KR102383139B1 - Industrial facility failure prediction diagnosis data learning labeling system and method based on artificial intelligence - Google Patents

Abstract

The present invention discloses a data learning labeling system for predicting and diagnosing an industrial facility failure based on artificial intelligence and a method thereof. The system includes a data labeling module that preprocesses, classifies, and labels data so that data measured from an industrial facility can be used as learning data for abnormal symptom detection or failure type classification.

Description

Data learning labeling system and method for predictive diagnosis of industrial equipment failure based on artificial intelligence

The present invention relates to a data learning labeling system and method for predictive diagnosis of industrial equipment failure based on artificial intelligence, and more particularly, data measured in industrial equipment can be used as learning data for detecting abnormal symptoms or classifying failure types. It is about technology for preprocessing, classifying, and labeling data so that

Data labeling refers to pre-processing, classifying, and processing various data required for machine learning, such as images, images, and text, according to the purpose so that the AI (Artificial Intelligence) module can learn.

Conventionally, it is possible to label water quality data for water quality analysis as in the patent literature described in the prior art, and to label measurement data of industrial facilities to detect abnormal signs of industrial facilities, and data labeling can be utilized for various fields. there is.

The conventional predictive maintenance system in the field of facility operation or maintenance uses AI-based machine learning to detect anomalies according to differences from normal data, and failure diagnosis to diagnose the cause according to the failure pattern in the event of a facility failure. Data is pre-processed and classified according to the purpose of the system, and the process of learning and utilizing data through labeling is developed as an algorithm and applied to facility management.

However, conventionally, in the case of equipment measurement data, the normal operation range is frequently changed depending on the external environment such as atmospheric temperature, atmospheric pressure, sea water temperature, and internal environment such as equipment deterioration, operation mode, and maintenance cycle. Additional data learning is required for this, and there is a problem in that it is not possible to learn and use data after the cause analysis in the event of a facility failure.

Korean Patent Publication No. 10-2021-0080772

In order to solve the above problems, the present invention labels normal operation learning data and learning data for each failure type, and through data labeling, artificial intelligence-based industry for diagnosing industrial equipment to secure data necessary for data learning and classification of failure types Provided are a data learning labeling system and method for predictive diagnosis of equipment failure.

The data learning labeling system for predicting and diagnosing the failure of industrial equipment based on artificial intelligence according to an embodiment of the present invention for the above-mentioned problem to be solved, the learning range of the learning data based on the normal operation period and the set output of the industrial equipment Normal operation data labeling that selects, filters unnecessary data of industrial equipment in the learning range using reference signals for data filtering, and labels normal operation learning data for creating a learning model based on the time-series characteristics of the filtered data Including the module 110 and the data labeling module 120 for each failure type, which sets the failure type and cause, and labels the learning data for each failure type for generating a learning model based on the characteristics of the cause for each failure type, through data labeling It is characterized by securing data necessary for data learning and classification of failure types.

The normal operation data labeling module determines the data similarity between the data for additional learning and the learning model during additional learning of the normal operation data, and labels only additional data above the set similarity level when determining the similarity, and the failure type data labeling module is an industry It can be characterized by determining the similarity between the previously learned model and additional data when an abnormal symptom or failure of the facility occurs, and performing labeling by classifying real-time data above the set similarity level at the time of similarity determination by failure type.

The normal operation data labeling module extracts time-series features of the normal operation learning data using a 1D-CNN algorithm, and the data labeling module for each failure type uses a 1D-CNN algorithm and a 2D-CNN algorithm of learning data for each failure type. It may be characterized by extracting time-series features and image features.

Data for artificial intelligence-based industrial equipment failure prediction diagnosis including a data labeling module 100 , a data learning module 200 , an anomaly detection module 300 and a failure type classification module 400 according to an embodiment of the present invention Method of operation of the learning labeling system, the data labeling module pre-processing, classifying and labeling the learning data for diagnosing the failure prediction of industrial equipment; generating, by the data learning module, a learning model by learning the labeled learning data; The abnormality detection module inputs real-time data of industrial equipment to a learning data-based learning model of the normal operation section, and detects abnormal symptoms of industrial equipment based on supervised learning that compares real-time data with learning data of the normal operation section, , early detection of abnormal signs of industrial equipment based on unsupervised learning that learns regularity of real-time data as input values, and when the failure type classification module detects abnormal signs Through data labeling, including the step of learning, generating future data to predict the operation of industrial facilities for a set future time after the detection of anomalies, and classifying the failure type by comparing the future data with the learning data for each failure type It is characterized by securing data necessary for data learning and classification of failure types.

The operation method of the data learning labeling system for predicting and diagnosing the failure of an artificial intelligence-based industrial facility according to an embodiment of the present invention defines the type of new data when a type of new data that does not match the learning data for each failure type occurs, and , it may be characterized by performing labeling and learning on new data.

The present invention automatically performs a series of processes divided into pre-processing-classification-labeling-learning of learning data when managing a data learning model for diagnosing abnormal signs and failure causes of equipment, thereby improving user convenience in managing learning data. Normal operation learning data for detecting abnormal symptoms and learning data for each failure type for equipment failure cause analysis can be obtained, and equipment failure cause analysis can be performed efficiently.

1 is a block diagram illustrating a data learning labeling system for predictive diagnosis of industrial equipment failure based on artificial intelligence for industrial equipment diagnosis according to an embodiment of the present invention.
FIG. 2 is a block diagram illustrating the data labeling module of FIG. 1 in detail.
3 is an example illustrating a process of automatically selecting a learning range of data.
4 is an example illustrating a process of filtering outlier data.
5 is an example illustrating an operation method of the 1D-CNN algorithm.
6 illustrates an example of labeling normal operation data.
7 is an example illustrating an operation method of the 2D-CNN algorithm.
8 is an example illustrating an STFT transformation method.
9 shows an example of labeling data for each failure type.
10 is a block diagram illustrating the data learning module of FIG. 1 in detail.
11 is a block diagram illustrating the anomaly detection module of FIG. 1 in detail.
12 is a block diagram illustrating in detail the failure type classification module of FIG. 1 .
13 is a diagram illustrating an example of classifying failure types with respect to temperature.
14 is a block diagram illustrating in detail the failure type classification module of FIG. 1 .
15 is an example illustrating a chart analysis screen.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings and contents described in the accompanying drawings, but the present invention is not limited or limited by the embodiments.

1 is a block diagram illustrating a data learning labeling system for an artificial intelligence-based industrial facility failure prediction diagnosis according to an embodiment of the present invention. The data learning labeling system 10 labels and learns data measured in industrial facilities. It creates data and builds a learning model by learning the learning data.

Industrial equipment data may include measurement elements related to the internal environment such as temperature, pressure, flow rate, vibration, sound and current of industrial equipment, and may include measurement elements related to the external environment, depending on the operating conditions of industrial equipment. data may be included. Data labeling means displaying similar data patterns as labels through comparison between input values or between input values and a learning model.

The data learning labeling system 10 detects abnormal signs of industrial equipment with artificial intelligence using real-time data and learning models measured in real time with respect to industrial equipment, and what type of failure occurs when detecting abnormal signs It is possible to determine the future state of industrial equipment when it proceeds to the identified failure. Artificial intelligence (AI) is the realization of human thinking and learning ability through computer programs, often called AI. The real-time data is data measured in real time with respect to industrial equipment.

Artificial intelligence technology may include deep learning and machine learning technology. Deep learning is to perform machine learning learning using an artificial neural network (ANN) with multiple layers, also called deep learning. It can be called a learning algorithm that can.

Machine learning is divided into supervised learning and unsupervised learning. Supervised learning is a method of finding output values corresponding to input values with reference to learning data, and unsupervised learning uses training data with only input values to find regularity of input values and outputs the found results.

The data learning labeling system 10 may include a data labeling module 100 , a data learning module 200 , an anomaly detection module 300 , a failure type classification module 400 , and a data visualization module 500 .

The present invention can be applied and utilized in various industrial facilities, and a combustor and a compressor among industrial facilities will be described as an example.

FIG. 2 is a block diagram illustrating the data labeling module of FIG. 1 in detail. The data labeling module 100 includes a normal operation data labeling module 110 and a data labeling module 120 for each failure type.

The normal operation data labeling module 110 selects the learning range of the learning data based on the normal operation period and the set output of the industrial equipment, and filters unnecessary data of the industrial equipment in the learning range using the reference signal for data filtering, , and label the normal driving learning data for generating a learning model based on the time-series characteristics of the filtered data.

3 is an example showing the process of automatically selecting the learning range of data, the normal operation data labeling module 110 can set the normal operation period as a condition, and can automatically filter the data according to the set condition, The learning range of the learning data can be selected based on the set output of the industrial equipment.

4 is an example illustrating a process of filtering outlier data. The normal operation data labeling module 110 filters unnecessary data of industrial equipment in the learning range using a reference signal or reference signal for data filtering. The reference signal is a signal related to unnecessary data of an industrial facility, and the unnecessary data of the industrial facility is, for example, data related to a stop section, O/H section, or test section of an industrial facility. The reference signal may be a digital signal or a current signal.

The normal operation data labeling module 110 may map the reference signal to the learning data in which the learning range is selected, and filter outlier data matching the reference signal.

5 is an example showing an operation method of the 1D-CNN algorithm. The normal driving data labeling module 110 labels normal driving learning data for generating a learning model based on the time-series characteristics of the filtered data. The normal operation data labeling module 110 may utilize the 1D-CNN algorithm in the labeling process. The normal operation data labeling module 110 may extract time-series characteristics of data through a 1D-CNN algorithm.

The original data is placed in two dimensions, a rectangular window called a filter (or feature detector) is covered, and then a new value is created by moving it. At this time, the moving process is called convolution, and if the filter moves in one direction, it is 1D-CNN, and if it moves in two directions, it is 2D-CNN.

6 shows an example of labeling the normal operation data, the normal operation data labeling module 110, in the case of the measurement data of the equipment, the external environment such as atmospheric temperature, atmospheric pressure, sea water temperature, equipment deterioration, operation mode, maintenance cycle Since the range of normal operation is frequently changed depending on the internal environment, etc., there may be situations in which additional data learning is required for the initially configured data learning model. In determining the similarity, only additional data above the set similarity level can be labeled. The similarity can be set by a user's input, and is a value set in % units to determine how similar the additional data is to the learning data of the learning model.

Normal operation learning data with leveling is to learn the normal operation section data to detect the failure prognosis, that is, abnormal signs of equipment when abnormal operation occurs. It is used to learn normal operation data.

The present invention can provide convenience for the user's learning data management by pre-processing, classifying and labeling data under complex conditions such as normal operation period, standard output of industrial equipment, data filtering, and labeling and similarity related to characteristics, Normal driving learning data for detecting abnormal symptoms can be secured.

Since the present invention labels learning data based on time-series characteristics, it should proceed in the order of normal operation period, standard output of industrial equipment, and data filtering.

The data labeling module 120 for each failure type sets the failure type and cause, and labels the learning data for each failure type for generating a learning model based on the characteristics of the cause for each failure type. The cause of the failure mode may vary by power facility, such as misalignment of the shaft, poor oil viscosity, and wear of bearings.

The labeling learning data for each failure type is used to analyze the types of failures when abnormal symptoms of equipment occur. used in the system.

7 is an example showing an operation method of the 2D-CNN algorithm, and the learning data for each failure type may include operation data and vibration/sound/current data related to the operation of power facilities. The data labeling module 120 for each failure type may utilize a 1D-CNN algorithm to extract characteristics of driving data, and may utilize a 2D-CNN algorithm to extract characteristics of vibration/sound/current data.

8 is an example showing the STFT conversion method, and the failure type data learning module 120 images vibration, sound and current signals related to industrial equipment from the failure type learning data through SFTF conversion, and using a 2D-CNN algorithm. In this way, the characteristics of the image data can be extracted.

In STFT (Spectrogram), the x-axis means the time axis (unit: frame), and the y-axis means the frequency. STFT is a Fourier transform with the passage of time that expresses a three-dimensional two-dimensional value by expressing a value of a frequency per time with a color according to the size of the value. STFT is widely used in speech applications with effects such as noise reduction, pitch detection, and pitch shift, and in the present invention, it is used for feature extraction using a 2D-CNN algorithm.

9 shows an example of labeling data for each failure type, the failure type data labeling module 120 determines the data similarity between the real-time data and the learning model generated in real time when an abnormal symptom or failure of the industrial equipment occurs, and , it is possible to perform labeling by classifying real-time data above the set similarity level by failure type when determining similarity.

Normal driving data labeling module 110 extracts time-series features of normal driving learning data using 1D-CNN algorithm

10 is a block diagram illustrating the data learning module of FIG. 1 in detail. The data learning module 200 includes an abnormal symptom data learning module 210 and a failure type data learning module 220 .

The abnormal symptom data learning module 210 may generate a GRU (Gated Recurrent Unit)-VAE learning model for abnormal symptom detection by learning the time-series characteristics of the normal driving learning data. The failure type data learning module 220 may generate a 1D-CNN learning model for failure type classification by learning the characteristics of the learning data for each failure type.

The GRU-VAE learning model is an unsupervised learning-based model, and the 1D-CNN learning model is a supervised learning-based model. The GRU-VAE learning model consists of an unsupervised learning-based model to detect abnormal symptoms early. The 1D-CNN learning model is a model built to classify failure types based on supervised learning using labeled learning data.

To apply the GRU-VAE learning model or algorithm, auto-encoder and variational auto-encoder technologies are required. An autoencoder is simply a neural network that copies its input to its output. In some ways, it looks like a simple neural network, but it makes it a difficult neural network by constraining the network in various ways. For example, as shown in the figure below, data is compressed (dimension reduced) by making the number of neurons in the hidden layer smaller than that of the input layer, or the original input can be restored after adding noise to the input data. There are various autoencoders, such as training the network to do this. These constraints prevent the autoencoder from simply copying the input directly to the output, and control it to learn how to represent the data efficiently.

The auto-encoder can detect an abnormal state based on the difference (reconstruction error) between the original data and the data restored through the encoder and the decoder for the input data. If the auto-encoder reduces the dimension of the input data itself and restores the reduced dimensional data, the transform auto-encoder's encoder can infer the characteristics of the data through the latent variable, the average and standard of the original data. Deviation is extracted, and it is assumed that this distribution follows a normal distribution. A latent variable is created by sampling a random value from a normal distribution and calculating it with the mean and standard deviation of the encoder's result. Since the data is restored through the decoder, the cluster strength is higher than that of the autoencoder's latent variable distribution, so the original data It may be more advantageous to create A latent variable refers to a variable that can only be measured indirectly through other variables because it cannot be measured directly.

11 is a block diagram illustrating the abnormality detection module of FIG. 1 in detail, wherein the abnormality detection module 300 inputs real-time data of industrial equipment to a learning data-based learning model of a normal operation section, real-time data and normal operation It detects abnormal signs of industrial facilities based on supervised learning that compares the learning data of each section, and detects abnormal signs of industrial facilities early based on unsupervised learning that learns the regularity of real-time data as input.

Anomaly detection module 300 can detect abnormal signs of industrial equipment early by using both supervised learning and unsupervised learning.

The anomaly detection module 300 can generate an expected value reflecting the trend according to the time-series characteristics of real-time data using the GRU-VAE learning model, and compare the residual between the real-time data and the real-time expected value to detect abnormal symptoms of industrial equipment early. can be detected in Data obtained by comparing the residuals between real-time data and real-time expected values can be used as input values for predicting future operation of industrial facilities.

12 is a block diagram illustrating the failure type classification module of FIG. 1 in detail. When an abnormal symptom is detected, the failure type classification module 400 learns past data before the abnormal symptom detection time, and after the abnormal symptom detection time It generates future data to predict the operation of industrial facilities for a set future time.

The failure type classification module 400 may generate future data for a future time (eg, about 3 to 4 hours) in which a preliminary action can be taken by using a Long Short-Term Memory (LSTM) algorithm.

The present invention can construct a model by grafting an LSTM model with improved limitations of RNN to an auto-encoder or a modified auto-encoder to generate an expected value reflecting the trend of the data due to the characteristics of the time series data. In addition, the GRU model, which has simplified the cell structure of the LSTM model, can be used instead of the LSTM model. Therefore, the following model can be constructed by applying the LSTM or GRU layer to the layer layer in the process of reducing (encoding) and restoring (decoding) the input data of the time series data.

In the present invention, an LSTM model can be used to accurately predict future motion, and a GRU model with a simplified cell structure of the LSTM model can be used to quickly and early detect anomalies.

When an abnormal symptom is detected, the failure type classification module 400 may utilize past data prior to the detection of the abnormal symptom as data comparing the residual between the real-time data and the real-time expected value. The failure type classification module 400 classifies the failure type by comparing the future data with the learning data for each failure type.

13 shows an example of classifying a failure type with respect to temperature, the failure type classification module 500 compares future data with data learned for each failure type and performs an algorithm such as 1D-CNN as shown in FIG. 14 It is possible to provide failure type analysis results and weight information through The weight information indicates in the form of information that a plurality of measurement elements are weighted as a factor of main failure.

The failure type classification module 500 may define a type of new data when a type of new data that does not match the learning data occurs, and may perform labeling and learning on the new data.

The present invention can further improve the reliability of learning data and improve the accuracy of industrial equipment diagnosis by providing additional data learning according to environmental changes of industrial equipment and data learning of new failure types.

14 is a block diagram showing the failure type classification module of FIG. 1 in detail, and the data visualization module 500 can provide information on the abnormal detection alarm history and failure type classification history of industrial facilities such as combustors and compressors, and , alarm data accumulation and statistical analysis functions, data analysis charts and visualization functions can be provided together.

15 is an example of a chart analysis screen, and the data visualization module 500 may provide various analysis tools for inquiring about calculated data, such as real-time data, real-time expected values, and future data.

The present invention can improve the efficiency of industrial facilities, create added value, and create new business through preemptive response to the 4th industrial revolution by reducing the loss cost of industrial facilities (facility defect loss, stop loss, excessive maintenance, etc.) indirect effects can be expected.

In the related art, when the normal operation range is changed, an inaccurate alarm may be generated due to inconsistency with respect to the learning data, and there is a problem in that the frequency of occurrence of a false alarm increases depending on the external environment. However, the present invention can solve the conventional problems by providing data labeling regarding changes in the normal operation range and changes in the external environment.

Conventionally, when equipment failure occurs, data of the corresponding failure occurrence section cannot be utilized, but the present invention can solve the conventional problems by providing a labeler of failure type classification for supervised learning of failure data.

10: data learning labeling system 100: data labeling module
110: normal operation data labeling module 120: data labeling module for each failure type
200: data learning module 210: abnormal symptom data learning module
220: failure type data learning module 300: anomaly detection module
400: failure mode classification module 500: data visualization module

Claims (5)

Select the learning range of the learning data based on the normal operation period and the set output of the industrial equipment, filter unnecessary data of the industrial equipment in the learning range using the reference signal for data filtering, and check the time series characteristics of the filtered data Normal driving data labeling module 110 for labeling normal driving learning data for creating a learning model based on
Including a data labeling module 120 for each failure type that sets the failure type and cause, and labels the learning data for each failure type for creating a learning model based on the characteristics of the cause for each failure type,
It is characterized by securing the data necessary for data learning and classification of failure types through data labeling,
The normal driving data labeling module determines the data similarity between the data for additional learning and the learning model when additional learning about the normal driving data, and labels only additional data above the set similarity when determining the similarity,
The data labeling module for each failure type determines the similarity between the previously learned model and additional data when an abnormal symptom or failure of industrial equipment occurs, and performs labeling by classifying real-time data above the set similarity level when determining similarity by failure type. with
The normal driving data labeling module extracts time-series features of normal driving learning data using 1D-CNN algorithm,
The data labeling module for each failure type uses a 1D-CNN algorithm and a 2D-CNN algorithm to extract time-series features and image features of learning data for each failure type. labeling system. delete delete Data labeling module 100, data learning module 200, anomaly detection module 300, and an operation method of a data learning labeling system for an artificial intelligence-based industrial equipment failure prediction diagnosis including a failure type classification module 400 in
Pre-processing, classifying, and labeling, by the data labeling module, learning data for diagnosing failure prediction of industrial equipment;
generating, by the data learning module, a learning model by learning the labeled learning data;
The abnormality detection module inputs real-time data of industrial equipment into a learning model based on learning data in the normal operation section, and detects abnormal signs of industrial equipment based on supervised learning that compares real-time data with learning data in the normal operation section, , early detection of abnormal signs of industrial facilities based on unsupervised learning that learns the regularity of real-time data as input; and
When the failure type classification module detects an abnormal symptom, it learns the past data before the abnormal symptom detection time, and generates future data for predicting the operation of industrial facilities for a set future time after the abnormal symptom detection time, future data and failure Including the step of classifying failure types by comparing the learning data by type,
An operation method of a data learning labeling system for predicting and diagnosing failures in industrial facilities based on artificial intelligence, characterized in that data required for data learning and classification of failure types is secured through data labeling. 5. The method of claim 4,
Artificial intelligence-based industrial equipment failure for industrial equipment diagnosis, characterized in that when a new data type that does not match the learning data for each failure type occurs, the new data type is defined, and labeling and learning are performed on the new data Operation method of data learning labeling system for predictive diagnosis.

Images