Heavy duty gas turbine rotor fault diagnosis method fusing field generalization mechanism and GCN

Abstract

The invention discloses a heavy duty gas turbine rotor fault diagnosis method fusing a field generalization mechanism and a GCN, and relates to the technical field of fault diagnosis, and the method comprises the following steps: collecting vibration data, and carrying out the preprocessing; generating sample data, and performing data enhancement on the sample data; combining the generated enhanced sample with adversarial training to expand target domain distribution, and performing feature extraction through a CNN model; gaussian similarity of target domain distribution samples is calculated, errors are calculated by using a cross entropy loss function, and parameters of the graph convolutional neural network model are optimized; and training the whole model, and observing whether the loss rate tends to converge or not. According to the heavy-duty gas turbine rotor fault diagnosis method fusing the field generalization mechanism and the GCN, sample diversity is enhanced through multiple convolution kernels of different sizes, cross-domain target domain distribution is generated in combination with the field generalization mechanism, an efficient graph convolution neural network is constructed by using Gaussian similarity and a label propagation mechanism, and the fault diagnosis efficiency is improved. And the precision and robustness of small sample fault diagnosis are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis of a heavy-duty gas turbine pull rod rotor system, and in particular to a heavy-duty gas turbine rotor fault diagnosis method that integrates a field generalization mechanism and GCN. Background Art

[0002] Heavy-duty gas turbines have been widely used in the aviation and power generation industries due to their superior efficiency and energy-saving benefits. The rod rotor is a key component of a heavy-duty gas turbine. Its structure is very complex and it works in a high-temperature, high-pressure and high-speed environment for a long time. Rod rotor failure is the most common failure of heavy-duty gas turbines. Once the gas turbine is shut down for maintenance due to a failure, it will affect the energy and power supply and cause huge economic losses. Therefore, how to achieve safe and stable operation of the rod rotor system of heavy-duty gas turbines is an urgent problem to be solved.

[0003] Due to different operating conditions, external noise interference, installation errors and other reasons, the rod rotor is prone to inconsistent distribution between the source domain and the target domain, which ultimately affects the diagnostic accuracy in the test phase. In addition, the labeled data of the target domain is often expensive, and many existing cross-domain diagnosis techniques mainly focus on aligning the feature distribution of the source domain and the target domain. For heavy-duty gas turbines, the data related to their domain is insufficiently available and the procurement cost is too high, resulting in data scarcity in this field. Therefore, it is of great engineering significance to use the cross-domain small sample method to study the fault diagnosis of the heavy-duty gas turbine rod rotor system. Summary of the invention

[0004] The purpose of the present invention is to provide a heavy-duty gas turbine rotor fault diagnosis method that integrates domain generalization mechanism and GCN, solves the problems raised in the above-mentioned background technology, effectively simulates extreme cases of data, and realizes cross-domain feature alignment through adversarial training to improve robustness and generalization ability.

[0005] To achieve the above object, the present invention provides a heavy-duty gas turbine rotor fault diagnosis method integrating domain generalization mechanism and GCN, comprising the following steps:

[0006] Step S1: collecting vibration data for preprocessing, and dividing the vibration data into a training set and a test set;

[0007] Step S2: generating sample data using convolution kernels of different sizes for the collected original vibration data, and performing data enhancement on the sample data;

[0008] Step S3: Combine the generated enhanced samples with adversarial training to expand the target domain distribution, and perform feature extraction through the CNN model;

[0009] Step S4: Calculate the Gaussian similarity of the target domain distribution samples, generate the adjacency matrix of the graph, propagate the label information through the normalized graph convolutional neural network, calculate the error using the cross-entropy loss function, and optimize the parameters of the graph convolutional neural network model;

[0010] Step S5: Train the entire model, observe whether the loss rate tends to converge. If it converges, output the fault detection result; if not, repeat steps S2 - S4;

[0011] In the said step S3, an adversarial training is carried out by adopting a domain generalization mechanism, and the core formula of the domain generalization mechanism is as follows:

[0012]

[0013] where, T 0 represents the source domain distribution, representing the distribution in the training data; T represents the potential virtual target domain distribution, constructed in the neighborhood of T 0 ; D(T, T 0 ) represents the distance between the source domain distribution and the target domain distribution, restricting the generation range of the virtual target domain distribution; ρ represents the distance limit parameter, controlling the generalization range; L(θ; (T, Y)) represents the adversarial training loss function; θ represents the model input parameter; Θ represents the filter; min represents minimization; sup represents the supremum; P represents the conditional threshold; E P represents the expected value under the condition of P; Y represents the distribution label vector.

[0014] Preferably, in the said step S1, the preprocessing of the collected vibration data specifically includes collecting the vibration data of different measuring points in the heavy-duty gas turbine tie-rod rotor system under normal and abnormal states, and preprocessing the missing values and abnormal values of the data.

[0015] Preferably, in the said step S2, data augmentation is carried out on the original vibration data by adopting various different convolution kernel sizes to generate more diverse and robust sample data, and the specific formula is:

[0016]

[0017] where, X * represents the augmented data, X represents the original data, k 1 represents the convolution kernel size, N represents the standard normal distribution, and v represents the sample conforming to the standard normal distribution.

[0018] Preferably, in the said step S3, adversarial samples are generated by combining adversarial training, and the generation formula of the adversarial samples is:

[0019]

[0020] where, X advdenotes the adversarial perturbation sample; β denotes the step size parameter, which controls the perturbation intensity; denotes the gradient of the input of the adversarial training loss function, which is used to calculate the perturbation direction in the worst case; Y 0 denotes the source domain distribution label vector.

[0021] In the optimization phase, it involves minimizing the gradients of the original samples and adversarial samples in the dataset, and the resulting virtual cross-domain target domain distribution is used to update the model parameters. The formula is as follows:

[0022]

[0023] where a denotes the learning rate; denotes the gradients of the original samples and adversarial samples;

[0024] Cross-domain feature alignment is achieved through adversarial training to improve the robustness and generalization ability of the model.

[0025] Preferably, the specific steps of step S4 are as follows:

[0026] Step S41: Calculate the edge weights between samples using Gaussian similarity;

[0027] Step S42: Normalize the graph convolutional neural network;

[0028] Step S43: Perform label propagation and feature aggregation based on the graph convolutional neural network, and embed the relationships between target domains into the model;

[0029] Step S44: Optimize the cross-entropy loss function through the adversarial samples and the label propagation results on the normalized graph convolutional neural network.

[0030] Preferably, the formula for calculating the edge weights in step S41 is as follows:

[0031]

[0032] where W ij denotes the Gaussian similarity matrix, d denotes the distance between samples, f φ (·) denotes the feature extraction function, σ denotes the degree of dispersion of the sample distribution, X i denotes the row input sample, X j denotes the column input sample, i denotes the row index value, and j denotes the column index value.

[0033] Preferably, the formula for normalization processing in step S42 is as follows:

[0034]

[0035] Among them, S represents the connectivity weight matrix of the normalized graph, D represents the degree matrix of nodes, D = diag(d i ), d i = ∑ j W ij , and W represents the Gaussian similarity matrix.

[0036] Preferably, in step S43, the relationship between target domains is embedded into the model, and the formula is as follows:

[0037] N t+1 = k 2 SN t + (1 - k 2 )Y;

[0038] Among them, N t+1 represents the label or feature matrix after the (t + 1)-th propagation; N t represents the label or feature matrix after the t-th propagation; k 2 represents the weight between the current target and the initial target; Y represents the distribution label vector.

[0039] Preferably, the cross-entropy loss function formula in step S44 is as follows:

[0040]

[0041] Among them, J(φ, θ) represents, φ represents the network input feature, θ represents the model input parameter, P<y = j|X i > represents the probability of converting the label prediction matrix into the membership degree of the sample in each category, Y ij represents the label vector, y represents the prediction vector, m represents the number of samples in the validation set, and N represents the total number of all categories.

[0042] Therefore, the heavy gas turbine rotor fault diagnosis method of the present invention adopting the above-mentioned fusion of domain generalization mechanism and GCN enhances the sample diversity through multiple convolution kernels of different sizes, generates the cross-domain target domain distribution in combination with the domain generalization mechanism, and constructs an efficient graph convolutional neural network using Gaussian similarity and label propagation mechanism, significantly improving the accuracy and robustness of small-sample fault diagnosis.

[0043] Next, through the drawings and embodiments, the technical solutions of the present invention will be further described in detail. Description of the Drawings

[0044] Figure 1 is a flowchart of an embodiment of the heavy gas turbine rotor fault diagnosis method of the present invention integrating the domain generalization mechanism and GCN;

[0045] Figure 2This is a comparison diagram of the effects of the embodiment of the heavy gas turbine rotor fault diagnosis method that integrates the domain generalization mechanism and GCN of the present invention. Among them, (a) is the sample enhancement effect diagram when the convolution kernel size is 3, (b) is the sample enhancement effect diagram when the convolution kernel size is 5, and (c) is the sample enhancement effect diagram when the convolution kernel size is 7;

[0046] Figure 3 This is a schematic diagram of the graph convolutional neural network of the embodiment of the heavy gas turbine rotor fault diagnosis method that integrates the domain generalization mechanism and GCN of the present invention;

[0047] Figure 4 This is a schematic diagram of the visualization of the S matrix graph structure of the embodiment of the heavy gas turbine rotor fault diagnosis method that integrates the domain generalization mechanism and GCN of the present invention;

[0048] Figure 5 This is a schematic diagram of the structure of the heavy gas turbine of the embodiment of the heavy gas turbine rotor fault diagnosis method that integrates the domain generalization mechanism and GCN of the present invention. Detailed implementation manners

[0049] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0050] Unless otherwise defined, the technical terms or scientific terms used in the present invention should have the ordinary meanings understood by those of ordinary skill in the field to which the present invention belongs. The "first", "second" and similar terms used in the present invention do not denote any order, quantity or importance, but are only used to distinguish different components. The terms such as "including" or "comprising" mean that the elements or objects appearing before this term cover the elements or objects listed after this term and their equivalents, without excluding other elements or objects. The terms such as "connected" or "coupled" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The terms such as "upper", "lower", "left" and "right" are only used to represent relative positional relationships, and when the absolute position of the object being described changes, the relative positional relationship may also change accordingly.

[0051] Embodiment

[0052] Please refer to Figures 1-5 , the present invention provides a heavy gas turbine rotor fault diagnosis method that integrates the domain generalization mechanism and GCN, including the following steps:

[0053] Step S1: Collect vibration data for preprocessing, specifically including collecting vibration data of different measuring points in the normal and abnormal states of the heavy gas turbine tie rod rotor system, and preprocessing the missing values and outliers of the data. After preprocessing, the vibration data is divided into a training set and a test set.

[0054] Step S2: Generate sample data from the collected original vibration data using convolution kernels of different sizes, and perform data augmentation on the sample data.

[0055] Please refer to Figure 2 , in step S2, multiple different convolution kernel sizes are used to perform data augmentation on the original vibration data to generate more diverse and robust sample data. The specific formula is as follows:

[0056]

[0057] Among them, X * represents the augmented data, X represents the original data, k 1 represents the convolution kernel size, N represents the standard normal distribution, and v represents the sample conforming to the standard normal distribution.

[0058] Figure 2 The results show that the convolution layers with different convolution kernel sizes retain the main features of the samples, while introducing random local feature perturbations, generating samples with similar global shapes but random local variations, effectively improving the diversity of the samples.

[0059] Step S3: Combine the generated augmented samples with adversarial training to expand the target domain distribution, and perform feature extraction through the CNN model.

[0060] Due to different operating conditions, the tie-rod rotor system of a heavy-duty gas turbine is prone to the situation where the source domain and target domain distributions are inconsistent, which in turn leads to the problem of poor generalization in fault diagnosis. Therefore, a domain generalization mechanism is proposed. By expanding the source domain distribution T 0 to simulate the potential target domain distribution T, the cross-task adaptation ability of the model is improved. The core formula of the domain generalization mechanism is as follows:

[0061]

[0062] Among them, T 0 represents the source domain distribution, representing the distribution in the training data; T represents the potential virtual target domain distribution, constructed in the neighborhood of T 0 ; D(T, T 0 ) represents the distance between the source domain distribution and the target domain distribution, restricting the generation range of the virtual target domain distribution; ρ represents the distance limit parameter, controlling the generalization range; L(θ; (T, Y)) represents the adversarial training loss function; θ represents the model input parameter; Θ represents the filter; min represents minimization; sup represents the supremum; P represents the conditional threshold; E P represents the expected value under the condition of P; Y represents the distribution label vector.

[0063] Step S3 uses adversarial training sample generation as a way to simulate the target domain distribution T, applies perturbations to the augmented data, and generates a more challenging virtual target domain distribution. The formula for generating adversarial samples is as follows:

[0064]

[0065] where X adv represents the adversarial perturbation sample; β represents the step size parameter that controls the perturbation intensity; represents the gradient of the input to the adversarial training loss function, which is used to calculate the worst-case perturbation direction; Y 0 represents the source domain distribution label vector;

[0066] The optimization phase involves minimizing the gradients of the original samples and adversarial samples in the dataset. The resulting virtual cross-domain target domain distribution is used to update the model parameters, and the formula is as follows:

[0067]

[0068] where a represents the learning rate; represents the gradient of the original sample and the adversarial sample.

[0069] Cross-domain feature alignment is achieved through adversarial training, improving the robustness and generalization ability of the model.

[0070] Step S4: Calculate the Gaussian similarity of the target domain distribution samples, generate the adjacency matrix of the graph, propagate the label information through the normalized graph convolutional neural network, calculate the error using the cross-entropy loss function, and optimize the model parameters of the graph convolutional neural network. Using the Gaussian similarity and label propagation mechanism to construct an efficient graph convolutional neural network significantly improves the accuracy and robustness of small-sample fault diagnosis. The specific steps are as follows:

[0071] Step S41: Calculate the edge weights between samples using Gaussian similarity, and the formula for calculating the edge weights is as follows:

[0072]

[0073] where W ij represents the Gaussian similarity matrix, d represents the distance between samples, f φ (·) represents the feature extraction function, σ represents the degree of dispersion of the sample distribution, X i represents the row input sample, X j represents the column input sample, i represents the row index value, and j represents the column index value.

[0074] Step S42: Normalize the graph convolutional neural network shown in Figure 3 as follows:

[0075]

[0076] Among them, S represents the connectivity weight matrix of the normalized graph, D represents the degree matrix of nodes, D = diag(d i ), d i = Σ j W ij , and W represents the Gaussian similarity matrix.

[0077] The schematic diagram of the graph structure visualization based on the S matrix is as shown in Figure 4 . Among them, the S matrix is the connectivity weight of the normalized graph, that is, the similarity weight between nodes. The highly regular graph structure means that the similarity pattern between the model nodes is more structured, which will lead to a faster convergence rate of predicting labels in the information propagation stage.

[0078] Step S43: Perform label propagation and feature aggregation based on the graph convolutional neural network, and embed the relationship between target domains into the model. The formula is as follows:

[0079] N t+1 = k 2 SN t + (1 - k 2 )Y;

[0080] Among them, N t+1 represents the label or feature matrix after the (t + 1)-th propagation; N t represents the label or feature matrix after the t-th propagation; k 2 represents the weight between the current target and the initial target; Y represents the distribution label vector.

[0081] Step S44: Optimize the cross-entropy loss function through the adversarial samples and the label propagation results on the normalized graph convolutional neural network. The formula is as follows:

[0082]

[0083] Among them, J(φ, θ) represents, φ represents the network input feature, θ represents the model input parameter, P<y = j|X i > represents the probability of converting the label prediction matrix into the membership degree of the sample in each category, Y ij represents the label vector, y represents the prediction vector, m represents the number of samples in the validation set, and N represents the total number of all categories.

[0084] Step S5: Train the entire model, observe whether the loss rate tends to converge. If it converges, output the fault detection result; if not, repeat steps S2 - S4.

[0085] Figure 5The figure shows a structural schematic diagram of a certain type of single-shaft heavy-duty gas turbine, and both its compressor and turbine rotors are of the disk-drum type. The tie-rod rotor is a key component of the heavy-duty gas turbine, and it is of great significance to study its fault diagnosis technology.

[0086] It can be seen from this embodiment that the method realizes cross-domain feature alignment through adversarial training, strengthens the robustness and generalization ability of the model, and uses a graph convolutional neural network to model the target domain relationship, improving the diagnostic accuracy of the model.

[0087] Therefore, the heavy-duty gas turbine rotor fault diagnosis method of the present invention that adopts the above-mentioned domain generalization mechanism and GCN enhances the sample diversity through multiple convolution kernels of different sizes, generates a cross-domain target domain distribution in combination with the domain generalization mechanism, and constructs an efficient graph convolutional neural network using Gaussian similarity and label propagation mechanism, significantly improving the accuracy and robustness of small-sample fault diagnosis.

[0088] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that they can still modify or equivalently replace the technical solutions of the present invention, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A heavy-duty gas turbine rotor fault diagnosis method integrating domain generalization mechanism and GCN is characterized by: The following steps are involved: Step S1: collecting vibration data for preprocessing, and dividing the vibration data into a training set and a test set; Step S2: generating sample data using convolution kernels of different sizes for the collected original vibration data, and performing data enhancement on the sample data; Step S3: Combine the generated enhanced samples with adversarial training to expand the target domain distribution, and perform feature extraction through the CNN model; Step S4: Calculate the Gaussian similarity of the target domain distribution samples, generate the adjacency matrix of the graph, propagate the label information through the normalized graph convolutional neural network, calculate the error using the cross entropy loss function, and optimize the graph convolutional neural network model parameters; Step S5: Train the entire model to observe whether the loss rate tends to converge. If it converges, output the fault detection result. If it does not converge, repeat steps S2-S4; In step S3, a domain generalization mechanism is used for adversarial training. The core formula of the domain generalization mechanism is as follows: Among them, T0 represents the source domain distribution, which represents the distribution in the training data; T represents the potential virtual target domain distribution, which is constructed in the neighborhood of T0; D(T,T0) represents the distance between the source domain distribution and the target domain distribution, which limits the generation range of the virtual target domain distribution; ρ represents the distance limit parameter, which controls the generalization range; L(θ; (T,Y)) represents the adversarial training loss function; θ represents the model input parameter; Θ represents the filter; min represents minimization; sup represents the supremum; P represents the conditional threshold; E P represents the expected value under the condition of P; Y represents the distribution label vector.

2. The heavy-duty gas turbine rotor fault diagnosis method integrating domain generalization mechanism and GCN according to claim 1 is characterized by: Collecting vibration data for preprocessing in step S1 specifically includes collecting vibration data of different measuring points in the heavy-duty gas turbine tie rod rotor system under normal and abnormal conditions, and preprocessing missing values ​​and abnormal values ​​of the data.

3. The heavy-duty gas turbine rotor fault diagnosis method based on the fusion domain generalization mechanism and GCN according to claim 2 is characterized in that: The step S2 uses a variety of different convolution kernel sizes to perform data enhancement on the original vibration data to generate more diverse and robust sample data. The specific formula is: Among them, X * represents the enhanced data, X represents the original data, k1 represents the convolution kernel size, N represents the standard normal distribution, and v represents the sample that conforms to the standard normal distribution.

4. The heavy-duty gas turbine rotor fault diagnosis method integrating domain generalization mechanism and GCN according to claim 3 is characterized in that: In step S3, adversarial samples are generated in combination with adversarial training, and the generation formula of adversarial samples is: X adv =X * +b·▽ X L(θ;X * ,Y0); Among them, X adv represents the adversarial perturbation sample; β represents the step size parameter, which controls the perturbation intensity; ▽ X L(θ;X * ,Y0) represents the gradient of the adversarial training loss function input, which is used to calculate the perturbation direction in the worst case; Y0 represents the source domain distribution label vector. The optimization phase involves minimizing the gradients of the original and adversarial samples in the dataset. The resulting virtual cross-domain target domain distribution is used to update the model parameters as follows: θ←θ-a▽ θ L(θ); Where a represents the learning rate; θ L(θ) represents the gradient of the original sample and the adversarial sample; Cross-domain feature alignment is achieved through adversarial training to improve model robustness and generalization ability.

5. The heavy-duty gas turbine rotor fault diagnosis method integrating domain generalization mechanism and GCN according to claim 4 is characterized in that: The specific steps of step S4 are as follows: Step S41: Calculate the edge weights between samples using Gaussian similarity; Step S42: normalizing the graph convolutional neural network; Step S43: label propagation and feature aggregation are performed based on the graph convolutional neural network to embed the relationship between the target domains into the model; Step S44: Optimize the cross entropy loss function by adversarial samples and label propagation results on the normalized graph convolutional neural network.

6. The heavy-duty gas turbine rotor fault diagnosis method integrating domain generalization mechanism and GCN according to claim 5 is characterized in that: The formula for calculating the edge weight in step S41 is as follows: Among them, W ij represents the Gaussian similarity matrix, d represents the distance between samples, and f φ (·) represents the feature extraction function, σ represents the discrete degree of sample distribution, X i represents the row input sample, X j Represents column input samples, i represents row index value, and j represents column index value.

7. The heavy-duty gas turbine rotor fault diagnosis method integrating domain generalization mechanism and GCN according to claim 6 is characterized in that: The formula for normalization processing in step S42 is as follows: Where S represents the connectivity weight matrix of the normalized graph, D represents the degree matrix of the node, and D = diag(d i ), d i =∑ j W ij , W represents the Gaussian similarity matrix.

8. The heavy-duty gas turbine rotor fault diagnosis method integrating domain generalization mechanism and GCN according to claim 7 is characterized in that: In step S43, the relationship between target domains is embedded into the model, and the formula is as follows: N t+1 =k2SN t +(1-k2)Y; Among them, N t+1 N represents the label or feature matrix after the t+1th propagation; t represents the label or feature matrix after t propagation; k2 represents the weight between the current target and the initial target; Y represents the distribution label vector.

9. The heavy-duty gas turbine rotor fault diagnosis method integrating domain generalization mechanism and GCN according to claim 8 is characterized in that: The cross entropy loss function formula in step S44 is as follows: Among them, J(φ,θ) represents, φ represents the network input feature, θ represents the model input parameter, P <y=j|X i > represents the conversion of the label prediction matrix into the probability of the sample's membership in each category, Y ij represents the label vector, y represents the prediction vector, m represents the sample size of the validation set, and N represents the total number of all categories.

Images