KR102340652B1 - Method, apparatus and program for acquiring failure prediction model - Google Patents

Abstract

A method for obtaining a failure prediction model is provided. The method includes, by a server, acquiring data from a plurality of sensors, by the server acquiring a plurality of features for each of the plurality of sensors based on the data, by the server, among the plurality of features obtaining, by the server, a first set of features from which features corresponding to invalid data are removed; , normalizing the second feature set, and dividing, by the server, the normalized data.

Description

Method, device and program for acquiring failure prediction model

The present invention relates to a failure prediction model acquisition method, apparatus and program.

The Industrial Internet of Things (IIoT) refers to devices such as sensors and equipment interconnected by networks with the industrial sector of computers, including manufacturing and energy management. These connections enable data collection, exchange, analysis, and facilitating improvements in production and efficiency, as well as other economic benefits. IIoT is an evolution of Distributed Control Systems (DCS), enabling a high degree of automation using cloud computing to improve process control.

However, since a large number of devices interact with an IoT system including the Industrial Internet of Things (IIoT), if any one device fails, the entire system may fail.

Therefore, there is a need for a method for predicting the failure of the IIoT system. More specifically, from the process of pre-processing data for a plurality of sensors included in the IIoT system, a failure prediction model is obtained based on the pre-processed data. There is a growing need for a method to do this.

Registered Patent Publication No. 10-1796583, 2017.11.06

The problem to be solved by the present invention is to provide a failure prediction model acquisition method, apparatus, and program.

The problems to be solved by the present invention are not limited to the problems mentioned above, and other problems not mentioned will be clearly understood by those skilled in the art from the following description.

A failure prediction model acquisition method according to an aspect of the present invention for solving the above-described problems, the server, the steps of acquiring data from a plurality of sensors; obtaining, by the server, a plurality of features for each of the plurality of sensors based on the data; obtaining, by the server, a first feature set from which a feature corresponding to invalid data is removed from among the plurality of features; obtaining, by the server, a second feature set in which data of a feature in which data is missing from among the first feature set is replaced; normalizing, by the server, the second set of features; and dividing, by the server, the normalized data.

Other specific details of the invention are included in the detailed description and drawings.

According to the above-described various embodiments of the present invention, there is a new effect of constructing an efficient and accurate failure prediction model in the IIoT environment through the failure prediction model obtained by repeatedly selecting various features.

Effects of the present invention are not limited to the effects mentioned above, and other effects not mentioned will be clearly understood by those skilled in the art from the following description.

1 is a system diagram illustrating an IIoT system according to an embodiment of the present invention.
2 is a block diagram illustrating a method for obtaining a failure prediction model according to an embodiment of the present invention.
3 is a flowchart illustrating a method for obtaining a failure prediction model according to an embodiment of the present invention.
4 to 9 are exemplary views for explaining the experimental results according to an embodiment of the present invention.
10 is a block diagram of an apparatus according to an embodiment of the present invention.

Advantages and features of the present invention and methods of achieving them will become apparent with reference to the embodiments described below in detail in conjunction with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below, but may be implemented in various different forms, and only the present embodiments allow the disclosure of the present invention to be complete, and those of ordinary skill in the art to which the present invention pertains. It is provided to fully understand the scope of the present invention to those skilled in the art, and the present invention is only defined by the scope of the claims.

The terminology used herein is for the purpose of describing the embodiments and is not intended to limit the present invention. In this specification, the singular also includes the plural unless specifically stated otherwise in the phrase. As used herein, “comprises” and/or “comprising” does not exclude the presence or addition of one or more other components in addition to the stated components. Like reference numerals refer to like elements throughout, and "and/or" includes each and every combination of one or more of the recited elements. Although "first", "second", etc. are used to describe various elements, these elements are not limited by these terms, of course. These terms are only used to distinguish one component from another. Accordingly, it goes without saying that the first component mentioned below may be the second component within the spirit of the present invention.

Unless otherwise defined, all terms (including technical and scientific terms) used herein will have the meaning commonly understood by those of ordinary skill in the art to which this invention belongs. In addition, terms defined in a commonly used dictionary are not to be interpreted ideally or excessively unless specifically defined explicitly.

As used herein, the term “unit” or “module” refers to a hardware component such as software, FPGA, or ASIC, and “unit” or “module” performs certain roles. However, "part" or "module" is not meant to be limited to software or hardware. A “unit” or “module” may be configured to reside on an addressable storage medium or to reproduce one or more processors. Thus, by way of example, “part” or “module” refers to components such as software components, object-oriented software components, class components, and task components, and processes, functions, properties, Includes procedures, subroutines, segments of program code, drivers, firmware, microcode, circuitry, data, databases, data structures, tables, arrays, and variables. Components and functionality provided within “parts” or “modules” may be combined into a smaller number of components and “parts” or “modules” or additional components and “parts” or “modules”. can be further separated.

Spatially relative terms "below", "beneath", "lower", "above", "upper", etc. It can be used to easily describe the correlation between a component and other components. Spatially relative terms should be understood as terms including different directions of components during use or operation in addition to the directions shown in the drawings. For example, when a component shown in the drawing is turned over, a component described as “beneath” or “beneath” of another component may be placed “above” of the other component. can Accordingly, the exemplary term “below” may include both directions below and above. Components may also be oriented in other orientations, and thus spatially relative terms may be interpreted according to orientation.

In this specification, a computer refers to all types of hardware devices including at least one processor, and may be understood as encompassing software configurations operating in the corresponding hardware device according to embodiments. For example, a computer may be understood to include, but is not limited to, smart phones, tablet PCs, desktops, notebooks, and user clients and applications running on each device.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings.

Each step described in this specification is described as being performed by a computer, but the subject of each step is not limited thereto, and at least a portion of each step may be performed in different devices according to embodiments.

1 is a system diagram illustrating an IIoT system according to an embodiment of the present invention.

As shown in FIG. 1 , the IIoT system may include a server 100 and a plurality of sensors 200 - 1 to 200 - 5 .

The server 100 may acquire sensing data from the plurality of sensors 200 - 1 to 200 - 5 . The server 100 may build a failure prediction model for predicting system failure based on the acquired sensing data and the plurality of sensors 200-1 to 200-5 corresponding to the sensing data.

2 is a block diagram illustrating a method for obtaining a failure prediction model according to an embodiment of the present invention.

The server 100 may obtain a failure prediction model based on a feature selected to maximize the prediction accuracy of the failure prediction model. To this end, the server 100 may repeatedly perform feature selection, and may acquire a plurality of failure prediction models based on the repeatedly selected feature. Thereafter, the server 100 may obtain a failure prediction model having the highest prediction accuracy among the plurality of failure prediction models.

Specifically, as shown in FIG. 2 , in order to obtain a server 100 failure prediction model, 1) a pre-processing step (Processing, 2) an importance measurement step (Importance Measurement), 3) a feature selection step (Feature Selection), 4 ) model building step (Model Building) and 5) model selection step (Model Selection) can be performed.

Specifically, in the pre-processing step, the server 100 may perform feature removal, missing data replacement, normalization, and data segmentation steps. Thereafter, the server 100 may acquire the importance of the feature obtained through the pre-processing process. Thereafter, the server 100 may repeatedly perform the feature selection and model building steps. Thereafter, the server 100 may build a failure prediction model through the SVM based on the selected feature in the model building step.

3 is a flowchart illustrating a method for obtaining a failure prediction model according to an embodiment of the present invention.

In step S110, the server 100 may acquire data from a plurality of sensors.

According to an embodiment, in various embodiments of the present invention, the term feature and the term sensor may be interpreted the same.

In step S120 , the server 100 may acquire a plurality of features for each of a plurality of sensors.

In step S130 , the server 100 may obtain a first feature set from which a feature corresponding to invalid data is removed from among a plurality of features.

In operation S140 , the server 100 may obtain a second feature set in which data of a feature in which data is missing from among the first feature set is replaced.

In step S150 , the server 100 may normalize the second feature set.

In step S160, the server 100 may divide the normalized data.

Specifically, according to an embodiment of the present invention, the server 100 may remove at least one feature from among a plurality of features corresponding to each of data obtained from a plurality of sensors.

Specifically, the server 100 may acquire invalid data among data acquired from each of a plurality of sensors.

In an embodiment, the server 100 may remove a feature corresponding to the invalid data when the ratio of the invalid data for the specific feature is greater than a preset validity coefficient in the range of [0,1]. .

In this case, the invalid data may mean data having no value among data collected from the feature. In one embodiment, when the feature is a temperature sensor, data in which the temperature is not sensed may be invalid data. Invalid data may be data corresponding to a row having no value for each feature in the collected data set (which may include a training data set and a test data set). In another embodiment, the server 100 may obtain a respective variance value for data obtained from each of a plurality of sensors. Thereafter, when the variance value is equal to or less than the preset value, the server 100 may remove features corresponding to the variance value equal to or less than the preset value.

Specifically, in the pre-processing step, the server 100 may perform feature removal, missing data replacement, normalization, and data segmentation steps. Specifically, the server 100 retrieves the invalid (NA) data of each feature to remove invalid features from the input data set (ie, data collected from multiple sensors), and calculates the variance of each feature. have.

According to an embodiment, the server 100 may remove the specific feature from the input data set when the ratio of the NA data of the specific feature is greater than the preset validity coefficient in the [0, 1] range.

As another embodiment, when the variance of a specific feature is less than or equal to a preset value, the server 100 may remove the corresponding feature from the input data set. Preferably, the closer the variance of a particular feature to zero, the higher the probability that the feature will be removed from the input data set.

On the other hand, when missing data exists (that is, when NA data exists in the features remaining after feature removal), the server 100 may replace the missing data with the average value of the non-missing data of each feature. .

Thereafter, the server 100 may normalize data for a plurality of features from which at least one feature is removed.

Specifically, the server 100 may perform a normalization process according to the data size of each feature. According to an embodiment, the server 100 may perform a normalization process of each feature based on Equation 1 below.

[Equation 1]

x _i '= (x _i -m) / s

In this case, x _i ' may be the (i + 1)-th normalized characteristic data, x _i may be the (i + 1)-th data of the feature, m may be the average of the feature, and s may be the standard deviation of the feature.

Meanwhile, the average m of the features and the standard deviation s of the features may be obtained based on Equations 2 and 3, respectively.

[Equation 2]

[Equation 3]

In this case, N may mean the number of data of the feature. Meanwhile, the average m of features may mean an average of data values collected from each feature (eg, a sensor).

As an embodiment, when the feature is a temperature sensor, the average of the features may mean an average value of a plurality of temperature data obtained by the temperature sensor. It is self-evident that the standard deviation of features and variance of features is also the standard deviation or variance of data values collected from each feature.

Thereafter, the server 100 may divide the normalized data.

Specifically, the server 100 may classify the normalized data set into a training data set and a test data set. The server 100 may use the training data set to obtain a failure prediction model, and the test data set to evaluate the performance of the predictive model.

Meanwhile, the server 100 may acquire the importance of the feature based on the preprocessed data.

According to an embodiment, the server 100 may acquire a training data set from among the data acquired through the pre-processing step. Thereafter, the server 100 may acquire the importance of a feature corresponding to the training data based on the random forest technique for the training data.

Specifically, the server 100 may acquire the importance of the feature acquired through the preprocessing process. The server 100 may obtain the importance of each feature by analyzing the training data set based on the random forest technique.

Specifically, the server 100 may analyze the relationship between each feature and Fail by comprehensively considering a plurality of decision trees created through the random forest technique and the created decision trees.

According to an embodiment, in the random forest technique performed by the server 100 , a plurality of subsets of the training data set may be generated to obtain a decision tree differently, respectively. Each of the plurality of subsets generated may consist of different data and features.

To this end, the server 100 may randomly select n data (ie, n rows) and mtry features (ie, mtry columns) from the training data set. The server 100 may repeat the random selection until the number of obtained subsets reaches ntree, a predefined number of decision trees.

Then, each decision tree can be built separately based on one of the obtained subsets. That is, after constructing an ntree decision tree, the importance of each feature is measured through the Mean Decrease Gini (MDG, Gini importance score), which can be an indicator of how much each feature affects accurate prediction results.

More specifically, in each decision tree, the server 100 may calculate the sum of differences in Gini impurities between a parent node using a specific feature and a child node of the parent node. Thereafter, the server 100 may calculate an average of all decision tree results. In this case, the decision tree may be composed of several nodes that make a decision using a threshold value of a specific feature. Each node can also contain Gini impurities to measure the probability of a wrong decision.

Meanwhile, the random forest technique is an ensemble method for learning a plurality of decision trees, and the random forest can be used for various problems such as detection, classification, and regression.

On the other hand, in the random forest according to the present invention, trees have slightly different characteristics from each other due to randomness, and this characteristic causes the predictions of each tree to be decorrelated, and as a result, generalization ), there is an effect that can improve the performance. Also, randomization can make the forest robust against data including noise. Randomization is performed in the training process of each tree, and bagging and randomized node optimization, an ensemble learning method using a random learning data extraction method, may be used.

Meanwhile, the server 100 may select a feature based on the obtained importance level and generate a plurality of models based on the selected feature.

The server 100 may repeatedly perform the feature selection and model building steps. Specifically, Algorithm 1 below shows the overall operation of the feature selection step and the model building step.

Meanwhile, in Algorithm 1, the importance set of _{features can be defined as IMP=[imp 0} ,..,imp _j ,..,imp _nf-1 ] , and the set of features is F=[f ₀ ,.., f _j ,..,f _nf-1 ] can be defined.

In this case, nf may be the number of features, and imp _j may be the importance of the (j + 1)-th feature.

In Algorithm 1, variables are initialized at the start, sf _cnt means a feature selected in the (cnt + 1)-th iteration, and pa _cnt means the failure prediction accuracy of the model built in the (cnt + 1)-th iteration, , cnt may mean a counter value that increases from 0 to 1. The server 100 may repeat Algorithm 1 until the importance of the feature newly selected in Algorithm 1 is less than the importance threshold. That is, the server 100 may repeat the feature selection step and the model building step until the importance of the newly selected feature is less than the importance threshold.

In this case, the maximum number of iterations (max) of Algorithm 1 may be set equal to the number of features whose importance is greater than the importance threshold. Therefore, when cnt reaches the maximum number of iterations (max), the server 100 may terminate the iteration of Algorithm 1. Specifically, while the feature selection step is repeated, according to Equation 4 below, the server 100 may select a new feature with the highest importance and add the selected feature to the feature set SF.

[Equation 4]

SF=[sf ₀ , sf ₁ ,..,sf _cnt ], cnt ∈ [0, max-1]

In conclusion, the number of features included in SF can increase by one as the number of iterations increases. Then, when the SF is updated, the server 100 may build a failure prediction model through the SVM based on the SF in the model building step.

Thereafter, the server 100 may update the prediction accuracy set PA expressed by Equation 5 below by evaluating the prediction accuracy of the built failure prediction model.

[Equation 5]

PA=[pa ₀ ,pa ₁ ,.. pa _cnt ,], cnt ∈ [0, max-1]

At this time, each index of SF and PA (that is, the index and prediction accuracy of each selected feature) may indicate the number of iterations. In Algorithm 1, modelBuildEvalFunction() is a function that builds and evaluates a predictive model, Prediction accuracy can be derived.

In this case, in Algorithm 1, modelBuildEvalFunction (SF) is a function for building and evaluating a prediction model, and of course, a conventional prediction evaluation function may be used.

For example, if cnt is 2 (3rd loop) and selected feature set SF = [f ₃₉ , f ₁₂ , f ₁₉ ] (ie, sf ₀ =f ₃₉ , sf ₁ =f ₁₂ , sf ₂ =f ₁₉ ) , the server 100 may generate an SVM-based prediction model with only the training data set of the three features. _{Thereafter, the server 100 may calculate pa 2} with the test data set for the three features and update the PA set. (That is, the number of pa in the PA set increases from 2 (PA=[pa ₀ , pa ₁ ]) to 3 (PA=[pa ₀ , pa ₁ , pa ₂ ]). When cnt reaches the max value, , of course, the task is terminated.

Meanwhile, according to an embodiment of the present invention, the SF may be a set of selected features, and may be arranged in order of increasing importance. That is, every time cnt increases by 1 in the Repeat loop of Algorithm 1, a feature having a large importance included may be added one by one.

The server 100 may derive prediction accuracy by taking SF as an input. When the iteration ends, the server 100 may derive the final SF and PA as a result of the operation.

That is, the server 100 can obtain prediction accuracy (PA) by using SF as an input value, and as the iteration of algorithm 1 is terminated, finally SF and PA derived as shown in Equations 6 and 7 below It can be calculated as a result.

[Equation 6]

SF=[sf ₀ ,..,sf _cnt ,..,sf _max-1 ],

[Equation 7]

PA=[pa ₀ ,..,pa _cnt ,..,pa _max-1 ]

Meanwhile, the server 100 may select at least one model from among a plurality of generated models.

Specifically, the server 100 may select a failure prediction model having the highest prediction accuracy with reference to the PA and the SF. In particular, in step S150 , the server 100 may search for the highest prediction accuracy in the PA using max(PA). In this case, max() may be a function that searches for the index having the maximum value in the specified set. Thereafter, the server 100 may obtain an index of max(PA) in the PA. In order to obtain the index of max(PA), the server 100 may compare max(PA) of each element of the PA, and the index of the element equal to max(PA) may be derived from the PA. Thereafter, the server 100 may select a failure prediction model in consideration of the indexes of max(PA) and SF.

More specifically, the server 100 may obtain elements having an index equal to or less than an index of max(PA) from the SF. Thereafter, the server 100 may select one predictive model from among the failure prediction models generated in the model generation step by comparing the features used in the step model generation step with the features extracted from the SF.

Let's look at the failure prediction model selection method in more detail. First, the server 100 may obtain max(PA) for _{pa i} having the highest prediction accuracy among _{pa 0} to pa _{max-1 obtained by Algorithm 1 .} Thereafter, the server 100 may obtain an index of max(PA) to obtain at least one feature having a prediction accuracy equal to or greater than a preset value. Thereafter, the server 100 may select the at least one model based on the max(PA) and at least one feature having a prediction accuracy equal to or greater than the preset value.

Thereafter, in the model selection step, the server 100 may select a failure prediction model having the highest prediction accuracy with reference to the PA. Specifically, the server 100 selects the highest prediction accuracy through max(PA) and derives the index of max(PA) to calculate the number of iterations with the highest prediction accuracy (ie, the number of selected features). . The server 100 may select a failure prediction model in consideration of the index and SF of max(PA). In this case, of course, max(PA) may mean the largest value among PAs (ie, PA sets) obtained through repetition of Algorithm 1.

In this case, the index of max(PA) may mean in which order the maximum accuracy is included in the PA set. That is, the index is an index indicating the highest accuracy when the value of cnt is what (ie, what number of iterations).

The index compares the pa values with max(PA) in the PA set, and the sequence number (index=cnt) of the same pa as max(PA) may be obtained from the PA set, and the server 100 may set max(PA) in the PA set. A predictive model can be constructed by obtaining the index (ie, cnt value) of (PA) and finally selecting features in the iterative trial.

For example, as shown in Figure 5, the index of max(PA) is 7, where the selected feature set is SF = [F60, F349, F41, F289, F427, F65, F66, F154] (i.e., 8 features in order of importance), and an SVM-based prediction model made from data of features included in the SF can be finally selected and used.

4 to 9 are exemplary views for explaining the experimental results according to an embodiment of the present invention.

In the present invention, an experimental implementation was performed to verify the validity of the proposed failure prediction model using the open source R version 3.4.3. For this purpose, the SECOM data set provided by the UCI repository was used. The data set consisted of 1567 data and 591 features, and data were collected from the semiconductor manufacturing process by monitoring sensors and process measurement points. Data for 590 features were measured from multiple sensors, and the data for the rest of the features are the results of the pedigree test indicated as Pass and Fail.

In this experiment, the effectiveness coefficient was set to 0.1 for feature removal. That is, features with NA data of 10% or more and features without variance were removed from the data set. It can be seen that the number of features decreased from 591 to 393 through the feature removal process described above.

The ratio of the training data set and the test data set was set to 7:3. Random forest technique and caret package were used to measure the importance of each feature. n, mtry, and ntree were set to 1000, 19 and 500, respectively. With this setup, 500 decision trees were built using a randomly generated 1000*19 matrix. The significance of each feature was measured through the MDG (mean Decrease Gini).

4 shows the importance of the top 30 features. The x-axis and y-axis represent MDG and features, respectively. In the figure, it can be seen that F60 has the highest MDG (1.52) among all features.

Meanwhile, in this experiment, in order to perform the iterative feature selection process, the importance threshold was set to 0.7. Therefore, (max) is determined to be 24, which means that the number of repetitions is 24.

The training data set contains 70 Fails and 1038 Passes. This imbalance of the training data set makes it difficult for the failure prediction model to predict failure cases, and to solve this problem, sampling was performed before creating the failure prediction model. decreased. As described above, a failure prediction model was built using the SVM through the feature selection results and the sampled training data set, and the failure prediction model was built using the e1071 package in R. The table shown in Fig. 5 shows the obtained SF and PA. It can be seen from the table of FIG. 5 that max(PA) is 0.72 and the index is 7. In conclusion, eight features (eg, F60, F349, F41, F289, F427, F65, F66, and F154) are selected.

For performance evaluation, the prediction accuracy of the proposed model was compared with that of the existing model. Specifically, we can consider a fixed number of features (e.g. 12 and 24 features) and three existing models built on the basis of all features. 6 to 9 show the ROC (receiver operating characteristic) curves for three failure prediction models using different numbers of features. The ROC curve is a measure of the performance of a predictive model, and represents a relationship between a true positive rate (TPR) and a false positive rate (FPR).

In this case, TPR and FPR may be expressed as in Equations 8 and 9 below.

[Equation 8]

[Equation 9]

In this case, TP, FN, FP, and TN may be true positive, false negative, false positive, and true negative, respectively.

On the other hand, in the ROC curve, the area under the curve (AUC) is used to evaluate the prediction accuracy of the model. Specifically, the larger the AUC, the higher the prediction accuracy. 6, it can be seen that the failure prediction model built based on iterative feature selection has the largest AUC compared to other models. This is because the model is iteratively built using a different number of features and one of the models with the highest prediction accuracy is selected.

Furthermore, as shown in FIGS. 7 and 8, when a fixed number of features is used to build a model, it can be confirmed that the prediction accuracy may be relatively deteriorated.

In other words, since it is difficult to build an accurate prediction model due to irrelevant features, prediction accuracy may be relatively poor when a fixed number of features are used to build the model.

In other words, if you build a model with more features that are not related to failure, the predictive accuracy of the model may decrease. Therefore, when all features of the data set are used as shown in Fig. 9, the AUC decreases significantly. Quantitatively, the proposed model obtains 14.3 and 22.0% higher AUCs compared to the case of fixed number of features and all features, respectively.

However, the present invention builds a failure prediction model iteratively using a different number of features, and furthermore, builds a failure prediction model by selectively acquiring failure-related features. can produce a better effect than

10 is a block diagram of an apparatus according to an embodiment of the present invention.

The processor 102 may include one or more cores (not shown) and a graphic processing unit (not shown) and/or a connection path (eg, a bus, etc.) for transmitting and receiving signals to and from other components. .

The processor 102 according to one embodiment performs the method described with respect to FIG. 3 by executing one or more instructions stored in the memory 104 .

On the other hand, the processor 102 is a RAM (Random Access Memory, not shown) and ROM (Read-Only Memory: ROM) for temporarily and / or permanently storing a signal (or, data) processed inside the processor 102. , not shown) may be further included. In addition, the processor 102 may be implemented in the form of a system on chip (SoC) including at least one of a graphic processing unit, a RAM, and a ROM.

The memory 104 may store programs (one or more instructions) for processing and controlling the processor 102 . Programs stored in the memory 104 may be divided into a plurality of modules according to functions.

The steps of a method or algorithm described in relation to an embodiment of the present invention may be implemented directly in hardware, as a software module executed by hardware, or by a combination thereof. A software module may include random access memory (RAM), read only memory (ROM), erasable programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), flash memory, hard disk, removable disk, CD-ROM, or It may reside in any type of computer-readable recording medium well known in the art to which the present invention pertains.

The components of the present invention may be implemented as a program (or application) to be executed in combination with a computer, which is hardware, and stored in a medium. Components of the present invention may be implemented as software programming or software components, and similarly, embodiments may include various algorithms implemented as data structures, processes, routines, or combinations of other programming constructs, including C, C++ , may be implemented in a programming or scripting language such as Java, assembler, or the like. Functional aspects may be implemented in an algorithm running on one or more processors.

As mentioned above, although embodiments of the present invention have been described with reference to the accompanying drawings, those skilled in the art to which the present invention pertains know that the present invention may be embodied in other specific forms without changing the technical spirit or essential features thereof. you will be able to understand Therefore, it should be understood that the embodiments described above are illustrative in all respects and not restrictive.

100: server
200-1 to 200-5: a plurality of IoT devices

Claims (9)

In the failure prediction model acquisition method,
obtaining, by the server, data from a plurality of sensors;
obtaining, by the server, a plurality of features for each of the plurality of sensors based on the data;
obtaining, by the server, a first feature set from which a feature corresponding to invalid data is removed from among the plurality of features;
obtaining, by the server, a second feature set in which data of a feature in which data is missing from among the first feature set is replaced;
normalizing, by the server, the second set of features; and
dividing, by the server, the normalized data; including,
The normalizing step is normalized based on Equation 1 below,
The failure prediction model acquisition method is,
obtaining, by the server, the importance of the feature based on the normalized data;
selecting, by the server, a feature based on the obtained importance level, and generating a plurality of models based on the selected feature; and
selecting, by the server, at least one model from among the plurality of generated models; including,
The step of generating the plurality of models comprises:
generated by the following algorithm 1,
The step of selecting the at least one model comprises:
acquiring a selection feature set (SF) while increasing the size of cnt until the value of cnt becomes max by the algorithm 1;
obtaining a prediction accuracy set (PA) by inputting the selected feature set as an input value into a failure prediction model; including,
The failure prediction model acquisition method is,
Obtaining max(PA) for _{pa i} having the highest prediction accuracy among _{pa 0} to pa _max-1 obtained by obtaining the selection feature set (SF) and obtaining the prediction accuracy set (PA) to do;
obtaining the index of max(PA) to obtain at least one feature having a prediction accuracy equal to or greater than a preset value; and
selecting the at least one model based on the max(PA) and at least one feature having a prediction accuracy equal to or greater than the preset value; A failure prediction model acquisition method comprising a.
[Equation 1]
x _i '= (x _i -m) / s
In this case, x _i ' is the (i + 1)-th normalized feature data, x _i is the (i + 1)-th data of the feature, m is the average of the feature, and s is the standard deviation of the feature.
[Algorithm 1]

where cnt is the counter value, SF is the selected feature, PA is the prediction accuracy, sf _cnt is the selected feature in cnt, and the importance set of the feature is IMP=[imp ₀ ,...,imp _j ,...,imp _{nf -1} ], the set of features is F=[f ₀ ,...,f _j ,...,f _nf-1 ], imp _j is the value of importance at j+1, nf is the number of features, modelBuildEvalFunction(SF) is an arbitrary function for building and evaluating a predictive model. delete According to claim 1,
The failure prediction model acquisition method is,
obtaining, by the server, learning data based on the normalized data;
obtaining, by the server, the importance of a feature corresponding to the learning data based on a random forest technique; A failure prediction model acquisition method comprising a. delete delete delete delete a memory storing one or more instructions; and
a processor executing the one or more instructions stored in the memory;
The processor by executing the one or more instructions,
An apparatus for performing the method of claim 1 . A computer program stored in a computer-readable recording medium in combination with a computer, which is hardware, to perform the method of claim 1 .

Images