KR102418117B1 - Method and apparatus for diagnosing equipment condition - Google Patents

Abstract

Disclosed are a facility state diagnosis method and a facility state diagnosis system. The facility state diagnosis method according to an embodiment of the present invention comprises the steps of: classifying, based on similarity between a plurality of pieces of learning facility state data, each correct answer value (GT) assigned to the plurality of pieces of learning facility state data into a plurality of classes; generating a classification model by training and processing a representative correct answer value representing a class into which each correct answer value is classified together with the pieces of learning facility state data; calculating a normalization value by applying a selected normalization technique to each correct answer value; training the calculated normalization value together with the pieces of learning facility state data to generate a regression model; and outputting, with respect to input facility state data of an object to be diagnosed, a final predicted value for diagnosing the facility state of the object to be diagnosed by using a first predicted value confirmed from the classification model and a second predicted value confirmed from the regression model.

Description

Equipment condition diagnosis method and facility condition diagnosis system

The present invention relates to a method and system for high-precision equipment state diagnosis using a deep learning model having dual outputs, and more specifically, based on two types of outputs output from the deep learning model, for equipment state data of a diagnosis target It is related to facility condition diagnosis technology that can accurately diagnose according to 'real value'.

This patent is a research conducted with the support of the Special Research and Development Zone Promotion Foundation with funding from the government (Ministry of Science and Technology, Information and Communication) in 2020 (2020-DD-UP-0348).

The technology underlying the present invention is disclosed in the following documents.
1) Korean Patent Registration 10-1936283 (2019.01.02), "Method for diagnosis and prediction of mechanical defects".
2) Japanese Patent Registration 6851558 (2021.03.11), "Abnormality diagnosis method, abnormality diagnosis device, and abnormality diagnosis program".
3) Korean Patent No. 10-1808461 (2017.12.06), "Method and Apparatus for Predicting Remaining Life of Machine".
Recently, deep learning models are being commercialized to perform classification, detection, regression, etc. of inspection objects in various fields. In particular, in the field of facility condition diagnosis for diagnosing facility conditions, facilities such as pressure, temperature, lifespan, etc. There is a growing demand for more precise diagnosis of the actual state of

However, designing a deep learning model that clearly classifies the equipment status is not as simple as you might think, and when designing a deep learning model with the existing learning method, each section to which the ground truth (GT) you want to diagnose belongs belongs. Since the similarity between the two is low, prediction accuracy may be lowered for the entire data.

Therefore, technology that can further improve the accuracy of prediction when diagnosing equipment status is required through the design of a deep learning model that can classify each section with different similarities and predict an accurate real value within each section. .

The present invention designs a deep learning model ('classification model' and 'regression model') having dual outputs, and combines two types of outputs ('first and second prediction values') output from the deep learning model. According to the predicted value ('real value'), the purpose of the diagnosis is to more precisely diagnose the equipment state data of the diagnosis target.

The present invention classifies the correct answer value (GT) given to a plurality of equipment state data for learning into a plurality of classes based on the similarity between the data, and then the representative correct value representing the class into which each equipment state data for learning is classified. A deep learning model with dual outputs is designed by generating a classification model by learning and generating a regression model by learning a normalized value obtained by normalizing the correct answer value (GT) given to each learning equipment state data. aim to do

The present invention obtains, as a first prediction value, a representative correct value of a class into which equipment state data of a diagnosis object is to be classified by the classification model, and normalizes the correct answer value to be given to equipment state data of the diagnosis object by the regression model A value is obtained as a second predicted value, and based on an integer value obtained by the first predicted value and a final predicted value of a real value obtained using a decimal value obtained by the second predicted value, the equipment state data of the diagnosis target is obtained The purpose is to accurately diagnose the condition.

A facility state diagnosis method according to an embodiment of the present invention includes classifying each correct answer value (GT) given to the plurality of equipment state data for learning into a plurality of classes based on the similarity between a plurality of equipment state data for learning Step, the step of generating a classification model by learning and processing the representative correct value representing the class in which each correct value is classified along with the learning equipment state data, and applying the selected regularization technique to each correct value to normalize generating a regression model by calculating a value, learning and processing the calculated normalized value together with the equipment state data for learning; The method may include outputting a final predicted value for diagnosing the equipment state of the diagnosis target by using the first predicted value and the second predicted value identified from the regression model.

In addition, the equipment state diagnosis system according to an embodiment of the present invention divides each correct answer value (GT) given to the plurality of equipment state data for learning into a plurality of classes, based on the degree of similarity between the plurality of equipment state data for learning. A pre-processing unit for classifying, a classification model generating unit for generating a classification model by learning and processing a representative correct value representing a class into which each correct value is classified, together with the learning equipment state data, and each correct value A normalization processing unit for calculating a normalized value by applying the selected normalization technique to , a regression model generation unit for generating a regression model by learning the calculated normalized value together with the learning equipment state data, and an input diagnosis A diagnostic processing unit for outputting a final predicted value for diagnosing the equipment state of the object to be diagnosed by using the first predicted value confirmed from the classification model and the second predicted value confirmed from the regression model with respect to the equipment state data of the object can do.

According to the final predicted value ('real value') obtained by combining two types of outputs output from the deep learning model (classification model, regression model) built by the learning method proposed in the present invention, diagnostic accuracy can be improved.

According to the present invention, instead of simply generating a deep learning model with a multi-layer structure to obtain a final predicted value for equipment state diagnosis, the correct answer value is classified into a plurality of sections (classes) in consideration of the similarity between a plurality of equipment state data for learning. After that, a deep learning model (a classification model and a regression model) having two types of outputs is generated by performing the learning process of the representative correct value for each section and the learning process of the normalized value normalizing the correct answer value, respectively, and the deep learning model By outputting the final predicted value for diagnosing the equipment state of a diagnosis target based on

According to the present invention, by using a data pre-processing method suitable for a network structure, various industrial fields that perform deep learning-based condition diagnosis, for example, state diagnosis using smart factories, which are intelligent production factories, medical equipment, inspection equipment, etc. It is possible to build a deep learning model suitable for

1 shows an example for explaining the existing product (equipment) state diagnosis method.
2 is a block diagram illustrating a schematic configuration of a facility state diagnosis system according to an embodiment of the present invention.
3 is a diagram illustrating a pre-processing process for classifying a correct answer value (GT) given to equipment state data for learning in the equipment state diagnosis system according to an embodiment of the present invention.
4 is a diagram illustrating a learning process of generating a deep learning model having two types of outputs by learning in the facility state diagnosis system according to an embodiment of the present invention.
5 is a diagram illustrating a normalization process for normalizing a correct answer value (GT) given to equipment state data for learning in the equipment state diagnosis system according to an embodiment of the present invention.
6 is a diagram illustrating an example of obtaining a final predicted value based on an output of a learned deep learning model in the facility state diagnosis system according to an embodiment of the present invention.
7 is a diagram illustrating an example of calculating the total loss of a deep learning model having two types of outputs in the facility state diagnosis system according to an embodiment of the present invention.
8 is a flowchart illustrating a procedure of a method for diagnosing a facility state according to an embodiment of the present invention.

Hereinafter, embodiments will be described in detail with reference to the accompanying drawings. However, since various changes may be made to the embodiments, the scope of the patent application is not limited or limited by these embodiments. It should be understood that all modifications, equivalents and substitutes for the embodiments are included in the scope of the rights.

The terms used in the examples are used for the purpose of description only, and should not be construed as limiting. The singular expression includes the plural expression unless the context clearly dictates otherwise. In this specification, terms such as "comprise" or "have" are intended to designate that a feature, number, step, operation, component, part, or a combination thereof described in the specification exists, but one or more other features It should be understood that this does not preclude the existence or addition of numbers, steps, operations, components, parts, or combinations thereof.

Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by one of ordinary skill in the art to which the embodiment belongs. Terms such as those defined in commonly used dictionaries should be interpreted as having a meaning consistent with the meaning in the context of the related art, and should not be interpreted in an ideal or excessively formal meaning unless explicitly defined in the present application. does not

In addition, in the description with reference to the accompanying drawings, the same components are given the same reference numerals regardless of the reference numerals, and the overlapping description thereof will be omitted. In describing the embodiment, if it is determined that a detailed description of a related known technology may unnecessarily obscure the gist of the embodiment, the detailed description thereof will be omitted.

1 shows an example for explaining the existing product (equipment) state diagnosis method.

Recently, various technologies for implementing a smart factory have been intensively studied. In order to implement a smart factory, process data is required, and the resulting product quality data must be converted into a DB (data base). The process data may be various data such as raw material amount, raw material quality, manufacturing process data (temperature, humidity, etc.) and facility operation conditions. That is, the smart factory should be able to predict quality by analyzing the correlation between input and output, prevent mass defects in advance, and predict equipment failure in advance to respond preemptively.

A state diagnosis system may be a core technology for implementing a smart factory. For example, the condition diagnosis system may be configured with a vision system, an optical module, a data processing module, and a man machine interface (MMI) module to inspect product defects in real time and determine product quality. In this case, the main element of the state diagnosis system may be a data processing module (or a defect recognition module) that determines the location and type of a defect in the image.

The vision system can inspect the surface of the product for defects. The optical module is composed of a camera and a light to take a picture of the product, and can acquire an image of the product. The data processing module may inspect the defect in the image acquired by the optical module. The MMI module can display (or display) and store images and status diagnosis results for products.

In particular, the recognition module that determines the location and type of the defect in the image is a major factor affecting the state diagnosis performance. In the past, the system was developed as an algorithm development method that extracts the main factors of features by relying on the rule-base method based on the developer's empirical judgment to determine normality and defects. Accordingly, the disadvantage of having to develop a major factor extraction algorithm each time and the resulting decrease in recognition performance appeared. Therefore, in recent years, there is a trend to develop recognition algorithms using deep learning technology that can extract features for defect detection and classification by itself.

As such, a deep learning method is widely used in a technology for diagnosing equipment status using machine vision and image data. Deep learning is a structure in which neural network layers are deeply stacked, and by using a learned neural network (or artificial neural network), it is possible to perform a high-speed diagnosis of the state of a test target. The trained neural network may be learned by expanding input data, expanding label data, and mixing the two to generate new data.

Neural networks (or artificial neural networks) can include statistical learning algorithms that mimic the neurons of biology in machine learning and cognitive science. A neural network may refer to an overall model having problem-solving ability by changing the bonding strength of synapses through learning in which artificial neurons (nodes) formed a network by bonding of synapses.

The neural network may include a deep neural network. Neural networks include Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), Perceptron, Feed Forward (FF), Radial Basis Network (RBF), Deep Feed Forward (DFF), Long Short Term Memory (LSTM), Gated Recurrent Unit (GRU), Auto Encoder (AE), Variational Auto Encoder (VAE), Denoising Auto Encoder (DAE), Sparse Auto Encoder (SAE), Markov Chain (MC), Hopfield Network (HN), Boltzmann Machine (BM) ), Restricted Boltzmann Machine (RBM), Deep Belief Network (DBN), Deep Convolutional Network (DCN), Deconvolutional Network (DN), Deep Convolutional Inverse Graphics Network (DCIGN), Generative Adversarial Network (GAN), Liquid State Machine (LSM) ), Extreme Learning Machine (ELM), Echo State Network (ESN), Deep Residual Network (DRN), Differential Neural Computer (DNC), Neural Turning Machine (NTM), Capsule Network (CN), Kohonen Network (KN), and AN (Attention Network) may be included.

On the other hand, it is not simple to design a regression model that clearly classifies the equipment state. For example, assuming that the ground truth (GT) to be diagnosed has a section between '0' and '29', the learning equipment state data with a correct value between '0' and '14' are mutually exclusive. The similarity between the two may be high, but the similarity may be low with the learning equipment state data having a correct value between '15' and '29'.

That is, the similarity between the different sections in the facility status data is low, but the similarity between the data within each section is high. Then, since the similarity between each section is low, it is difficult to accurately predict the entire data, and thus the accuracy may be lowered.

Therefore, the present invention is to design a deep learning model having dual outputs capable of predicting accurate real values within each section while classifying each section with a different degree of similarity, and based on the double output of the deep learning model, the equipment By diagnosing the state, it is possible to increase the accuracy of prediction with respect to the entire data when diagnosing the equipment state.

2 is a block diagram illustrating a schematic configuration of a facility state diagnosis system according to an embodiment of the present invention.

Referring to FIG. 2 , the facility state diagnosis system 200 according to an embodiment of the present invention may include a preprocessor 210 , a classification model generator 220 , and a diagnosis processor 250 . . In addition, according to an embodiment, the facility state diagnosis system 200 may be configured by adding a normalization processing unit 230 , a regression model generation unit 240 , a loss value calculating unit 260 , and an update processing unit 270 , respectively. .

The preprocessor 210 functions to classify each correct answer value (Ground Truth, GT) given to the plurality of equipment state data for learning into a plurality of classes, based on the degree of similarity between the plurality of equipment state data for learning.

The equipment state data for learning is a physical quantity indicating the state of various equipment, such as pressure, temperature, lifespan, and the like, and may be a real value indicating an actual state.

The correct answer value GT may be given in advance to the learning equipment state data for learning processing according to the supervised learning method. In the present specification, it is assumed that the correct answer is, for example, an integer value between '0' and '29'.

For the generation of a classification model to be described later, the preprocessor 210 calculates the entire section ('0' to '29') of the correct value given to each learning equipment state data based on the similarity analysis between the data in a plurality of sections (classes). ), and data preprocessing can be performed in which the correct answer values (GTs) included in each section are replaced with representative correct answer values representing the corresponding correct values. Through this, it is possible to create a classification model by learning based on the preprocessed data.

Specifically, the preprocessor 210 performs a similarity analysis between the data for each of a plurality of learning equipment state data, sets a correct answer range for each class based on the similarity analysis result, and sets the correct answer value range Taking this into consideration, the correct answer value given to each learning equipment state data may be classified to belong to any one of a plurality of classes. Also, the preprocessor 210 may designate a representative correct answer value representing a class into which each correct answer value is classified.

Hereinafter, the preprocessor 210 will be described with reference to FIG. 3 .

3 is a diagram illustrating a pre-processing process for classifying a correct answer value (GT) given to equipment state data for learning in the equipment state diagnosis system according to an embodiment of the present invention.

Referring to FIG. 3 , the pre-processing unit 210 analyzes the similarity between data when a plurality of (eg, 30) learning equipment state data to which a correct value 310 between '0' and '29' is given is input. By doing so, it is possible to group learning equipment state data having a high degree of similarity, and to group learning equipment state data having a similarity level different by more than a selected level into different groups.

The preprocessor 210 may provide a plurality of classes 320 as many as the number of groups of equipment state data for learning, and classify the class of the correct answer value 310 given to each equipment state data for learning.

For example, when the preprocessor 210 groups 30 pieces of equipment state data for learning into groups A, B, and C according to the degree of similarity between the data, the class 320 to classify the correct answer value 310 according to the first The range of the correct answer value of each class 320 is determined by using the minimum correct value and the maximum correct value among the correct answer values 310 provided to the class to the third class and assigned to the learning equipment state data belonging to each group, as shown in FIG. As shown in the table 330, the first class ('0 to 9'), the second class ('10 to 19'), and the third class ('20 to 29') may be set.

If 30 pieces of equipment state data for learning are grouped into groups A, B, C, and D, the pre-processing unit 210 classifies the correct answer value 310 into the class 320 according to the number of groups according to the first class to the second 4 classes 320 are provided, and the correct value range of each class 320 is, for example, the first class ('0 to 8'), the second class ('9 to 16'), the third class ( '17~24') and the 4th class ('25~29') can be changed and set.

With reference to the correct value range of the table 330, the pre-processing unit 210 classifies the correct answer value 310 given to each learning equipment state data into any one of a first class, a second class, and a third class, and each The representative correct answer values representing the correct answer values 310 classified into the class 320 can be designated as the first class ('0'), the second class ('1'), and the third class ('2'). have.

Through the above-described pre-processing process, the pre-processing unit 210 sets the correct answer value 310 given to the 30 pieces of equipment state data for learning, based on the similarity between the equipment state data for learning, any one of the first class to the third class. Pre-processing of classifying into class 320 may be performed. In this case, the preprocessor 210 may designate the representative correct answer value representing each class as an integer greater than or equal to '0', such as '0', '1', or '2'.

The classification model generating unit 220 functions to generate a classification model by learning the representative correct value representing the class into which each correct value is classified, along with the learning equipment state data.

As an example, the classification model generation unit 220 learns the equipment state data for learning and the representative correct answer value, which are data preprocessed by the preprocessor 210, in a neural network such as a convolutional neural network (CNN), A classification model can be created as a neural network.

Here, the classification model may be a model that outputs a representative correct answer value of a class in which the equipment state data of the diagnosis object is to be classified among the plurality of classes (first to third classes) when equipment state data of a diagnosis target is input have. For example, a classification model is constructed in a structure in which basic blocks composed of a convolution layer, a batch normalization layer, a max pooling layer, and an activation function are connected in multiple layers. can get

Prior to the learning process for generating a regression model to be described later, through the process of normalizing the correct answer value (GT) given to each learning equipment state data, the range of data is matched or the distribution is adjusted in processing a large amount of data It is required to increase the reliability of the data by making them similar.

To this end, the normalization processing unit 230 performs a function of calculating a normalized GT by applying a selected normalization technique to each correct answer value for each class.

That is, the normalization processing unit 230 may normalize the correct answer values classified for each class in the data pre-processing process described above using information of each class, for example, a minimum value and a maximum value.

The normalization technique is variously applicable to normalization using an average value, normalization using a median value, quantile normalization, Min-Max Normalization, and Z-Score Normalization, etc., but will be described later. In FIG. 5, calculation of a normalized value for each correct value according to Min-Max Normalization will be described as an example.

Min-max normalization is, for all features, the minimum value is converted to '0', the maximum value is converted to '1', and other values are between '0' and 1' (i.e. decimal values). As a technique for normalizing by transforming into can be converted to '0.5'.

Hereinafter, the normalization processing unit 230 will be described with reference to FIG. 5 .

5 is a diagram illustrating a normalization process for normalizing a correct answer value (GT) given to equipment state data for learning in the equipment state diagnosis system according to an embodiment of the present invention.

Referring to FIG. 5 , the normalization processing unit 230 applies the minimum-maximum normalization according to the following equation to the correct answer value (GT) 510 between '0 to 29' given to each learning equipment state data, thereby '0' It can be normalized to a decimal value between ' and '1'.

(X - MIN) / (MAX - MIN)

In FIG. 5 , a normalized value 520 obtained by performing Min-max normalization of a correct answer value (GT) 510 is shown for each class.

First, for '0 to 9' classified as the first class among the correct answer values 510, the normalization processing unit 230 converts the minimum value '0' of the first class into '0', and the maximum value of the first class '9' is converted to '1', and the correct value between '0 to 9' can be substituted for X of (X - 0) / (9 - 0) to calculate a normalized value.

Also, with respect to '10 to 19' classified as the second class among the correct values 510, the normalization processing unit 230 converts the minimum value of '10' of the second class into '0', and the maximum value of the second class '19' is converted to '1', and the correct answer value between '10 to 19' can be substituted for X of (X - 10) / (19 - 10) to calculate a normalized value.

In addition, with respect to '20 to 29' classified into the third class among the correct values 510, the normalization processing unit 230 converts the minimum value of '20' of the third class into '0', and the maximum value of the third class '29' is converted to '1', and the correct answer value between '20 to 29' can be substituted for X of (X - 20) / (29 - 20) to calculate a normalized value.

If the correct answer value is '15', the correct answer value belongs to the range '10 to 19', and thus is classified into the second class designated as the representative correct value '1'. In this case, the normalization processing unit 230 substitutes the minimum value ('10') and the maximum value ('19') among the correct value range ('10' to '19') of the second class into the above equation, (15 A normalized value is calculated as -10) / (19-10) = 0.55, and the correct answer value '15' is converted into the calculated normalized value '0.55' to be normalized.

In this way, the normalization processing unit 230 sets the remaining correct answer values (GTs) 510 excluding the maximum and minimum values for each class between '0' and '1' through the normalization process for the correct answer value (GT) 510 . can be converted to a decimal value of

The regression model generating unit 240 functions to generate a regression model by learning the calculated normalized value together with the learning equipment state data.

As an example, the regression model generator 240 performs learning processing on a neural network such as a convolutional neural network (CNN), which is data normalized by the regularization processor 230, for learning equipment state data and a normalized value, and the learned neural network A regression model can be created as

Here, the regression model may be a model that outputs a correct value of a normalized state to be given to the equipment state data of the diagnosis target, that is, a normalized value when the equipment state data of the diagnosis target is input. This regression model is also made into a structure in which basic blocks composed of, for example, a convolution layer, a batch normalization layer, a max pooling layer, an activation function, etc. are connected in multiple layers. can get

Hereinafter, the classification model generator 220 and the regression model generator 240 for building the deep learning model 400 having dual outputs will be described with reference to FIG. 4 .

4 is a diagram illustrating a learning process of generating a deep learning model having two types of outputs by learning in the facility state diagnosis system according to an embodiment of the present invention.

Referring to FIG. 4 , in the facility state diagnosis system of the present invention, an output terminal for the regression model 420 is additionally designed in addition to the output terminal from the classification model 410 , and in one neural network, the It is possible to build a deep learning model 400 having a neural network structure in the form of branching to an output end of the regression model 420 and an output end of the regression model 420 .

That is, the deep learning model of the present invention may be designed as a deep learning model 400 having a dual output structure having both an output terminal from the classification model 410 and an output terminal from the regression model 420 . Accordingly, in the diagnostic processing unit 250 to be described later, it is possible to more precisely diagnose the equipment state according to the final predicted value obtained based on the double output of the deep learning model 400 .

Specifically, in the classification model generating unit 220, the learning equipment state data and the representative correct value, which are pre-processed data, are trained in the neural network to generate the classification model 410 as a learned neural network. , the learning equipment state data and the normalized value, which are normalized data, may be trained in a neural network to generate a regression model 420 as a learned neural network.

Here, when the equipment state data of the diagnosis target is input, the classification model 410 calculates a representative correct answer value representing any one of the first to third classes to which the equipment state data of the diagnosis target is classified by deep learning. It may be a model that is output as the first prediction value of the model 400 .

In addition, the regression model 420 sets the correct value of the normalized state to be given to the equipment state data of the diagnosis target when the equipment state data of the diagnosis target is input, that is, the normalized value of the deep learning model 400 . 2 It may be a model that outputs as a predicted value.

Representative correct values '0', '1', and '2', which are the outputs (first predicted values) of the classification model 410 , are '0', ' Since an integer greater than or equal to '0', such as 1' or '2', is designated, it can be used to determine the integer part of the final predicted value, which will be described later.

In addition, the normalized value that is the output (the second predicted value) of the regression model 420 is normalized by the normalization processing unit 230, the correct answer values given to the learning equipment state data, for each class in which the correct answer value is classified. Since the other correct values except for the minimum and maximum values of stars are converted to decimal values, they can be used to determine the fractional part of the final predicted value, which will be described later.

As described above, according to the present invention, after classifying the correct answer value (GT) given to a plurality of equipment state data for learning into a plurality of classes based on the degree of similarity between the data, each equipment state data for learning represents the classified class. The classification model 410 is generated by learning the representative correct answer value to By doing so, through the design of the deep learning model 400 having a dual output, the problem that the accuracy of prediction decreases when the similarity between data in the entire section in the existing learning method tends to be low is solved, and when diagnosing the equipment condition Prediction accuracy can be further improved.

For the deep learning model 400 consisting of the classification model 410 and the regression model 420, a weight applied to each parameter of the data used for each learning is stored in a database (not shown) can be

According to an embodiment, the facility state diagnosis system 200 may include a loss value calculating unit 260 and an update processing unit 270 .

The loss value calculating unit 260 functions to obtain respective loss values for the classification model and the regression model by using a loss function.

As an example, the loss calculation unit 260 includes a loss function exemplified by Cross Entropy (CE), Mean Squared Error (MSE), and Mean Absolute Error (MAE), etc. function), it is possible to obtain respective loss values for the classification model 410 and the regression model 420 .

Here, the loss function is a function used to measure the degree to which a value predicted by the deep learning model 400 is close to an actual value, and may represent an error with respect to a correct value numerically. That is, the loss function calculates a larger value as the value predicted by the deep learning model 400 is closer to an incorrect answer, and on the contrary, a smaller value is calculated as it is closer to the correct value.

As the sum total of losses by summing the respective loss values increases, the error between the predicted value and the actual value increases. Therefore, a process of correcting the error by performing backpropagation based on the total loss is required.

To this end, the update processing unit 270 performs backpropagation based on the total loss obtained by summing the respective loss values to apply a weight to each parameter of the next learning equipment state data. function to update.

That is, the update processing unit 270 reads a weight for each parameter of the deep learning model 400 previously stored in the database from the database, and each loss function for the classification model and the regression model. ) can be used to update based on the total loss obtained.

Hereinafter, a process of calculating the total loss in the loss value calculating unit 260 will be described with reference to FIG. 7 .

7 is a diagram illustrating an example of calculating the total loss of a deep learning model having two types of outputs in the facility state diagnosis system according to an embodiment of the present invention.

Referring to FIG. 7 , the loss value calculating unit 260, as the equipment state data of the diagnosis target is input to the deep learning model 400 , the loss with respect to the first predicted value output from the classification model among the deep learning models 400 . Calculate the value according to the cross entropy (CE), calculate the loss value for the second predicted value output from the regression model among the deep learning model 400 according to the mean square error (MSE), and sum the respective loss values The total loss of the entire deep learning model 400 can be obtained through the linear combination.

Then, the update processing unit 270, while performing backpropagation based on the total loss obtained from the loss value calculating unit 260, updating the weight of the parameter stored after the learning process of the deep learning model 400 Accordingly, the error with the actual value can be minimized.

That is, the update processing unit 270 backpropagates one deep learning model 400 composed of a classification model and a regression model by using two types of loss values obtained from the loss function of the classification model and the loss function of the regression model. It can be updated by re-learning by

The diagnosis processing unit 250 is configured to use the first predicted value confirmed from the classification model and the second predicted value confirmed from the regression model for the input equipment state data of the diagnosis target for diagnosing the equipment state of the diagnosis target. It functions to output the final predicted value.

For example, the diagnosis processing unit 250 may generate, from the classification model, a representative correct answer value (eg, '0', '1', '2 ', etc.) as the first predicted value, and from the regression model, normalized values (eg, '0', '1', '0' and ' 1') as the second predicted value, the final predicted value may be output using the first and second predicted values identified from the classification model and the regression model.

That is, the diagnosis processing unit 250 identifies the correct answer range set in the class to which the equipment state data of the diagnosis target is to be classified, using the first predicted value (representative correct value), and the second predicted value (normalized value) When a value obtained by inversely applying the selected normalization technique to ? is included in the correct value range, the obtained value may be output as the final predicted value.

Since the second predicted value output from the regression model is a normalized value, the diagnostic processing unit 250 may inversely apply the normalization technique to the second predicted value to restore the second predicted value to the value of the actual predicted scale. have. For example, if the class output (representative correct value) of the classification model is '2' and the regression output (normalized value) of the regression model is '0.4', the diagnosis processing unit 250 performs the inverse application of min-max normalization by The actual value can be restored to 0.4 * (29-20) + 20 = 23.6. Since the restored value '23.6' is a value belonging to the range ('20' to '29') in which the representative correct value '2' is specified, the diagnosis processing unit 250 uses the restored value '23.6' as the final predicted value. can be printed out.

In this case, the first predicted value is used as a reference for confirming whether the value restored according to the inverse application of the normalization technique of the second predicted value (normalized value) is restored to the original scale, and the final predicted value itself is the second predicted value ( It can be obtained by inversely applying the normalization technique to the normalized value).

In another example, the diagnostic processing unit 250 determines the first predicted value as an integer part of the final predicted value, determines the second predicted value as a fractional part of the final predicted value, and divides the integer part and the fractional part Including, the final predicted value may be output.

Hereinafter, the diagnosis processing unit 250 will be described with reference to FIG. 6 .

6 is a diagram illustrating an example of obtaining a final predicted value based on the output of the learned deep learning model 400 in the facility state diagnosis system according to an embodiment of the present invention.

Referring to FIG. 6 , the diagnosis processing unit 250 uses two types of predicted values ('6' and '0.02') output from one deep learning model 400, and a final predicted value ('6.02') for equipment state diagnosis. ') can be printed.

For example, the diagnosis processing unit 250 determines the first predicted value ('6') from the classification model in the deep learning model 400 as an integer part of the final predicted value, and the deep learning model 400 from the regression model The second predicted value ('0.02') may be determined as a fractional part of the final predicted value, and the final predicted value ('6.02') may be output by adding the first and second predicted values.

As such, the diagnosis processing unit 250 combines the dual outputs output from one deep learning model 400 and outputs the final predicted value for equipment state diagnosis in the form of a real value, based on the deep learning model 400, A problem in which the accuracy of prediction is lowered when the equipment state data for various physical quantities such as temperature, humidity, and lifespan can be diagnosed more precisely, and the similarity between data tends to be low in the entire section in the existing learning method can solve

8 is a flowchart illustrating a procedure of a method for diagnosing a facility state according to an embodiment of the present invention.

The facility condition diagnosis method according to the present embodiment may be performed by the facility condition diagnosis system 200 described above.

Referring to FIG. 8 , in step 810 , the facility state diagnosis system 200 performs a similarity analysis between data on each of a plurality of equipment state data for learning.

In step 820 , the equipment state diagnosis system 200 classifies each correct answer value GT given to a plurality of equipment state data for learning into a plurality of classes based on the similarity analysis result.

In step 830 , the facility state diagnosis system 200 generates a classification model by learning and processing the representative correct value representing the class into which each correct value is classified together with the equipment state data for learning.

In step 840, the facility state diagnosis system 200 calculates a normalized value by applying the selected normalization technique to each correct answer value.

In step 850, the equipment state diagnosis system 200 generates a regression model by learning the calculated normalized value together with the equipment state data for learning.

In steps 860 to 870 , the facility state diagnosis system 200 uses a first predicted value identified from the classification model and a second predicted value identified from the regression model with respect to the input equipment state data of the diagnosis target. Thus, the final predicted value for diagnosing the equipment state of the diagnosis target is output.

As an example of outputting the final predicted value by the combination of the first and second predicted values, in the facility state diagnosis system 200, the class output (representative correct value) of the classification model is '2', and the regression output (normalized value) of the regression model ) is '0.4', the actual value is restored to 0.4 * (29-20) + 20 = 23.6 by inverse application of min-max normalization, and the restored value '23.6' is the correct answer with the representative correct value '2' Since the value belongs to the value range ('20' to '29'), the restored value '23.6' may be output as the final predicted value.

As another example of outputting the final predicted value by the combination of the first and second predicted values, referring to FIG. 6 , the facility state diagnosis system 200 calculates the first predicted value ('6') from the classification model in the deep learning model. Determine as the integer part of the final predicted value, determine the second predicted value ('0.02') from the regression model in the deep learning model as the fractional part of the final predicted value, and by summing the first and second predicted values, the final predicted value (' 6.02') can also be printed.

In this way, according to the final predicted value ('real value') obtained by combining two types of outputs output from the deep learning model (classification model, regression model) built by the learning method proposed in the present invention, the equipment of the diagnosis target It is possible to further improve the diagnostic accuracy of the state data.

The method according to the embodiment may be implemented in the form of program instructions that can be executed through various computer means and recorded in a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, etc. alone or in combination. The program instructions recorded on the medium may be specially designed and configured for the embodiment, or may be known and available to those skilled in the art of computer software. Examples of the computer-readable recording medium include magnetic media such as hard disks, floppy disks and magnetic tapes, optical media such as CD-ROMs and DVDs, and magnetic such as floppy disks. - includes magneto-optical media, and hardware devices specially configured to store and execute program instructions, such as ROM, RAM, flash memory, and the like. Examples of program instructions include not only machine language codes such as those generated by a compiler, but also high-level language codes that can be executed by a computer using an interpreter or the like. The hardware devices described above may be configured to operate as one or more software modules to perform the operations of the embodiments, and vice versa.

Software may comprise a computer program, code, instructions, or a combination of one or more thereof, which configures a processing device to operate as desired or is independently or collectively processed You can command the device. The software and/or data may be any kind of machine, component, physical device, virtual equipment, computer storage medium or apparatus, to be interpreted by or to provide instructions or data to the processing device. , or may be permanently or temporarily embody in a transmitted signal wave. The software may be distributed over networked computer systems and stored or executed in a distributed manner. Software and data may be stored in one or more computer-readable recording media.

As described above, although the embodiments have been described with reference to the limited drawings, those skilled in the art may apply various technical modifications and variations based on the above. For example, the described techniques are performed in an order different from the described method, and/or the described components of the system, structure, apparatus, circuit, etc. are combined or combined in a different form than the described method, or other components Or substituted or substituted by equivalents may achieve an appropriate result.

Therefore, other implementations, other embodiments, and equivalents to the claims are also within the scope of the following claims.

200: equipment status diagnosis system
210: preprocessor
220: classification model generation unit
230: normalization processing unit
240: regression model generator
250: diagnostic processing unit
260: loss value calculation unit
270: update processing unit

Claims (10)

classifying each correct answer value (GT) given to the plurality of equipment state data for learning into a plurality of classes based on the similarity between the plurality of equipment state data for learning;
generating a classification model by learning and processing a representative correct value representing a class in which each correct value is classified, together with the learning equipment state data;
calculating a normalized value by applying a selected normalization technique to each correct answer value;
generating a regression model by learning the calculated normalized value together with the learning equipment state data; and
Regarding the equipment status data to be inputted,
outputting a final predicted value for diagnosing the equipment state of the diagnosis target by using the first predicted value confirmed from the classification model and the second predicted value confirmed from the regression model
A facility condition diagnosis method comprising a. According to claim 1,
The step of classifying into the plurality of classes,
performing similarity analysis between the data for each of the plurality of learning equipment state data;
setting a correct answer range for each of the plurality of classes based on the similarity analysis result; and
Classifying the correct answer value given to each learning equipment state data to belong to any one class among the plurality of classes in consideration of the correct answer value range
A facility condition diagnosis method comprising a. According to claim 1,
identifying, from the classification model, a representative correct value representing a class to which the equipment state data of the diagnosis target is to be classified, as the first predicted value, among the plurality of classes; and
confirming, from the regression model, a normalized value for a correct answer value to be given to the equipment state data of the diagnosis target as the second predicted value;
further comprising,
Outputting the final predicted value comprises:
identifying a correct answer value range set in a class to which the equipment state data of the diagnosis target is to be classified, using the first predicted value; and
outputting the obtained value as the final predicted value when a value obtained by inversely applying the selected normalization technique to the second predicted value is included in the correct value range
A facility condition diagnosis method comprising a. According to claim 1,
identifying, from the classification model, a representative correct value representing a class to which the equipment state data of the diagnosis target is to be classified, as the first predicted value, among the plurality of classes; and
confirming, from the regression model, a normalized value for a correct answer value to be given to the equipment state data of the diagnosis target as the second predicted value;
further comprising,
Outputting the final predicted value comprises:
determining the first predicted value as an integer part of the final predicted value;
determining the second predicted value as a fractional part of the final predicted value; and
outputting the final predicted value including the integer part and the fractional part
A facility condition diagnosis method comprising a. According to claim 1,
For the classification model and the regression model, a cross entropy (CE), a mean squared error (MSE), and a mean absolute error (MAE) are selected as any one of a loss function (loss). function) to obtain each loss value; and
A step of updating the weight to be applied to each parameter of the next equipment state data for learning by performing backpropagation based on the total loss obtained by adding up the respective loss values.
Equipment condition diagnosis method further comprising a. a preprocessor for classifying each correct answer value (GT) given to the plurality of equipment state data for learning into a plurality of classes based on the similarity between the plurality of equipment state data for learning;
a classification model generation unit for generating a classification model by learning and processing a representative correct value representing a class into which each correct value is classified together with the learning equipment state data;
a normalization processing unit for calculating a normalized value by applying a selected normalization technique to each correct answer value;
a regression model generator for generating a regression model by learning and processing the calculated normalized value together with the learning equipment state data; and
Regarding the equipment status data to be inputted,
A diagnosis processing unit for outputting a final predicted value for diagnosing the equipment state of the diagnosis target by using a first predicted value confirmed from the classification model and a second predicted value confirmed from the regression model
A facility condition diagnosis system comprising a. 7. The method of claim 6,
The preprocessor is
For each of the plurality of learning equipment state data, a similarity analysis between data is performed, and based on the similarity analysis result, a correct answer value range is set for each of the plurality of classes, and the correct answer value range is taken into account for each learning Classifying the correct answer value given to the equipment state data to belong to any one of the plurality of classes
Equipment health diagnosis system. 7. The method of claim 6,
The diagnostic processing unit,
From the classification model, a representative correct value representing a class in which the equipment state data of the diagnosis target is to be classified among the plurality of classes is identified as the first predicted value;
From the regression model, a normalized value for a correct answer value to be given to the equipment state data of the diagnosis target is confirmed as the second predicted value,
Identify the correct answer value range set in the class to which the equipment state data of the diagnosis target is to be classified using the first predicted value,
When a value obtained by inversely applying the selected normalization technique to the second predicted value is included in the correct value range, the obtained value is output as the final predicted value
Equipment health diagnosis system. 7. The method of claim 6,
The diagnostic processing unit,
From the classification model, a representative correct value representing a class in which the equipment state data of the diagnosis target is to be classified among the plurality of classes is identified as the first predicted value;
From the regression model, a normalized value for a correct answer value to be given to the equipment state data of the diagnosis target is confirmed as the second predicted value,
determining the first predicted value as an integer part of the final predicted value, determining the second predicted value as a fractional part of the final predicted value, and including the integer part and the fractional part to output the final predicted value
Equipment health diagnosis system. 7. The method of claim 6,
For the classification model and the regression model, a loss function selected as any one of cross entropy (CE), mean squared error (MSE), and mean absolute error (MAE) function), a loss value calculating unit that calculates each loss value; and
An update processing unit that updates the weight to be applied to each parameter of the next learning equipment state data by performing backpropagation based on the total loss obtained by adding up each loss value
Equipment condition diagnosis system further comprising a.

Images