System for generating learning data

Abstract

System for generating learning data, which includes: a data integration part (100) configured to generate new data by filtering the plant data based on a warning condition in order to integrate the plant data (X) for a plant configuration into existing learning data (Z), wherein the data integration part (100) includes a data selector (110) wherein a user can select plant data for a desired period or parameter to generate a plant model by means of the data selector (110); and a training data generation part (200) configured to differentiate between a weight applied to the new data (Xs3) and a weight applied to the existing training data (Z) by comparing the number of new data (Xs3) and the number of existing training data (Z), and to generate new training data (Xm) by combining the new data (Xs3) with the existing training data (Z) to which the weight is applied, wherein the training data generation part (200) is configured to determine whether the number of existing training data (Z) is equal to, greater than, or less than a first target value, and to change the weight applied to the new data (Xs3) or the existing training data (Z) according to the number of new data (Xs3) and the number of existing training data (Z), wherein, in response to the determination that the number of existing training data (Z) is less than the first setpoint, the training data generation part (200) is configured to sample at least one of the existing training data (Z) and the new data (Xs3) such that the sum of the number of existing training data (Z) and the number of new data (Xs3) becomes the first setpoint, wherein, in response to the determination that the number of existing training data (Z) is equal to or greater than the first setpoint, the training data generation part (200) is configured to sample at least one of the existing training data (Z) and the new data (Xs3) by differentiating the weight applied to the existing training data (Z) and the new data (Xs3) respectively, based on a ratio between the existing training data (Z) and the new data (Xs3), wherein in response to the determination that the ratio is equal to or less than 0.25, the training data generation part (200) is configured to sample only the existing training data (Z) and to generate the new training data (Xm) by combining the sampled existing training data (Z) with the new data (Xs3), and wherein the system for generating the training data is configured to recommend the new training data (Xm) to the user.

Description

BACKGROUND area

[0001] Apparatus and methods consistent with exemplary embodiments relate to a system for generating learning data, and more particularly to a system for generating new learning data by differentiating a weight applied to existing learning data and new data, respectively. Description of the state of the art

[0002] Generally, large facilities such as power generation or chemical plants are operated by complexly interconnected various types of machinery and electrical equipment. These large facilities should always monitor for abnormal symptoms that lead to accidents to ensure reliable power supply. Therefore, a monitoring device is used to detect whether a key component of a facility is damaged in real time and to generate an alarm for an operator when an abnormal symptom is detected in the component.

[0003] This means that a failure of a process in the plant can cause undesirable performance due to damage to the plant. If the plant is destroyed, it can also cause injury or death to a person and also cause environmental problems. Therefore, an early warning system capable of detecting the failure early is essential.

[0004] An early warning system that provides early notification of a failure or the possibility of a failure must receive and store a real-time observation signal from the plant and, based on the above, distinguish anomalous signals to provide advance notification. Therefore, the most important component of a failure detection system may be a failure prediction model capable of predicting failure early. The failure prediction model can accurately predict the steady-state state of the plant and predict the occurrence of a failure when an observation model that differs from the steady-state state is received.

[0005] For the failure prediction model to work accurately, high-quality learning data is required. This means that steady-state learning can be performed using high-quality learning data, thereby increasing the accuracy of the failure prediction model. Therefore, ensuring high-quality learning data is essential for an accurate failure prediction model.

[0006] However, if an untrained person performs predictive modeling, the learning data may be generated incorrectly due to incorrect data preprocessing and other processing. Consequently, prediction accuracy is significantly reduced, thereby degrading the solution's performance.

[0007] US 9 454 733 B1 relates to a method for training a machine learning model based on a first data set taken from a complete data set.

[0008] The paper "RUSBoost: A hybrid approach to alleviating class imbalance" by Seiffert, Chris, et al., IEEE transactions on systems, man, and cybernetics—part A: systems and humans, 2009, Vol. 40, No. 1, pp. 185–197, discusses a boosting classifier optimized for datasets with class imbalance. SUMMARY

[0009] At least one of the needs described above is met by the system for generating learning data according to independent claim 1.

[0010] Aspects of one or more exemplary embodiments provide a system for generating learning data that extracts only the data suitable for generating learning data by an early warning condition even if a user selects arbitrary data, and generates new learning data with high reliability by adjusting a reflection ratio of new data and existing learning data.

[0011] Additional aspects will be set forth in part in the description which follows and in part will be obvious from the description, or may be learned by practice of the example embodiments.

[0012] According to one aspect of an exemplary embodiment, a system for generating learning data is provided, the system comprising: a data incorporation part configured to generate new data by filtering plant data based on a warning condition to incorporate the plant data for a configuration of a plant into existing learning data; and a learning data generation part configured to differentiate a weight applied to the new data or the existing learning data by comparing the number of new data and the number of existing learning data, and generate new learning data by combining the new data with the existing learning data to which the weight is applied, wherein the learning data generation part determines whether the number of existing learning data is equal to, greater than, or less than a first target value, and a weight applied to the new data or the existing learning data.the existing learning data is applied, changes according to the number of new data and the number of existing learning data.

[0013] According to one aspect of an exemplary embodiment, the learning data generation system can generate high-reliability learning data by determining a reflection ratio between the existing learning data and the new data and recommend it to the user. Furthermore, the learning data generation system can classify the data used in the prediction model and the data not used in the prediction model to perform the filtering process twice, thereby automatically filtering low-reliability data even if an untrained person erroneously selects data.

[0014] In some embodiments, in response to determining that the number of existing learning data is equal to or greater than the first target value, the

[0015] The learning data generating part is configured to sample at least one of the existing learning data and the new data by differentiating the weight applied to the existing learning data and the new data, respectively, based on a ratio between the existing learning data and the new learning data.

[0016] In some embodiments, in response to determining that the ratio is equal to or less than 0.25, the learning data generating part is configured to sample only the existing learning data and generate the new learning data by combining the sampled existing learning data with the new data.

[0017] In some embodiments, in response to determining that the ratio is greater than 0.25, the learning data generating part is configured to sample the existing learning data such that the number of existing learning data corresponds to a first setting value, and sample the new data such that the number of new data is a number obtained by subtracting the first setting value from the first target value, and generate the new learning data by combining the sampled existing learning data with the new data.

[0018] In some embodiments, when the number of existing learning data is equal to or greater than the first target value, the first setting value is the number of existing learning data included in the new learning data.

[0019] In some embodiments, in response to determining that the number of existing learning data is less than the first target value, the learning data generation part is configured to sample at least one of the existing learning data and the new learning data such that the sum of the number of existing learning data and the number of new data becomes the first target value.

[0020] In some embodiments, in response to determining that the number of existing learning data exceeds a first setting value, the learning data generation part is configured to differentiate the weight applied to the existing learning data and the new data, respectively, by determining whether the number of new data is equal to or greater than or less than a second target value.

[0021] In some embodiments, the first setting value is a number less than the first target value and a number greater than the second target value.

[0022] In some embodiments, in response to determining that the number of new data is greater than the second target value, the learning data generation part is configured to sample the existing learning data so that the number of existing learning data corresponds to the first setting value, and to sample the new data so that the number of new data is a number obtained by subtracting the first setting value from the first target value.

[0023] In some embodiments, in response to determining that the number of existing learning data is equal to or less than a first setting value and the number of new data is greater than a second target value, the learning data generation part is configured to sample the new data such that the number of new data is a number obtained by subtracting the number of existing learning data from the first target value.

[0024] In some embodiments, in response to determining that the number of existing learning data is less than the first target value and the sum of the number of existing learning data and the number of new learning data is less than the first target value, the learning data generation part is configured to generate the new learning data by combining the existing learning data and the new data.

[0025] In some embodiments, the learning data generating part is configured to sample the new data such that the number of new data is equal to or less than the first target value, and generate the new learning data using the sampled new data.

[0026] In some embodiments, the data incorporation portion comprises a data processor configured to filter the asset data based on a first warning condition; and a data incorporation determiner configured to generate the new data by filtering the filtered asset data based on a second warning condition.

[0027] In some embodiments, the data processor includes: a condition-based data processing section configured to filter the asset data based on a feature condition and the first warning condition; and an outlier processing section configured to remove the data that deviates from a normal distribution formed by the filtered asset data.

[0028] Preferably, the feature condition is defined on the basis of a parameter for a configuration of the system.

[0029] Preferably, the first warning condition is a condition for filtering data outside the allowable range of a measured value based on the measured value of the plant data.

[0030] In some embodiments, the data incorporation determiner includes a data type determination section configured to determine whether the asset data is data used in a predictive model; and an incorporation logic section configured to filter the asset data used in the predictive model based on the second warning condition.

[0031] Preferably, the prediction model is a model generated based on the existing learning data.

[0032] Preferably, the incorporation logic section defines a value obtained by subtracting a predicted value from a measured value as a remainder by defining the plant data as a measured value and defining the data obtained by the prediction model as a predicted value.

[0033] Preferably, the second warning condition is a condition to filter the data outside the allowable range of the remainder.

[0034] Preferably, the data incorporation part further comprises a learning data candidate group operator configured to generate a learning data candidate group, and wherein the learning data generation part generates the new learning data based on the learning data candidate group.

[0035] Preferably, the data incorporation part comprises a filter configured to filter the existing learning data based on the first warning condition to provide the filtered existing learning data to the learning data candidate group operator.

[0036] Preferably, the learning data generation part comprises an existing learning data presence determination device configured to determine whether the existing learning data is present; a comparator configured to compare whether the number of existing learning data is equal to, greater than, or less than the first target value; a calculator configured to calculate a ratio of the new data to the existing learning data and to calculate a weight to be applied to the new data and the existing learning data, respectively, based on the ratio; and a data sampling device configured to sample the new data and the existing learning data based on the ratio and the weight.

[0037] Preferably, the data sampling device samples such that the sum of the number of new data and the number of existing learning data becomes the first target value.

[0038] Preferably, when the existing learning data presence determining means determines that the existing learning data is not present, the data sampling means samples the new data so that the number of the new data becomes equal to or less than the first target value.

[0039] Preferably, the calculator calculates the weight applied to the new data as 1 in response to determining that the number of new data is equal to or less than a second target value, wherein the data sampling means uses all of the new data to generate the new learning data without sampling the new data.

[0040] Preferably, the second setpoint is a number that is smaller than the first setpoint.

[0041] Preferably, a maximum value of the weight applied to the new data and the existing learning data is 1. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The above and other aspects will become more apparent from the following description of the exemplary embodiments with reference to the accompanying drawings, in which: Fig. 1 is a block diagram illustrating a learning system having a system for generating learning data according to an exemplary embodiment; Fig. 2 is a block diagram illustrating a data integration portion according to an exemplary embodiment; Fig. 3 is a block diagram showing a conditional data processing section of Fig. 2 represents; Fig. 4 is a block diagram illustrating a learning data generation part according to an exemplary embodiment; and Fig. 5 is a diagram for explaining a procedure for generating learning data by the learning data generating part of Fig. 4 is. DETAILED DESCRIPTION OF EXEMPLARY EMBODIMENTS

[0043] The inventive concept will be described in detail below with reference to the accompanying drawings, in which exemplary embodiments thereof are shown. However, the inventive concept should not be limited to the embodiments set forth below, but may be implemented in many different forms. Rather, the exemplary embodiments are provided only to complete the description of the present disclosure and to fully convey the scope of the inventive concept to those skilled in the art to which the present disclosure pertains. The same reference numerals designate the same elements throughout the specification.

[0044] Fig. 1 is a block diagram illustrating a learning system having a system for generating learning data according to an example embodiment.

[0045] With reference to Fig. 1, a learning system may comprise a system for generating learning data 10 and a modeling part 20.

[0046] The learning data generation system 10 may include a data integration part 100 and a learning data generation part 200. The learning data generation system 10 may generate data for generating a plant prediction model and generate new learning data Xm based on measured data for various parameters of a plant configuration (e.g., the amount of power of the plant, the temperature of a plant configuration, and the like). The learning data generation system 10 may generate the new learning data Xm with high reliability by appropriately determining a reflection ratio between existing learning data and new data and recommend it to a user.The system for generating the learning data 10 can be designed based on a reflection relationship between the existing learning data and the new learning data and an algorithm for different scenarios, and can generate the learning data with high reliability even if the user is an unskilled person in the task of generating the learning data.

[0047] The data integration part 100 may filter plant data to integrate the plant data X for a configuration of the plant into the existing learning data. The data integration part 100 may filter, for example, anomalous data, no / zero data, low-reliability data, and the like present in the plant data X. The anomalous data may be corrupted data, the no / zero data may be data with a value of 0, and the low-reliability data may be a pattern that deviates from a common pattern, i.e., a pattern in which most of the data is distributed. The data integration part 100 may generate a learning data candidate group Xtp by combining the existing learning data with the filtered plant data. The learning data candidate group Xtp may be transferred to the learning data generation part 200.

[0048] The learning data generation part 200 can generate the new learning data Xm by sampling the learning data candidate group Xtp. The learning data generation part 200 can differentiate a weight applied to the filtered plant data or the existing learning data and generate the new learning data Xm. The plant data X through which the filtering was completed can be new data, which will be described later.

[0049] The modeling part 20 can generate a plant prediction model by modeling the generated new learning data Xm. The modeling part 20 can update an existing plant prediction model by receiving the new learning data Xm at a predetermined cycle and performing plant modeling. Furthermore, the modeling part 20 can generate a new plant prediction model based on the new learning data Xm received from the system to generate the learning data 10.

[0050] Fig. 2 is a block diagram illustrating the data integration part 100 according to an exemplary embodiment, and Fig. 3 is a block diagram showing a conditional data processing section of Fig. 2 represents.

[0051] With reference to Fig. 1 to 3, the data integration part 100 may include a data selector 110, a data processor 120, a data integration determiner 130, a learning data candidate group operator 140, and a filter 150.

[0052] The data selector 110 can select data measured for a specific period from the plant data X. For example, the data selector 110 can select the plant data X from January to March within a year. Furthermore, the data selector 110 can select some of the parameters that should be considered primarily to generate the plant model of the plant. For example, the data selector 110 can select some of the parameters such as a minimum output, a maximum output, a turbine speed, a temperature, a humidity, and a pressure. That is, the learning data generation system 10 can generate learning data based on the data selected by the data selector 110. The plant data X selected by the data selector 110 is defined as the first plant data Xs1.The data selector 110 can be controlled by the user and the user can select the data for a desired period or parameter.

[0053] The data processor 120 may include a condition-based data processing section 121 and an outlier processing section 123. The data processor 120 may filter the first asset data Xs1, and the filtered first asset data Xs1 is defined as the first sub-asset data Xs11. The data processor 120 may remove the no / zero data included in the asset data X.

[0054] The condition-based data processing section 121 may include a feature condition-based processor 121a and a first warning condition-based processor 121b. The plant data X from which the no / zero data has been removed may be filtered by the feature condition-based processor 121a and the first warning condition-based processor 121b.

[0055] The feature condition-based processor 121a may filter the first plant data Xs1 based on a parameter indicating a feature of the plant. For example, the feature condition may be data equal to or less than a specific output value of a plant configuration, and the feature condition-based processor 121a may filter data indicating a case where the output value is equal to or greater than a specific reference. That is, the feature condition-based processor 121a may use only the data that satisfy the feature condition to generate learning data. The first warning condition-based processor 121b may filter the first plant data Xs1 using a first warning condition.Specifically, based on the first warning condition, the processor 121b may filter the data outside the allowable range of an actually measured value based on the actually measured value of the first system data Xs1. That is, the first warning condition may be a condition for filtering the data outside the allowable range of the actually measured value based on the actually measured value of the first system data Xs1. The actually measured value may be measured data for a parameter (e.g., the amount of power of the system, the temperature of a configuration of the system, or the like) for a configuration of the system, and the allowable range of the actually measured value may be the data with a value between an upper limit of the actually measured value and a lower limit of the actually measured value.Based on the first warning condition, the processor 121b may issue a warning if the first system data Xs1 exceeds the allowable range permitted by the first warning condition. The data after the warning is issued is determined to be of low reliability and cannot be used to generate the learning data.

[0056] The outlier processing section 123 can filter out highly problematic data to be used to generate the learning data. The outlier processing section 123 can filter out data that has not been conditionally filtered by the data processing section 121 but deviates from the normal distribution formed by most of the data. That is, the outlier processing section 123 can remove an outlier, which is data that is normal data but deviates from the general data distribution. The first sub-plant data Xs11 from which the outlier was removed is defined as the second plant data Xs2.

[0057] The data incorporation determiner 130 may include a data type determination section 131 and an incorporation logic section 133. The data incorporation determiner 130 may determine whether the second asset data Xs2 is used by the prediction model, and filter the second asset data Xs2 based on a second warning condition when the second asset data Xs2 is used by the prediction model.

[0058] The data type determination section 131 may determine whether the second plant data Xs2 is used by the prediction model. The data used by the prediction model among the second plant data Xs2 is defined as prediction model data Xp, and the data not used by the prediction model among the second plant data Xs2 is defined as non-prediction model data Xnp. The prediction model data Xp may be the data included in the existing learning data, as data with prediction information. Here, the prediction model may be a model generated based on the existing learning data or a model previously constructed by a designer.For example, if the data selector 110 selects data from January to June among 12 months and uses data from April to June when generating the forecast model, the data from January to March is the non-forecast model data Xnp and the data from April to June is the forecast model data Xp.

[0059] The incorporation logic section 133 may filter the prediction model data Xp based on the second warning condition. The incorporation logic section 133 may define the remainder obtained by subtracting a predicted value from an actually measured value, defining the prediction model data Xp as the actually measured value, and defining the data obtained by the prediction model as the predicted value. The incorporation logic section 133 may filter the data outside the allowable range of the remainder. The allowable range of the remainder may be the data with a value between an upper limit of the remainder and a lower limit of the remainder. The incorporation logic section 133 may issue a warning when the prediction model data Xp exceeds the allowable range permitted by the second warning condition.Generally, the warning condition can change over time, and the second warning condition can be the current warning condition. For example, the warning conditions for January and June can be different from each other. If the forecast model data Xp is the data from January to June, and it is currently July, the second warning condition can be a warning condition set based on July. That is, the second warning condition can be a condition that filters the data outside the allowable range of the residual based on the residual of the forecast model data Xp, reflecting the current warning condition. The current warning condition can be changed by the designer and the current time.

[0060] The prediction model data Xp filtered by the integration logic section 133 and the non-prediction model data Xnp may be combined to generate new data Xs3. The new data Xs3 may be data used to generate the new learning data Xm as data through which the filtering has been completed.

[0061] The learning data candidate group operator 140 can generate the learning data candidate group Xtp for generating the new learning data Xm by combining the existing learning data Z with the new data Xs3. At this time, some data of the existing learning data Z is filtered by the filter 150 to be transmitted to the learning data candidate group operator 140. The filter 150 can filter the existing learning data Z outside the allowable range of the actually measured value of the new data Xs3. The filter 150 can filter the existing learning data Z to adjust the first warning condition of the new data Xs3 to combine the existing learning data Z with the new data Xs3.For example, since some data of the existing learning data Z may have data deviating from the general distribution of the new data Xs3, only the existing learning data Z with the distribution similar to the new data Xs3 obtained by the procedure for filtering the existing learning data Z can be used as the learning data candidate group Xtp. Ideally, the filter 150 may filter so that the number of existing learning data Z becomes 2000. However, if the number of existing learning data Z is less than 2000, the number of existing learning data Z included in the learning data candidate group Xtp may be less than 2000. That is, the number of existing learning data Z included in the learning data candidate group Xtp may also be 2000.

[0062] The data integration part 100 according to an exemplary embodiment can filter the plant data X using the first warning condition based on the actual measured value and the second warning condition based on the residual. The data integration part 100 can filter the plant data X based on the allowable range of the actual measured value (i.e., the first warning condition) and filter only the prediction model data Xp used in the prediction model of the allowable range of the residual (i.e., the second warning condition). Therefore, the data integration part 100 can distinguish between the data used by the prediction model and the data not used by the prediction model, even if the user selects the plant data X measured in a specific period.The data incorporation part 100 can refilter the prediction model data Xp using the predicted value of the data used by the prediction model. Therefore, the new data Xs3 can be generated with high reliability. That is, the data incorporation part 100 can be a universal warning algorithm.

[0063] Fig. 4 is a block diagram illustrating a learning data generation part according to an exemplary embodiment.

[0064] With reference to Fig. 4, the learning data generation part 200 may include a determination device 210 for the presence of existing learning data, a comparator 230, a calculator 250, and a data sampling device 270. Although the learning data generation part 200 is described by dividing it into different configurations for convenience of explanation, the learning data generation part 200 may be understood as an algorithm for generating the learning data.

[0065] The existing learning data presence determination unit 210 may determine whether the existing learning data Z exists. If the existing learning data Z exists, the learning data generation unit 200 may combine the existing learning data Z with the new data Xs3 by differentiating a weight applied to the existing learning data Z and the new data Xs3, respectively, based on a ratio between the existing learning data Z and the new data Xs3. If the existing learning data Z does not exist, the learning data generation unit 200 may generate the new learning data Xm using only the new data Xs3.

[0066] The comparator 230 can compare whether the number of existing learning data Z is equal to, greater than, or less than a first target value. The first target value can be the number of ideal learning data determined by the designer. Therefore, it may be preferable that the number of existing learning data Z satisfies the first target value and the number of generated new learning data Xm satisfies the first target value. The first target value can be, for example, 2000.

[0067] The calculator 250 can calculate a ratio S of the new data Xs3 to the existing learning data Z and calculate a weight applied to the new data Xs3 and the existing learning data Z, respectively, based on the ratio S. The calculator 250 can calculate a size (the number of data) of all the new data Xs3 and the existing learning data Z to calculate the ratio S and the weight. The ratio S can be S=the size of Xs3the size of Z The weights applied to the new data Xs3 and the existing learning data Z may be different from each other according to the ratio S. The calculator 250 may determine the weights applied to the new data Xs3 and the existing learning data Z using the size of all of the new data Xs3 and the existing learning data Z and the ratio S. However, if the number of new data Xs3 is equal to or less than a second target value, the weight applied to the new data Xs3 may be calculated as 1. The second target value may be, for example, 500. In an exemplary embodiment, a maximum value of the weight applied to the data is defined as 1. That is, if the number of new data Xs3 is equal to or less than the second target value, the calculator 250 calculates the weight applied to the new data Xs3 as 1 regardless of the ratio S.

[0068] The data sampler 270 can sample the new data Xs3 and the existing learning data Z based on the ratio S and the weight. Ideally, the data sampler 270 can sample the new data Xs3 and the existing learning data Z such that the number of existing learning data Z applied to the new learning data Xm is greater than the number of new data Xs3 applied to the new learning data Xm. However, if the sum of the number of new data Xs3 and the number of existing learning data Z is equal to or less than the first target value, the data sampler 270 cannot perform the sampling procedure.

[0069] For example, if the number of existing learning data Z is 2000, the new data Xs3 and the existing learning data Z can be sampled based on whether the ratio S is equal to, greater than, or less than 0.25. If the ratio S is 0.25, the data sampler 270 can sample the existing learning data Z so that the number of existing learning data Z becomes 1500, and cannot sample the new data Xs3. As a result of the sampling, the number of existing learning data Z becomes 1500 and the number of new data Xs3 becomes 500. That is, the data sampler 270 can sample so that the sum of the number of existing learning data Z and the number of new data Xs3 becomes a number equal to the first target value. At this time, the calculator 250 can calculate the weight applied to the new data Xs3 as 1. The data sampling device 270 can reflect all new data Xs3 and sample only the existing learning data Z.

[0070] In another example, when the number of existing learning data Z is less than 2000, the new data Xs3 and the existing learning data Z may be sampled so that the ratio S satisfies 0.25. When the number of existing learning data Z is 1600 and the number of new data Xs3 is 600, the data sampler 270 may sample the existing learning data Z so that the number of existing learning data Z becomes 1500 and sample the new data Xs3 so that the number of new data Xs3 becomes 500. However, when the number of new data Xs3 is less than the second target value, the data sampler 270 may only sample the existing learning data Z without sampling the new data Xs3.

[0071] In yet another example, when the number of existing learning data Z is less than the first setting value, the data sampler 270 may perform the sampling process based on whether the number of new data Xs3 is equal to, greater than, or less than the second target value, without considering the ratio S. For example, the first setting value may be 1500. When the number of existing learning data Z is the first target value, the first setting value is defined as a value when the ratio of the first target value to the existing learning data Z is 3 / 4. When the number of existing learning data Z is 1000 and the number of new data Xs3 is 600, the data sampler 270 may not sample the existing learning data Z and the new data Xs3. This is because the sum of the number of existing learning data Z and the number of new data Xs3 is less than the first target value.If the number of existing learning data Z is 1400 and the number of new data Xs3 is 800, the calculator 250 may calculate the weight applied to the existing learning data Z as 1. The data sampling device 270 may sample the new data Xs3 so that the number of new data Xs3 becomes 600 without sampling the existing learning data Z.

[0072] There may be different scenarios in the procedure of calculating the weight by the computer 250 and the procedure of scanning by the data scanning device 270. These are described in Fig. 5 described in detail.

[0073] Fig. 5 is a diagram for explaining a procedure for generating learning data by the learning data generating part 200 of Fig. 4. In an exemplary embodiment according to Fig. 5, the first setpoint may be 2000, the second setpoint may be 500, and the first setting value may be 1500. In the exemplary embodiment of Fig. 5 the size of the data can be the number of data.

[0074] With reference to Fig. 4 and Fig.5, the learning data generation part 200 may differentiate the weight applied to the new data Xs3 or the existing learning data Z by comparing the number of new data Xs3 with the number of existing learning data Z, and generate the new learning data Xm by combining the new data Xs3 with the existing learning data Z to which the weights have been applied. The learning data generation part 200 may determine whether the number of existing learning data Z is equal to, greater than, or less than the first target value, and change the weight applied to the new data Xs3 or the existing learning data Z according to the number of new data Xs3 and the number of existing learning data Z.

[0075] The existing learning data presence determining means 210 receiving the learning data candidate group Xtp may determine whether the existing learning data Z included in the learning data candidate group Xtp is present (operation S100).

[0076] Comparator 230 may determine whether the number of existing learning data Z is equal to the first target value (operation S210). The first target value may be the number of ideal learning data determined by the designer.

[0077] If the number of existing learning data Z is equal to the first target value, the calculator 250 can calculate the number of new data Xs3 and the number of existing learning data Z, and calculate the ratio S between the existing learning data Z and the new data Xs3 (operation S230). The calculator 250 can differentiate the weight applied to the new data Xs3 and the existing learning data Z according to the ratio S (operation S250).

[0078] If the ratio S is greater than 0.25, the calculator 250 may calculate a weight such that the number of existing learning data Z becomes the first setting value and the number of new data Xs3 becomes the second target value. The data sampler 270 may sample the existing learning data Z such that the number of existing learning data Z becomes the first setting value and sample the new data Xs3 such that the number of new data Xs3 becomes the second target value. The data sampler 270 may generate the new learning data Xm by combining the sampled existing learning data Z with the sampled new data Xs3. If the number of existing learning data Z is equal to or greater than the first target value and the ratio S is greater than 0.25, the first setting value may be equal to the number of existing learning data Z included in the new learning data Xm (operation S270).

[0079] If the ratio S is equal to or less than 0.25, the calculator 250 may define the weight applied to the new data Xs3 as 1. That is, all the new data Xs3 may be applied to generate the new learning data Xm. In other words, the data sampler 270 may sample only the existing learning data Z without sampling the new data Xs3, and generate the new learning data Xm by combining the sampled existing learning data Z with the new learning data Xs3. If the number of existing learning data Z is equal to or greater than the first target value and the ratio S is equal to or less than 0.25, the first setting value may be less than the number of existing learning data Z included in the new learning data Xm (operation S275).

[0080] If the number of existing learning data Z is less than the first target value, the calculator 250 may calculate whether the ratio S of the first target value to the existing learning data Z is 3 / 4. That is, the calculator 250 may calculate whether the number of existing learning data Z is greater than the first setting value (operation S310).

[0081] If the number of existing learning data Z is greater than the first setting value, the calculator 250 may differentiate the weight applied to the existing learning data Z and the new data Xs3 by determining whether the number of new data Xs3 is greater than the second target value. The first setting value is defined as the number less than the first target value and the number greater than the second target value (operation S330). If the number of new data Xs3 is greater than the second target value, the data sampler 270 may sample the existing learning data Z so that the number of existing learning data Z corresponds to the first setting value and sample the new data Xs3 so that the number of new data Xs3 is a number obtained by subtracting the first setting value from the first target value (operation S350).If the number of new data Xs3 is equal to or less than the second target value, the computer 250 may determine whether the sum of the number of existing learning data Z and the number of new data Xs3 is greater than the first target value (operation S360). If the sum of the number of existing learning data Z and the number of new data Xs3 is greater than the first target value, the data sampler 270 may sample only the existing learning data Z so that the sum of the number of existing learning data Z and the number of new data Xs3 becomes the first target value (operation S370). If the sum of the number of existing learning data Z and the number of new data Xs3 is equal to or less than the first target value, the data sampler 270 may not perform the sampling process.That is, when the sum of the number of existing learning data Z and the number of new data Xs3 is equal to or less than the first target value, the new learning data Xm can be generated by combining the existing learning data Z with the new data Xs3 without any sampling procedure.

[0082] If the number of existing learning data Z is less than the first target value and the ratio S of the first target value to the existing learning data Z is equal to or less than 3 / 4, the calculator 250 may determine whether the number of new data Xs3 is greater than the second target value (operation S410). If the number of new data Xs3 is equal to or less than the second target value, the data sampler 270 may not perform the sampling process. That is, the new learning data Xm may be generated by combining the existing learning data Z with the new data Xs3 without the sampling process. If the number of new data Xs3 is greater than the second target value, the calculator 250 may determine whether the sum of the number of existing learning data Z and the number of new data Xs3 is greater than the first target value (operation S430).If the sum of the number of existing learning data Z and the number of new data Xs3 is greater than the first set value, the data sampler 270 can only sample the new data Xs3. In this case, since the number of existing learning data Z is less than the first set value, all of the existing learning data Z can be applied to the new learning data Xm. That is, the weight applied to the existing learning data Z can be greater than the weight applied to the new data Xs3. Specifically, the weight applied to the existing learning data Z can be 1 (operation S450). If the sum of the number of existing learning data Z and the number of new data Xs3 is equal to or less than the first set value, the data sampler 270 cannot perform the sampling process. That is, the new learning data Xm can be generated by combining the existing learning data Z with the new data Xs3 without the sampling process.

[0083] If the existing learning data Z is not used to generate the new learning data Xm, or if the existing learning data Z is not present, the new learning data Xm can be generated using only the new data Xs3. The calculator 250 can determine whether the number of new data Xs3 is greater than the first target value (operation S510). If the number of new data Xs3 is greater than the first target value, the data sampler 270 can filter the new data Xs3 so that the number of new data Xs3 is equal to the first target value (operation S530). If the number of new data Xs3 is equal to or less than the first target value, the data sampler 270 cannot perform the sampling procedure.

[0084] Although not limited thereto, an exemplary embodiment may be embodied as computer-readable code on a computer-readable recording medium. The computer-readable recording medium is any data storage device capable of storing data that can subsequently be read by a computer system. Examples of the computer-readable recording medium include read-only memory (ROM), random access memory (RAM), CD-ROMs, magnetic tapes, floppy disks, and optical data storage devices. The computer-readable recording medium may also be distributed across computer systems coupled to a network so that the computer-readable code is stored and executed in a distributed manner. An exemplary embodiment may also be written as a computer program that can be transmitted over a computer-readable transmission medium, such as a hard disk.a carrier wave is transmitted and received and implemented in general-purpose or special-purpose digital computers that execute the programs. Furthermore, it should be understood that in exemplary embodiments, one or more units of the devices and apparatus described above may include or be implemented by circuitry, a processor, a microprocessor, etc., and may execute a computer program stored on a computer-readable medium.

Claims

[1] System for generating learning data, comprising: a data integration part (100) configured to generate new data by filtering the plant data based on a warning condition to integrate the plant data (X) for a configuration of a plant into existing learning data (Z), wherein the data integration part (100) comprises a data selector (110), wherein a user can select plant data for a desired period or a desired parameter through the data selector (110) to generate a plant model; and a learning data generation part (200) configured to differentiate a weight applied to the new data (Xs3) and a weight applied to the existing learning data (Z) by comparing the number of new data (Xs3) and the number of existing learning data (Z), and to generate new learning data (Xm) by combining the new data (Xs3) with the existing learning data (Z) to which the weight is applied, wherein the learning data generation part (200) is configured to determine whether the number of existing learning data (Z) is equal to or greater than or less than a first target value, and to change the weight applied to the new data (Xs3) or the existing learning data (Z) according to the number of new data (Xs3) and the number of existing learning data (Z), wherein, in response to determining that the number of existing learning data (Z) is less than the first target value, the learning data generation part (200) is configured to sample at least one of the existing learning data (Z) and the new data (Xs3) such that the sum of the number of existing learning data (Z) and the number of new data (Xs3) becomes the first target value, wherein, in response to determining that the number of existing learning data (Z) is equal to or greater than the first target value, the learning data generation part (200) is configured to sample at least one of the existing learning data (Z) and the new data (Xs3) by differentiating the weight applied to the existing learning data (Z) and the new data (Xs3), respectively, based on a ratio between the existing learning data (Z) and the new data (Xs3), wherein, in response to determining that the ratio is equal to or less than 0.25, the learning data generating part (200) is configured to sample only the existing learning data (Z) and generate the new learning data (Xm) by combining the sampled existing learning data (Z) with the new data (Xs3), and wherein the system for generating the learning data is configured to recommend the new learning data (Xm) to the user. [2] System for generating the learning data according to claim 1, wherein, in response to the determination that the ratio is greater than 0.25, the learning data generation part (200) is configured to sample the existing learning data (Z) such that the number of existing learning data (Z) corresponds to a first setting value, and to sample the new data (Xs3) such that the number of new data (Xs3) is a number obtained by subtracting the first setting value from the first target value, and to generate the new learning data (Xm) by combining the sampled existing learning data (Z) and the sampled new data (Xs3), and / or wherein, when the number of existing learning data (Z) is equal to or greater than the first target value, the first setting value is the number of existing learning data (Z) included in the new learning data (Xm). [3] System for generating the learning data according to claim 1, wherein, in response to the determination that the number of existing learning data (Z) exceeds a first setting value, the learning data generation part (200) is configured to differentiate the weight applied to the existing learning data (Z) and the new data (Xs3) by determining whether the number of new data (Xs3) is equal to or greater than or less than a second target value, and wherein the first setting value is a number less than the first set value and a number greater than the second set value. [4] The system for generating the learning data according to claim 3, wherein, in response to determining that the number of new data (Xs3) is greater than the second target value, the learning data generating part (200) is configured to sample the existing learning data (Z) so that the number of existing learning data (Z) corresponds to the first setting value, and to sample the new data (Xs3) so that the number of new data (Xs3) is a number obtained by subtracting the first setting value from the first target value. [5] The system for generating the learning data according to claim 1, wherein, in response to determining that the number of existing learning data (Z) is equal to or less than a first setting value and the number of new data (Xs3) is greater than a second target value, the learning data generating part (200) is configured to sample the new data (Xs3) such that the number of new data (Xs3) is a number obtained by subtracting the number of existing learning data (Z) from the first target value. [6] The system for generating the learning data according to any one of the preceding claims, wherein, in response to determining that the number of existing learning data (Z) is less than the first target value and the sum of the number of existing learning data (Z) and the number of new data (Xs3) is less than the first target value, the learning data generating part (200) is configured to generate the new learning data (Xm) by combining the existing learning data (Z) and the new learning data (Xs3). [7] The system for generating the learning data according to any one of the preceding claims, wherein the learning data generating part (200) is configured to sample the new data (Xs3) so that the number of the new data (Xs3) becomes equal to or less than the first target value, and to generate the new learning data (Xm) using the sampled new data (Xs3). [8] A system for generating the learning data according to any one of the preceding claims, wherein the data integration part (100) comprises: a data processor (120) configured to filter the system data based on a first warning condition; and a data incorporation determination device (130) configured to generate the new data (Xs3) by filtering the filtered plant data based on a second warning condition. [9] A system for generating the learning data according to claim 8, wherein the data processor (120) comprises: a condition-based data processing section (121) configured to filter the plant data based on a feature condition and the first warning condition; and an outlier processing section (123) configured to remove the data that deviates from a normal distribution formed by the filtered plant data, wherein the feature condition is determined on the basis of a parameter for a configuration of the system, and / or wherein the first warning condition is a condition for filtering data outside the allowable range of a measured value based on the measured value of the plant data. [10] A system for generating the learning data according to claim 8 or 9, wherein the data integration determining means (130) comprises: a data type determination section (131) configured to determine whether the plant data is data used in a prediction model; and an integration logic section (133) configured to filter the plant data used in the predictive model based on the second warning condition, and where the prediction model is a model generated based on the existing learning data. [11] The system for generating the learning data according to claim 10, wherein the incorporation logic section (133) is configured to define a value obtained by subtracting a predicted value from a measured value by defining the plant data as the measured value and defining the data obtained by the prediction model as the predicted value as the remainder, and / or wherein the second warning condition is a condition for filtering the data outside the allowable range of the remainder. [12] A system for generating the learning data according to any one of the preceding claims, wherein the data integration part (100) further comprises: a learning data candidate group operator (140) configured to generate a learning data candidate group, and wherein the learning data generating part (200) is configured to generate the new learning data (Xm) based on the learning data candidate group. [13] The system for generating the learning data according to claim 12, wherein the data integration part (100) further comprises: a filter (150) configured to filter the existing learning data (Z) based on the first warning condition to provide the filtered existing learning data (Z) to the learning data candidate group operator (140). [14] A system for generating the learning data according to any one of the preceding claims, wherein the learning data generating part (200) comprises: an existing learning data presence determination device (210) configured to determine whether the existing learning data is present; a comparator (230) configured to compare whether the number of existing learning data (Z) is equal to or greater than or less than the first target value; a calculator (250) configured to calculate a ratio of the new data (Xs3) to the existing learning data and to calculate a weight applied to the new data (Xs3) or the existing learning data based on the ratio; and a data sampling device (270) configured to sample the new data (Xs3) and the existing learning data based on the ratio and the weight. [15] The system for generating the learning data according to claim 14, wherein the data sampling means (270) is configured to sample such that the sum of the number of new data (Xs3) and the number of existing learning data (Z) becomes the first target value. [16] The system for generating the learning data according to claim 14 or 15, wherein, when the existing learning data presence determination means (210) determines that the existing learning data (Z) is not present, the data sampling means (270) is configured to sample the new data (Xs3) so that the number of the new data (Xs3) becomes equal to or less than the first target value. [17] The system for generating the learning data according to claim 14, 15 or 16, wherein the calculator (250) is configured to calculate the weight applied to the new data (Xs3) as 1 in response to determining that the number of new data (Xs3) is equal to or less than a second target value, wherein the data sampling device (270) is configured to use all new data (Xs3) to generate the new learning data without sampling the new data (Xs3), where the second setpoint is a number that is smaller than the first setpoint, and where a maximum value of the weight applied to the new data (Xs3) and the existing learning data is 1.

Images