ C and σ C The reference value update unit 106A updates the reference value when the acquired data is, for example, "μ A ±2σ A If it is within the range of μ A and σ A The same applies to the reference value update units 106B and 106C.

[0087] The above-mentioned standard values ​​are calculated based on different collection periods of data, and therefore the values ​​differ accordingly. For example, when making inferences for facilities that operate year-round, μ A , σ A is calculated using all data used for inference in the past, and μ B , σ B is calculated using the data used for inference over the last three months, and μ C , σ Cmay be calculated using data used in the most recent week of inference. In this case, μ A and σ A , μ B and σ B , μ C and σ C If the numbers of data used in the calculation are N1, N2, and N3, respectively, then N1>N2>N3.

[0088] In this way, by varying the collection periods of the data used to calculate μ and σ, it becomes possible for inference models 111A to 111C to output inference results adapted to the collection periods. For example, inference model 111C is updated to output inference results adapted to short-term changes in the data. Meanwhile, inference model 111B is updated to output inference results adapted to medium-term changes in the data, and inference model 111C is updated to output inference results adapted to long-term changes in the data.

[0089] For example, in a waste incineration facility, if a disaster or the like occurs, the type of waste to be incinerated may change temporarily, which may cause a temporary change in the trend of data measured by sensors 4 or the like installed in the facility. For this reason, data measured for a period of time after the occurrence of a disaster or the like (called the disaster-affected period) is considered to be "μ A ±σ A " or "μ B ±σ B Therefore, it is difficult to update the inference models 111A and 111B using data measured during the disaster impact period. C and σ C Since the characteristics of the data during the disaster-affected period are strongly reflected in μ C ±σ C " range. Therefore, the inference model 111C is likely to be updated using data measured during the disaster-affected period, which makes it easier for the inference model 111C to output inference results that are adapted to the disaster-affected period. In other words, it is possible to improve the accuracy of inference during the disaster-affected period.

[0090] Furthermore, for example, if the trends in the data measured change from season to season, an inference model corresponding to each season may be prepared, and the reference values ​​for determining whether or not to update the inference model may be calculated from the data acquired in the corresponding season, thereby improving the inference accuracy for each season.

[0091] For example, the need for updating a summer inference model may be determined using μ and σ calculated from data measured in the summer (e.g., June to August). This allows the summer inference model to be updated using data that is an outlier from the trends of data for other seasons or throughout the year, as long as it falls within a normal range for summer data (i.e., the μ±σ range for summer data). This allows the summer inference model to be updated so that inference results suited to summer data can be output. Note that in this case, it is sufficient to change the inference model used for each season; there is no need to integrate and output the inference results of the inference models for each season. Alternatively, the inference model may be updated using data measured over the past few months (e.g., three months). This type of update also allows the inference model to adapt to seasonal changes.

[0092] The learning unit 107 included in the information processing device 1A performs the following process for each of the multiple inference models 111A to 111C, each having a different data collection period, to update the inference model based on the data acquired by the data acquisition unit 101 when the value of the data is within a normal range represented by a reference value calculated from multiple data collected during a data collection period corresponding to the inference model. This allows the inference models 111A to 111C to adapt to the corresponding data collection period, thereby improving the inference accuracy during that period.

[0093] Furthermore, the reference value update unit 106A updates the reference value μ calculated from a plurality of data collected during a data collection period corresponding to the inference model 111A when the value of the data acquired by the data acquisition unit 101 is A and σA Within the normal range expressed using the formula (e.g., "μ A ±2σ A " range), the reference value μ A and σ A is updated based on the data. This makes it possible to update the reference value to an appropriate value according to the data collection period and appropriately determine whether or not the inference model 111A needs to be updated during that period. The same applies to the reference value update unit 106B and the reference value update unit 106C.

[0094] [Modification] The execution entity of each process described in the above-described embodiment may be any entity and is not limited to the above-described examples. That is, the functions of the information processing device 1 or 1A can be realized by a plurality of information processing devices (which may also be called processors) that can communicate with each other. For example, each process described in the flowchart of FIG. 3 can be shared and executed by a plurality of information processing devices. That is, the execution entity of the learning method in the above-described embodiment may be one information processing device or a plurality of information processing devices.

[0095] [Software implementation example] The functions of the information processing device 1 can be realized by a program (learning program) that causes a computer to function as the information processing device 1 and causes a computer to function as each control block of the device (particularly each part included in the control unit 10).

[0096] In this case, the information processing device 1 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 each function described in the above embodiment.

[0097] The program may be stored non-transitory on one or more computer-readable storage media. The storage media may or may not be included in the information processing device 1. In the latter case, the program may be supplied to the device via any wired or wireless transmission medium.

[0098] 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.

[0099] 〔summary〕 An information processing device according to aspect 1 of the present invention includes a data acquisition unit that sequentially acquires data for machine learning of an inference model, which is generated sequentially, and a learning unit that updates the inference model based on the acquired data when the acquired data satisfies predetermined conditions for determining the normality of the data.

[0100] In an information processing device according to aspect 2 of the present invention, in aspect 1, the predetermined condition includes a condition that the value of the acquired data is within the range of (μ-Xσ) to (μ+Xσ), where μ is the average value of multiple data acquired in the past, σ is the standard deviation of the multiple data, and X is a positive number that can be set by a user of the information processing device.

[0101] An information processing device according to aspect 3 of the present invention is the same as in aspect 2, and includes a reference value update unit that updates the mean value and the standard deviation using the acquired data when the value of the data is within the range of (μ-Yσ) to (μ+Yσ), where Y can be set by a user of the information processing device and is a positive number greater than X.

[0102] An information processing device according to aspect 4 of the present invention is, in any of aspects 1 to 3, an information processing device in which the inference model outputs inference results regarding a specified facility based on data acquired at the facility, and is equipped with an anomaly detection unit that detects abnormalities at the facility, and the specified condition includes a condition that the data acquired by the data acquisition unit was acquired at the facility during a period in which the anomaly detection unit did not detect any abnormalities.

[0103] In the information processing device of aspect 5 of the present invention, in any of aspects 1 to 4, the specified condition includes a condition that the value of the data acquired by the data acquisition unit is a numerical value within a specified normal range set based on the design value of the measuring device that measured the data.

[0104] An information processing device according to aspect 6 of the present invention is, in any of aspects 1 to 5, characterized in that the inference model outputs a predicted value of the amount of change in specified data at a specified facility based on specified data generated at a certain point in time at the facility.

[0105] An information processing device according to aspect 7 of the present invention, in any one of aspects 1 to 6, is provided with a reset unit that returns the inference model to its state before updating.

[0106] An information processing device according to aspect 8 of the present invention is, in any of aspects 1 to 7, characterized in that the inference model outputs a predicted value for a specified facility based on data measured at the facility, and is equipped with a training data generation unit that generates training data by associating actual measured values ​​measured at a specified time as correct answer data with input data input to the inference model to calculate a predicted value at the specified time.

[0107] In an information processing device according to aspect 9 of the present invention, in any of aspects 1 to 8, the learning unit performs the process of updating the inference model based on the acquired data if the value of the acquired data is within a normal range represented by a reference value calculated from multiple data collected during a data collection period corresponding to the inference model, for each of multiple inference models each having a different data collection period.

[0108] A learning method according to aspect 10 of the present invention is a learning method executed by one or more information processing devices, and includes a data acquisition step of sequentially acquiring data for machine learning of an inference model, which is generated sequentially, and a learning step of updating the inference model based on the data when the acquired data satisfies predetermined conditions for determining the normality of the data.

[0109] A learning program according to an eleventh aspect of the present invention is a learning program for causing a computer to function as the information processing device according to the first aspect, and causes the computer to function as the data acquisition unit and the learning unit.

[0110] 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]

[0111] 1. Information processing equipment 101 Data Acquisition Unit 106 Reference value update unit 107 Learning Department 108 Abnormality detection unit 109 Reset section 111 Inference Model 4. Sensors (measuring devices) 1A Information processing equipment 106A~106C Reference value update section 111A~111C Inference Model

Claims

1. a data acquisition unit that sequentially acquires data for machine learning of the inference model, which is sequentially generated; An information processing device comprising: a learning unit that updates the inference model based on the acquired data when the acquired data satisfies predetermined conditions for determining the normality of the data.

2. 2. The information processing device according to claim 1, wherein the predetermined condition includes a condition that the value of the acquired data is within a range of (μ-Xσ) to (μ+Xσ), where μ is an average value of a plurality of data acquired in the past, σ is a standard deviation of the plurality of data, and X is a positive number that can be set by a user of the information processing device.

3. 3. The information processing device according to claim 2, further comprising a reference value update unit that updates the average value and the standard deviation using the acquired data when the value of the acquired data is within a range from (μ-Yσ) to (μ+Yσ), wherein Y is settable by a user of the information processing device and is a positive number greater than X.

4. the inference model outputs inference results regarding a predetermined facility based on data acquired at the facility; an abnormality detection unit that detects an abnormality in the facility; 4. The information processing device according to claim 1, wherein the predetermined condition includes a condition that the data acquired by the data acquisition unit is acquired in the facility during a period in which the anomaly detection unit does not detect an anomaly.

5. 4. The information processing device according to claim 1, wherein the predetermined condition includes a condition that the value of the data acquired by the data acquisition unit is a numerical value within a predetermined normal range set based on a design value of the measuring device that measured the data.

6. An information processing device described in any one of claims 1 to 3, wherein the inference model outputs a predicted value of the amount of change in specified data at a specified facility based on specified data generated at a certain point in time at the specified facility.

7. An information processing device as described in any one of claims 1 to 3, comprising a reset unit that returns the inference model to its state before updating.

8. The inference model outputs a predicted value for a specified facility based on data measured at the facility, An information processing device described in any one of claims 1 to 3, comprising a training data generation unit that generates training data by corresponding actual measured values ​​measured at a specified time to input data input into the inference model to calculate a predicted value at the specified time as correct answer data.

9. An information processing device as described in any one of claims 1 to 3, wherein the learning unit performs the process of updating the inference model based on the data if the value of the acquired data is within a normal range represented by a reference value calculated from multiple data collected during a data collection period corresponding to the inference model, for each of multiple inference models each having a different data collection period.

10. A learning method executed by one or more information processing devices, a data acquisition step of sequentially acquiring data for machine learning of an inference model, which is sequentially generated; A learning method including a learning step of updating the inference model based on the acquired data if the acquired data satisfies predetermined conditions for determining the normality of the data.

11. 2. A learning program for causing a computer to function as the information processing device according to claim 1, the learning program causing a computer to function as the data acquisition unit and the learning unit.

Images