Gas turbine baseline modeling method based on regression Kriging

Abstract

The invention relates to the technical field of running state monitoring of gas turbines for power generation, in particular to a gas turbine baseline modeling method based on regression Kriging, which comprises the following steps: S1, collecting running data of a gas turbine, and extracting data features; s2, carrying out map division, normalization and data set division processing on the collected data; s3, combining a nonlinear autoregressive neural network (NARX) and a regression Kriging method to construct a noise space field prediction model; s4, training the noise space field prediction model through the training data divided from the data set in the step S2, judging whether the trained noise space field prediction model meets requirements or not through a test result of a test set divided from the data set, and storing the noise space field prediction model meeting the requirements; a noise space field prediction model can be constructed according to the noise data of the gas turbine, and the health condition of the gas turbine is monitored.

Description

Technical Field

[0001] This invention relates to the field of gas turbine operating status monitoring technology for power generation, and more specifically to a baseline modeling method for gas turbines based on regression kriging. Background Technology

[0002] Gas turbines, as an important type of power machinery, are characterized by their compact structure, stable operation, and high thermal efficiency, leading to their increasingly wide range of applications. The safe and reliable operation of gas turbines requires high standards. Monitoring the health of the unit during daily operation and analyzing and detecting potential anomalies can help prevent or facilitate the timely handling of major gas turbine failures.

[0003] When a gas turbine is running, the rotating machinery will generate vibrations and push the gas, which will further generate noise. In order to effectively judge the operating status of the gas turbine from these vibration and sound pressure signals, a gas turbine baseline modeling method is needed. Summary of the Invention

[0004] The purpose of this invention is to provide a baseline modeling method for gas turbines based on regression Kriging. This method can construct a noise spatial field prediction model based on gas turbine noise data, monitor the health of the gas turbine, analyze and detect various possible abnormalities, and avoid or facilitate timely handling of potential gas turbine failures.

[0005] The objective of this invention is achieved through the following technical solution:

[0006] A baseline modeling method for gas turbines based on regression kriging, comprising the following steps:

[0007] Step S1: Collect gas turbine sound pressure data and speed signals, and extract data features;

[0008] Step S2: Perform graph partitioning, normalization, and dataset partitioning on the collected data;

[0009] Step S3: Combine nonlinear autoregressive neural network (NARX) and regression kriging method to construct a noise space field prediction model;

[0010] Step S4: Train the noise spatial field prediction model using the training data divided from the dataset in step S2. Determine whether the trained noise spatial field prediction model meets the requirements based on the test results of the test set divided from the dataset, and save the noise spatial field prediction model that meets the requirements.

[0011] The sound pressure data acquisition is achieved by placing a sound pressure sensor 1 meter away from the compressor side of the gas turbine to acquire the sound pressure data of the gas turbine; the speed signal acquisition is achieved by storing the gas turbine speed sequentially and using the first-in-first-out principle to obtain the gas turbine speed sequence.

[0012] In step S1, data feature extraction involves performing a Fourier transform on the sound pressure data based on the gas turbine speed sequence to extract the time-domain and frequency-domain features of the noise signal.

[0013] The time-domain features include root mean square value and maximum value, and the frequency-domain signals include frequency conversion, frequency conversion second harmonic, frequency conversion third harmonic and blade passage frequency;

[0014] The gas turbine noise spectrum M is calculated using Fast Fourier Transform. i ;

[0015] Calculate the rotational frequency f based on the rotational speed data n. r :

[0016] f r =n / 60

[0017] After obtaining the frequency, a range of f is performed in the vicinity of that frequency. r ±f r A peak search of 0.1 yields a frequency f that is doubled. r The frequency domain characteristics are calculated in the following intervals using a 1x turnaround:

[0018] Frequency amplitude: P1 = max(M fr-fr*0.1 M fr+fr*0.1 )

[0019] Amplitude of double frequency: P2 = max(M 2fr-fr*0.1 M 2fr+fr*0.1 )

[0020] Amplitude of three times the rotation frequency: P3 = max(M 3fr-fr*0.1 M 3fr+fr*0.1 )

[0021] Based on the number of blades from stage 1 to stage n of the compressor, calculate the passing frequency of each stage of blades at a specific speed, and obtain the amplitude corresponding to the passing frequency of each stage of blades;

[0022] The method for selecting the amplitude corresponding to the frequency of each blade stage is as follows:

[0023]

[0024] Map partitioning involves dividing the collected data according to the operational map, while normalization involves performing maximum and minimum normalization on the collected input and output data.

[0025] The dataset partitioning involves classifying all graphs, using a single graph as the basic unit, selecting 90% of the graphs as the training set and 10% as the test set.

[0026] The normalization is performed on each data feature, and the formula is as follows:

[0027]

[0028] Step S3 specifically includes the following steps:

[0029] Step S31: Define the NARX model, with specific parameters including: output delay order, input delay order, neural network hidden layer structure, activation function, optimizer, and number of training rounds;

[0030] Step S32: Combine the processed characteristic frequencies with the high-pressure compressor speed and low-pressure compressor speed collected from the gas turbine to train the NARX model and obtain the NARX fitting function. The NARX formula is: y(t)=f(y(t-1),y(t-2),…,y(t-ny),u(t-1),u(t-2),…,u(t-nu))

[0031] Where y(t) is the current output, y(ti) is the previous n historical outputs, and u(tj) is the previous n external inputs;

[0032] Step S33: Taking the predicted residual as the optimization objective, the window width, sliding step size, delay order, and hidden layer structure in Step 1-2 are globally optimized using the Bayesian optimization algorithm to obtain the hyperparameter combination corresponding to the minimum residual.

[0033] Step S34: Utilize the optimized NARX model to output the residuals, and construct the residual spatial distribution field using Kriging interpolation. The formula is:

[0034]

[0035] Where ∈ represents the residual, and λ i These are the weighting coefficients;

[0036] Step S35: Use K-fold cross-validation to evaluate the model stability, and select the optimal model with a mean squared error (MSE) lower than a preset threshold as the final modeling output;

[0037] In step S4, the identification result is measured by comparing the predicted results of the test set with the actual values ​​of the gas turbine and performing error calculation and analysis, using the average relative error.

[0038]

[0039] Where n is the number of samples, yi It is the true value of the i-th sample. The predicted value of the i-th sample is denoted as . When the average relative error of each prediction parameter meets the requirements, the parameters of the spatial field prediction model are saved.

[0040] The beneficial effects of this invention are as follows:

[0041] By combining nonlinear autoregressive neural networks (NARX) and regression kriging, a more accurate noise spatial field prediction model can be constructed. This hybrid architecture retains the modeling advantages of NARX for temporal nonlinearity and quantifies spatial correlation through kriging. Finally, the fully connected layer fuses the trend and residual to achieve unbiased estimation of the noise spatial field, providing a high-precision reference surface for noise anomaly detection.

[0042] By constructing a more accurate noise spatial field prediction model to monitor the health of the unit and analyzing and detecting various possible abnormalities, major gas turbine failures can be avoided or facilitated in a timely manner. Attached Figure Description

[0043] The present invention will now be described in further detail with reference to the accompanying drawings and specific implementation methods.

[0044] Figure 1 This is a schematic block diagram of the steps of the gas turbine baseline modeling method based on regression Kriging of the present invention;

[0045] Figure 2 This is a noise data diagram of a gas turbine in an embodiment of the present invention.

[0046] Figure 3 This is a peak noise characteristic data diagram of a gas turbine in an embodiment of the present invention.

[0047] Figure 4 This is a graph showing the root mean square noise data of a gas turbine in an embodiment of the present invention.

[0048] Figure 5 This is a graph showing the frequency characteristics of gas turbine noise in an embodiment of the present invention.

[0049] Figure 6 This is a frequency characteristic data diagram of the gas turbine noise blades in an embodiment of the present invention.

[0050] Figure 7 This is a schematic diagram of the NARX model architecture of the present invention;

[0051] Figure 8 This is a schematic diagram of the Bayesian optimization process for the NARX model of the present invention;

[0052] Figure 9This is a schematic diagram of the regression kriging model using the NARX model as the regression model of the present invention;

[0053] Figure 10 This is a schematic diagram of regression prediction using the NARX model as the regression model in this invention. Detailed Implementation

[0054] The present invention will now be described in further detail with reference to the accompanying drawings.

[0055] like Figures 1 to 10 As shown, in order to achieve the technical effect of "building a spatial field prediction model based on gas turbine operation data, monitoring the health of the unit, analyzing and detecting various possible abnormalities, and avoiding or facilitating timely handling of major gas turbine failures", the steps and functions of a gas turbine baseline modeling method based on regression Kriging are explained in detail below.

[0056] A baseline modeling method for gas turbines based on regression kriging, comprising the following steps:

[0057] Step S1: Collect gas turbine sound pressure data and speed signals, and extract data features;

[0058] Step S2: Perform graph partitioning, normalization, and dataset partitioning on the collected data;

[0059] Step S3: Combine nonlinear autoregressive neural network (NARX) and regression kriging method to construct a noise space field prediction model;

[0060] Step S4: Train the noise spatial field prediction model using the training data divided from the dataset in step S2. Determine whether the trained noise spatial field prediction model meets the requirements based on the test results of the test set divided from the dataset, and save the noise spatial field prediction model that meets the requirements.

[0061] In step S1, the sound pressure data acquisition is achieved by placing a sound pressure sensor 1 meter away from the gas turbine compressor side to acquire gas turbine sound pressure data; the speed signal acquisition involves storing the gas turbine speed sequentially and obtaining the gas turbine speed sequence using a first-in-first-out principle.

[0062] In step S1, data feature extraction involves performing a Fourier transform on the sound pressure data based on the gas turbine speed sequence to extract the time-domain and frequency-domain features of the noise signal.

[0063] The time-domain features include root mean square value and maximum value, and the frequency-domain signals include frequency conversion, frequency conversion second harmonic, frequency conversion third harmonic and blade passage frequency;

[0064] The gas turbine noise spectrum M is calculated using Fast Fourier Transform. i ;

[0065] Calculate the rotational frequency f based on the rotational speed data n. r :

[0066] f r =b / 60

[0067] After obtaining the frequency, a range of f is performed in the vicinity of that frequency. r ±f r A peak search of 0.1 yields a frequency f that is doubled. r The frequency domain characteristics are calculated in the following intervals using a 1x turnaround:

[0068] Frequency amplitude: P1 = max(M fr-fr*0.1 M fr+fr*0.1 )

[0069] Amplitude of double frequency: P2 = max(M 2fr-fr*0.1 M 2fr+fr*0.1 )

[0070] Amplitude of three times the rotation frequency: P3 = max(M 3fr-fr*0.1 M 3fr+fr*0.1 )

[0071] Based on the number of blades from stage 1 to stage n of the compressor, calculate the passing frequency of each stage of blades at a specific speed, and obtain the amplitude corresponding to the passing frequency of each stage of blades;

[0072] The method for selecting the amplitude corresponding to the frequency of each blade stage is as follows:

[0073]

[0074] In step S2, the map division is to divide the collected data according to the running map; the normalization is to perform maximum and minimum normalization processing on the collected input and output data.

[0075] The dataset splitting involves classifying all graphs, using a single graph as the basic unit, selecting 90% of the graphs as the training set and 10% as the test set.

[0076] The normalization is performed on each data feature, and the formula is as follows:

[0077]

[0078] Step S3 specifically includes the following steps:

[0079] Step S31: Define the NARX model, with specific parameters including: output delay order, input delay order, neural network hidden layer structure, activation function, optimizer, and number of training rounds;

[0080] Step S32: Combine the processed characteristic frequencies with the high-pressure compressor speed and low-pressure compressor speed collected from the gas turbine to train the NARX model and obtain the NARX fitting function. The NARX formula is: y(t)=f(y(t-1),y(t-2),…,y(t-ny),y(t-1),u(t-2),…,u(t-nu))

[0081] Where y(t) is the current output, y(ti) is the previous n historical outputs, and u(tj) is the previous n external inputs;

[0082] Step S33: Taking the predicted residual as the optimization objective, the window width, sliding step size, delay order, and hidden layer structure in Step 1-2 are globally optimized using the Bayesian optimization algorithm to obtain the hyperparameter combination corresponding to the minimum residual.

[0083] Step S34: Utilize the optimized NARX model to output the residuals, and construct the residual spatial distribution field using Kriging interpolation. The formula is:

[0084]

[0085] Where ∈ represents the residual, and λ i These are the weighting coefficients;

[0086] Step S35: Use K-fold cross-validation to evaluate the model stability, and select the optimal model with a mean squared error (MSE) lower than a preset threshold as the final modeling output;

[0087] In step S4, the identification result is measured by comparing the predicted results of the test set with the actual values ​​of the gas turbine and performing error calculation and analysis, using the average relative error.

[0088]

[0089] Where n is the number of samples, y i It is the true value of the i-th sample. The predicted value of the i-th sample is denoted as . When the average relative error of each prediction parameter meets the requirements, the parameters of the spatial field prediction model are saved.

Claims

1. A baseline modeling method for gas turbines based on regression kriging, characterized in that: The method includes the following steps: Step S1: Collect gas turbine operating data and extract data features; Step S2: Perform graph partitioning, normalization, and dataset partitioning on the collected data; Step S3: Combine nonlinear autoregressive neural network (NARX) and regression kriging method to construct a noise space field prediction model; Step S4: Train the noise spatial field prediction model using the training data divided from the dataset in Step S2. Determine whether the trained noise spatial field prediction model meets the requirements by using the test results of the test set divided from the dataset, and save the noise spatial field prediction model that meets the requirements.

2. The method for baseline modeling of gas turbines based on regression kriging according to claim 1, characterized in that: In step S1, the data collected from the operation of the gas turbine includes sound pressure data and speed signals.

3. The method for baseline modeling of gas turbines based on regression kriging according to claim 2, characterized in that: The sound pressure data acquisition is achieved by placing a sound pressure sensor 1 meter away from the compressor side of the gas turbine to acquire the sound pressure data of the gas turbine; the speed signal acquisition is achieved by storing the gas turbine speed sequentially and using the first-in-first-out principle to obtain the gas turbine speed sequence.

4. The method for baseline modeling of gas turbines based on regression kriging according to claim 3, characterized in that: In step S1, data feature extraction involves performing a Fourier transform on the sound pressure data based on the gas turbine speed sequence to extract the time-domain and frequency-domain features of the noise signal.

5. The method for baseline modeling of gas turbines based on regression Kriging according to claim 4, characterized in that: The time-domain features include root mean square value and maximum value, and the frequency-domain signals include frequency conversion, frequency conversion second harmonic, frequency conversion third harmonic and blade passage frequency; The gas turbine noise spectrum M is calculated using Fast Fourier Transform. i ; Calculate the rotational frequency f based on the rotational speed data n. r : f r =n / 60 After obtaining the frequency, a range of f is performed in the vicinity of that frequency. r ±f r A peak search of 0.1 yields a frequency f that is doubled. r The frequency domain characteristics are calculated in the following intervals using a 1x turnaround: Frequency amplitude: P1 = max(M fr-fr*0.1 M fr+fr*0.1 ) Amplitude of double frequency: P2 = max(M 2fr-fr*0.1 M 2fr+fr*0.1 ) Amplitude of three times the rotation frequency: P3 = max(M 3fr-fr*0.1 M 3fr+fr*0.1 ) Based on the number of blades from stage 1 to stage n of the compressor, calculate the passing frequency of each stage of blades at a specific speed, and obtain the amplitude corresponding to the passing frequency of each stage of blades; The method for selecting the amplitude corresponding to the frequency of each blade stage is as follows:

6. The method for baseline modeling of gas turbines based on regression kriging according to claim 4, characterized in that: In step S2, the map division is to divide the collected data according to the running map, and the normalization is to perform maximum and minimum normalization processing on the collected input and output data.

7. The method for baseline modeling of gas turbines based on regression Kriging according to claim 6, characterized in that: The dataset partitioning involves classifying all graphs, using a single graph as the basic unit, selecting 90% of the graphs as the training set and 10% as the test set.

8. The method for baseline modeling of gas turbines based on regression kriging according to claim 6, characterized in that: The normalization is performed on each data feature, and the formula is as follows:

9. The method for baseline modeling of gas turbines based on regression kriging according to claim 5, characterized in that: Step S3 specifically includes the following steps: Step S31: Define the NARX model, with specific parameters including: output delay order, input delay order, neural network hidden layer structure, activation function, optimizer, and number of training rounds; Step S32: Combine the processed characteristic frequencies with the high-pressure compressor speed and low-pressure compressor speed collected by the gas turbine to train the NARX model and obtain the NARX fitting function. The NARX formula is: y(t)=f(y(t-1),y(t-2),…,y(t-ny),u(t-1),u(t-2),…,u(t-nu)) where y(t) is the current output, y(ti) is the previous n historical outputs, and u(tj) is the previous n external inputs; Step S33: Taking the predicted residual as the optimization objective, the window width, sliding step size, delay order, and hidden layer structure in Step 1-2 are globally optimized using the Bayesian optimization algorithm to obtain the hyperparameter combination corresponding to the minimum residual. Step S34: Utilize the optimized NARX model to output the residuals, and construct the residual spatial distribution field using Kriging interpolation. The formula is: Where ∈ represents the residual, and λ i These are the weighting coefficients; Step S35: Use K-fold cross-validation to evaluate the model stability, and select the optimal model with a mean squared error (MSE) lower than a preset threshold as the final modeling output.

10. The method for baseline modeling of gas turbines based on regression kriging according to claim 1, characterized in that: In step S4, the identification result is measured by comparing the predicted results of the test set with the actual values ​​of the gas turbine and performing error calculation and analysis, using the average relative error. Where n is the number of samples, y i It is the true value of the i-th sample. The predicted value of the i-th sample is denoted as . When the average relative error of each prediction parameter meets the requirements, the parameters of the spatial field prediction model are saved.

Images