KR20230107519A - System and Method for detecting tube leakage using state signal of power generation plant - Google Patents

Abstract

A tube leakage detection system capable of monitoring a tube leakage and estimating the location using acoustic emission signals and/or operation signals of a power generation plant is disclosed. The tube leakage detection system comprises a sensor block for detecting status signals of the power generation plant, a collection unit for collecting the status signals, a calculation unit for calculating a probability distribution for each sensor using the collected status signals, and an analysis unit for estimating whether there is a leakage in a tube of the power generation plant according to the probability distribution for each sensor.

Description

System and method for detecting tube leakage using state signal of power generation plant

The present invention relates to a tube leak detection technology, and more particularly, to a system and method for monitoring and estimating the location of a tube leak using an acoustic emission signal and/or an operating signal of a boiler.

Coal-fired power plants are composed of boilers, steam turbines, main generators, and auxiliary equipment such as fuel supply, condensate, and water supply systems to operate them. Among them, a boiler is a very important power generation facility that burns fuel to generate high-temperature and high-pressure steam and sends the steam to a steam turbine to drive a generator.

Inside the boiler, heat exchange tubes for generating steam are arranged in a very complicated form. If this tube breaks and leaks, the plant must be shut down. If tube leakage is not detected early, it is very important to detect leakage early because the leakage expands, resulting in a longer maintenance period and increased losses due to plant shutdown.

For this reason, BTLD (Boiler Tube Leak Detection) systems are installed in many power plants to detect boiler tube leaks at an early stage. The BTLD system installs an acoustic sensor on the outer wall of the boiler to determine whether or not there is a leak through an acoustic emission signal, and indicates the approximate location of the leak.

The existing BTLD system has two major limitations. First, when determining the leak location, the location can be expressed only when clear signals are generated from three or more sensors. In other words, when the leakage position is calculated by simple trilateration, it is difficult to express the tube leakage position because the unique position is not determined when leakage signals are generated only by two sensors.

Therefore, since the existing boiler tube leakage localization system expresses the solution by deterministically calculating the leakage position by formula, there is a problem in that the position can be expressed only when signals are clearly received from three or more sensors.

Second, since the leakage is determined only by the sound signal excluding the operation signal, it is difficult to be certain that leakage actually occurred in the field. In other words, since operation signals such as temperature and pressure are not considered, physical changes in the boiler operating state cannot be considered. Therefore, it is not easy to determine whether a leak alarm generated in the system is actually a leak or a false alarm.

Reliability of the leakage alarm can be increased by comprehensively determining the leakage by considering the operation signal and the sound signal at the same time. However, it is not easy to determine the leakage signal by considering all of these characteristics because the operating signals of the power plant have correlations between signals and time-series characteristics at the same time. There have been attempts to detect abnormalities in these time-series signals through auto-encoders, but basic auto-encoders are structurally difficult to capture all of the complex characteristics of power generation facilities. In general, an auto-encoder consists of an input layer, a hidden layer, and an output layer.

This is because each node of the auto-encoder cannot reflect the time-series characteristics of the signal. Most of the time series signals have a trend over time. Trend refers to how a signal has a relationship with a previous point in time, such that data may show an increasing trend, decrease or oscillate over time. Since a typical autoencoder sees each data point as a piece of independent information, it does not reflect such a trend over time well.

1. Korean Registered Patent No. 10-1107261 (registration date: January 11, 2012)

The present invention has been proposed to solve the problems caused by the above background art, and provides a tube leakage detection system and method capable of monitoring and estimating the position of tube leakage using an acoustic emission signal and / or operation signal of a power generation facility. But it has a purpose.

In addition, another object of the present invention is to provide a tube leak detection system and method capable of detecting and expressing a location even without using only three or more sensors.

In order to achieve the object presented above, the present invention provides a tube leak detection system capable of monitoring and estimating the location of a tube leak using an acoustic emission signal and/or an operation signal of a power generation facility.

The tube leak detection system,

A sensor block for detecting a state signal of a power generation facility;

a collection unit that collects the state signal;

a calculation unit that calculates a probability distribution for each sensor using the collected state signals; and

and an analyzer for estimating leakage from the tube of the power generation facility according to the probability distribution for each sensor.

At this time, the sensor block is characterized in that it consists of at least two sensors for detecting the state signal.

In addition, the state signal is an acoustic signal, and the at least two sensors are AE (Acoustic Emission) sensors.

The calculation unit may calculate the distance from each sensor to the leak location by using a time difference between at least two sensors when the signal characteristics of the state signal change.

The calculation unit may calculate a leak probability for each sensor location using a normal probability distribution having an average of distances to the leak location for each of the at least two sensors.

In addition, the calculation unit calculates a probability distribution by location for all the at least two sensors using the leakage probability by sensor location, and estimates a leak location in a tube of the power generation facility based on the probability distribution by location. to be characterized

(여기서,

: i번째 센서의 가중치 파라미터, Likelihood는 위치별 누설 가능도이다)로 정의되는 것을 특징으로 한다.In addition, the leakage probability for each sensor location is

(here,

: a weight parameter of the i-th sensor, and Likelihood is a leak probability by location).

(여기서,

이고, In addition, the possibility of leakage for each location is expressed by Equation

(here,

ego,

: i번째 센서와 j번째 센서의 불확실성합)로 정의되는 것을 특징으로 한다.m _i = u × t _{i ,} (x _i ,y _i ,z _i ): the position of the ith sensor, u: the wave propagation speed, t _ij : the difference in arrival time from the i-th sensor to the j-th sensor,

: It is characterized in that it is defined as the uncertainty sum of the i-th sensor and the j-th sensor).

In addition, the calculation unit receives the operation signal of the power generation facility and learns local characteristics and continuous characteristics of the operation signal by using a CNN (Convolutional Neural Networks)-LSTM (Long Short-Term Memory models)-AE (Auto Encoder) algorithm. It is characterized by carrying out learning.

In addition, the driving signal is partially overlapped at regular time intervals through a windowing technique and converted into a primary sequence signal having a sequence number, a sequence length, and a tag.

In addition, the primary sequence signal is divided into subsequences to be used as an input signal of a convolutional neural network (CNN) and converted into an input sequence signal having the number of sequences, the number of subsequences, the length of subsequences, and a tag. to be

In addition, the analyzer may calculate a difference between an output sequence signal reconstructed through a latent variable and the driving signal, and detect an abnormality when the difference exceeds a predetermined boundary value.

In addition, the analyzer may generate leakage alarm information when leakage is estimated according to the presence or absence of leakage.

*At this time, the leakage alarm information is characterized in that it is a combination of voice, graphic, and text.

On the other hand, another embodiment of the present invention includes (a) detecting, by a sensor block, a state signal of a power generation facility; (b) collecting the status signal by a collection unit; (c) calculating a probability distribution for each sensor using the state signals collected by a calculator; and (d) estimating, by an analysis unit, whether there is leakage in the tubes of the power generation facilities according to the probability distribution for each sensor.

At this time, in the step (c), the calculation unit receives the operation signal of the power generation facility and calculates the operation signal by using a CNN (Convolutional Neural Networks)-LSTM (Long Short-Term Memory models)-AE (Auto Encoder) algorithm. and performing local feature learning and continuous feature learning.

According to the present invention, the estimated position of tube leakage is probabilistically expressed according to the position, so that even if only two acoustic emission sensors respond, the leakage probability distribution for each position can be expressed.

In addition, as another effect of the present invention, tube leakage can be determined through operation signals (temperature, pressure) of power generation facilities, solving the problem of existing anomaly detection methods that learn only some characteristics of data, and power plant data are mostly time series. Multi-channel data in the form is collected at the same time, and it is possible to identify and diagnose information in real time about the collected big data.

In addition, as another effect of the present invention, although most of the data acquired from the power plant is mixed with a lot of noise and normal/abnormal information is not displayed, the auto-encoder-based anomaly detection method is resistant to noise and does not require labeling of data. Because it is an unsupervised learning method, it can effectively process power plant data.

In addition, as another effect of the present invention, in the case of the signal processing-based state diagnosis method, since frequency analysis is performed by analyzing and reading only data corresponding to a specific acquisition section, it is difficult to diagnose by reflecting the history information that has occurred in the past. For example, it is possible to determine the abnormality and the normal state by learning the previous history data.

In addition, another effect of the present invention is that the reliability of the tube leakage alarm can be raised by simultaneously considering the sound emission signal and the operation signal of the tube.

1 is a block diagram of a tube leakage detection system according to an embodiment of the present invention.
2 is a flowchart showing a tube leak detection process according to an embodiment of the present invention.
3 is a conceptual diagram showing the structure of a Convolutional Neural Networks (CNN)-Long Short-Term Memory models (LSTM)-Auto Encoder (AE) algorithm according to an embodiment of the present invention.
4 and 5 are graphs showing leakage probability distributions for each sensor location using only two sensors.
6 is a diagram comparing a final probability model and an existing final probability model according to an embodiment of the present invention.
7 is a flowchart of the CNN-LSTM-AE algorithm shown in FIG. 3;
8 is a flowchart showing a pre-processing process of the CNN-LSTM-AE algorithm shown in FIG.
9 is a block diagram showing the encoder structure of the CNN-LSTM-AE algorithm shown in FIG.
10 is a block diagram showing the decoder structure of the CNN-LSTM-AE algorithm shown in FIG.
11 is a flowchart showing a tube leak diagnosis process according to an embodiment of the present invention.
12 is a configuration block diagram of a power plant model for verifying a tube leakage detection system according to an embodiment of the present invention.
13 and 14 are graphs showing the result of applying the algorithm of the super heater 2 metal temperature signal.
15 and 16 are graphs showing the result of applying the algorithm to the super heater 1 injection header temperature signal.
17 is a view of installing a sound emission sensor in a power plant according to an embodiment of the present invention.
18 is a diagram showing an example of a time difference in which signal characteristics between sensors change according to an embodiment of the present invention.
19 is a graph expressing a cumulative probability distribution for each sensor position according to an embodiment of the present invention.

Advantages and features of the present invention, and methods for achieving them, will become clear with reference to the embodiments described below in detail in conjunction with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below, but will be implemented in a variety of different forms, only the present embodiments make the disclosure of the present invention complete, and those skilled in the art in the art to which the present invention belongs It is provided to fully inform the person of the scope of the invention, and the invention is only defined by the scope of the claims. The sizes and relative sizes of components shown in the drawings may be exaggerated for clarity of explanation.

Like reference numbers throughout the specification indicate like elements, and “and/or” includes each and every combination of one or more of the recited items.

Terms used in this specification are for describing embodiments and are not intended to limit the present invention. In this specification, singular forms also include plural forms unless specifically stated otherwise in a phrase. The referenced elements, steps, operations and/or elements that “comprise” and/or “comprise” as used in the specification do not exclude the presence or addition of one or more other elements, steps, operations and/or elements. .

Although used to describe various components such as first and second, these components are not limited to these terms, of course. These terms are only used to distinguish one component from another. Accordingly, it goes without saying that the first component mentioned below may also be the second component within the technical spirit of the present invention.

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

Hereinafter, a tube leakage detection system and method using a state signal of a power generation facility according to an embodiment of the present invention will be described in detail with reference to the accompanying drawings.

1 is a block diagram of a tube leak detection system 100 according to an embodiment of the present invention. Referring to FIG. 1, the tube leakage detection system 100 includes a sensor block 110 that detects a state signal of a power generation facility 10, a collection unit 120 that collects the state signal, and a sensor using the collected state signal. A calculation unit 130 that calculates the probability distribution for each sensor, an analysis unit 140 that estimates whether or not there is a leak and the location of the leak according to the probability distribution for each sensor, and an output unit 150 that outputs whether or not there is a leak and the location of the leak as notification information, etc. It is characterized by including.

Power plant 10 generally means a boiler. The power plant 10 generally refers to a power plant and auxiliary facilities for converting energy sources such as hydropower, thermal power, and nuclear power into electric power. In one embodiment of the present invention, a tube installed inside and outside the boiler will be described.

The sensor block 110 is installed inside and outside the power generation facility 10 to perform a function of detecting a state signal of the tube. To this end, the sensor block 110 includes first to nth sensors 111-1 to 111-n. The first to nth sensors 111-1 to 111-n may be Acoustic Emission (AE) sensors that detect acoustic signals. AE sensors are classified into a resonant type that becomes highly sensitive (narrowband type) at a certain frequency and a wideband type that maintains constant sensitivity over a wide frequency range.

The collection unit 120 functions to collect state signals from the sensor block 110 . To this end, the collection unit 120 may include a communication circuit, a digital signal processor (DSP), and the like.

The calculation unit 130 measures the wave propagation speed from the acoustic signal to obtain the distance between the sensors 111-1 to 111-n and the leak location. Then, the time difference between the points at which the signal characteristics between the sensors change is calculated. The distance from each sensor to the leak location can be calculated using the measured wave propagation speed and the time difference between the sensor signal characteristics change. In addition, a normal probability distribution having an average of the distances to the leak location for each sensor is generated to calculate the leak probability for each location in each sensor.

The analysis unit 140 performs a function of obtaining a probability distribution for each location for all sensors and expressing a final leak probability for each location.

The output unit 150 performs a function of displaying information processed by the collection unit 120, the calculation unit 130, the analysis unit 140, and the like. Accordingly, the output unit 150 may output information in a combination of text, voice, and graphics. To this end, the output unit 150 may include a display, a sound system, and the like.

The display may be a liquid crystal display (LCD), a light emitting diode (LED) display, a plasma display panel (PDP), an organic LED (OLED) display, a touch screen, a cathode ray tube (CRT), a flexible display, or the like. In the case of a touch screen, it can function as an input means.

The storage unit 160 performs a function of storing software, command sets, or data for operating the system 100 . To this end, the storage unit 160 includes a random access memory (RAM), a magnetic disk, a flash memory, a static random access memory (SRAM), a read only memory (ROM), ROM), electrically erasable programmable read only memory (EEPROM), programmable read only memory (PROM), etc., but is not limited thereto.

The calculation unit 130 and the analysis unit 140 shown in FIG. 1 refer to units that process at least one function or operation, and may be implemented in software and/or hardware. In hardware implementation, ASIC (application specific integrated circuit), DSP (digital signal processing), PLD (programmable logic device), FPGA (field programmable gate array), processor, microprocessor, other It may be implemented as an electronic unit or a combination thereof.

In software implementation, software component components (elements), object-oriented software component components, class component components and task component components, processes, functions, properties, procedures, subroutines, segments of program code, drivers, firmware, microcode, data , databases, data structures, tables, arrays, and variables. Software, data, etc. may be stored in memory and executed by a processor. The memory or processor may employ various means well known to those skilled in the art.

2 is a flowchart showing a tube leak detection process according to an embodiment of the present invention. Referring to FIG. 2 , the wave propagation speed is measured from the acoustic signal to obtain the distance between the sensors 111-1 to 111-n and the leak location. Then, the calculation unit 130 calculates a time difference at a point in time when the signal characteristics between the sensors change (step S210). The distance from each sensor to the leak location can be calculated using the measured wave propagation speed and the time difference between the sensor signal characteristics change.

Then, the calculation unit 130 calculates the leak probability for each sensor location by generating a normal probability distribution taking the average of the distances to the leak location for each sensor (step S220). The leak probability for each location becomes a Gaussian probability distribution.

Thereafter, the calculation unit 130 obtains a probability distribution for each location for all sensors using the leak probability for each sensor location, and expresses the final leak probability for each location (step S230).

Therefore, if the location estimation method disclosed in FIG. 2 is introduced, since the leak location is estimated with a probability distribution for each location rather than deriving a determined solution, even if leakage signals are received from less than three sensors, the leak estimation location can be expressed. there is.

3 is a conceptual diagram showing the structure of a Convolutional Neural Networks (CNN)-Long Short-Term Memory models (LSTM)-Auto Encoder (AE) algorithm according to an embodiment of the present invention. In general, tube leakage is determined only by the acoustic emission signal. Therefore, in order to increase the reliability of the leakage alarm in the field, leakage detection through an operation signal is required. It is difficult to learn all the characteristics of power plant operation signals with the existing auto-encoder neural network structure. The operation signal is a physical quantity such as temperature and pressure of a power generation facility (particularly, a boiler may be mentioned), and is a parameter for knowing the overall state of operation.

To this end, in one embodiment of the present invention, an auto encoder comprising a convolutional neural network (CNN) and a long and short-term memory (LSTM) neural network in order to learn both local and continuous characteristics of the power plant's operation signal An artificial neural network is implemented.

Prior to learning CNN-LSTM-AE (Convolutional Neural Networks-Long Short-Term Memory models-Auto Encoder) data, a pre-processing process of input data (i.e., driving signals) for learning characteristics is required. The input data is divided into units called input sequences 310 through a windowing technique in the preprocessing process, and this sequence is further divided into units called subsequences to become input sequence signals 310, and the encoder 320 is entered into The encoder 320 is sequentially composed of a 1D-convolutional neural network (1D-CNN) and a long short-term memory models (LSTM) neural network. Here, 1D-CNN refers to the structure of a CNN algorithm for one-dimensional input data.

Local features are learned in the convolutional neural network (1D-CNN) in units of this input subsequence, and then continuous features of data are learned in units of sequences in the long and short-term memory model neural network (LSTM). When model learning is completed, the input multi-channel data is reconstructed through encoding and decoding processes based on the learned weights. To this end, a latent variable block 330 and a decoder 340 are configured. Unlike the encoder, the decoder 340 is sequentially composed of long short-term memory models (LSTM) neural networks and ID-convolutional neural networks (1D-CNN). Accordingly, the output sequence 350 is finally output through the decoder 340 . By comparing the reconstructed signal with the original signal, an abnormality in the power generation facility is detected and an abnormal tag is selected. This will be described later with reference to FIG. 7 .

On the other hand, in general, the tube leakage position determination can be expressed only when clear signals are generated from three or more sensors. However, since only a small number of actual acoustic sensors can respond in some cases, it is difficult to express the position in the conventional method. One embodiment of the present invention uses Maximum Likelihood Estimation (MLE) to solve this problem.

Maximum Likelihood Estimation (MLE) is a method of obtaining the parameters of a random variable based on a sample extracted from that random variable. It is a method of selecting a parameter that maximizes the likelihood of a desired value when given a certain parameter. By applying this method to the problem of estimating the location of leaks in boiler tubes, it is possible to express the probability that leaks occur at all locations inside the boiler.

The distance between the sensor and the leak location can be measured through the wavelength and time lag of the acoustic emission signal from each sensor. Assuming that the leak probability at the corresponding location follows a normal distribution with an average of the distances to the measured leak location, the likelihood of leakage at each location due to each sensor can be expressed as the following equation.

이고, here,

ego,

m _i = u × t _i

(x _i ,y _i ,z _i ): the position of the ith sensor

u: wavelength propagation speed

t _ij : difference in arrival time from i-th sensor to j-th sensor

: i번째 센서와 j번째 센서의 불확실성합

: Sum of uncertainty of i-th sensor and j-th sensor

With the sum of the products of the probabilities of all sensors at all locations as the denominator, and the product of the probabilities of all sensors at that location as the numerator, the leakage probability for each sensor location when leakage occurred at that location can be expressed as the following equation there is.

: i번째 센서의 가중치 파라미터here,

: Weight parameter of the ith sensor

Examples of leak probabilities for each sensor location disclosed above are illustrated in FIGS. 4 and 5 .

4 and 5 are graphs showing leakage probability distributions for each sensor location using only two sensors. Referring to FIGS. 4 and 5 , each probability distribution for four sensors is obtained. That is, the leak probability distribution 410 based on the time difference between the first sensor and the second sensor, the leak probability distribution 420 based on the time difference between the first sensor and the third sensor, and the time arrival between the first sensor and the fourth sensor Difference-based leak probability distribution 430, leak probability distribution based on the time difference between the second sensor and the third sensor 510, leak probability distribution based on the time difference between the second sensor and the fourth sensor 520, and the third sensor A leak probability distribution 530 based on the time arrival difference between the sensor and the fourth sensor is shown.

For visibility, you can represent 95% area. Even if only two sensors react, it is possible to express the area where leakage occurs by limiting only a specific area as shown in the figure above. Since the position was estimated with only two sensors, the position with a high leakage probability comes out as a plane between the two sensors. Here, if the additional sensor reacts, the section with a high leak probability will be concentrated in one place, but even if it is not, as shown in the drawing, the plane can be specified to some extent, which can help find the actual leak point.

6 is a diagram comparing a final probability model 610 and an existing final probability model 620 according to an embodiment of the present invention. That is, FIG. 6 is a diagram showing a final leakage probability distribution considering all sensors. Referring to FIG. 6 , it is possible to probabilistically limit and express an area where leakage is likely to occur. In addition, as shown in FIG. 6, it can be seen that much higher accuracy can be obtained based on the Most Probable Point (MPP) than the result obtained using the linear/nonlinear hyperbolic equation using the existing TDOA (Time difference of arrival). can 6 shows a sensor position 601 , a target position 602 , a linear target estimate 603 , a non-linear target estimate 604 and a stochastic target estimate 605 .

FIG. 7 is a flowchart of the CNN-LSTM-AE (Convolutional Neural Networks-Long Short-Term Memory models-Auto Encoder) algorithm shown in FIG. 3; 7, in a large view, 1) receiving signals of multiple channels and generating sequence input data (ie, operation signals) in real time so that sequential time-series data can be applied to deep learning, 2) localization of signals The continuous and continuous characteristics are learned through deep learning, and 3) the signal is reconstructed through latent variables. 4) Finally, based on the difference between the reconstructed signal and the original signal, abnormal detection and abnormal tag selection are performed.

Referring to FIG. 7 , driving signals acquired from multiple channels are acquired in real time in a time-series form (step S710). A pre-processing process is required to apply the multi-channel data to the deep learning algorithm to use.

Therefore, data must be converted into a form suitable for learning both the local and continuous characteristics of signals. To this end, the acquired multi-channel data (i.e., the moving signal) is divided into regular time intervals through a windowing technique (step S720). At this time, the time interval is partially overlapped. These divided time intervals are called primary sequence signals. This converts the data type from (time)*(tags) to (sequence count)*(sequence length)*(tags).

The primary sequence signal converted in this way is further divided into subsequences to be used as an input signal of a convolutional neural network (CNN) for local feature learning, and (number of sequences) * (number of subsequences) * (length of subsequences) )*(tag) form, the input sequence signal 310 is generated through a preprocessing process (step S730).

Thereafter, local sequence signal learning through a convolutional neural network (CNN) and continuous feature learning through long short-term memory models neural networks are performed (steps S740 and S750).

Thereafter, signal reconstruction is performed through latent variables (step S760). Therefore, the original signal is reconstructed by performing decoding in the opposite process to the encoding process from the data where necessary characteristics are compressed and stored. Latent variables are variables whose constructs cannot be directly observed or measured.

A latent variable is data that is stored after compressing the characteristics of an input value through an encoder. In the algorithm according to an embodiment of the present invention, correlations between signals and time-series characteristics are all stored in latent variables through CNN and LSTM, and these latent variables are reconstructed through a decoder to obtain correlations between signals and time-series characteristics of signals. It outputs a signal with characteristics.

Thereafter, a statistical abnormal boundary value is generated, and it is checked whether an abnormal value is detected (steps S761 and S770). In other words, the difference between the output sequence signal 350 reconstructed through the decoder 340 and the original signal is calculated, and when this difference exceeds a predetermined boundary value by applying a statistical technique to the steady state data, an abnormality is detected ( S780, S790). Since the difference value can be calculated for each signal, the tag with the greatest difference can be selected.

Unlike this, if the predetermined boundary value is exceeded, it is detected as normal (step S771).

8 is a flowchart showing a pre-processing process of the CNN-LSTM-AE algorithm shown in FIG. Referring to FIG. 8 , the calculator 130 receives a driving signal 811 that is a multivariate signal (step S810). Thereafter, the input driving signal 811 is divided into predetermined time intervals through a windowing technique to generate the primary sequence signal 821 (step S820).

Thereafter, the primary sequence signal 821 converted in this way is further divided into subsequences to be used as an input signal of a convolutional neural network (CNN) for local feature learning to become an input sequence signal 831 (step S830).

9 is a block diagram showing the encoder structure of the CNN-LSTM-AE algorithm shown in FIG. Referring to FIG. 9 , the input sequence signal 831 converted through the preprocessing process passes through an encoder algorithm composed of a convolutional neural network 910 and a short-term memory neural network 920 . Normal data is used for learning, and local feature learning of the signal is performed by using subsequences of the input sequence signal 831 as inputs to a convolutional neural network (CNN).

Then, the weights of the model are adjusted by learning the features of the long sequence using a long short-term memory (LSTM) neural network. When the preprocessed data is received after the learning of the model is finished, the dimensions are reduced through the encoder, and only necessary encoding features 930 are stored in each node.

10 is a block diagram showing the decoder structure of the CNN-LSTM-AE algorithm shown in FIG. Referring to FIG. 10, the original signal is reconstructed by performing decoding in a process opposite to the encoding process from data in which necessary characteristics are compressed and stored. The decoder structure is symmetrical with the encoder and inputs the encoded encoding features 930 to a long short-term memory (LSTM) neural network and finally reconstructs the output sequence signal 1010 through a convolutional neural network (CNN).

11 is a flowchart showing a tube leak diagnosis process according to an embodiment of the present invention. Referring to FIG. 11, the operation signal is a physical quantity such as temperature and pressure inside the boiler, and the general state of boiler operation can be known. This driving signal is pre-processed and a steady-state driving signal is learned using the CNN-LSTM-AE algorithm developed above (steps S1101, S1102, and S1103). Then, after learning, whether the driving signal is normal/abnormal is determined in real time (step S1104).

At the same time, in the existing Boiler Tube Leak Detection (BTLD) system, tube leakage is detected using the sound emission signal of the boiler (steps S110, S1120, and S1121). When both of the two systems determine leakage, it is determined to be a sure leakage and leakage alarm information is generated (steps S1122 and S1123). Leakage alarm information can be a combination of voice, graphics, and text. Of course, leakage alarm information may be generated even if leakage is determined in one system.

When a leakage alarm occurs, the acoustic emission signal is received and the estimated leak location in the boiler is expressed as a probability distribution for each location through a maximum likelihood estimation (MLE) based location estimation algorithm (steps S1130, S1131, and S1132).

In the embodiment of the present invention, if anomaly is detected in CNN-LSTM-AE based on the operation signal and leakage is detected in the BTLD system at the same time, it is finally determined as leakage, and the MLE-based tube leak estimation position is expressed as a probability distribution for each position characterized by giving To this end, the algorithm was applied to continuous multi-channel data acquired in time-series form, such as the temperature signal and pressure signal of the boiler.

A thermocouple, which is a temperature sensor, is attached to the tube of the boiler, and the temperature of the tube is measured through the thermocouple. If leakage occurs, a signal having a different characteristic from a normal operating situation is acquired. The algorithm learns the characteristics of signals in normal driving conditions and then diagnoses the leakage situation through the generated model.

A list of tags used in the learning model is as follows.

facility tag type number of tags Superheater 1 injection header temperature 3 metal temperature 6 Superheater 1 steam temperature 2 metal temperature 5 Superheater 3 steam temperature 3 metal temperature 5 Heater 1 metal temperature 6 steam discharge temperature One Heater 1 metal temperature 6

The data used are the temperature data of 37 tubes in the furnace for 3 months before shutdown due to leakage as shown in Table 1. About 20 days of the first 27 days that are judged to be normal are set as a validation data set for about 7 days of training, and the data including normal and leaky data after that are used as a test set (Test set) ), the algorithm was learned and verified. The positions of the superheater and reheater are shown in FIG. 12. FIG. 12 is a configuration block diagram of a power plant model for verifying a tube leakage detection system according to an embodiment of the present invention. Referring to FIG. 12, the power plant model includes a coal bunker 1210 for loading coal, a grinder 1220 for pulverizing coal, an air cooler 1230, a super heater 1220 for generating high temperature, and preheating air 13 and 14 are graphs showing results of applying the super heater 2 metal temperature signal algorithm. In particular, FIG. 13 is a graph showing the result of applying the Superheater2 metal temperature signal algorithm among the tags included in the model. In the graph shown in FIG. 13, the blue part represents the actual input data, and the yellow part represents the average value (predicted value) of the original sequences reconstructed by the algorithm. It shows the residual value indicating whether or not it is flying. Looking at the results, it can be seen that the algorithm constructs the original data well until about 50 days, but after 50 days, there is a large gap between the reconstructed data and the actual data. This is because the characteristics of the data at that time are different from the characteristics of the learned data. That is, the characteristics of the data changed from around the 50th day of the data start date, and it can be estimated that the abnormality occurred from that point on.

15 and 16 are graphs showing the result of applying the algorithm to the super heater 1 injection header temperature signal. In particular, FIGS. 15 and 16 are graphs showing results of applying the algorithm of the injection header temperature signal of Superheater 1 among the tags included in the model. Likewise, it can be seen that the algorithm constructs the original data well until about 50 days, but after 50 days there is a large difference between the reconstructed data and the actual data. As a result of this algorithm, leakage was detected about 39 days earlier than the starting point.

If a leak alarm occurs at the same time in the CNN-LSTM-AE algorithm described above and the existing BTLD system, it can be estimated with a high probability that it is a leak. . The following is an example of finding an estimated leakage position with an acoustic emission signal for the same leakage case during the same period as before.

17 is a view of installing an acoustic emission sensor in a power plant according to an embodiment of the present invention. Referring to FIG. 17, in the 'A' power plant, a sound emission sensor for detecting boiler leakage is installed on the furnace front wall as follows. Two sensors per DAQ are connected to a total of six DAQs (Data Acquisition), so sound signals can be acquired from a total of 12 sensors. As a result of frequency analysis among a total of 12 sensors, the leakage estimated location was shown using the data of 6 sensors in which leakage signals were detected.

18 is a diagram showing an example of a time difference in which signal characteristics between sensors change according to an embodiment of the present invention. Referring to FIG. 18, in order to estimate the leak location, a high-sampling signal must first be analyzed to obtain a time difference between sensors whose signal characteristics change. As shown in FIG. 18, it is a result of analyzing a 500 kHz acoustic emission signal using a Short Time Fourier Transform (STFT) technique of two sensors. It can be seen that there is a difference in the time when the energy changes in the frequency band of 0 to 100 kHz.

19 is a graph expressing a cumulative probability distribution for each sensor position according to an embodiment of the present invention. Referring to FIG. 19 , if the leakage location is expressed by applying the data obtained in FIG. 17 to Equations 1 and 2, the probability distribution for each location shown in FIG. 19 can be expressed. The point with the highest leakage probability is expressed as MPP (Most Probable Point), so that the point to be checked first during maintenance can be indicated.

In addition, the steps of a method or algorithm described in connection with the embodiments disclosed herein are implemented in the form of program instructions that can be executed through various computer means such as a microprocessor, processor, CPU (Central Processing Unit), etc. It can be recorded on any available medium. The computer readable medium may include program (instruction) codes, data files, data structures, etc. alone or in combination.

The program (command) code recorded on the medium may be specially designed and configured for the present invention, or may be known and usable to those skilled in computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks and magnetic tapes, optical media such as CD-ROMs, DVDs, and Blu-rays, and ROMs and RAMs ( A semiconductor storage element specially configured to store and execute program (instruction) codes such as RAM), flash memory, or the like may be included.

Here, examples of the program (command) code include high-level language codes that can be executed by a computer using an interpreter, as well as machine language codes such as those produced by a compiler. The hardware devices described above may be configured to act as one or more software modules to perform the operations of the present invention, and vice versa.

10: power plant
100: tube leak detection system
110: sensor block
111-1 to 111-n: first to nth sensors
120: collection unit
130: calculation unit
140: analysis unit
150: output unit
160: storage unit

Claims (1)

A sensor block 110 for detecting a state signal of the power generation facility 10;
a collection unit 120 that collects the state signal;
a calculator 130 that calculates a probability distribution for each sensor using the collected state signals; and
An analysis unit 140 for estimating whether there is leakage from the tube of the power generation facility 10 according to the probability distribution for each sensor; includes,
The sensor block 110 includes at least two sensors 111-1 to 111-n that detect the state signal,
The calculation unit 130 determines the leakage position from each sensor 111-1 to 111-n by using the time difference between the at least two sensors 111-1 to 111-n when the signal characteristics of the state signal change. calculate the distance to
The calculator 130 calculates the distance between the sensor and the leak location using the measured wave propagation speed and the time difference,
The operation signal of the power generation facility 10 is partially overlapped at regular time intervals through a windowing technique and converted into a primary sequence signal 821 having a sequence number, a sequence length, and a tag. Tube leak detection system using status signal.

Images