JP7219430B2 - LEARNING PROCESSING APPARATUS, METHOD AND PROGRAM - Google Patents

Description

The present invention relates to a learning processing device, method and program.

Recently, techniques have been proposed for predicting the degree of deterioration or the degree of freshness of a device or a component that constitutes the device (hereinafter referred to as a "prediction target") using machine learning. During this prediction, a learner is used that receives as input a set of state values indicating the state of the object to be predicted (hereinafter referred to as a "state value set") and outputs a label value indicating the degree of deterioration or freshness. .

Patent Literature 1 discloses a deterioration diagnosis system that acquires a series of measured data from normal to deteriorated equipment through a deterioration acceleration experiment and creates learning data from the series of measured data.

International Publication No. 2018/159169 pamphlet

By the way, in order to obtain a learner with high prediction accuracy, it is desirable to sequentially collect state value sets through a network from operating devices installed in various places and accumulate more learning data. However, in actual operation, priority is given to system maintenance, and equipment is often replaced or removed before a failure occurs. For example, a manufacturer or a maintenance company may provide a service life (that is, a set life span) shorter than the actual life span, and replace the device with a new one after the set life span has passed.

As a result, it is difficult to calculate the correct label value because the actual degree of deterioration or freshness cannot be identified in the case of a device that has finished its operation without completing its actual life. On the other hand, it is conceivable to generate training data by extracting only the state value set of the equipment that has completed the actual life, but the amount of training data that can be collected will be reduced by that amount, and the decrease in prediction accuracy due to insufficient learning will be avoided. I invite you. The same applies to the case where the device is replaced or removed for some reason before the set life has passed.

The present invention has been made in view of such a problem, and its object is to provide a method for identifying label value inaccuracies even when using a state value set obtained from a prediction target that has not reached its actual life or set life. An object of the present invention is to provide a learning processing device, method, and program capable of suppressing deterioration in prediction accuracy caused by determinism.

A learning processing device according to a first aspect of the present invention includes a state acquisition unit that acquires a state value set indicating a state of a prediction target that is a device or a component that constitutes the device; A population of learning data consisting of a label calculation unit that assigns or corrects a label value that indicates , the state value set acquired by the state acquisition unit, and the label value assigned or corrected by the label calculation unit and a learning processing unit that performs learning processing on a learning device that inputs the state value set and outputs the label value using the learning data set that is the same prediction target. A plurality of label values respectively corresponding to the plurality of state value sets are assigned or corrected collectively with respect to the plurality of state value sets obtained from the object in time series.

In the learning processing device according to the second aspect of the present invention, the label value is a value obtained by normalizing the degree of deterioration, and the label calculation unit applies a correction amount that increases as the label value increases to the original label value. The plurality of label values are collectively corrected by adding them respectively.

In the learning processing device according to the third aspect of the present invention, the correction amount is a correction coefficient obtained by dividing a learning error related to learning data, which is the maximum value among the plurality of label values, by the maximum value. It is the value multiplied by the label value.

The learning processing device according to the fourth aspect of the present invention uses a learning device subjected to learning processing by the learning processing unit to exclude learning data included in the learning data set or correct the label value of the learning data. It further includes a data determination unit that determines whether or not.

In the learning processing device according to the fifth aspect of the present invention, the learning processing section repeats re-learning or additional learning for the learning device using the learning data set whose label values are corrected by the label calculating section.

In the learning processing method according to the sixth aspect of the present invention, one or more computers acquire a state value set indicating the state of a prediction target that is a device or a component that constitutes the device; Learning data consisting of a calculating step of assigning or correcting a label value indicating the degree of deterioration or freshness of the data, the state value set obtained in the obtaining step, and the label value assigned or corrected in the calculating step a learning step of performing a learning process on a learner that inputs the state value set and outputs the label value, using a learning data set that is a population of A plurality of label values respectively corresponding to the plurality of state value sets are collectively assigned or corrected for the plurality of state value sets obtained from the prediction object in time series.

In the learning processing program according to the seventh aspect of the present invention, one or more computers are configured to acquire a state value set indicating the state of a prediction target that is a device or a component that constitutes the device; label calculation means for assigning or correcting a label value indicating the degree of deterioration or freshness of the label; the state value set acquired by the state acquisition means; and the label value assigned or corrected by the label calculation means. Using a learning data set, which is a population of learning data, to function as learning processing means for performing learning processing on a learning device that inputs the state value set and outputs the label value, and the same prediction object A plurality of label values respectively corresponding to the plurality of state value sets obtained in time series from is assigned or corrected collectively.

According to the present invention, even when using a state value set acquired from a prediction target that has not reached its actual life or set life, it is possible to suppress deterioration in prediction accuracy due to uncertainty in label values. can.

1 is an overall configuration diagram of a device diagnosis system incorporating a learning processing device according to an embodiment of the present invention; FIG. 2 is a functional block diagram of the processor shown in FIG. 1; FIG. 4 is a flow chart relating to the operation of the learning processing device; FIG. 4 is a diagram showing an example of a data structure of IoT data; FIG. 10 is a diagram illustrating an example of a method of assigning label values corresponding to multiple sets of state values; 3 is a diagram showing an example of a network structure of the learning device of FIG. 2; FIG. It is a figure which shows an example of the determination method of learning data. FIG. 10 is a diagram illustrating an example of a method of correcting label values corresponding to multiple sets of state values; FIG. 11 is a first graph showing the effect of improving prediction accuracy by collectively correcting label values; FIG. FIG. 11 is a second graph showing the effect of improving prediction accuracy by batch correction of label values; FIG.

Embodiments of the present invention will be described below with reference to the accompanying drawings. In order to facilitate understanding of the description, the same constituent elements in each drawing are denoted by the same reference numerals as much as possible, and overlapping descriptions are omitted.

[Configuration of Device Diagnosis System 10]
FIG. 1 is an overall configuration diagram of a device diagnosis system 10 incorporating a learning processing device 30 according to one embodiment of the present invention. The device diagnosis system 10 is configured to be able to provide a device diagnosis service for estimating the device age of a prediction target and diagnosing the degree of deterioration or freshness of the prediction target. The "prediction target" is the device 20 or the part 22 that constitutes the device 20, and its use may be for business or home use. “Equipment age” is an index that quantitatively indicates the state of the equipment 20 from the start of operation to the time of failure, and the maximum value corresponds to the actual life of the equipment 20 .

This equipment diagnosis system 10 specifically includes one or more equipment 20 , a learning processing device 30 , and an age prediction device 40 . The equipment 20, the learning processing device 30 and the age prediction device 40 can communicate with each other through the network NT.

The device 20 realizes various functions by causing a plurality of parts 22 to work together to perform a series of operations. A sensor 24 that detects a physical quantity indicating the state of the device 20 or the component 22 is built in or externally attached to the device 20 . The device 20 transmits information generated by itself to the outside as IoT (Internet Of Things) data D1. Thereby, the learning processing device 30 or the age prediction device 40 can collect the IoT data D1 from each device 20 .

The learning processing device 30 is a computer that executes learning processing using the IoT data D1 collected from each device 20 and generates model data D2 that can specify the calculation rule of the learning device LN (FIGS. 2 and 6). be. This learning processing device 30 includes a communication I/F 32 , a processor 34 , a memory 36 and a storage device 38 .

The communication I/F 32 is an interface that transmits and receives electrical signals to and from an external device. Thereby, the learning processing device 30 can receive the IoT data D1 from the device 20 and can transmit the model data D2 generated by itself to the age prediction device 40 . The processor 34 may be a general-purpose processor including a CPU (Central Processing Unit), or a dedicated processor including an FPGA (Field Programmable Gate Array) and a GPU (Graphics Processing Unit).

The memory 36 is a non-transitory storage medium and stores programs and data necessary for the processor 34 to control each component. The storage device 38 is a non-transitory storage medium including, for example, a hard disk drive (HDD) and a solid state drive (SSD).

The age prediction device 40 is a computer having substantially the same device configuration as the learning processing device 30 . The age prediction device 40 acquires the model data D2 provided by the learning processing device 30 to construct a learned learner LN. Then, the age prediction device 40 receives the IoT data D1 from the device 20, which is the prediction target, and predicts the device age of the prediction target using the learner LN.

FIG. 2 is a functional block diagram of processor 34 shown in FIG. The processor 34 functions as a state acquisition unit 50 , a label calculation unit 52 , a learning processing unit 54 , a data determination unit 56 and an age prediction unit 58 . A learning data set 60 and a learning parameter group 62 are stored in the storage device 38 .

The state acquisition unit 50 acquires a set of state values (hereinafter referred to as a “state value set”) indicating the state of the device 20 or the component 22 (that is, the prediction target) that constitutes the device 20 . Examples of state values include: [1] Measured amount related to prediction target (e.g., current, voltage, temperature, amount of light, etc.), [2] Measured amount related to external environment (e.g., time, location, external temperature and humidity, etc.), [3] control command values for driving the component 22; [4] log data of the device 20 (normal/abnormal status information, etc.);

The label calculator 52 assigns or corrects a label value indicating the degree of deterioration or the degree of freshness of the prediction object. The label calculation unit 52 collectively assigns or corrects a plurality of label values respectively corresponding to the plurality of state value sets obtained from the same prediction object in time series. . Here, "assigning" means newly associating a label value with a state value set for which labeling has not yet been completed. Also, "correction" means changing the label value corresponding to the state value set that has already been labeled to a more appropriate value.

First, the label calculator 52 may assign each label value by normalizing the key state value within a predetermined range. Here, the key state value is a state value having a corresponding correlation with the device age, and is used for assigning the label value. Examples of key state values include various parameters (specifically, the number of times of operation, the operating time, the number of days installed, etc.) that can be used as the basis for determining the set life. Although the key state value may be included in the state value set, it is desirable to exclude it from the state value set so as not to excessively depend on learning.

Second, if the label value is a value obtained by normalizing the degree of deterioration of the prediction object and there is learning data with a large learning error among the plurality of sets of learning data, the label calculation unit 52 sets the label value to A plurality of label values may be collectively corrected by adding correction amounts that increase as they increase to the original label values. This correction amount is a value obtained by multiplying the original label value by a correction coefficient obtained by dividing the learning error associated with the learning data, which is the maximum value among the plurality of label values, by the maximum value.

Thirdly, if the label value is a value obtained by normalizing the freshness of the prediction target object and there is learning data with a large learning error among the plurality of sets of learning data, the label calculation unit 52 sets the label value to A plurality of label values may be collectively corrected by adding correction amounts that decrease as they increase to the original label values. This correction amount is a value obtained by multiplying the original label value by a correction coefficient obtained by dividing the learning error associated with the learning data, which is the minimum value among the plurality of label values, by the maximum value.

The learning processing unit 54 determines each value of the learning parameters forming the learning parameter group 62 by performing various learning processes (here, supervised learning) on the learning device LN. The learner LN consists of a learning model that inputs a state value set and outputs a label value. The learning processing unit 54 repeats re-learning or additional learning for the learning device LN using the learning data set 60 whose label values have been corrected by the label calculation unit 52 . This learning method can be applied to batch learning, mini-batch learning, and online learning.

The data determination unit 56 uses the learner LN subjected to the learning process by the learning processing unit 54 to determine whether or not to exclude learning data included in the learning data set 60 or correct the label value of the learning data. do. Specifically, the data determination unit 56 performs threshold determination between learning errors (for example, L1 error, L2 error) regarding the learning data to be determined. For example, the data determination unit 56 maintains the learning data as it is when [1] the learning error is less than the first threshold, [2] when the learning error is greater than or equal to the first threshold and less than the second threshold [3] If the learning error is equal to or greater than the second threshold value, the corresponding learning data is deleted.

The age prediction unit 58 predicts the degree of deterioration or the degree of freshness of the prediction target using the learner LN that has undergone the learning process. An example of the degree of deterioration is age of equipment, and an example of degree of freshness is Remaining Useful Life. Note that the age prediction unit 58 may be executed by the learning processing device 30 that performs learning processing, or may be executed by another device (specifically, the age prediction device 40 in FIG. 1).

The learning data set 60 is a collection of learning data composed of state value sets and label values. The learning data includes [1] “training data” used to update the learning parameter group 62, [2] “verification data” used to evaluate generalization performance by the learning parameter group 62, and [3] learning It is classified into three types of "test data" used for evaluating the prediction performance of the learned learner LN.

The learning parameter group 62 is a collection of learning parameters used for computation of the learner LN. The learning parameter group 62 is composed of "variable parameters" to be updated during learning and "fixed parameters" (so-called hyperparameters) not to be updated during learning. Examples of variable parameters include coefficients that describe the activation functions of operational units, coupling strengths between operational units, and the like. Examples of fixed parameters include the number of arithmetic units, the number of hidden layers, the kernel size of convolution operations, and the like.

[Operation of learning processing device 30]
The device diagnosis system 10 incorporating the learning processing device 30 in this embodiment is configured as described above. Next, the learning operation by the learning processing device 30 will be described with reference to the flowchart of FIG. 3 and FIGS. 4 to 10. FIG.

In step SP10 of FIG. 3, the state acquisition unit 50 acquires a set of state values indicating the state of the prediction target (that is, a state value set) by performing analysis and preprocessing on the IoT data D1. .

FIG. 4 is a diagram showing an example of the data structure of the IoT data D1. This IoT data D1 is data in a table format showing the correspondence between the "device ID", which is the identification information of the device 20, and the (N+1) state values. The (N+1) state values are classified into one "key state value" (here, the number of times the device 20 has been activated) and N "state value sets".

At step SP12, the label calculator 52 assigns a label value corresponding to the state value set obtained at step SP10. Here, the label calculation unit 52 collectively assigns label values corresponding to a plurality of sets of state values obtained from the same prediction object at a plurality of times.

FIG. 5 is a diagram showing an example of a method of assigning label values corresponding to a plurality of status value sets. The horizontal axis of the graph indicates the number of operations (unit: times), and the vertical axis of the graph indicates the device age (unit: years). For convenience of explanation, only 5 plots are shown. The number of status value sets may be four or less, or six or more.

For example, assume that the number of actuations corresponding to the last obtained state value set is N times, that is, the maximum value of the number of actuations is "N". Depending on the circumstances under which the device 20 was replaced or removed, the value of N may correspond to the actual life span, or may be smaller than the actual life span. Here, the label calculator 52 assigns each label value by normalizing the number of operations within the range of [0, 100]. Specifically, when the number of times of operation is 0, the age is set to "0", and when the number of times of operation is N, the age is set to "100".

At step SP14 in FIG. 3, the processor 34 causes the storage device 38 to store the state value set obtained at step SP10 and the label value generated at step SP12 in a state of being associated with each other. The learning data set 60 is updated by accumulating new learning data.

At step SP16, the learning processing unit 54 confirms whether or not it is necessary to perform learning processing on the learning device LN. If execution of the learning process is not required (step SP16: NO), return to step SP10 and repeat steps SP10 to SP16 until execution becomes necessary. On the other hand, if execution of the learning process is necessary (step SP16: YES), the process proceeds to the next step SP18.

At step SP18, the learning processing unit 54 uses the learning data set 60 to perform learning processing on the learning device LN. Here, it is assumed that new learning is performed when a sufficient number of sets of learning data is accumulated.

FIG. 6 is a diagram showing an example of the network structure of the learning device LN of FIG. 2. In FIG. A learner LN is a hierarchical neural network consisting of an input layer, an intermediate layer and an output layer. The input layer consists of N arithmetic units for inputting N state values. The hidden layer consists of M arithmetic units for generating dimensionally compressed M (1<M<N) dimensional feature vectors. The output layer consists of 21 operation units for outputting label values.

In the example of this figure, the device age range of 0 to 100 years is divided into five-year intervals, and a softmax function is applied in the output layer. Thus, the learner LN outputs probabilities of being 0, 5, 10, . For example, the label value with the maximum probability in the output probability distribution is obtained as the predicted value of the device age.

Various methods can be adopted for the learning algorithm including the batch selection rule for learning data, the objective function equation including the loss term, and the learning parameter update rule. As a result of repeatedly executing this learning process, a learning parameter group 62 that satisfies a predetermined convergence condition is determined.

In step SP20 of FIG. 3, the data determination unit 56 determines the learning data included in the learning data set 60 using the learning device LN that has undergone the learning process in step SP18. Specifically, the data determination unit 56 evaluates the learning error (here, L1 error) between the "predicted value" obtained from the state value data and the "teaching value" corresponding to this state value data.

FIG. 7 is a diagram illustrating an example of a learning data determination method. The axis of the graph indicates the device age (unit: years). Here, it is assumed that the label value (or teaching value) is 100 years old. [Condition A] If the learning error is less than 15, the data determination unit 56 determines that the learning data should be maintained as is because the assigned label value is appropriate. [Condition B] If the learning error is 15 or more and less than 30, the data determination unit 56 determines that the label value should be corrected because the assigned label value is not appropriate. [Condition C] When the learning error is 30 or more, the data determination unit 56 determines that the learning data is excluded because the learning data itself is inappropriate.

At step SP22 in FIG. 3, the data determination unit 56 checks whether or not there is learning data that needs to be changed, based on the determination result at step SP20. If there is no learning data to be changed (step SP22: NO), the flowchart ends. On the other hand, if there is learning data to be changed (step SP22: YES), the process proceeds to the next step SP24.

In step SP24, the data determination unit 56 issues an instruction to update the learning data set 60 stored in the storage device 38 ("batch correction of label values" or "delete learning data"). Here, the label calculation unit 52 collectively corrects label values corresponding to a plurality of sets of state values obtained from the same prediction object at a plurality of times.

FIG. 8 is a diagram showing an example of a method of correcting label values corresponding to a plurality of status value sets. The horizontal axis of the graph indicates age before correction (unit: years), and the vertical axis indicates age after correction (unit: years). As in the case of FIG. 5, the number of status value sets is five.

For example, assume that the maximum value among five label values is 100 years old and the corresponding learning error is 20 years old. In this case, the correction coefficient C is obtained as (100-20)/100=0.8. Each label value is linearly corrected by multiplying the original label value by the correction coefficient C=0.8. By collectively correcting the specific label values in this way, it becomes easier to maintain the relationship between the state value sets acquired in time series.

After that, the process returns to step SP18, and steps SP18 to SP24 are sequentially repeated until it is no longer necessary to change the learning data. Note that after the learning data is changed in step SP24, the learning processing unit 54 may initialize the learning parameter group 62 and then re-learn the learning device LN. Alternatively, the learning processing unit 54 may perform additional learning on the learner LN that is in the process of learning, using only the multiple sets of learning data changed in step SP24.

FIG. 9 is a first graph showing the effect of improving prediction accuracy by batch correction of label values. More specifically, FIG. 9A shows a "comparative example" in which label values are not collectively corrected, and FIG. 9B shows an "embodiment" in which label values are collectively corrected. As in the case of FIG. 5, the horizontal axis of the graph indicates the number of operations (unit: times), and the vertical axis of the graph indicates the device age (unit: years).

Here, all time-series IoT data D1 are collected from the devices 20 that have normally reached the end of their lifespan, and a graph is created by predicting device ages. It should be noted that since the device 20 continues to operate while the component 22 is periodically replaced, the device age graph varies microscopically. That is, there is a tendency that the age of the device is high immediately before the replacement of the component 22, and that the age of the device is low immediately after the replacement of the component 22.

In the “comparative example” of FIG. 9( a ), the age of the device increases linearly from a region with a small number of times of operation to an intermediate region. However, in a region where the number of times of operation is large (that is, a region close to the actual life span), the clipping phenomenon of the device age occurs, and the prediction accuracy is reduced accordingly. The reason for this is that since the set life (time to replace the device) is determined by subtracting a margin from the actual life, it is not possible to predict subtle differences in device age within the margin region. On the other hand, in the "embodiment" of FIG. 9(b), the clipping phenomenon described above is suppressed through batch correction of label values, so the entire range of the number of operations (that is, the entire range of device age) can be predicted with high accuracy. can do.

FIG. 10 is a second graph showing the effect of improving the prediction accuracy by batch correction of label values. More specifically, FIG. 9A shows a "comparative example" in which label values are not collectively corrected, and FIG. 9B shows an "embodiment" in which label values are collectively corrected. Here, the entire time-series IoT data D1 is collected from a device 20 that has failed for some reason before completing its actual life, and the age of each device is predicted to create a graph.

In the “comparative example” of FIG. 10( a ), the device age is expected to be substantially constant regardless of the number of times of operation, so it is difficult to detect signs of state change of the device 20 . On the other hand, in the "embodiment" of FIG. 10B, as in the case of FIG. It can be predicted with high accuracy.

[Summary of embodiment]
As described above, the learning processing device 30 in this embodiment includes the state acquisition unit 50 that acquires the state value set indicating the state of the prediction target that is the device 20 or the component 22 that configures the device 20, and the prediction target A label calculation unit 52 that assigns or corrects a label value indicating the degree of deterioration or freshness of the label, a state value set acquired by the state acquisition unit 50, and a label value assigned or corrected by the label calculation unit 52. A learning processing unit 54 that performs learning processing on a learning device LN that inputs a state value set and outputs a label value using a learning data set 60 that is a population of learning data. Then, the label calculation unit 52 collectively assigns or assigns a plurality of label values respectively corresponding to the plurality of state value sets to the plurality of state value sets obtained from the same prediction object in time series. to correct.

Further, in the learning processing method and program in this embodiment, one or a plurality of computers obtains a state value set indicating the state of the prediction object that is the device 20 or the component 22 that constitutes the device 20 (SP10 ), a calculation step (SP12, SP24) for assigning or correcting a label value indicating the degree of deterioration or freshness of the prediction object, an acquired state value set, and an assigned or corrected label value. A learning step (SP18) of performing learning processing on the learning device LN that inputs the state value set and outputs the label value is executed using the learning data set 60, which is the population of data. Then, in the calculating step (SP24), a plurality of label values corresponding to the plurality of state value sets are collectively given to the plurality of state value sets obtained in time series from the same prediction object. Or correct.

In this way, with respect to a plurality of sets of state values obtained in time series from the same prediction object, a plurality of label values respectively corresponding to the plurality of sets of state values are collectively assigned or corrected. A plurality of label values are set while maintaining the relative relationship between the plurality of sets of state values. In other words, when the actual degree of deterioration or freshness cannot be specified, it is possible to set an appropriate label value that reflects the prediction of the actual lifespan through batch assignment or correction. As a result, even when using a state value set obtained from a prediction target that has not reached the end of its actual life or set life, it is possible to suppress deterioration in prediction accuracy due to uncertainty in label values.

Further, when the label value is a value obtained by normalizing the degree of deterioration of the prediction target object and there is learning data with a large learning error among the plurality of sets of learning data, the label calculation unit 52 calculates A plurality of label values may be collectively corrected by adding a larger correction amount to each of the original label values.

Further, the correction amount may be a value obtained by multiplying the original label value by a correction coefficient obtained by dividing the learning error related to the learning data, which is the maximum value among the plurality of label values, by the maximum value. As a result, the learning error associated with the learning data having the maximum label value can be made smaller.

In addition, the learning processing device 30 determines whether or not to exclude learning data included in the learning data set 60 or correct the label value of the learning data using the learning device LN subjected to the learning processing by the learning processing unit 54. You may further provide the data determination part 56 which determines.

Further, the learning processing unit 54 may repeat re-learning or additional learning for the learning device LN using the learning data set 60 whose label values have been corrected by the label calculation unit 52 . As a result, the learning data is cleansed sequentially, and the prediction accuracy of the learning device LN is improved.

[Modification]
It goes without saying that the present invention is not limited to the above-described embodiments, and can be freely modified without departing from the gist of the present invention. Alternatively, each configuration may be arbitrarily combined as long as there is no technical contradiction.

In the above embodiment, the learning processing device 30 is shown as a single computer, but instead of this, the learning processing device 30 may be a group of computers constructing a distributed system. The learning processing device 30 may be a cloud server (so-called cloud server) or an on-premise server.

In the above-described embodiment, the learner LN (see FIG. 6) is constructed using a hierarchical neural network, but the machine learning technique is not limited to this. For example, various techniques may be employed including support vector machines, decision trees (eg, random forests), boosting methods (eg, gradient boosting methods).

In the above-described embodiment, the device age is classified by a plurality of categories or frequencies (that is, discrete values), and the label value with the maximum probability is calculated as the device age. It is not limited to this. For example, the device age may be the sum of products of frequencies and probabilities (that is, expected values). Also, the number of operation units provided in the output layer of the learning device LN may be reduced to one, and the device age normalized to [0, 1] or the like may be output as one continuous value.

DESCRIPTION OF SYMBOLS 10... Equipment diagnostic system 20... Equipment (prediction target object) 22... Part (prediction target object) 30... Learning processing apparatus 50... State acquisition part 52... Label calculation part 54... Learning processing part 56... data determination unit 60 learning data set LN learning device

Claims (6)

a state acquisition unit that acquires a state value set indicating the state of a prediction target that is a device or a component that constitutes the device;
a label calculation unit that assigns or corrects a label value that is a value obtained by normalizing the degree of deterioration of the prediction target;
The state value set is obtained by using a learning data set, which is a population of learning data including the state value set obtained by the state obtaining unit and the label values assigned or corrected by the label calculation unit. a learning processing unit that performs learning processing on a learning device that receives an input and outputs the label value;
with
The learning data set includes time-series learning data composed of a plurality of sets of state values obtained in time-series from the same prediction object and a plurality of label values assigned by the label calculation unit. includes
If the time-series learning data include learning data in which the learning error calculated during the learning process is larger than a threshold value, the label calculation unit calculates a correction amount that increases as the label value increases. A learning processing device that collectively corrects a plurality of label values in the time-series learning data by adding each to an original label value . The correction amount is a value obtained by multiplying the original label value by a correction coefficient obtained by dividing a learning error related to learning data, which is the maximum value among the plurality of label values, by the maximum value.
2. The learning processing device according to claim 1 . further comprising a data determination unit that determines whether to exclude learning data included in the learning data set or correct the label value of the learning data using the learning device that has been subjected to the learning processing by the learning processing unit. ,
3. The learning processing device according to claim 1 or 2 . The learning processing unit repeats re-learning or additional learning for the learning device using the learning data set whose label value is corrected by the label calculation unit.
The learning processing device according to any one of claims 1 to 3 . one or more computers
an acquisition step of acquiring a state value set indicating the state of a prediction target object that is a device or a component that constitutes the device;
a calculation step of assigning or correcting a label value that is a value obtained by normalizing the degree of deterioration of the prediction target;
inputting the state value set using a learning data set that is a population of learning data including the state value set obtained in the obtaining step and the label values assigned or corrected in the calculating step; a learning step of performing a learning process on a learner that outputs the label value by
and run
The learning data set includes time-series learning data composed of a plurality of sets of state values obtained in time-series from the same prediction object and a plurality of label values assigned in the calculating step. included,
In the calculating step, when there is learning data in the time-series learning data in which the learning error calculated during the learning process is larger than a threshold value, a correction amount that increases as the label value increases is calculated. A learning processing method for collectively correcting a plurality of label values in the time-series learning data by adding each to the label values of . one or more computers,
State acquisition means for acquiring a state value set indicating the state of a prediction target that is a device or a component that constitutes the device;
label calculation means for assigning or correcting a label value that is a value obtained by normalizing the degree of deterioration of the prediction object;
The state value set is obtained by using a learning data set which is a population of learning data including the state value set obtained by the state obtaining means and the label values assigned or corrected by the label calculation means. learning processing means for performing learning processing on a learner that inputs and outputs the label value;
function as
The learning data set includes time-series learning data composed of a plurality of sets of state values obtained in time-series from the same prediction object and a plurality of label values assigned by the label calculating means. includes
The label calculation means calculates a correction amount that increases as the label value increases when there is learning data with a learning error calculated during the learning process larger than a threshold in the time-series learning data. A learning processing program for collectively correcting a plurality of label values in the time-series learning data by adding each to an original label value .

Images