KR102256980B1 - Gas turbine combustion instability diagnosis system based on machine learning and method thereof - Google Patents

Abstract

A gas turbine combustion instability diagnosis system and a gas turbine combustion instability diagnosis method based on machine learning are disclosed. The gas turbine combustion instability diagnosis system according to an embodiment of the present invention includes a combustion unit including a sensor unit installed inside the combustion unit to measure combustion data inside the combustion unit, and features that can reflect the combustion state. Select and acquire fully developed stable combustion data and fully developed unstable combustion data from several combustors, calculate selected features from the obtained combustion data, and use the calculated features as learning data. Using as, an artificial neural network model for diagnosis of combustion instability can be trained, and the combustion instability state of the combustor can be determined using the artificial neural network model.

Description

Gas turbine combustion instability diagnosis system based on machine learning and method thereof

The present invention relates to a technology for diagnosing combustion instability during combustion of a gas turbine, and more specifically, a combustion state in real time by learning an artificial neural network model for determining whether combustion instability is based on combustion data measured inside the combustor. It relates to a technology that enables quick and accurate diagnosis of gas turbine combustion instability.

In general, a gas turbine is composed of a compressor, a combustor, a turbine, and a generator, and when the high-pressure air by the compressor and the fuel gas which has become hot through a heat exchanger are sent to the combustor and burned, the turbine is driven by the combustion gas.

In such a gas turbine, the combustion temperature in the combustor rises to about 1500°C, and pressure fluctuations or flame position fluctuations due to unstable combustion occur inside, resulting in local stress concentration and thermal cycle fluctuations, causing cracks. have. When cracks or breakages occur in the combustor, there is a problem in that the amount of air introduced into the combustor deviates from the plan, and combustion abnormality occurs and power generation efficiency decreases, or in some cases, the broken piece enters the turbine and damages the blades. Therefore, it is an important task to detect cracks and various abnormalities early.

For monitoring abnormalities in plants such as gas turbines, for example, a temperature detector is installed in the exhaust chamber, and the pattern characteristic of the exhaust temperature distribution in the cross-section is obtained from the temperature distribution in the plane orthogonal to the exhaust gas flow obtained by the installed temperature detector. A gas turbine combustion monitoring device for determining the cause has been proposed.

In addition, a gas turbine combustor monitoring device may be proposed that provides a temperature detector for detecting a surface temperature distribution in the combustor and determines the cause of an abnormality in the combustor based on the surface temperature distribution obtained by the temperature detector.

In addition, since the performance deterioration or failure of the gas turbine is precisely detected, operation data of each part of the gas turbine is detected, the detected operation data is standardized and sampled over a predetermined period of time, and then the sampled data is calculated as a moving average. An apparatus technology for diagnosing the operating state of a gas turbine that is processed to diagnose the operating state of the gas turbine based on the data may be presented.

However, this technique has a problem in that a plurality of temperature detectors are required in one combustor because the characteristic of the temperature distribution pattern is obtained from the temperature distribution obtained by the temperature detector provided in the exhaust chamber or the combustor.

The present invention was invented to solve the above problems, and by diagnosing combustion instability by applying combustion data measured in real time inside the combustor to a learned artificial neural network model, pressure fluctuations due to unstable combustion of a gas turbine, To provide a gas turbine combustion instability diagnosis system capable of detecting local stress concentration due to fluctuations in flame position and cracks due to heat cycle fluctuations in advance, and a gas turbine combustion instability diagnosis method using the same.

In addition, by fine-tuning the learned artificial neural network model to determine combustion instability, gas turbine combustion instability diagnosis, which determines whether the gas turbine is combustion instability more quickly and accurately, and at the same time informs it so that the corresponding measures can be taken quickly. It is to provide a system and a gas turbine combustion instability diagnosis method using the same.

In the gas turbine combustion instability diagnosis system according to an embodiment of the present invention, a combustion unit including a sensor unit mounted inside the combustion unit to measure combustion data inside the combustion unit, features that can reflect the combustion state Is selected, fully developed stable combustion data and fully developed unstable combustion data are obtained from various combustors, and selected features are calculated from the obtained combustion data, and the calculated features are learned. It may include a diagnostic unit for learning an artificial neural network model for diagnosis of combustion instability using the data, and a diagnosis unit for determining the combustion instability of the combustor using the artificial neural network model and a combustion control unit for controlling the operation of the combustion unit according to the determination of the diagnosis unit. have.

In addition, in the gas turbine combustion instability diagnosis system according to an embodiment of the present invention, features are the magnitude of the dynamic pressure signal, the rate of change in the magnitude of the dynamic pressure signal, and time, which are features that can be obtained from the dynamic pressure signal. Distribution within a unit frame in a time domain, a rate of change in distribution within a unit frame in a time domain, a frequency distribution within a unit frame in a frequency domain, and a rate of change in a frequency distribution within a unit frame in a frequency domain It may include at least one of.

In addition, in the gas turbine combustion instability diagnosis system according to an embodiment of the present invention, features are, a difference in intensity or contrast between an image of a previous time and an image of a current time within a specific area, It may include at least one of a distance difference between a feature point (Corner, Edge) between an image of time and an image of current time, and Scale Invariant-Scale Invariant Feature Transform (SIFT).

In addition, in the gas turbine combustion instability diagnosis system according to an embodiment of the present invention, features may include at least one of a size of a temperature and a rate of change of a temperature.

In addition, in the gas turbine combustion instability diagnosis system according to an embodiment of the present invention, features may include Dynamic Pressure Energy and Relative Entropy of Energy.

In addition, in the gas turbine combustion instability diagnosis system according to an embodiment of the present invention, the diagnosis unit performs fine tuning on the learned artificial neural network model by using combustion data of a transient process as additional training data. can do.

In addition, in the gas turbine combustion instability diagnosis system according to an embodiment of the present invention, the artificial neural network model may be a single layer perceptron model.

The method for diagnosing combustion instability in a gas turbine according to an embodiment of the present invention includes the steps of selecting features that can reflect a combustion state, combustion data in a fully developed stable state from various combustors, and an unstable state. Acquiring (Fully developed unstable) combustion data, calculating selected features from the obtained combustion data, learning an artificial neural network model for combustion instability diagnosis using the calculated features as training data, and artificial neural networks It may include the step of determining the combustion instability state of the combustor using the model.

In addition, in the gas turbine combustion instability diagnosis method according to an embodiment of the present invention, features are the magnitude of the dynamic pressure signal, which are features that can be obtained from the dynamic pressure signal, the rate of change in the magnitude of the dynamic pressure signal, and time. Distribution within a unit frame in a time domain, a rate of change in distribution within a unit frame in a time domain, a frequency distribution within a unit frame in a frequency domain, and a rate of change in a frequency distribution within a unit frame in a frequency domain It may include at least one of.

In addition, in the gas turbine combustion instability diagnosis method according to an embodiment of the present invention, features are, a difference in intensity or contrast between an image of a previous time and an image of a current time within a specific area, It may include at least one of a distance difference between a feature point (Corner, Edge) between an image of time and an image of current time, and Scale Invariant-Scale Invariant Feature Transform (SIFT).

In addition, in the method for diagnosing combustion instability of a gas turbine according to an embodiment of the present invention, features may include at least one of a size of a temperature and a rate of change of a temperature.

In addition, in the gas turbine combustion instability diagnosis method according to an embodiment of the present invention, features may include Dynamic Pressure Energy and Relative Entropy of Energy.

In addition, in the gas turbine combustion instability diagnosis method according to an embodiment of the present invention, the step of learning an artificial neural network model for diagnosis of combustion instability by using the calculated features as learning data includes: It may further include performing fine tuning on the learned artificial neural network model using the additional training data.

In addition, in the gas turbine combustion instability diagnosis method according to an embodiment of the present invention, the artificial neural network model may be a single layer perceptron model.

Meanwhile, as an embodiment of the present invention, a computer-readable recording medium in which a program for executing the above-described method on a computer is recorded may be provided.

According to a gas turbine combustion instability diagnosis system according to an embodiment of the present invention and a gas turbine combustion instability diagnosis method using the same, combustion data in a fully developed stable state and combustion in a fully developed unstable state from various combustors An artificial neural network model for determining whether combustion instability is determined may be trained by acquiring data and extracting training data from the acquired combustion data.

According to a gas turbine combustion instability diagnosis system according to an embodiment of the present invention and a gas turbine combustion instability diagnosis method using the same, a previously learned artificial neural network model is further elaborated by using combustion data of a transition process as additional training data. You can learn to do it.

According to the gas turbine combustion instability diagnosis system according to an embodiment of the present invention and a gas turbine combustion instability diagnosis method using the same, an artificial neural network model is applied as a method for determining whether combustion instability is used, so that it is possible to diagnose combustion instability more accurately and quickly. Do.

According to the gas turbine combustion instability diagnosis system according to an embodiment of the present invention and a gas turbine combustion instability diagnosis method using the same, an alarm for a combustion instability state can be displayed in real time by enabling rapid diagnosis of combustion instability. By quickly reducing the load on the gas turbine, safety accidents can be prevented in advance.

1 is a schematic configuration diagram of a gas turbine combustion instability diagnosis system according to an embodiment of the present invention.
FIG. 2 is an enlarged view of an area A of FIG. 1.
3 is a diagram showing the magnitude of a dynamic pressure signal according to a combustion state. FIG. 3A is a diagram showing the magnitude of the dynamic pressure signal in a stable state, and FIG. 3B is a diagram showing the magnitude of the dynamic pressure signal in an unstable state.
4 is a diagram showing a frequency distribution in a unit frame according to a combustion state in a frequency domain. FIG. 4A is a diagram showing a frequency distribution in a unit frame in a stable state, and FIG. 4B is a diagram showing a frequency distribution in a unit frame in an unstable state.
5 is a view showing an image measured during the combustion process. FIG. 5A is a view showing a flame in a stable state, and FIG. 5B is a view showing a flame in an unstable state.
6 is a diagram illustrating three features calculated using combustion data and shown in a feature space.
7 is a diagram illustrating a method of generating a hyperplane that separates true data and false data in a feature space using a binary classification model.
8 is a diagram showing a Perceptron model among artificial neural network models.
9 is a diagram showing a deep neural network model among artificial neural network models.
10 is a diagram schematically illustrating a Perceptron model.
11 is a single layer perceptron model using dynamic pressure energy (DPE) and relative energy entropy (ReEoE) in a gas turbine combustion instability diagnosis system based on machine learning according to an embodiment of the present invention. It is a diagram showing a result of selecting as features of and using it as learning data on a feature space.
12 is a diagram illustrating a method of fine tuning a previously learned perceptron model using transient combustion data (dynamic pressure signal) according to an embodiment of the present invention.
13 is a diagram showing an example of application of empirical data to a gas turbine combustion instability diagnosis system based on machine learning according to an embodiment of the present invention.
14 is a flowchart illustrating a method for diagnosing combustion instability in a gas turbine according to an embodiment of the present invention.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings so that those of ordinary skill in the art can easily implement the present invention. However, the present invention may be implemented in various different forms and is not limited to the embodiments described herein. In the drawings, parts irrelevant to the description are omitted in order to clearly describe the present invention, and similar reference numerals are attached to similar parts throughout the specification.

The terms used in the present specification will be briefly described, and the present invention will be described in detail.

Terms used in the present invention have selected general terms that are currently widely used as possible while taking functions of the present invention into consideration, but this may vary according to the intention or precedent of a technician working in the field, the emergence of new technologies, and the like. In addition, in certain cases, there are terms arbitrarily selected by the applicant, and in this case, the meaning of the terms will be described in detail in the description of the corresponding invention. Therefore, the terms used in the present invention should be defined based on the meaning of the term and the overall contents of the present invention, not a simple name of the term.

When a part of the specification is said to "include" a certain component, it means that other components may be further included rather than excluding other components unless specifically stated to the contrary. In addition, terms such as "... unit" and "module" described in the specification mean a unit that processes at least one function or operation, which may be implemented as hardware or software or a combination of hardware and software. . In addition, when a part is said to be "connected" with another part throughout the specification, this includes not only the case of being "directly connected", but also the case of being connected "with another element in the middle."

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

1 is a schematic configuration diagram of a gas turbine combustion instability diagnosis system according to an embodiment of the present invention, and FIG. 2 is an enlarged view of area A of FIG. 1.

1 and 2, the gas turbine combustion instability diagnosis system according to an embodiment of the present invention detects combustion instability of various combustors such as a gas turbine, and utilizes this to reduce combustion instability of the gas turbine. It enables stable and efficient operation. To this end, the gas turbine combustion instability diagnosis system may include a combustion unit 10, a control processor 20, and the like.

The gas turbine combustion instability diagnosis system of this configuration detects combustion instability of various combustors such as gas turbines and controls the load in response to the detected combustion instability, thereby enabling stable and efficient operation of the gas turbine.

Specifically, referring to the configuration of the gas turbine combustion instability diagnosis system according to an embodiment of the present invention, compressed air and fuel for combustion may be supplied to the combustion unit 10. That is, the gas burned in the combustion unit 10 may become a power source capable of driving the gas turbine. In the combustion unit 10, combustion instability may occur due to various causes such as compressed air, insufficient fuel, and a mixing ratio during the operation of the gas turbine.

In addition, the combustion unit 10 may include a sensor unit (not shown) that is mounted inside the combustion unit to obtain combustion data inside the combustion unit.

The sensor unit (not shown) includes a temperature sensor (not shown) that measures the temperature inside the combustor 10, an image sensor (not shown) that can acquire an image inside the combustor 10, and a dynamic pressure sensor that measures dynamic pressure ( 100) may be included. For example, the dynamic pressure sensor 100 is mounted in the combustion unit 10 and may be connected through a conduit connected to the combustion unit 10. That is, the dynamic pressure sensor 100 is for measuring combustion instability, and may measure the combustion dynamic pressure inside the combustion unit 10 to measure vibration and resonance phenomena inside the combustion unit 10 during the combustion operation.

For example, the dynamic pressure sensor 100 may be a sensor capable of measuring a dynamic pressure inside the combustion unit 10. In addition, the combustion unit 10 is adjacent to the dynamic pressure sensor 100, a pressure sensor 120 for measuring a dynamic pressure signal in the combustion process of the combustion unit 10 and a signal measurement for measuring a signal generated from the pressure sensor 120 It may include a unit 140 and the like.

In this way, the combustion instability inside the combustion unit 10 measured through the dynamic pressure sensor 100 may be referred to as a kind of resonance phenomenon. In other words, it means a phenomenon in which a specific frequency component predominately increases due to a resonance phenomenon caused by combustion instability.

The dynamic pressure signal generated in the combustion process of the combustion unit 10 through the dynamic pressure sensor 100 and the signal measuring unit 120 of the combustion unit 10 may be transmitted to the control processor 20. That is, the dynamic pressure signal transmitted to the control processor 20 may be used as data that can determine whether the current combustion operation is unstable or stable.

Referring to FIG. 1, when the control processor 20 is described in detail, the control processor 20 may include a sensitivity analysis unit 22, a diagnosis unit 24, a combustion control unit 26, and the like.

The sensitivity analysis unit 22 is a device that analyzes sensitivity in order to determine the number of frames of the combustion dynamic pressure signal or a specific length of the frame according to the combustion dynamic pressure measured by the dynamic pressure sensor 100.

Specifically, sensitivity is the degree to which the value of the discrimination factor (the reliability of the discrimination factor) changes according to the number of calculated combustion dynamic pressure signals when the combustion instability discrimination factor is calculated based on the dynamic pressure signal measured inside the combustion unit 10. Means. That is, when obtaining desired information (combustion instability discrimination factor) from the dynamic pressure signal measured during a combustion operation for a certain period of time, it is possible to select and calculate only a frame having a specific length of time, not the entire area of the specific signal.

For example, if the dynamic pressure sensor 100 measures the dynamic pressure signal within the combustion unit at a 10 kHz period, 10,000 combustion dynamic pressure signals may be obtained for 1 second. Here, in order to determine whether the combustion instability is unstable, it is possible to calculate the discriminant factor using 10,000 dynamic pressure signals while holding a frame for 1 second, but the amount of calculation is too large. Therefore, if each frame (a specific length to be calculated) is set to a specific length of 0.001 to 0.1 second interval, 10 to 1000 measurement data are calculated, so the amount of calculation can be reduced.

At this time, what should be considered in order to select a specific length is whether a sufficient signal (frequency) is included in a specific frame, and for this purpose, a sensitivity analysis is performed to determine the number of signals in a specific frame. That is, the sensitivity analysis unit 22 may determine the number of analysis data that can be analyzed in one frame based on the sensitivity analysis result.

However, if the number of analysis data included in one frame is too small, the amount of calculation may be reduced, but the number of analysis data is small, so an error may be included in the combustion instability determination result or an incorrect result may be displayed. Therefore, the determination of the number of data that can be analyzed in one frame must be determined to have a sufficient number in consideration of the possibility of errors in the result of determination of combustion instability and the amount of computation.

Therefore, due to this sensitivity analysis, the number of data for determination and diagnosis of combustion instability can be specified. That is, through sensitivity analysis, the number of data used in one calculation (one frame) is selected, and only the selected number of data is calculated through this, so that the combustion state can be immediately determined. Through this sensitivity analysis, the amount of computation can be reduced by reducing the number of analysis data used in one calculation, and the resolution of the time series can be improved by selecting a sufficient number of analysis data.

The diagnosis unit 24 according to an embodiment of the present invention may select features that can reflect a combustion state in order to determine combustion instability.

In the combustion process, basic information reflecting the corresponding combustion state may appear as pressure, light, and heat. Based on this, combustion data that can be measured inside the combustor 10 during the combustion process may include a dynamic pressure (DP), an image, a temperature, and the like.

That is, combustion data that can be measured during the combustion process, such as a dynamic pressure signal, an image, and a temperature, may be basic data for diagnosing combustion instability.

Feature is a term commonly used in machine learning or pattern recognition, and refers to a characteristic or measurable property found in an observation object (burning data). Therefore, selecting a feature that is informative, discriminating, and independent from the object to be observed can have a significant influence on the problem-solving algorithm. Therefore, it is necessary to select a feature that can reflect an appropriate combustion state from the combustion data.

The problem of classifying and determining whether the data is true or false by receiving data, that is, the problem of classifying data into two states is called a binary classification model. That is, since it is a problem of determining whether the combustion is stable (true) or unstable (false) from the combustion data that can be measured during the combustion process, it can be determined by applying a binary classification model.

In general, the methods commonly used in binary classification models are Decision trees, Random forests, Bayesian networks, Clustering, and Kernel support vector machines. Support Vector Machine), Neural Network. Logistic regression, Probit model, etc. may be included.

In order to apply a binary classification model in an artificial neural network, it must be sufficiently learned with a limited number of training data, the classification (discrimination stability, whether there is instability) must be fast, and the classification accuracy must be high.

Therefore, in order to apply a binary classification model as a model for diagnosing combustion instability, it must be sufficiently learned with limited training data, fast classification must be possible, and classification accuracy must be high.

In other words, in order to apply a binary classification model, the classification accuracy is high, the classification can be made clearly, so the classification speed can be achieved quickly, and the combustion data that can be sufficiently learned and features that can be calculated correspondingly Features should be selected.

So, let's look at the general facts known about the state of thermo-acoustic combustion instability.

It is known that thermo-acoustic combustion instability occurs when the phase of heat releases and acoustic wave inside the combustor coincide with the occurrence of resonance, resulting in an increase in the internal pressure.

For example, a pressure in a stable state inside the combustor shows a small value (± 0.05 [kPa]), but a pressure in an unstable state appears as a large value (about ± 0.8 [kPa]).

In addition, the pressure in a stable state inside the combustor has a constant and flat frequency component distribution, but the pressure in an unstable state is an excitation frequency component in which'the phase of heat releases and the acoustic wave inside the combustor are matched'. It appears to have a dominant frequency distribution.

In general, in the combustion process, basic information reflecting the combustion state may appear as pressure, light, and heat. Combustion data capable of measuring the combustion process based on this may include a dynamic pressure (DP), an image, a temperature, and the like.

That is, features that can be obtained or calculated from combustion data in order to determine whether combustion is unstable will be values that can be calculated or obtained from a dynamic pressure signal (Dynamic Pressure, DP), an image, and a temperature.

For example, features that can be obtained from the dynamic pressure signal as combustion data are the magnitude of the dynamic pressure signal, the rate of change in the magnitude of the dynamic pressure signal, the distribution within the unit frame in the time domain, and the time domain. It may include a distribution change rate within a unit frame, a frequency distribution within a unit frame in a frequency domain, a frequency distribution change rate within a unit frame in a frequency domain, and the like.

Let's look at how the features obtained from the dynamic pressure signal can reflect the stability and instability of combustion.

3 is a diagram showing the magnitude of a dynamic pressure signal according to a combustion state. FIG. 3(a) is a diagram showing the magnitude of the dynamic pressure signal in a stable state, and FIG. 3(b) is a diagram showing the magnitude of the dynamic pressure signal in an unstable state.

Referring to FIG. 3, if the combustion is maintained to be stable (in the case of (a) of FIG. 3), the magnitude of the internal dynamic pressure signal does not exceed ±0.05 [kPa]. However, if the combustion remains unstable (in the case of Fig. 3(b)), the magnitude of the internal dynamic pressure signal when the limit cycle is reached becomes about ±0.8 [kPa].

Therefore, if the dynamic pressure signal is measured inside the combustor and the magnitude is within ±0.05[kPa], the combustion state is stable, and if the magnitude is ±0.8[kPa], the combustion state can be determined as an unstable state. In other words,'the magnitude of the dynamic pressure signal', which is a characteristic acquired from the dynamic pressure signal, can reflect the state (information) of whether combustion is stable or unstable.

4 is a diagram showing a frequency distribution in a unit frame according to a combustion state in a frequency domain. FIG. 4A is a diagram showing a frequency distribution in a unit frame in a stable state, and FIG. 4B is a diagram showing a frequency distribution in a unit frame in an unstable state.

Referring to FIG. 4, when combustion is maintained to be stable (in the case of (a) of FIG. 4), the dynamic pressure signal shows a constant and flat distribution in the frequency component, but the combustion becomes unstable (in the case of (b) of FIG. 4). If maintained, it can be seen that the frequency of the dynamic pressure signal shows a distribution with a large magnitude only at the excitation frequency.

Therefore, if the dynamic pressure signal is measured inside the combustor and the frequency distribution is constant and shows a flat distribution, the combustion state is stable, and if the frequency distribution is narrow, the combustion state can be determined as an unstable state. That is, the'frequency distribution within a unit frame in the frequency domain', which is a characteristic acquired from the dynamic pressure signal, can reflect the state (information) of whether combustion is stable or unstable.

In addition, the features that can be calculated or obtained from the image acquired inside the combustion unit are the difference in intensity or contrast between the image at the previous time and the image at the current time within a specific area, and the image at the previous time. And the distance difference between feature points (Corner, Edge) between the image and the current time, Scale Invariant-Scale Invariant Feature Transform (SIFT), and the like.

Let's look at how the features acquired from the image inside the combustion section can reflect the stability and instability of combustion.

5 is a view showing an image measured during the combustion process. FIG. 5A is a view showing a flame in a stable state, and FIG. 5B is a view showing a flame in an unstable state.

Referring to FIG. 5, looking at the difference in intensity or contrast between an image of a previous time and an image of a current time within a specific area, when the combustion is maintained to be stable (in the case of (a) of FIG. 5), the flame ( flame) has a constant intensity over time. However, if the combustion is maintained unstable (in the case of (b) of FIG. 5), it can be seen that the intensity of the flame is not constant over time.

Therefore, by measuring images inside the combustor in real time, if the brightness or contrast of a specific region is constant over time, the combustion state is stable, and if the brightness or contrast of a specific region is not constant over time, the combustion state is unstable. It can be judged by the state. In other words,'the difference in intensity or contrast between the image of the previous time and the image of the current time', which is a feature acquired from an image, reflects the state (information) of whether the combustion is stable or unstable. It can be done.

In addition, features that can be obtained or calculated from the temperature inside the combustion unit may include a temperature size, a rate of change of temperature, and the like.

Next, let's look at the feature space.

The feature space refers to a space in which features obtained through combustion data are used as respective axes. If there are n features calculated from the combustion data, a feature space generated based on this is an n-dimensional space.

The roles and features of the feature space are as follows.

When determining combustion instability by calculating various features from the measured combustion data, the feature space expresses the relationship between the various features on the space and indicates the combustion state of the measured combustion data. Can help you understand.

6 is a diagram illustrating three features calculated using combustion data and shown in a feature space.

Specifically, the dynamic pressure signal and image are measured inside the combustor, and'the size of the dynamic pressure signal','the frequency distribution within a unit frame in the frequency domain','the brightness between the image at the previous time and the image at the current time within a specific area. This is a diagram in which three features of'(intensity) or contrast difference' are calculated and expressed in a feature space.

Referring to FIG. 6, since each feature reflects a state of stable or unstable combustion, it can be seen that they are expressed on similar regions. That is, it can be seen that features acquired in a stable combustion process are separated from features acquired in an unstable combustion process to form a space.

However, when determining whether combustion is unstable or not, when determining the combustion state using only a single feature from the measured combustion data, stability and instability cannot be clearly identified.

For example, suppose that a dynamic pressure signal is measured inside a combustor and combustion instability is determined by using the'magnitude of dynamic pressure signal' which is a characteristic acquired from the dynamic pressure signal. At this time, when the magnitude of the dynamic pressure signal is more than a certain amount, the criterion for determining that'the combustion process is unstable' becomes the most important problem. For reasons such as measurement error (outlier) of the sensor itself and fluctuation of the dynamic pressure signal caused by the complexity of the combustion process, the determination of instability and stability with only a single feature, the'dynamic pressure signal size' (single feature) It can be difficult to set standards clearly.

In general, if combustion is stable, the magnitude of the dynamic pressure signal is within ±0.05 [kPa]. However, referring to (a) of FIG. 3, due to a measurement error of the sensor itself or fluctuation occurring in the combustion process, data such as 0.35, 0.43, and 0.82 seconds deviate from the magnitude of the dynamic pressure signal in a generally stable state. Can be confirmed.

Therefore, when determining whether combustion is unstable, it is possible to clearly determine whether the combustion state is unstable by using various features.

For example, a dynamic pressure signal is measured inside the combustor, and two features of'the magnitude of the dynamic pressure signal' and'frequency distribution within a unit frame in the frequency domain' are obtained from the dynamic pressure signal. Suppose you use features to determine combustion instability. The biggest difference between determining combustion instability using two features and using a single feature is that there are two criteria for determining combustion instability.

As described above, when the combustion state is maintained stably, the magnitude of the dynamic pressure signal is small and the frequency distribution (frequency distribution in the unit frame in the frequency domain) of the corresponding signal appears constant and flat. In addition, when the combustion state is maintained unstable, the magnitude of the dynamic pressure signal is large, and the frequency distribution of the signal appears narrow.

That is, if there are these two criteria, it can be a solution to a problem that may occur when a single feature is used (measurement error of the sensor itself or fluctuation occurring in the combustion process). This is because the size of the dynamic pressure signal exceeds the reference value due to measurement error or fluctuation of the sensor, but if the frequency distribution is kept constant and flat, the combustion state can be determined to be stable. In other words, it is possible to clearly determine whether combustion is unstable through two criteria (features).

That is, when determining whether combustion instability is detected, if several features are used, it is possible to more accurately and clearly determine whether combustion instability is used than in the case of using a single feature.

For example, by measuring a dynamic pressure signal inside the combustor, and obtaining two features of the measured dynamic pressure signal: the magnitude of the dynamic pressure signal and the frequency distribution within a unit frame in the frequency domain. Suppose you want to determine whether combustion is unstable.

Here, in determining whether combustion is unstable,'the magnitude of the dynamic pressure signal' and'the frequency distribution in a unit frame in the frequency domain' have different criteria, so the timing of discrimination of instability may be expressed differently. . In general, in the case of discrimination based only on the'size of the dynamic pressure signal', since it takes time for the size to become sufficiently large, the timing at which it is determined that combustion is unstable may be expressed late. However, by considering the'frequency distribution in a unit frame in the frequency domain' together, even if the size is not large enough, if the frequency distribution of the dynamic pressure signal is narrowed, it can be determined that combustion is unstable, so the discrimination time is late. You can compensate for the shortcomings.

In addition, combustion instability can be more accurately determined by integrating combustion data (dynamic pressure signal, image, temperature) that can be obtained during the combustion process.

In the combustion process, basic information reflecting the corresponding combustion state may appear as pressure, light, and heat. Although it is a thermo-acoustic combustion instability phenomenon, if several combustion data (dynamic pressure signal, image, temperature) representing this are considered together, combustion instability can be more accurately determined. That is, a feature space can be created from multiple features of a single combustion data (e.g., only dynamic pressure signals are considered), but calculated from multiple combustion data (dynamic pressure signals, images, temperatures). A feature space may be created through several features.

7 is a diagram illustrating a method of generating a hyperplane that separates true data and false data in a feature space using a binary classification model.

Referring to FIG. 7, a hyperplane for separating true data and false data in a feature space may be generated using features obtained from various combustion data.

That is, if the data of the stable combustion process and the features acquired from the data of the unstable combustion process are expressed in the feature space, the data area of the true (stable state) and the data of the false (the unstable state area) are expressed. ). In this case, a hyperplane separating the true data from the false data may be created in the feature space.

Through this, if the features obtained from the combustion data measured in real time exist at the bottom of the hyperplane in the feature space, it is true (stable combustion process), and if it exists at the top of the hyperplane, it is false (unstable combustion process). I can judge.

A typical machine learning method that performs binary classification can be performed using an artificial neural network method that simulates a human brain, which is a set of neurons.

In the artificial neural network, a layer that performs a convolution operation may be added according to the characteristics of the input data, and if there are many characteristics to be reflected, the hidden layer of the neural network may be deepened. Here, the deep learning method is dealing with an artificial neural network with a deep neural network.

8 is a diagram illustrating a Perceptron model among artificial neural network models, and FIG. 9 is a diagram illustrating a deep neural network model among artificial neural network models.

As described above, the method of determining whether combustion instability should be sufficiently learned with limited learning data, fast classification should be possible, and the conditions of high classification accuracy should be satisfied.

In order to review artificial neural network models that satisfy the above conditions, a Perceptron model and a deep neural network model among artificial neural network models may be compared and examined.

Referring to FIG. 8, in the case of the Perceptron model, since the layer is a single layer, the number of weights to be obtained through training is small (learning is possible with less training data), and the structure is simple, so fast classification is performed. It is possible. In addition, the classification accuracy can be solved through fine tuning.

Referring to FIG. 9, in the case of a deep neural network model, since the layers are multiple layers, the number of weights to be obtained through training is large (it is impossible to learn with little training data), and the structure is complex, so fast classification is impossible. Do.

Therefore, since the Perceptron model is simpler than the deep neural network model, the number of weights to be trained is small, and it is possible to quickly discriminate.

In other words, in the case of a perceptron model, the number of weights is smaller in the case of a single layer, so it is expressed more simply than a deep neural network model.

The Perceptron model is an artificial neural network devised by Frank Rosenblatt and can be understood as a linear classifier that divides input vectors into two classes.

10 is a diagram schematically illustrating a Perceptron model.

Referring to FIG. 10, x _{i on the} left represents a value of input data, and w _i represents a weight. The bias input value is expressed as x ₀ , the bias slope is expressed as w ₀ , and f indicates the activation function.

The definition of terms used in the Perceptron model is as follows.

Threshold refers to a minimum value for activating a certain value, and weight refers to a value indicating a direction or form in which learning is performed. That is, since the purpose of the Perceptron model is to find a linear boundary (Hyperplane) for linearly classifying the training data (vector) into two classes, the weight means a value representing the direction or shape of such a linear boundary. .

Bias is a value representing the intercept of a linear boundary, and in the case of a straight line, it represents the y-intercept (= w ₀ *x ₀ ,).

The net value is a value obtained by summing the product of input data and weight. When geometrically analyzed, it is the same as the equation of a linear boundary (hyperplane), and can be expressed by the following <Equation 1>.

[Equation 1]

The activation function means a function that outputs 1 when the calculated net value is greater than or equal to the threshold value, and outputs 0 when it is less than the threshold value, and may be expressed by the following <Equation 2>.

[Equation 2]

However, this definition is only valid in a single-layer perceptron model, and a different type of activation function is used in the multi-layer perceptron model.

That is, a step function may be used as the most basic active function used in the perceptron model. That is, the activation function may activate the output value (output 1) when the net value is greater than or equal to the threshold value, and deactivate the output value (output 0) when the net value is less than the threshold value.

Therefore, the learning algorithm of the perceptron model can be expressed by the following <Equation 3>.

[Equation 3]

In addition, fine tuning refers to a method of finely adjusting the weight of the model by inputting additional data to the previously learned perceptron model.

For example, dynamic pressure energy (DPE) can be calculated when a fully developed stable state is maintained using a dynamic pressure signal, and dynamic pressure when a fully developed unstable is maintained. Each energy can be calculated.

In other words, if the perceptron model is trained using the dynamic pressure energy when the fully developed stable and fully developed unstable states are maintained, a certain degree of generality can be obtained, but combustion data measured in real time. If the dynamic pressure energy (DPE) value calculated in is measured as an intermediate value between the stable state and the unstable state, it is difficult to accurately determine whether the corresponding signal is combustion unstable.

Therefore, in order to accurately determine whether combustion instability or not, the weight of the previously learned model is used as additional training data using the transitional combustion data between the fully developed stable and the fully developed unstable state. Fine tuning to finely adjust (Weight) can be performed.

11 is a single layer perceptron model using dynamic pressure energy (DPE) and relative energy entropy (ReEoE) in a gas turbine combustion instability diagnosis system based on machine learning according to an embodiment of the present invention. It is a diagram showing a result of selecting as features of and using it as learning data on a feature space.

Dynamic pressure energy (DPE) can be expressed as in Equation 4 below, and relative energy entropy ReEoE can be expressed as in Equation 5 below.

[Equation 4]

[Equation 5]

Referring to Equation 5, P(x) may be a probability value. Here, P(x): The probability of the dynamic pressure signal energy may be a probability value as to whether the magnitude of the signal changes frequently within a specific unit frame.

Here, n is the number of signals in the frame, 1 frame is divided into 0.02 seconds, measured using the number of signals for 0.02 seconds, and the corresponding dynamic pressure energy (DPE) and relative energy entropy (ReEoE) are calculated. Thus, it is shown in FIG. 11.

The x-axis of FIG. 11 represents dynamic pressure energy (DPE), and the y-axis represents relative energy entropy (ReEoE). The graph of FIG. 11B is a diagram showing the graph of FIG. 11A converted from the y-axis to a logarithmic scale.

Referring to FIG. 11, it can be seen that a data area in a state in which combustion is stably maintained and a data area in an unstable state are clearly separated. In particular, it can be seen that in the graph of FIG. 11 (b) represented by a log scale, it is more clearly distinguished.

That is, as a result of learning a perceptron model using dynamic pressure energy (DPE) and relative energy entropy (ReEoE), a two-dimensional feature space is created, so that a hyperplane for determining whether combustion instability is in a straight line (Fig. 11).

Referring to FIG. 11, when combustion data (dynamic pressure signal) is input in real time, the combustion is in a stable state (Ture = 1) when it is located under a red straight line, and when it is located on a red straight line, the corresponding combustion Can be determined as an unstable state (False = 0).

Next, let's look at a method of performing fine tuning in the perceptron model for determining whether combustion instability is present.

12 is a diagram illustrating a method of fine tuning a previously learned perceptron model using transient combustion data (dynamic pressure signal) according to an embodiment of the present invention.

Referring to FIG. 12, a new perceptron model may be trained by adding features acquired from transition combustion data to a previously learned perceptron model as new training data. In this case, as shown in (b) of FIG. 12, a new hyperplane for determining whether combustion is unstable may be generated.

That is, the perceptron model undergoes new learning through fine tuning, and a new hyperplane corresponding thereto is generated so that combustion instability can be more accurately determined.

13 is a diagram showing an example of application of empirical data to a gas turbine combustion instability diagnosis system based on machine learning according to an embodiment of the present invention.

Referring to FIG. 13, a solid blue line indicates combustion dynamic pressure data 210, a green dotted line indicates a relative energy entropy (ReEoE) 220, and a black triangle indicates an index perceptron 230.

That is, in Fig. 13, each of the six graphs (Fig. 13 (a) to (f)) is the dynamic pressure energy (DPE) and relative energy, which are characteristics of the dynamic pressure signal from the combustion dynamic pressure data 210 measured inside the combustor. It is a diagram showing the result of obtaining the index perceptron 230 by determining combustion instability using the learned perceptron model after calculating entropy (ReEoE).

Referring to FIG. 13, the learned perceptron model is a case where combustion is determined to be stable when the value of the index perceptron 230 is 1, and the value of the index perceptron 230 is 0. In the case of, it can be confirmed that the combustion is determined to be unstable.

The gas turbine combustion instability diagnosis system based on machine learning according to an embodiment of the present invention determines whether combustion instability is determined using the learned perceptron model, and operates the combustion unit 10 based on whether combustion instability is determined. Can be controlled.

For example, if the combustion control unit 26 determines that the combustion instability state (index perceptron = 0) is 0.06 seconds or more using the perceptron model learned by the diagnostic unit 24, an alarm for the combustion instability state is generated. It can be displayed, and if it is determined that the combustion unstable state has been maintained for more than 0.1 seconds, it can be controlled to reduce the load of the gas turbine.

14 is a flowchart illustrating a method for diagnosing combustion instability in a gas turbine according to an embodiment of the present invention.

Referring to FIG. 14, in step S10, the diagnosis unit 24 may select features that can reflect the combustion state.

As described above, in order to apply an artificial neural network model (eg, a perceptron model) to determine whether combustion instability has occurred, features that can be used as training data must be selected in advance.

That is, the diagnosis unit 24 may preselect learning data for determining whether combustion instability is determined as a feature. Features that can be obtained or calculated from combustion data to determine whether combustion is unstable will be values that can be calculated or obtained from a dynamic pressure signal (Dynamic Pressure, DP), an image, and a temperature.

For example, features that can be selected from the dynamic pressure signal as combustion data are the magnitude of the dynamic pressure signal, the rate of change in the magnitude of the dynamic pressure signal, the distribution within the unit frame in the time domain, and the time domain. The distribution change rate in the unit frame, the frequency distribution in the unit frame in the frequency domain, and the frequency distribution change rate in the unit frame in the frequency domain may be included.

In addition, the features that can be selected from the image acquired inside the combustion unit are the difference in intensity or contrast between the image of the previous time and the image of the current time, the image of the previous time and the present. It may include a distance difference between feature points (Corner, Edge) between images in time, Scale Invariant-Scale Invariant Feature Transform (SIFT), and the like.

In addition, features that can be selected from the temperature inside the combustion unit may include the size of the temperature, the rate of change of the temperature, and the like.

In addition, features previously selected by the diagnostic unit 24 may include dynamic pressure energy and relative energy entropy.

That is, the diagnosis unit 24 may pre-select features to be selected as learning data, combine the features, or limit the learning data to a predetermined number of features in order to accurately and quickly determine whether combustion instability is present.

In step S20, the sensor unit (not shown) may acquire fully developed stable combustion data and fully developed unstable combustion data from various combustors.

That is, the gas turbine combustion instability diagnosis system must measure combustion data, which is the basic data to be used as learning data, and the gas turbine combustion instability diagnosis system is a fully developed stable state measured by various combustors through a sensor unit (not shown). ) And combustion data of a fully developed unstable state can be obtained.

Here, the stable state (Fully developed stable) means a state in which the stable state is maintained sufficiently and the stable state is maintained, and the unstable state (Fully developed unstable) refers to the state in which the unstable state remains sufficiently maintained and it means.

For example, the gas turbine combustion instability diagnosis system provides a dynamic pressure signal (Dynamic pressure, DP), image inside the combustor through the dynamic pressure sensor 100 in the fully developed and unstable states of various combustors. An image inside the combustor may be measured through a sensor (not shown), and a temperature inside the combustor may be measured through a temperature sensor (not shown).

In step S30, the diagnosis unit 24 may calculate selected features from the combustion data obtained through the sensor unit.

That is, the diagnosis unit 24 may calculate or obtain values of the selected features from the obtained combustion data by matching the acquired combustion data to the features selected in step S10.

For example, if the diagnostic unit 24 selects dynamic pressure energy and relative energy entropy as the features to be used as learning data in step S10, the obtained through the dynamic pressure sensor 100 By applying the dynamic pressure signal to <Equation 4> and <Equation 5>, values of features to be used as learning data, such as dynamic pressure energy and relative energy entropy, can be obtained. I can.

In step S40, the diagnosis unit 24 may train an artificial neural network model for diagnosis of combustion instability by using the calculated features as training data.

For example, the diagnostic unit 24 may determine the dynamic pressure energy and relative energy entropy of the dynamic pressure signals obtained in the fully developed and unstable states of various combustors. ) Can be calculated as features, and these features can be input to an artificial neural network as training data to train an artificial neural network model for diagnosis of combustion instability.

Referring to FIG. 11, the diagnosis unit 24 may train a tomographic perceptron model by using dynamic pressure energy (DPE) and relative energy entropy (ReEoE) as training data of the tomographic perceptron model. In addition, when a hyperplane for determining whether combustion instability is generated in a linear shape on a feature space, it may be determined that the artificial neural network model has been trained.

Here, as a result of learning the perceptron model using dynamic pressure energy (DPE) and relative energy entropy (ReEoE), a two-dimensional feature space is generated, so that a hyperplane for determining whether combustion instability is in a linear form ( 11).

However, the hyperplane generated in step S40 is a perceptron model that is trained using dynamic pressure energy when fully developed stable and fully developed unstable states are maintained, and has a certain degree of generality. ) Can be obtained, but if the dynamic pressure energy (DPE) value and relative energy entropy (ReEoE) value calculated from the combustion data measured in real time are measured as intermediate values between the steady state and the unstable state, it can accurately determine whether the corresponding signal is combustion unstable. It becomes difficult to do.

Therefore, in order to determine whether combustion instability is accurate, fine tuning may be performed on the learned artificial neural network model by using the combustion data of the transition process as additional training data.

Referring to FIG. 12, a new perceptron model may be trained by adding features acquired from transition combustion data to a previously learned perceptron model as new training data. As shown in (b), a new hyperplane for determining whether combustion is unstable may be created.

That is, the perceptron model undergoes new learning through fine tuning, and a new hyperplane corresponding thereto is generated so that combustion instability can be more accurately determined.

In step S50, the diagnosis unit 24 may determine a combustion instability state of the combustor in real time using the learned artificial neural network model.

For example, the diagnostic unit 24 may learn a perceptron model to generate a hyperplane as a criterion for determination, calculate features from combustion data measured in real time, and use the calculated features into a feature space. ), it is possible to determine the combustion instability state in real time corresponding to the position with the hyperplane.

Referring to FIGS. 12 and 13, when the calculated features exist on a hyperplane, it is determined as a combustion instability state, and the value of the index perceptron is 0 as shown in FIG. It is expressed as (Index perceptron = 0), and if it exists under the hyperplane, it is identified as a stable combustion state, and the value of the index perceptron is expressed as 1 (Index perceptron = 1).

In addition, when the diagnosis unit 24 determines whether combustion instability through the artificial neural network model is determined, the combustion control unit 26 controls the operation of the combustion unit 10 based on whether the diagnosis unit 24 determines combustion instability. I can.

For example, if the diagnosis unit 24 determines that the combustion instability state is 0.06 seconds or more, the combustion control unit 26 may display an alarm for the combustion instability state, and if it is determined that the combustion instability state is maintained for 0.1 second or more, It can be controlled to reduce the load on the gas turbine.

In relation to the gas turbine combustion instability diagnosis method according to an embodiment of the present invention, the contents of the gas turbine combustion instability diagnosis system described above may be applied. Accordingly, with respect to the gas turbine combustion instability diagnosis method, descriptions of the same contents as those of the gas turbine combustion instability diagnosis system described above have been omitted.

An embodiment of the present invention may also be implemented in the form of a recording medium including instructions executable by a computer, such as a program module executed by a computer. Computer-readable media can be any available media that can be accessed by a computer, and includes both volatile and nonvolatile media, removable and non-removable media. Further, the computer-readable medium may include all computer storage media. Computer storage media includes both volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data.

The above description of the present invention is for illustrative purposes only, and those of ordinary skill in the art to which the present invention pertains will be able to understand that other specific forms can be easily modified without changing the technical spirit or essential features of the present invention. will be. Therefore, it should be understood that the embodiments described above are illustrative and non-limiting in all respects. For example, each component described as a single type may be implemented in a distributed manner, and similarly, components described as being distributed may also be implemented in a combined form.

The scope of the present invention is indicated by the claims to be described later rather than the detailed description, and all changes or modified forms derived from the meaning and scope of the claims and their equivalent concepts should be construed as being included in the scope of the present invention. do.

10: combustion section
22: sensitivity analysis unit
24: diagnostic unit
26: combustion control unit
100: dynamic pressure sensor
120: pressure sensor
140: signal measuring unit
210: combustion dynamic pressure data
220: Relative Entropy of Energy (ReEoE)
230: Index perceptron

Claims (15)

In the gas turbine combustion instability diagnosis system,
A combustion unit including a sensor unit mounted inside the combustion unit and measuring combustion data inside the combustion unit;
Selecting a plurality of independent features that can reflect the combustion state, obtaining fully developed stable combustion data and fully developed unstable combustion data from various combustors, and the Calculate the selected features from the obtained combustion data, learn an artificial neural network model for combustion instability diagnosis using the calculated features as learning data, and determine the combustion instability state of the combustor using the artificial neural network model. Diagnostic unit; And
And a combustion control unit for controlling the operation of the combustion unit according to the determination of the diagnosis unit,
The features are,
Gas turbine combustion instability diagnosis system including Dynamic Pressure Energy and Relative Entropy of Energy.
The method of claim 1,
The features are,
The characteristics that can be obtained from the dynamic pressure signal are the size of the dynamic pressure signal, the rate of change in the size of the dynamic pressure signal, the time-series pattern of the dynamic pressure signal, the distribution within a unit frame in the time domain, and time. The rate of change of distribution within the unit frame in the time domain, the frequency distribution within the unit frame in the frequency domain and the rate of change in the frequency distribution within the unit frame in the frequency domain, the acoustic driving/damping capacity of the combustion system ( Damping rate / growth rate) gas turbine combustion instability diagnostic system comprising at least one or more.
The method of claim 1,
The difference in intensity or contrast between the image of the previous time and the image of the current time within a specific area, which are features that can be obtained from the flame image, and the time series pattern of the brightness of the image during a specific time. -series pattern), the distance difference between the feature points (Corner, Edge) between the image of the previous time and the image of the current time, and Scale Invariant-A gas turbine combustion instability diagnosis system including at least one of Scale Invariant Feature Transform (SIFT).
The method of claim 1,
The features are,
A gas turbine combustion instability diagnosis system comprising at least one of a temperature magnitude and a rate of change of temperature, which are features that can be obtained from a temperature sensor.
delete The method of claim 1,
The diagnostic unit,
Gas turbine combustion instability diagnosis system that performs fine tuning on the learned artificial neural network model by using the combustion data of the transient process as additional training data.
The method of claim 1,
The artificial neural network model,
Gas turbine combustion instability diagnostic system, a single-layer perceptron model.
In the gas turbine combustion instability diagnosis method,
(i) selecting features that can reflect the combustion state;
(ii) obtaining fully developed stable combustion data and fully developed unstable combustion data from several combustors;
(iii) calculating the selected features from the obtained combustion data;
(iv) learning an artificial neural network model for diagnosis of combustion instability by using the calculated features as training data; And
(v) determining a combustion instability state of the combustor using the artificial neural network model,
The features are,
Gas turbine combustion instability diagnosis method including dynamic pressure energy and relative energy entropy.
The method of claim 8,
The features are,
The characteristics that can be obtained from the dynamic pressure signal are the size of the dynamic pressure signal, the rate of change in the size of the dynamic pressure signal, the time-series pattern of the dynamic pressure signal, the distribution within a unit frame in the time domain, and time. The rate of change of distribution within the unit frame in the time domain, the frequency distribution within the unit frame in the frequency domain and the rate of change in the frequency distribution within the unit frame in the frequency domain, the acoustic driving/damping capacity of the combustion system ( Damping rate / growth rate) gas turbine combustion instability diagnosis method comprising at least one or more.
The method of claim 8,
The features are,
The difference in intensity or contrast between the image of the previous time and the image of the current time within a specific area, which are features that can be obtained from the flame image, and the time series pattern of the brightness of the image during a specific time. -series pattern), a difference in distance of feature points (Corner, Edge) between an image of a previous time and an image of the current time, and a gas turbine combustion instability diagnosis method comprising at least one or more of Scale Invariant-Scale Invariant Feature Transform (SIFT).
The method of claim 8,
The features are,
Gas turbine combustion instability diagnosis method comprising at least one of a temperature magnitude and a rate of change of temperature, which are features that can be obtained from a temperature sensor.
delete The method of claim 8,
The step of learning an artificial neural network model for diagnosis of combustion instability by using the calculated features of step (iv) as training data,
A gas turbine combustion instability diagnosis method further comprising the step of performing fine tuning on the learned artificial neural network model by using combustion data of a transient process as additional training data.
The method of claim 8,
The artificial neural network model,
A method for diagnosing combustion instability in gas turbines, which is a single layer perceptron model.
A computer-readable recording medium in which a program for implementing the method of any one of claims 8, 9, 10, 11, 13, and 14 is recorded.

Images