KR101842347B1 - Plant system, and fault detecting method thereof - Google Patents

Abstract

An automatic learning system and method for modeling a plant with automatic learning for plant fault detection is disclosed.
An automatic learning system for detecting abnormality of a plant includes a data collection unit for collecting sensor data measured in real time through each measurement sensor with respect to a plant; A learning data generation unit that learns steady state data from the collected sensor data to generate learning data; A plant model unit for modeling the generated learning data to generate a plant model; And a learning determination unit for determining whether or not to learn the currently measured sensor data through the measurement sensor. The learning data generation unit generates a learning data based on the currently measured sensor data, The plant modeling unit receives the new learning data at a predetermined cycle and performs plant modeling to update the plant model.
According to the present invention, it is possible to automatically detect the measured data of the plant and detect the failure (abnormality) of the plant.

Description

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an automatic learning system and method for detecting a plant anomaly,

The present invention relates to an automatic learning system and method for detecting a plant anomaly, and more particularly, to an automatic learning system and method for detecting an anomaly of a plant by determining whether to learn sensor data measured at a current point of time, The present invention relates to an automatic learning system and method for detecting a plant anomaly, which can generate new learning data by combining new learning data and model new learning data at a predetermined cycle.

Generally, large plants such as power plants or chemical plants are operated in complex operation with hundreds of machines and electrical equipments of various kinds and operated in a central control room called main control room. The number of people working here is 2-10, which is decreasing due to the recent improvement in corporate competitiveness and productivity. In particular, the number of facilities to be managed and controlled per worker has increased dramatically in the case of thermal power plants where dozens of workers have been working in the past. Accordingly, the operation and operation method has been developed from the operation method of the local panel to the operation method of the main control panel in the past. Recently, the operation method using the computer-based MMI (Man-Machine Interface) Respectively.

However, in the case of skilled workers with decades of work experience, they are familiar with various situations and understand the complex internal control logic, but most of the inexperienced workers are often referred to as related reference books (Operating procedures, supplier design data, internal books, etc.). This type of operating environment is extremely vulnerable to the operation of plants that require quick and stable action.

In addition, it also displays basic operation information in a driving operation console (console) for displaying driving information, and does not sufficiently display information necessary for actual operation in various situations.

Therefore, there is a need for a system that can provide appropriate information related to operation outside the operation console of the main control panel providing only basic information, and provide appropriate information according to each situation.

In other words, there is a need for a technique to warn a dangerous state when a plant-related operation-related variable deviates from a normal operation state, and to take prompt action.

Korean Registered Patent No. 1065767 (registered on September 09, 2011)

According to an aspect of the present invention, there is provided an apparatus and method for determining whether to learn sensor data measured at a current point of time, The present invention also provides an automatic learning system and method for detecting anomaly of a plant by modeling new learning data at a predetermined cycle to perform automatic learning.

According to another aspect of the present invention, there is provided an automatic learning system for detecting a plant abnormality, comprising: a data collection unit for collecting sensor data measured in real time through each measurement sensor with respect to a plant; A learning data generation unit that learns steady state data from the collected sensor data to generate learning data; A plant model unit for modeling the generated learning data to generate a plant model; And a learning determination unit for determining whether or not to learn the currently measured sensor data through the measurement sensor. The learning data generation unit generates a learning data based on the currently measured sensor data, The plant modeling unit receives the new learning data at a predetermined cycle and performs plant modeling to update the plant model.

The sensor data measured in real time is used as an actual measurement value. The learning data is used as a predictive value. A value obtained by subtracting a predictive value from an actual measurement value is generated as a residual. When the generated residual value exceeds an allowable value, And an alarm unit for outputting an early warning by the alarm unit.

The learning determination unit may be configured to determine whether the sensor data is measured through each measurement sensor when the production power of the plant is greater than or equal to a specific power or when the early warning is not output from the alarm unit according to the alarm logic, And determines to learn the currently measured sensor data when an alarm is not generated for each measurement sensor on a predetermined time basis.

The learning data generation unit may generate the steady state data by deleting the abnormal data through the pre-processing process on the collected sensor data, and may generate the learning data by learning the generated steady state data.

A diagnostic database storing the failed cases in the plant as diagnostic data; And a failure diagnosis unit for diagnosing a failure of the plant using the diagnostic logic based on the diagnostic database and for tracking the root cause.

According to another aspect of the present invention, there is provided an automatic learning method for detecting a plant abnormality, comprising the steps of: (a) collecting sensor data measured through each measurement sensor for a data collection unit; (b) the learning data generation unit learns steady state data from the collected sensor data to generate learning data; (c) generating a plant model by modeling the generated learning data by a plant model unit; (d) determining whether the learning determination unit learns the currently measured sensor data through the measurement sensor; (e) generating a new learning data by combining the currently measured sensor data with existing learning data if it is determined that the learning data is to be learned as a result of the determination by the learning determination unit; And (f) the plant model unit receives the new learning data at a predetermined cycle and performs plant modeling to update the plant model.

The step (c) may further include the steps of: generating sensor data measured in real time as an actual value, generating a value obtained by subtracting a predicted value from a measured value as a residual, with the learning data as a predicted value, And outputting an early warning by the alarm logic in the event of a failure.

In the step (d), when the learning data is measured by each of the measurement sensors when the production power of the plant is higher than or equal to a specific power, or when the early warning according to the alarm logic is output Or if the alarm is not generated on the basis of a predetermined time basis for each of the measurement sensors, it is determined to learn the currently measured sensor data.

In the step (b), the learning data generation unit may generate steady state data by deleting the abnormal data through a pre-processing process on the collected sensor data, and may learn the generated steady state data, Thereby generating learning data.

In the step (c), the fault diagnosis unit diagnoses the fault of the plant using the diagnosis logic based on the diagnosis database storing the faulty cases as the diagnosis data in the plant, and tracks the root cause And < / RTI >

According to the present invention, it is possible to acquire sensor data of various periods occurring in a plant in real time, and to automatically update the plant model by learning sensor data for prediction of anomalous indications.

In addition, we can capitalize diagnostic know-how using diagnostic logic and database, and provide diagnostic guide system.

In addition, prediction accuracy can be improved by always maintaining a reliable prediction model.

In addition, it is possible to provide an infrastructure system for capturing plant anomalies, to provide a diagnostic system for anomalous indications regardless of the operator's expertise, and to maximize operational efficiency by reducing the worker's time and effort by hand have.

1 is a block diagram schematically showing functional blocks of a plant anomaly detection system according to an embodiment of the present invention.
FIG. 2 is a flowchart illustrating an automatic learning method for detecting a plant anomaly according to an embodiment of the present invention. Referring to FIG.
FIG. 3 is a diagram illustrating a process of outputting an alarm from sensor data of a plant according to an embodiment of the present invention, and an automatic learning process for detecting a plant anomaly.
FIG. 4 is a diagram illustrating an ensemble learning process based on the model of the model and the model of non-parametricity according to the present invention.
FIG. 5 is a diagram illustrating an example of performing ensemble learning by combining an aqueous model and a non-parametric model according to an embodiment of the present invention.
6 is a diagram illustrating an example of performing data processing on sensor data measured at present according to an embodiment of the present invention.
7 is a diagram illustrating an example of generating new learning data based on a characteristic by combining sensor data currently measured and existing learning data according to an embodiment of the present invention.
FIG. 8 is a diagram showing a result of calculating a predicted value with high accuracy according to a combination of an aqueous model and a non-parametric model according to an embodiment of the present invention.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings, which will be readily apparent to those skilled in the art to which the present invention pertains. The present invention may be embodied in many different forms and is not limited to the embodiments described herein.

In order to clearly illustrate the present invention, parts not related to the description are omitted, and the same or similar components are denoted by the same reference numerals throughout the specification.

Throughout the specification, when a part is referred to as being "connected" to another part, it includes not only "directly connected" but also "electrically connected" with another part in between . Also, when an element is referred to as "comprising ", it means that it can include other elements as well, without departing from the other elements unless specifically stated otherwise.

If any part is referred to as being "on" another part, it may be directly on the other part or may be accompanied by another part therebetween. In contrast, when a section is referred to as being "directly above" another section, no other section is involved.

The terms first, second and third, etc. are used to describe various portions, components, regions, layers and / or sections, but are not limited thereto. These terms are only used to distinguish any moiety, element, region, layer or section from another moiety, moiety, region, layer or section. Thus, a first portion, component, region, layer or section described below may be referred to as a second portion, component, region, layer or section without departing from the scope of the present invention.

The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the invention. The singular forms as used herein include plural forms as long as the phrases do not expressly express the opposite meaning thereto. Means that a particular feature, region, integer, step, operation, element and / or component is specified and that the presence or absence of other features, regions, integers, steps, operations, elements, and / It does not exclude addition.

Terms indicating relative space such as "below "," above ", and the like may be used to more easily describe the relationship to other portions of a portion shown in the figures. These terms are intended to include other meanings or acts of the apparatus in use, as well as intended meanings in the drawings. For example, when inverting a device in the figures, certain portions that are described as being "below" other portions are described as being "above " other portions. Thus, an exemplary term "below" includes both up and down directions. The device can be rotated by 90 degrees or rotated at different angles, and terms indicating relative space are interpreted accordingly.

Unless otherwise defined, all terms including technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. Commonly used predefined terms are further interpreted as having a meaning consistent with the relevant technical literature and the present disclosure, and are not to be construed as ideal or very formal meanings unless defined otherwise.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings so that those skilled in the art can easily carry out the present invention. The present invention may, however, be embodied in many different forms and should not be construed as limited to the embodiments set forth herein.

1 is a block diagram schematically showing functional blocks of a plant anomaly detection system according to an embodiment of the present invention.

1, an automatic learning system 100 for detecting a plant abnormality according to the present invention includes a data collection unit 110, a learning data generation unit 120, a plant model unit 130, a learning determination unit 140 An alarm unit 150, and a diagnostic database 160. The diagnosis unit 170 includes a failure diagnosis unit 170,

The data collecting unit 110 collects sensor data measured in real time through each measurement sensor for the plant.

The learning data generation unit 120 learns steady state data from the collected sensor data to generate learning data.

The plant modeling unit 130 generates a plant model by modeling the generated learning data.

The learning determination unit 140 determines whether to learn the currently measured sensor data through the measurement sensor.

The learning data generation unit 120 generates new learning data by combining the currently measured sensor data with existing learning data and transmits the new learning data to the plant model unit 130 as a result of the determination by the learning determination unit 120, The unit 130 receives new learning data at a predetermined cycle and performs plant modeling to update the existing plant model.

The alarm unit 150 generates sensor data measured in real time as measured values, generates a value obtained by subtracting the predicted value from the measured value as the predicted value, and outputs the alarm logic (Alarm) when the generated residual value exceeds the allowable value. Logic) to output an early warning.

If the sensor data is measured through each measurement sensor when the production power of the plant is higher than a specific power or the early warning is not output according to the alarm logic from the alarm unit 150 Or if the alarm is not generated for each measurement sensor on a predetermined time basis, it is determined to learn the currently measured sensor data.

In addition, the learning data generation unit 120 may generate the steady state data by deleting the abnormal data through the pre-processing process on the collected sensor data, and may generate the learning data by learning the steady state data.

The diagnostic database 160 stores malfunctioning cases as diagnostic data in the plant.

The failure diagnosis unit 170 diagnoses the failure of the plant using the diagnosis logic based on the diagnosis database 160 and tracks the root cause.

The automatic learning system 100 for detecting a plant abnormality according to the present invention may be configured to store, in addition to the components shown in FIG. 1, sensor data collected by sensing from respective sensors as original data, And a plant database for storing the steady state data in which the abnormal data is deleted through the process as reference data or modeling the learning data and storing the model data as model data. The system may further include a prediction database (Prediction DB) for storing the plant model generated by modeling the learning data as a prediction model.

FIG. 2 is a flowchart illustrating an automatic learning method for detecting a plant anomaly according to an embodiment of the present invention. Referring to FIG.

Referring to FIG. 2, the automatic learning system 100 for detecting a plant abnormality according to the present invention first acquires sensor data measured by each measurement sensor with respect to the plant by the data collection unit 110 (S210).

That is, the data collecting unit 110 collects corresponding sensor data, for example, every 5 minutes through each measurement sensor installed in each area of the plant (Data Collecting), and collects the sensor data from each sensor according to a batch scheduler Sensor data is sequentially received every five minutes and stored in the database as original data (raw data). Here, the batch scheduler is an apparatus for arranging operation sequences for operating the measurement sensors installed in the respective regions of the plant sequentially from the measurement sensors in a specific region, and for collecting sensor data from the respective measurement sensors .

Next, the learning data generation unit 120 learns steady state data from the collected sensor data to generate learning data (S220).

That is, the learning data generation unit 120 generates steady state data by deleting the abnormal data through a pre-processing process on the sensor data collected in real time, and automatically learns the preprocessed steady state data Learning is performed to generate learning data.

Here, in the pre-processing step, the plant data collected in real time are taken as measured values, the learned steady-state data is used as a predicted value, and a value obtained by subtracting the predicted value from the measured value is generated as residual, The abnormal plant data in which an alarm is generated is deleted according to the alarm logic and the plant data in the normal state is generated. That is, the preprocessing process refers to a process of generating only normal plant data from which an abnormal state data is removed.

For the preprocessed steady state data, the learning data generation unit 120 generates learning data by executing a learning operation for identifying and discriminating patterns. For example, the learning data generation unit 120 compares and classifies one or more steady state data with each other, and learns each other's characteristics and learns oneself. Unlike Supervised Learning, which learns one steady-state data, it is called Unsupervised Learning that learns the characteristics of each data by itself for various other steady-state data.

The learning data generation unit 120 generates a feature map by searching for a certain pattern based on the steady state data and extracting a very small feature to a large feature. At this time, the learning data generation unit 120 can learn difficult contents as it goes up to the upper layer through, for example, CNN (Convolutional Neural Network) algorithm that extracts features through various stages.

The learning data generation unit 120 may learn new steady state data in a short period of time through a process of searching for a pattern through a feature map in a given steady state data and classifying the data.

Next, the plant modeling unit 130 models the generated learning data to generate a plant model (S230).

That is, the plant modeling unit 130 generates normal state data in which the abnormal data is deleted from the real-time sensor data collected in real time by the data collecting unit 110 every five minutes, for example, every predetermined period through a batch scheduler The learning data obtained by learning is modeled and generated as a plant model, and the plant model is stored in the plant database (DB) as reference data (Reference).

In addition, the plant modeling unit 130 can generate a plant model as a predictive model by modeling the learning data into a parametric model or a non-parametric model. That is, the plant modeling unit 130 uses a parametric model and a non-parametric model to predict a learning model based on a model having different characteristics as shown in FIG. 4 Perform modeling.

In addition, in order to select functions corresponding to the merits of specific single models in the model model and the non-parametric model and to supplement the functions corresponding to the weak points, the plant model unit 130 includes a plurality of prediction models Can be combined to perform modeling.

That is, as shown in FIG. 4, the plant modeling unit 130 generates a plurality of models, such as a Multiple Linear Regression Model (MLRM), a Partial Least Squares (PLS), a Dynamic Principal Component Analysis (DPCA), and a Nearest Neighbor It is possible to perform modeling by combining a plurality of prediction models having different characteristics.

4, the MLRM model includes an ARX (Auto Regressive eXogeneous) model and an MLM (Maximum Likelihood Method) model. The PLS model includes a GLM (Generalized Linear Model) model and an MLM DPCA model includes ARX model and MLM model, and K-NN model can include non-parametric model and RBF (Radial Basis Function) kernel.

Table 1 below shows the advantages and disadvantages of the model of the model and the model of non-parametricity applied to the present invention.

Model Advantages Disadvantages Mother water
Model ARX Model ㅇ Sophisticated model design possible
ㅇ Very high performance
ㅇ Famous model in engineering field
ㅇ High performance
ㅇ Easy model interpretation
ㅇ Nonlinear modeling possible ㅇ Model considering the result, not the process of system state
ㅇ Difficulty modeling MIMO system
ㅇ Very poor expressiveness
ㅇ Use of video and speech recognition Nonmonopoly
Model Non
Parametric
Model ㅇ Excellent performance when applied with technologies such as k-NN ㅇ DB establishment for highly sophisticated learning data
O Predictive performance is highly dependent on the sophistication of the model CNN Model ㅇ Excellent expressiveness
ㅇ Excellent performance compared to general NN model
ㅇ Excellent performance for image processing
ㅇ A useful model when model structure design is difficult
ㅇ Excellent for modeling MIMO system ㅇ Model interpretation is impossible
O Local Optimization
ㅇ There is no guarantee of making a sophisticated model.

Therefore, the plant modeling unit 130 can perform modeling by combining a plurality of prediction models having different characteristics through the examination result as shown in the following Table 2, taking into account the advantages, disadvantages, and performance constraints of the respective models described above have.

Model Performance constraints Review the results NN based
HTM O Optimized by model
ㅇ It is difficult to optimize with 50 setting variables.
No formalized design method
ㅇ Not recognized in academia
ㅇ License problem ㅇ Non-conforming (However, it can be used in case of precise detection of specific fault) DPCA
PLS
MLRM ㅇ Performance change due to DPCA time setting. ㅇ Highly suitable for optimization algorithm ㅇ (common) Multivariate Regression method, so when an abnormality is detected in specific sensor, error is detected in other sensor at that point. ㅇ Highly suitable for performance and performance k-NN
(VBM) ㅇ Because the performance of the model depends on the sophistication of the model, there is a possibility that the accuracy of prediction may be lowered when a signal of a pattern that has not existed comes in. ㅇ Fit in terms of implementation and performance. Especially if you can create sophisticated models, you can make very accurate predictions. NN based
DBN O Optimized by model
ㅇ Local Optimization Difficult to Optimize
No formalized design method
O vulnerable to dynamic data ㅇ Not enough due to lack of performance. SVM based
SVDD
SVR ㅇ Data sampling effect is significant
ㅇ Severe performance change due to kernel and variable setting ㅇ Not enough due to lack of performance.

That is, the plant modeling unit 130 selects the K-NN model when it is necessary to precisely detect a specific failure, and when the optimization is required for each model due to a severe performance change due to time setting, the DPCA model, the PLS model, You can choose a model.

(k)을 얻을 수 있다.5, the plant modeling unit 130 combines the MLRM model using the ARX model and the LS (Least Square) method as the modeling model and the k-NN model using the RBF kernel as the non-parametric model can do. FIG. 5 is a diagram illustrating an example of performing ensemble learning by combining an aqueous model and a non-parametric model according to an embodiment of the present invention. Therefore, when the plant model unit 130 receives the data (k), the plant model unit 130 generates the estimated value mlm (k) as x (1) through the MLRM model selected by the learning data generation unit 120 as shown in FIG. ) / Residual_mlrm (k), acquires an estimated value x (2) (Estimated Value_kNN (k) / Residual_kNN (k)) through the k-NN model, and selects the most optimal one among these estimated values The ensemble learning is performed by a bagging method and the ensemble learning is performed to obtain an estimated value and an optimal estimated value as shown in Equation 1 below:

(k).

Accordingly, the plant modeling unit 130 can perform ensemble learning on the steady state data through the learning model to output the most accurate prediction data (Optimal_Estimated Value) as shown in FIG. Accordingly, the plant modeling unit 130 transmits the optimal predicted data (Optimal_Estimated Value) calculated according to the ensemble learning to the alarm logic (alarm logic) so that it can be used for alarming the plant abnormality detection. FIG. 8 is a diagram showing a result of calculating a predicted value with high accuracy according to a combination of an aqueous model and a non-parametric model according to an embodiment of the present invention. 8, the plant model unit 130 outputs the Proposed Method ^* having an accuracy of 98.6% and 95.1% as a result of ensemble learning using the MLRM, PLS, DPCA, and k-NN models. 8, the plant modeling unit 130 uses the MLRM, PLS, DPCA, and k-NN models as well as the ensemble learning using the Auto-Learning Algorithm 98.6%, and outputs the Proposed Method ^** value with an accuracy of 97.9%.

At this time, the alarm unit 150 sets the sensor data measured in real time as the sensor data, sets the learning data as the estimated data, subtracts the predicted value from the measured value as shown in FIG. 3, (Residual Generation), and when the generated residual value exceeds the allowable value, an alarm is output by the alarm logic. In addition, the alarm unit 150 generates an early alarm when the residual value exceeds an allowable value (specific value), but may generate an early alarm for a plant failure even when the measured value exceeds an allowable value (specific value). FIG. 3 is a diagram illustrating a process of outputting an alarm from sensor data of a plant according to an embodiment of the present invention, and an automatic learning process for detecting a plant anomaly.

3, the failure diagnosis unit 170 diagnoses the failure of the plant using the diagnosis logic based on the diagnosis database 160 storing the failure cases in the plant as the diagnosis data , And the root cause is tracked according to the diagnosis result (Diagnosis Result).

Here, the data on the failure cases stored in the diagnostic database 160 are obtained by taking the plant data measured in real time as actual values, using the learned steady state data as the predicted values, and subtracting the predicted value from the actual values, In the case where the generated residual value exceeds the allowable value, the abnormal state plant data in which an alarm is generated according to the alarm logic and the normal state plant data in which the abnormal state of the plant data are removed These data are stored separately. For example, data of abnormality state exceeding the allowable value by vibration, pressure, temperature, flow rate, etc. of the corresponding device through correlation with the operating variable. At this time, the meaning exceeding the allowable value indicates that the upper limit and the lower limit are out of a predetermined reference range.

3, when the plant data in which the alarm is generated is inputted, the failure diagnosis unit 170 determines whether or not the diagnosis logic (algorithm) is executed based on the data of the failure cases stored in the diagnosis database 160 The diagnostic result is output based on whether the vibration exceeds the allowable value, whether the pressure exceeds the allowable value, whether the temperature exceeds the allowable value, and the like. (Temperature, vibration, pressure, and the like) as to why an alarm occurs for some reason when an alarm occurs in a part of the plant data collected in real time through the diagnosis result outputted from the failure diagnosis unit 170 It will be possible to do.

Meanwhile, the learning determination unit 140 determines (determines) whether to learn the sensor data measured at the current point through each measurement sensor as described above (S240).

Here, if the sensor data is measured through each measurement sensor when the production power of the plant is greater than or equal to a specific power, or if the early warning according to the alarm logic is not output from the alarm unit 150 Or to learn (measure) the currently measured sensor data when an alarm is not generated on a predetermined time basis for each measurement sensor.

If the learning determination unit 140 determines to learn about the currently measured sensor data (S250-YES), the learning data generation unit 120 generates a combination of the currently measured sensor data and existing learning data And generates new learning data (S260).

3, if the alarm is not generated based on the alarm information input from the alarm unit 150, that is, if the alarm data is not generated based on the sensor data measured at the current time 6, data processing is performed on the sensor data inputted from the data collecting unit 110 to obtain existing learning data, that is, modeling data, (Additional Data) to generate new learning data. 6 is a diagram illustrating an example of performing data processing on sensor data measured at present according to an embodiment of the present invention. Here, the existing modeling data is the data (2,000 pieces) for the plant model generated through the learning data obtained by learning the steady state data in the previous cycle, and the additional data is the sensor data (288 pieces) collected in the current cycle. An example of generating new learning data by applying additional data to existing data will be described in detail below.

First, the learning data generation unit 120 combines the currently measured sensor data (288) with the existing learning data set (2,000) to extract new data based on a specific feature. That is, the learning data generation unit 120 sets the magnitude and the angle of the characteristic of the data to two, and sets the learning data of each data set based on the technology such as N1 Norm, N2 Norm, and N-Infinity Norm. Obtain the magnitude and angle (Angle). For example, in the 1 Dimension Method, the magnitude is 1, in the 2 Dimension Method, in the Magnitude and Angle 1, in the 3 Dimension Method, in the Dimension Method, Magnitude and Angle, and Magnitude and Angle N-1 in the N Dimension Method are set as characteristics of data and calculation is performed.

When the number of the plant data is 2,000 and the alarm information is input from the alarm unit 150 as the alarm is output from the alarm unit 150 and the sensor data currently measured and inputted is 288, The data processing unit 120 performs data processing based on the characteristic data as shown in FIG. 3 for 2,288 pieces of total data (Xm_PoolOn) obtained by adding 288 pieces of the sensor data indexAddData currently measured to 2,000 pieces of the existing plant data (Xm_Pool) ). Here, the maximum number of additional data indexAddData is 288, and a value of 288 or less per day is input to the learning data generation unit 120 by automatic learning.

6, when the value of the additional data (indexAddData) is 0 to 288 in 2,000 existing learning data (Xm_Pool), the learning data generating unit 120 generates 2,288 pieces of total data (number XmPoolOn) And performs sorting on the obtained whole data. That is, the learning data generation unit 120 classifies the 2,288 data, which is the sum of the existing data 2,000 and the currently collected sensor data 288, into a specific data interval and discards less than a decimal point to generate new learning data.

First, the learning data generation unit 120 divides the total number of data (2,288) by the number of plant data (2,000) to obtain 1.144 for the data interval (interval) according to the following equation (1).

Accordingly, the learning data generation unit 120 may divide the entire data of 2,288 magnitudes into 1.144 intervals, which results in 1.14, 2.28, 6.864, and 8.008, 8 is obtained. At this time, the seventh is missing.

7, the learning data generation unit 120 combines the currently measured sensor data Xm_PoolOn (288) with the existing learning data Xm_Pool (2,000), and deletes the abnormal data 2,000 pieces of new learning data Xm are generated. 7 is a diagram illustrating an example of generating new learning data based on a characteristic by combining sensor data currently measured and existing learning data according to an embodiment of the present invention. Accordingly, the learning data generation unit 120 transmits the new learning data generated in this process to the plant model unit 130. [

Then, the plant modeling unit 130 receives the new learning data at a predetermined cycle and performs plant modeling to update the existing plant model (S270).

For example, the plant modeling unit 130 applies 2,000 new learning data to an existing plant model after performing a predetermined cycle, for example, once a day (once) Is newly updated.

Therefore, by accurately calculating and using the predictive data used for the abnormality detection of the plant, the abnormality of the plant can be accurately detected and alarmed early.

As described above, according to the present invention, it is determined whether to learn sensor data measured at the present time. If it is determined that learning is to be performed, new learning data is generated by combining existing learning data and measured sensor data, It is possible to realize an automatic learning system and method for plant anomaly detection that allows automatic learning to be performed by modeling new learning data in a cycle.

It will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the spirit and scope of the present invention as defined by the following claims and their equivalents. Only. It is intended that the present invention covers the modifications and variations of this invention provided they come within the scope of the appended claims and their equivalents. .

100: automatic learning system 110: data collection unit
120: learning data generation unit 130: plant model unit
140: learning determination unit 150: warning unit
160: Diagnosis DB 170:

Claims (10)

A data collection unit for collecting sensor data for a plant;
A learning determination unit for determining whether to learn the collected sensor data;
A learning data generation unit for generating learning data based on the collected sensor data according to a determination result of the learning determination unit; And
A plant model unit for generating a plant model by modeling based on the generated learning data;
Lt; / RTI >
Wherein the learning data generation unit generates new learning data by combining the sensor data and existing learning data and transmits the new learning data to the plant model unit,
The plant model unit updates the plant model by modeling based on the new learning data,
Wherein the learning data generation unit sorts and sorts all the data obtained by combining the collected sensor data and the existing learning data on the basis of the size and obtains the total number of the data and uses the total number of the data in the plant model unit And dividing the interval by the number of learning data to multiply the interval by a natural number from 1 to the number of learning data and then discarding the natural numbers to determine natural numbers for selection, Selecting the corresponding data as the new learning data,
Automatic learning system for plant fault detection.
The method according to claim 1,
An alarm unit which takes the collected sensor data as measured values, generates a value obtained by subtracting a predicted value from a measured value as a predicted value, and outputs an early warning when the generated residual value exceeds an allowable value;
Further comprising an automatic learning system for plant anomaly detection.
3. The method of claim 2,
The learning determination unit may be configured to determine whether the sensor data is measured through each measurement sensor when the production power of the plant is greater than or equal to a specific power or when the early warning is not output from the alarm unit, And determines to learn the collected sensor data when an alarm is not generated.
delete The method according to claim 1,
A diagnostic database storing the failed cases in the plant as diagnostic data; And
A failure diagnostic unit for diagnosing a failure of the plant based on the diagnostic database and for tracking a root cause;
Further comprising an automatic learning system for plant anomaly detection.
(a) collecting sensor data measured by each measurement sensor for a data collection unit plant;
(b) generating a learning data based on the steady state data among the collected sensor data;
(c) generating a plant model by modeling the plant model based on the generated learning data;
(d) determining whether the learning determination unit learns the currently measured sensor data through the measurement sensor,
(e) generating a new learning data by combining the currently measured sensor data with existing learning data if it is determined that the learning data is to be learned as a result of the determination by the learning determination unit; And
(f) modeling the plant model based on the new learning data and updating the plant model,
The step (e)
Sorting and sorting the total data including the currently measured sensor data and the existing learning data by a size reference;
Obtaining the total number of data;
Dividing the total number of data by the number of learning data used in the plant model unit to obtain an interval;
Multiplying the interval by a natural number from 1 to the number of learning data, and discarding the natural number to determine natural numbers for selection; And
Selecting data corresponding to a natural number for the selection from the sorted total data as the new learning data;
And an automatic learning method for plant anomaly detection.
The method according to claim 6,
In the step (c), the alarm unit generates sensor data measured in real time as measured values, generates a residual value by subtracting the predicted value from the measured value as the predicted value, and if the generated residual value exceeds the allowable value Further comprising outputting an early warning to the plant.
8. The method of claim 7,
Wherein the step (d) includes the steps of: when the learning data is measured through each measurement sensor when the production power of the plant is higher than a specific power, or when the early warning is not output from the alarm unit, And determining to learn the currently measured sensor data when an alarm is not generated for each measurement sensor on a predetermined time basis.
The method according to claim 6,
In the step (b), the learning data generation unit generates steady state data by deleting the abnormal data through a pre-processing process on the collected sensor data, and learns the generated steady state data, The method comprising the steps of: 8. The method of claim 7,
Wherein the step (c) further comprises: diagnosing a failure of the plant based on a diagnostic database storing failure cases in the plant as diagnostic data, and tracking the root cause. Automatic learning method.

Images