A gas turbine fault prediction diagnosis method based on deep transfer learning

Abstract

The application relates to the technical field of gas turbine fault diagnosis, in particular to a gas turbine fault prediction and diagnosis method based on deep migration learning, S1: dividing a source domain data set and a target domain data set of deep migration learning of a gas turbine; S2: analyzing the migratability of the source domain data set and the target domain data set; S3: training a deep neural network by using the source domain data set; S4: adjusting the deep neural network by using the target domain data set; S5: realizing gas turbine fault prediction and diagnosis by using the adjusted network; through deep migration learning, labeled fault data of a historical gas turbine unit is used to migrate fault knowledge to a newly-operated gas turbine unit; meanwhile, new fault categories possibly occurring in the new unit are considered; known non-covered labeled fault data in the newly-operated gas turbine unit is used to expand the fault knowledge of the newly-operated gas turbine unit; and the fault degree of the non-labeled data in the newly-operated unit is quantitatively described.

Description

Technical Field

[0001] This invention relates to the field of gas turbine fault diagnosis technology, and more specifically to a gas turbine fault prediction and diagnosis method based on deep transfer learning. Background Technology

[0002] Gas turbines are important pieces of equipment widely used in power generation, aviation, and marine industries. They mainly consist of compressors, combustion chambers, and turbines, and their performance is crucial to operational efficiency and safety. During operation, gas turbines are subjected to harsh conditions such as high temperature, high pressure, and high speed, as well as external influences. As operating time increases, their main components (such as compressors and turbines) will experience various modes of failure.

[0003] Current gas turbine maintenance strategies typically rely on prior knowledge to formulate maintenance methods, such as the method for optimizing gas turbine maintenance activities (patent application number CN201911163311.9) and the analytical method for optimizing gas turbine maintenance and repair schemes (patent application number CN201510240953.X). However, this method may reduce equipment reliability and increase operation and maintenance costs. Fault prediction and diagnosis can diagnose the gas turbine status in real time and provide early warnings before major failures occur. Therefore, research on gas turbine fault prediction and diagnosis is of great significance.

[0004] Currently, most gas turbine fault prediction and diagnosis methods typically employ two approaches: (1) Threshold diagnosis: setting a health threshold and predicting whether the gas turbine condition exceeds the threshold to determine if a fault has occurred. However, this method cannot quantitatively describe the degree of gas turbine faults, relying solely on the threshold to determine whether a fault has occurred. Examples include patent application number CN202010786693.7, "Gas Turbine Fault Prediction Method Based on Rank-Related Causal Structure Graph," and patent application number CN202110576134.8, "A Gas Turbine Rotor Fault Early Warning Method Based on Hybrid Prediction." (2) Data-driven approach: utilizing historical gas turbine unit fault data and employing a data-driven method to quantitatively describe the degree of faults in newly commissioned units. However, most of these methods do not consider the possibility of new fault categories in newly commissioned gas turbine units. Examples include patent application number CN202310640945.9, "Federated Knowledge Fusion and Fault Diagnosis Method for Gas Turbine Power Generation Equipment," and patent application number CN202010925183.3, "A Gas Turbine Gas Path Fault Prediction and Diagnosis Method Based on Knowledge + Data." Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by providing a scientific, reasonable, adaptable, and highly valuable method for predicting and diagnosing gas turbine faults.

[0006] This method can transfer fault knowledge to newly commissioned gas turbine units by using tagged fault data from historical gas turbine units through deep transfer learning. At the same time, it considers that new fault categories may occur in the new units and expands the fault knowledge of the newly commissioned gas turbine units by using known uncovered tagged fault data in the newly commissioned gas turbine units, and quantitatively describes the degree of faults without tagged data in the newly commissioned units.

[0007] The objective of this invention is achieved through the following technical solution:

[0008] A gas turbine fault prediction and diagnosis method based on deep transfer learning, comprising the following steps:

[0009] S1: Divide the source domain dataset and target domain dataset for deep transfer learning of gas turbines;

[0010] S2: Transferability analysis of source and target domain datasets;

[0011] S3: Train a deep neural network using the source domain dataset;

[0012] S4: Adjust the deep neural network using the target domain dataset;

[0013] S5: Utilize the adjusted network to achieve gas turbine fault prediction and diagnosis;

[0014] S1 specifically includes the following steps:

[0015] S101: Select labeled fault data from historical gas turbine unit operation data as the source domain dataset;

[0016] S102: Select the operating data of newly commissioned gas turbine units as the target domain dataset. The target domain dataset contains known uncovered label fault data and unlabeled data.

[0017] In S2, since the gas turbine gas circuit fault mechanisms are similar and there is a certain degree of similarity between different fault modes, deep transfer learning can utilize the similarity between the source domain dataset and the target domain dataset to make the source domain dataset and the target domain dataset transferable.

[0018] In S2, the maximum mean discrepancy (MMD) is used to describe the similarity between the source domain dataset and the target domain dataset. The maximum mean discrepancy is a non-parametric method for measuring the distribution difference between two datasets.

[0019] If the source domain dataset Ds and the target domain dataset Dt follow probability distributions p and q respectively, then MMD is defined as follows:

[0020]

[0021] In Equation (1), xi comes from the source domain dataset Ds, xj comes from the target domain dataset Dt, sup(·) is the maximum value of the input data, H represents the reproducing kernel Hilbert space (RKHS), and φ(·) represents the nonlinear mapping from the original feature space to RKHS; if RKHS is rich enough, a suitable distance metric can be found.

[0022] S3 specifically includes the following steps:

[0023] S301: Combine the gas turbine boundary condition vector and the gas path measurable parameter vector. As network input, the health parameter vector of the gas path components As the network output, a deep neural network is built;

[0024] S302: Hyperparameter optimization, using orthogonal experimentation to set the hyperparameters of the deep neural network. Hyperparameters include: learning rate, learning rate reduction factor, batch size, maximum number of training iterations, number of channels in the convolutional layer, and random inactivation rate.

[0025] S303: Train a deep neural network using a source domain dataset;

[0026] The boundary condition vector and the gas path measurable parameter vector Includes: a boundary condition vector consisting of the gas turbine ambient temperature T0, ambient pressure P0, and power generation P. The gas path measurable parameter vector consists of the gas turbine generator speed Ng, compressor outlet pressure P3, power turbine outlet temperature T6, and fuel quantity Gf.

[0027] S4 specifically includes the following steps:

[0028] S401: Train a deep neural network for the target domain using data with known uncovered labels in the target domain dataset;

[0029] S402: Saves the output layer structure parameters of the target domain deep neural network;

[0030] S403: Freeze the source domain dataset to train deep neural networks, excluding the output layer;

[0031] S404: Adjust the output layer structure parameters of the deep neural network trained on the source domain dataset to the output layer structure parameters of the deep neural network in the target domain.

[0032] S405: Obtain the target domain fault prediction and diagnosis network model;

[0033] S5 specifically includes the following steps:

[0034] S501: Combine the boundary condition vector and gas path measurable parameter vector of the unlabeled data in the target domain. Input into the target domain fault prediction and diagnosis network model;

[0035] S502: Obtain the health parameter vector of the network output result from the unlabeled data. To achieve predictive diagnosis of gas turbine faults;

[0036] The health parameter vector Including: Gas turbine compressor flow performance index (SF6) C,FC Turbine Flow Performance Index (SF) T,FC Flow performance index SF PT,FC Compressor efficiency performance index SF C,EFF Turbine efficiency performance index SF T,EFF And the power turbine efficiency performance index SF PT,EFF ;

[0037] The compressor flow performance index SF C,FC Turbine Flow Performance Index SF T,FC Flow performance index SF PT,FC Compressor efficiency performance index SF C,EFF Turbine efficiency performance index SF T,EFF And the power turbine efficiency performance index SF PT,EFF The definition is as follows:

[0038] SF C,FC =G C,cor / G0 C,cor (2)

[0039] SF T,FC =G T,cor / G0 T,cor (3)

[0040] SF PT,FC =G PT,cor / G0 PT,cor (4)

[0041] SF C,EFF =η C / η0 C (5)

[0042] SF T,EFF =η T / η0 T (6)

[0043] SF PT,EFF =η PT / η0 PT (7)

[0044] The beneficial effects of this invention are as follows:

[0045] First, the fault knowledge of historical gas turbine units with tagged fault data is transferred to newly commissioned gas turbine units. At the same time, the fault knowledge of newly commissioned gas turbine units is expanded by utilizing known uncovered tagged fault data in the newly commissioned gas turbine units. This solves the problem that the original diagnostic methods are not very applicable because new fault modes may occur in newly commissioned gas turbine units.

[0046] 2. Using the Maximum Mean Discrepancy (MMD), we achieved transferability analysis between the source and target domains;

[0047] III. Boundary condition vectors and gas path measurable parameter vectors were established. and health parameter vector The mapping relationship between them is represented by a health parameter vector. Quantitatively describe the degree of failure in a gas turbine unit;

[0048] Fourth, by adjusting the network structure using target domain data, the network can be made more suitable for fault prediction and diagnosis of newly commissioned gas turbine units, which is beneficial to improving model performance.

[0049] Fifth, in scenarios where newly commissioned gas turbine units have known uncovered and untagged fault data, this method can achieve fault prediction and diagnosis, improving the accuracy of fault diagnosis. It is scientifically sound, highly applicable, and has high engineering application value, providing a reference for subsequent gas turbine fault diagnosis, operation, and maintenance. Attached Figure Description

[0050] The present invention will now be described in further detail with reference to the accompanying drawings and specific implementation methods.

[0051] Figure 1 This is a schematic diagram of the gas turbine fault prediction and diagnosis method based on deep transfer learning of the present invention;

[0052] Figure 2 This is a schematic diagram of the relative bias results of the training set of the source domain training network of the present invention;

[0053] Figure 3 This is a schematic diagram of the relative deviation results of the source domain training network test set according to the present invention;

[0054] Figure 4 This is a schematic diagram of the adjusted network relative deviation results of the present invention. Detailed Implementation

[0055] The present invention will now be described in further detail with reference to the accompanying drawings.

[0056] like Figures 1 to 4 As shown, in order to provide a scientific, reasonable, adaptable and valuable method for predicting and diagnosing gas turbine faults, the steps and functions of a gas turbine fault prediction and diagnosis method based on deep transfer learning are explained in detail below.

[0057] This invention provides a gas turbine fault prediction and diagnosis method based on deep transfer learning, such as... Figure 1 As shown, the method includes the following steps:

[0058] S1: Divide the source domain dataset and target domain dataset for deep transfer learning of gas turbines;

[0059] S101: Select labeled fault data from historical gas turbine unit operation data as the source domain dataset;

[0060] S102: Select the operating data of newly commissioned gas turbine units as the target domain dataset. The target domain dataset contains known uncovered label fault data and unlabeled data.

[0061] Historical gas turbine unit failure data was selected as the source domain dataset, and newly commissioned gas turbine unit failure data was selected as the target domain dataset.

[0062] Table 1. Sample size and characteristics of source and target domains:

[0063]

[0064] S2: Source domain to target domain transferability analysis;

[0065] Since the fault mechanisms of gas turbine gas circuits are similar, there is a certain degree of similarity between different fault modes. Transfer learning can utilize the similarity between the source domain and the target domain to make the source domain and the target domain transferable.

[0066] Maximum Mean Discrepancy (MMD) is a non-parametric method for measuring the difference in distributions between two datasets, used to describe the similarity between a source domain and a target domain. If the source domain Ds and the target domain Dt follow probability distributions p and q, respectively, then MMD is defined as follows:

[0067]

[0068] In equation (1), xi comes from the source domain dataset Ds, xj comes from the target domain dataset Dt, sup(·) is the maximum value of the input data, H represents the reproducing kernel Hilbert space (RKHS), and φ(·) represents the nonlinear mapping from the original feature space to the RKHS. With a sufficiently rich RKHS, a suitable distance metric can be found.

[0069] Table 2 shows the results of the maximum mean difference between the source and target domains:

[0070]

[0071] The small difference between the maximum mean of the source domain and the target domain proves that the source domain and the target domain have a certain degree of similarity and strong transferability.

[0072] S3: Train a deep neural network using the source domain dataset;

[0073] S301: Combine the gas turbine boundary condition vector and the gas path measurable parameter vector. As network input, the health parameter vector of the gas path components As the network output, a deep neural network is built;

[0074] The boundary condition vector consisting of the gas turbine ambient temperature T0, ambient pressure P0, and power generation P

[0075] The gas path measurable parameter vector consists of the gas turbine generator speed Ng, compressor outlet pressure P3, power turbine outlet temperature T6, and fuel quantity Gf.

[0076] S302: Hyperparameter optimization, using orthogonal experimentation to set the hyperparameters of the deep neural network. Hyperparameters include: learning rate, learning rate reduction factor, batch size, maximum number of training iterations, number of channels in the convolutional layer, and random inactivation rate.

[0077] Table 3 Horizontal settings for each hyperparameter

[0078]

[0079] The orthogonal experiment method mainly relies on conducting experiments sequentially based on orthogonal arrays to obtain the optimal results.

[0080] In this invention, a 7-factor, 4-level orthogonal array was used, and the final experiment consisted of a total of 16 groups.

[0081] Table 4 Orthogonal Experiment Table

[0082]

[0083]

[0084] Orthogonal experiments were conducted according to Table 4.

[0085] RMSE was used as a network evaluation metric.

[0086] RMSE is defined as follows:

[0087]

[0088] In formula (2) and These are the actual and predicted values ​​obtained from the k-th iteration during the hyperparameter optimization process, respectively.

[0089] 80% of the source domain data is randomly selected as the training set, and the remaining 20% ​​is used as the test set.

[0090] To avoid the influence of network randomness, each experiment was conducted 5 times, and the average value was used as the final RMSE result.

[0091] Table 5. RMSE Results of Orthogonal Experiments

[0092]

[0093]

[0094] The optimal experiment is selected from the RMSE results of the training set and the RMSE results of the test set as the hyperparameter optimization result.

[0095] Experiments 10 and 13-16 are test results with high accuracy.

[0096] Experiment 10 showed the highest accuracy, but the RMSE on the training and test sets differed significantly, indicating poor network stability. Experiments 13-16 showed higher accuracy and some stability.

[0097] Therefore, this invention selects Experiment 14 as the result of hyperparameter optimization.

[0098] S303: Train a deep neural network using a source domain dataset;

[0099] The relative bias results of the training set of the source domain training network of the present invention are as follows: Figure 2 As shown;

[0100] The relative bias results of the source domain training network test set of this invention are as follows: Figure 3 As shown.

[0101] The source domain training network has a high accuracy rate in fault diagnosis. It learns various fault mode characteristics well, not only on the training set but also on the test set, and has a certain degree of stability, which makes subsequent transfer tasks easier to achieve.

[0102] S4: Adjust the network structure using the target domain dataset;

[0103] S401: Train a deep neural network for the target domain using data with known uncovered labels in the target domain dataset;

[0104] S402: Saves the output layer structure parameters of the target domain deep neural network;

[0105] S403: Freeze the source domain dataset to train deep neural networks, excluding the output layer;

[0106] S404: Adjust the output layer structure parameters of the deep neural network trained on the source domain dataset to the output layer structure parameters of the deep neural network in the target domain.

[0107] S405: Obtain the target domain fault prediction and diagnosis network model;

[0108] S5: Utilize the adjusted network to achieve gas turbine fault prediction and diagnosis;

[0109] S501: Combine the boundary condition vector and gas path measurable parameter vector of the unlabeled data in the target domain. Input into the target domain fault prediction and diagnosis network model;

[0110] S502: Obtain the health parameter vector of the network output result from the unlabeled data. To achieve predictive diagnosis of gas turbine faults;

[0111] The gas turbine compressor flow performance index SF C,FC Turbine Flow Performance Index SF T,FC Flow performance index SF PT,FC Compressor efficiency performance index SF C,EFF Turbine efficiency performance index SF T,EFF And the power turbine efficiency performance index SF PT,EFF Composition of health parameter vector

[0112] The compressor flow performance index SF C,FC Turbine Flow Performance Index (SF) T,FC Flow performance index SF PT,FC Compressor efficiency performance index SF C,EFF Turbine efficiency performance index SF T,EFF And the power turbine efficiency performance index SF PT,EFF The definition is as follows:

[0113] SF C,FC =G C,cor / G0 C,cor (3)

[0114] SF T,FC =G T,cor / G0 T,cor (4)

[0115] SF PT,FC =G PT,cor / G0 PT,cor (5)

[0116] SF C,EFF =η C / η0 C (6)

[0117] SF T,EFF =η T / η0 T (7)

[0118] SF PT,EFF =η PT / η0 PT (8)

[0119] The present invention provides the following results regarding the adjusted relative bias of the network: Figure 4 As shown.

[0120] Table 6. Comparison of RMSE between the adjusted network and the source domain-trained network.

[0121]

[0122]

[0123] Compared to networks trained in the source domain, the adjusted network using deep transfer learning achieves higher accuracy in fault diagnosis prediction.

[0124] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A gas turbine fault prediction and diagnosis method based on deep transfer learning, characterized in that: The method includes the following steps: S1: Divide the source domain dataset and target domain dataset for deep transfer learning of gas turbines; S2: Transferability analysis of source and target domain datasets; S3: Train a deep neural network using the source domain dataset; S4: Adjust the deep neural network using the target domain dataset; S5: Utilize the adjusted network to achieve gas turbine fault prediction and diagnosis; S3 specifically includes the following steps: S301: Combine the gas turbine boundary condition vector and the gas path measurable parameter vector. , As network input, the health parameter vector of the gas path components will be used. As the network output, a deep neural network is built; S302: Hyperparameter optimization, using orthogonal experimentation to set the hyperparameters of the deep neural network. Hyperparameters include: learning rate, learning rate reduction factor, batch size, maximum number of training iterations, number of channels in the convolutional layer, and random inactivation rate. S303: Train a deep neural network using a source domain dataset; S4 specifically includes the following steps: S401: Train a deep neural network for the target domain using data with known uncovered labels in the target domain dataset; S402: Saves the output layer structure parameters of the target domain deep neural network; S403: Freeze the source domain dataset to train deep neural networks, excluding the output layer; S404: Adjust the output layer structure parameters of the deep neural network trained on the source domain dataset to the output layer structure parameters of the deep neural network in the target domain. S405: Obtain the target domain fault prediction and diagnosis network model; S5 specifically includes the following steps: S501: Combine the boundary condition vector and gas path measurable parameter vector of the unlabeled data in the target domain. , Input into the target domain fault prediction and diagnosis network model; S502: Obtain the health parameter vector of the network output result from the unlabeled data. This enables the prediction and diagnosis of gas turbine faults.

2. The gas turbine fault prediction and diagnosis method based on deep transfer learning according to claim 1, characterized in that: S1 specifically includes the following steps: S101: Select labeled fault data from historical gas turbine unit operation data as the source domain dataset; S102: Select the operating data of newly commissioned gas turbine units as the target domain dataset. The target domain dataset contains known uncovered label fault data and unlabeled data.

3. The gas turbine fault prediction and diagnosis method based on deep transfer learning according to claim 1, characterized in that: In S2, deep transfer learning can utilize the similarity between the source domain dataset and the target domain dataset to make the source domain dataset transferable to the target domain dataset.

4. The gas turbine fault prediction and diagnosis method based on deep transfer learning according to claim 3, characterized in that: In S2, the maximum mean difference is used to describe the similarity between the source domain dataset and the target domain dataset. The maximum mean difference is a non-parametric method for measuring the distribution difference between two datasets.

5. The gas turbine fault prediction and diagnosis method based on deep transfer learning according to claim 1, characterized in that: The boundary condition vector and the gas path measurable parameter vector [ , This includes: a boundary condition vector consisting of the gas turbine ambient temperature T0, ambient pressure P0, and power generation P. The gas path measurable parameter vector consists of the gas turbine generator speed Ng, compressor outlet pressure P3, power turbine outlet temperature T6, and fuel quantity Gf. .

6. The gas turbine fault prediction and diagnosis method based on deep transfer learning according to claim 1, characterized in that: The health parameter vector Including: Gas turbine compressor flow performance index Turbine flow performance index Dynamic flow performance index Compressor efficiency performance index Turbine efficiency performance index and power turbine efficiency performance index .

Images