Federal knowledge fusion and fault diagnosis method for gas turbine power generation equipment

Abstract

The application discloses a federal knowledge fusion and fault diagnosis method for gas turbine power generation equipment. A conditional generative adversarial network is established for edge fault diagnosis on each power plant edge, and then the established model is uploaded to the cloud for federal knowledge fusion. In the federal knowledge fusion process, each edge uploaded model is regarded as a teacher network, multi-level fault knowledge in the teacher network is extracted, and the dynamically weighted and averaged multi-level fault knowledge is distilled into the global model trained in the cloud, so that the construction of the unbiased global model is realized. Finally, the fused global model is deployed to each edge to update the local model of each edge. The application effectively realizes the diagnosis of all faults by fusing the knowledge of multiple power plant fault diagnosis models, realizes privacy protection, and greatly improves the fault detection performance.

Description

Technical Field

[0001] This invention belongs to the field of fault diagnosis of gas turbine power generation equipment, and particularly relates to a privacy-preserving statistical heterogeneous federated learning method in a cloud-edge collaborative scenario. Background Technology

[0002] In recent years, various sensors, such as temperature sensors and pressure sensors, have been widely used in industry, and the complexity of industrial systems has increased rapidly. As a result, data-driven models have attracted widespread research attention. Most data-driven modeling processes use centralized training with a centralized dataset. However, the equipment failure information collected by a single factory is always limited. In particular, some equipment failure categories may never have occurred in the historical operation of a particular factory. Therefore, a single factory cannot guarantee that the collected equipment failure category information is complete.

[0003] Today, cloud-edge collaboration has emerged as a new paradigm for intelligent manufacturing to address the problem of data silos between factories. Cloud-edge collaboration is a product of the complementary advantages of edge computing and cloud computing. Cloud servers can aggregate more information and handle more complex, computationally intensive tasks, such as training large models; while edge nodes are responsible for data collection and transmission, model deployment and inference, and training smaller models. Therefore, cloud-edge collaboration can aggregate equipment fault information from multiple factories, breaking down data silos and establishing a more generalizable and diagnostically robust global model. Federated learning, proposed by Google in 2016, is the most typical collaborative training method. Federated learning considers data privacy restrictions; different edge nodes cannot share data, but they can share model parameters and other relevant information. Information aggregation is performed through cloud servers, thereby achieving distributed deep model training.

[0004] Deep learning models have been at the forefront of data-driven fault diagnosis due to their remarkable ability to capture the nonlinear characteristics behind complex, high-dimensional process data. However, a common assumption of deep learning models is that the training data is class-balanced. This assumption is often not met in power plants due to various factors, such as the root causes of different faults, maintenance strategies for different facilities, and varying frequencies of different faults. This directly leads to significant differences in the amount of data for different faults. During power generation, different power plants have different production plans and operating environments, resulting in varying data imbalances. Furthermore, due to the different operating conditions of different power plants, the types of faults included in the historical equipment data of different plants are not entirely the same. Both the imbalance in fault data volume and the differences in fault categories lead to statistical heterogeneity, also known as data heterogeneity. This heterogeneity problem is particularly important in federated fault diagnosis because centralized data methods can always greatly alleviate this problem through data aggregation. However, in federated fault diagnosis, due to privacy considerations, the fault data of each edge node cannot leave its local environment. As a result, the fault diagnosis model trained by each edge node using only its own local data often deviates significantly from the global optimal solution.

[0005] However, existing methods for addressing statistical heterogeneity, including typical federated learning models such as SCAFFOLD, FedNova, FedProx, and FedDF, are not suitable for cloud-edge collaborative scenarios. In cloud-edge collaborative scenarios, each factory is considered an edge node, and the central server shared by all factories is considered the cloud. Existing heterogeneous federated learning methods deploy the training task entirely to each edge node with limited computing resources, while the central server, with its complex computing capabilities, only performs simple operations such as averaging or initialization. Therefore, this leads to a conflict between cloud and edge computing resources and task complexity. Summary of the Invention

[0006] The purpose of this invention is to address the statistical heterogeneity problem and privacy protection requirements of multi-gas turbine power generation equipment data in cloud-edge collaborative systems by providing a federated knowledge fusion and fault diagnosis method for gas turbine power generation equipment. Since the model continuously extracts knowledge from the data during training, this invention iteratively integrates knowledge from the edge models into the global model in the cloud while protecting data privacy. This effectively solves the statistical heterogeneity problem and avoids the conflict between computational resources and task complexity.

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

[0008] A federated knowledge fusion and fault diagnosis method for gas turbine power generation equipment is as follows:

[0009] Real-time acquisition of process variable data of gas turbine power generation equipment operation;

[0010] The process variable data is input into the trained fault diagnosis model to obtain the fault diagnosis results;

[0011] The trained fault diagnosis model is obtained based on a cloud-edge collaborative scenario, which includes K edge devices and a cloud platform, where each edge device corresponds to a power plant. The training method is as follows:

[0012] Each edge constructs a training dataset based on the fault data of the local gas turbine power generation equipment; each sample in the training dataset contains a segment of process variable data and the corresponding fault label;

[0013] Each edge constructs an edge-based conditional generative adversarial network (GAN), and the cloud constructs a global conditional GAN. Each edge-based GAN includes an edge feature extractor, an edge feature generator, and an edge predictor. The edge feature extractor extracts true fault features based on the process variable data of the input samples. The edge feature generator maps the concatenated features constructed based on the input samples into generated fault features. The concatenated features are a one-dimensional vector randomly sampled from a standard normal distribution and concatenated with the true fault labels in the input samples. The edge predictor outputs corresponding fault diagnosis results and authenticity judgment results for the true and generated fault features. The global conditional GAN ​​includes a global feature extractor, a global feature generator, and a global predictor.

[0014] Each edge inputs each sample from the training dataset into its corresponding edge conditional generative adversarial network (GAN) to train by minimizing the first loss function. During training, each edge uploads its GAN to the cloud at fixed intervals. The cloud performs parameter averaging on the edge GAN's model parameters to obtain the initial model parameters of the global GAN. The cloud data is then input into both the global GAN ​​and each edge GAN, and the global GAN ​​is trained by minimizing the second loss function. The trained global GAN ​​is then redistributed to each edge. Upon receiving the global GAN, each edge uses its global GAN's model parameters as the basis for the final parameters of its edge GAN. The training continues from the starting point; this process is repeated until the preset number of rounds is reached, completing the training of the edge-conditional generative adversarial network and the global conditional generative adversarial network for each edge. The trained edge feature extractor and edge predictor are the fault diagnosis models for the corresponding edges. The cloud data is a concatenation of several one-dimensional vectors sampled from a standard normal distribution with the same number of preset fault labels. The first loss function includes the classification loss of the edge predictor and the adversarial loss between the edge feature generator and the edge predictor regarding feature authenticity. The second loss function includes the alignment loss between the feature map knowledge of the global feature generator and the feature map knowledge of each edge feature generator, the alignment loss between the output class probabilities of the global predictor and each edge predictor, and the classification loss of the global predictor.

[0015] Furthermore, the classification loss of the edge predictor is:

[0016]

[0017] In the formula, D2 represents the mapping function of the second layer output of the k-th edge predictor, which outputs the fault diagnosis result. This represents the i-th real fault feature at edge k. This represents the i-th generated fault feature of edge k. The fault label represents the i-th real fault feature of edge k. This represents the fault label for the i-th generated fault feature at edge k; n represents the model parameters of the edge predictor for edge k. k This represents the total number of samples at edge k.

[0018] Furthermore, the adversarial loss between the edge feature generator and the edge predictor regarding feature authenticity is expressed as follows:

[0019]

[0020] In the formula, D1 represents the mapping function of the first layer output of the edge predictor, which outputs the authenticity judgment result; This represents the set of true fault features at edge k. Y represents the set of generation fault features of edge k. k This represents the set of fault labels for edge k. This represents the expectation of the loss function on the true fault characteristics. This represents the expectation of the loss function on the generated fault features. This represents the model parameters of the edge feature generator for edge k.

[0021] Furthermore, the first loss function also includes regularization loss, expressed as follows:

[0022]

[0023] f, g, and p represent the feature extractor, feature generator, and predictor, respectively. i = {f, g, p} represents the model parameters of the edge feature extractor, edge feature generator, and edge predictor for each edge, the subscript s represents the model parameters of the global conditional generative adversarial network distributed from the cloud in the previous round, and ||*|| represents the regularization function.

[0024] Furthermore, the loss for each sample in the alignment loss of the output class probabilities of the global predictor and each edge predictor is expressed as:

[0025]

[0026]

[0027]

[0028] Where p is the output class probability of the cloud predictor corresponding to sample n, p k This represents the output class probability of the k-th edge predictor corresponding to sample n, where C represents the fault type, and the superscript s indicates that it has passed through a softmax layer with an increased temperature T. This represents the c-th element of the softmax layer output probability of the cloud predictor when the temperature T increases. α represents the c-th element of the softmax layer output class probability of the k-th edge predictor corresponding to sample n when the temperature T is increased. k The weight of the dark knowledge of the k-th edge predictor corresponding to sample n is calculated from the class probabilities output by each edge predictor corresponding to sample n.

[0029] Furthermore, the classification loss of the global predictor is the cross-entropy loss.

[0030] Furthermore, the error for each sample in the error between the feature map knowledge of the global feature generator and the feature maps of each edge feature generator is expressed as:

[0031]

[0032] in This represents the feature map output by each layer of the global feature generator network corresponding to sample n, with a total of l layers; α represents the feature map output by each layer of the network in the k-th edge feature generator corresponding to sample n; k The weight represents the feature map knowledge of the k-th edge feature generator corresponding to sample n. This weight is calculated from the class probabilities output by each edge predictor corresponding to sample n. ||*||2 represents the L2 norm.

[0033] Furthermore, the learning rate for parameter updates during training of both the global conditional generative adversarial network and the edge conditional generative adversarial network is a dynamic learning rate, as shown in the following formula:

[0034] lr f =lr p =10 -3 ·γ r

[0035] lr g =10 -2 ·γ r

[0036] lr cloud =10 -3 ·γ r

[0037] lr f ,lr p The learning rate represents the learning rate of the edge feature extractor and the edge predictor; lr g The learning rate of the edge feature generator; lr cloud γ represents the learning rate of the global conditional generative adversarial network; γ is a hyperparameter used to make the learning rate decay with each iteration; r is the number of the current iteration.

[0038] The beneficial effects of this invention are as follows: This invention discloses a federated knowledge fusion and fault diagnosis method for gas turbine power generation equipment. This invention addresses the statistical heterogeneity problem caused by inconsistent fault categories of gas turbine power generation equipment in different power plants, under the premise of protecting data privacy. Due to the sporadic nature of faults, a single power plant cannot collect data on all fault categories, and the different operating conditions of each power plant lead to different fault categories for their gas turbine power generation equipment. This invention utilizes historical data from actual gas turbine power generation processes. First, a conditional generative adversarial network is established at each power plant's edge for edge fault diagnosis. Then, the established model is uploaded to the cloud for federated knowledge fusion. During the federated knowledge fusion process, to alleviate the model drift problem caused by statistical heterogeneity, a knowledge distillation method reflecting multi-level fault knowledge representation that reflects fault semantic information is constructed. Each model uploaded from the edge is considered a teacher network, and multi-level fault knowledge is extracted from the teacher network. To prevent the fused global model from being biased towards edge models with larger sample sizes, a dynamic weighting mechanism for sample-level diagnostic confidence is designed to dynamically average the multi-level fault knowledge of each teacher network. This method distills multi-level fault knowledge, obtained through dynamic weighted averaging, into a global model trained in the cloud, thus constructing an unbiased global model. Finally, the fused global model is deployed to each edge device to update the local model at each edge. By fusing knowledge from multiple power plant fault diagnosis models, this method effectively diagnoses all faults, achieving not only privacy protection but also significantly improving fault detection performance. Attached Figure Description

[0039] Figure 1 This is an overall framework diagram of the federated knowledge fusion and fault diagnosis method for gas turbine power generation equipment according to the present invention.

[0040] Figure 2 Flowchart of global model training based on knowledge fusion. Detailed Implementation

[0041] The present invention will be further described in detail below with reference to the accompanying drawings and specific examples.

[0042] This invention uses historical operating data and fault data of a power plant in Zhejiang as an example to illustrate the process. The generator set includes typical equipment such as a compressor, combustion chamber, turbine, etc., as well as auxiliary equipment.

[0043] Data generated during the operation of gas turbine power generation equipment includes multiple process variables, involving measurement points such as temperature, pressure, water level, flow rate, valve opening, and vibration. Since gas turbine power generation equipment is a complex and critical piece of equipment, intentionally operating it in a faulty state to collect fault data would result in significant economic losses. Therefore, it is necessary to integrate diverse diagnostic knowledge from different power plants and study the problem of fault diagnosis in a cloud-edge collaborative scenario with heterogeneous data.

[0044] like Figure 1 and Figure 2 As shown, this invention is a federated knowledge fusion and fault diagnosis method for gas turbine power generation equipment, specifically as follows:

[0045] Real-time acquisition of process variable data of gas turbine power generation equipment operation;

[0046] The process variable data is input into the trained fault diagnosis model to obtain the fault diagnosis results;

[0047] The trained fault diagnosis model is obtained based on a cloud-edge collaborative scenario, which includes K edge devices and a cloud platform, where each edge device corresponds to a power plant. The training method is as follows:

[0048] (1) Each edge constructs a training dataset based on the fault operation data of the local gas turbine power generation equipment: Each power plant is taken as an edge, and the original process variable data corresponding to the operation process of the fault label is selected from the historical database of the gas turbine power generation equipment recorded by each power plant. These data include real-time detected temperature, pressure, flow rate, etc. The selected original process variable data is preprocessed by outlier cleaning and other preprocessing operations. The preprocessed process variable data is then divided into sliding windows to obtain the time-series extended process variable data.

[0049]

[0050] Where w is the length of the sliding window, and w = 30 is taken as a sliding window, the step size of the sliding window is 1, k represents the k-th edge, and n k This represents the total number of samples at edge k. This represents the i-th sample at the k-th edge, obtained through the i-th sliding window partition, containing process variable data at 30 time points. This represents the process variable data at time 1 of the i-th sliding window, including temperature, pressure, flow rate, etc. at that time. K represents the total number of edges; here, three edges are set, i.e., K = 1, 2, 3. Each sample is mapped one-to-one with a fault label to obtain the training dataset.

[0051] (2) Establish the edge conditional generative adversarial network for each edge: Taking the k-th edge as an example, the specific steps are as follows:

[0052] (2.1) Initialize the parameters of the edge conditional generative adversarial network. The edge conditional generative adversarial network consists of three parts: edge feature extractor, edge feature generator and edge predictor.

[0053] (2.2) The edge feature extractor is used to extract real fault features based on the process variable data of the input samples. Specifically, it extracts the process variable data x of all samples. k Input the edge feature extractor to obtain the extracted true fault features:

[0054]

[0055] Where F represents the mapping function of the edge feature extractor. In particular, the LSTM algorithm is used as the first layer of the feature extractor to extract the temporal features within each sliding window. This represents the model parameters of the edge feature extractor for the k-th edge. This represents the real fault features extracted from the real fault data of the k-th edge.

[0056] (2.3) The edge feature generator is used to map the concatenated features constructed based on the input samples into generated fault features. Specifically, one-dimensional vectors with the same number as the training dataset are randomly sampled from the standard normal distribution, and then concatenated with the fault labels of the real data samples in the training dataset one-to-one to obtain the concatenated features, which are then input into the edge feature generator to obtain the generated fault features:

[0057]

[0058] Where G is the mapping function of the edge feature generator. This represents the model parameters of the edge feature generator for the k-th edge. Y represents the generation fault characteristic of the k-th edge. k These are fault category labels corresponding to actual fault data, typically encoded using one-hot encoding, Z. n These are sample points randomly sampled from a normal distribution.

[0059] (2.4) The edge predictor is used to adversarially test the extracted real and generated fault features. Specifically, the real and generated fault features are input into the edge predictor. The edge predictor has two outputs: one outputs the classification result of the category to which the real fault feature belongs, and the other outputs the confidence level of the feature's authenticity. The edge predictor should, on the one hand, classify the input features as accurately as possible, and on the other hand, distinguish between the real and generated fault features as much as possible.

[0060] (3) Each edge inputs each sample from the training dataset into the corresponding edge conditional generative adversarial network, calculates the first loss function of the edge conditional generative adversarial network, obtains the gradient, and backpropagates to train the model parameters. The first loss function generally includes the classification loss of the edge predictor and the adversarial loss between the edge feature generator and the edge predictor regarding the authenticity of features. For example, it is expressed as follows:

[0061]

[0062]

[0063] in The loss represents the adversarial loss between the k-th edge feature generator and the edge predictor regarding feature authenticity. The edge feature generator should generate fault features that can confuse the edge predictor as much as possible, while the edge predictor should identify the fault features generated by the edge feature generator as false features and identify the real fault features extracted from the real data by the edge feature extractor as real features. D1 represents the classification loss of the k-th edge predictor, which aims to obtain the most accurate fault classification result for the input true fault features; D2 represents the mapping function of the first layer output of the edge predictor, which outputs the truth judgment result; D3 represents the mapping function of the second layer output of the k-th edge predictor, which outputs the fault diagnosis result. This represents the expectation of the loss function on the true fault characteristics. Y represents the expectation of the loss function on the generated fault features. k Indicates fault label, This represents the model parameters of the predictor for edge k. This represents the i-th real fault feature at edge k. This represents the i-th generated fault feature of edge k. The fault label represents the i-th real fault feature of edge k. Let represent the fault label of the i-th generated fault feature at edge k. Furthermore, this first loss function also includes a regularization loss. It is a method to make the model parameter θ k Compared to the model parameters θ obtained in the cloud in the previous round s Regularization constraints that do not allow for large changes:

[0064]

[0065] ||*|| represents the regularization function.

[0066] (4) During the training process, each edge uploads its conditional generative adversarial network to the cloud at fixed intervals, merges the edge models, and then inputs the cloud data into the global conditional generative adversarial network and each edge's conditional generative adversarial network. The training of the global conditional generative adversarial network is completed by minimizing the second loss function. The trained global conditional generative adversarial network is then redistributed to each edge. After receiving the global conditional generative adversarial network, each edge uses the model parameters of the global conditional generative adversarial network as the starting point for subsequent training of its edge conditional generative adversarial network. This process is repeated until the preset number of rounds is reached, completing the training of the edge conditional generative adversarial network and the global conditional generative adversarial network for each edge. The structure of the global conditional generative adversarial network is consistent with that of the edge conditional generative adversarial network, including a global feature extractor, a global feature generator, and a global predictor. This step includes the following sub-steps:

[0067] (4.1) Upload all edge condition generative adversarial networks to the cloud and average the model parameters:

[0068]

[0069] Where f, g, and p represent the feature extractor, feature generator, and predictor, respectively. This represents the model parameters for the edge feature extractor, edge feature generator, and edge predictor at each edge. This represents the initial model parameters for the corresponding global feature extractor, global feature generator, and global predictor.

[0070] (4.2) N one-dimensional vectors are sampled from a standard normal distribution and concatenated with the same number of preset fault labels as cloud data. This data is then input into the global conditional generative adversarial network (GAN) and each edge conditional GAN. The sampled points must ensure class balance; generally, the data volume is a multiple of the number of classes. In this embodiment, the number of classes is 9, and 8640 samples are taken. The cloud data is input into the averaged global feature generator to obtain the same number of generated fault features. These generated fault features are then input into the averaged global predictor to obtain the prediction result. Similarly, the cloud data is input into each edge feature generator to obtain the same number of generated fault features. These generated fault features are then input into the edge predictor to obtain the prediction result. During the forward propagation of data through the model to obtain the prediction result, the feature generator stores knowledge using feature maps, while the predictor stores knowledge using dark knowledge.

[0071] (4.3) Construct a second loss function, which includes the alignment loss between the feature map knowledge of the global feature generator and the feature maps of each edge feature generator, the alignment loss between the output class probabilities of the global predictor and each edge predictor, and the classification loss of the global predictor.

[0072] (4.3.1) In this embodiment, the alignment loss between the feature map knowledge of the global feature generator and the feature maps of each edge feature generator is calculated. To simplify the notation, the loss function for a single sample n is shown as follows:

[0073]

[0074] in Let l represent the feature map output by each layer of the global feature generator network corresponding to sample n, with a total of l layers, and in this case l = 3; This represents the l-layer feature map knowledge of the k-th edge feature generator corresponding to sample n; α k The weight represents the feature map knowledge of the k-th edge feature generator corresponding to sample n. This weight is calculated from the class probabilities output by each edge predictor corresponding to sample n. ||*||2 represents the L2 norm.

[0075] (4.3.2) In this embodiment, the dark knowledge of the global predictor is aligned with the dark knowledge of each edge predictor, and the alignment loss of the output class probabilities of the global predictor and each edge predictor is calculated. For the sake of simplifying the notation, the loss function of a single sample n is shown as follows:

[0076]

[0077]

[0078]

[0079] Where p is the output class probability of the cloud predictor corresponding to sample n, p k is the output class probability of the k-th edge predictor corresponding to sample n, C represents the fault type, and the superscript s indicates that it has passed through a softmax layer with an increased temperature T, where T is a temperature parameter, which is set to 5 to smooth the class probability calculated by the traditional softmax layer. This represents the c-th element of the softmax layer output class probability of the cloud predictor for the temperature rise T corresponding to sample n. The calculation of each element in the class probability is shown below. Since there are C = 9 types of faults, the dimension is 9:

[0080]

[0081] s c s represents the c-th element of the output vector of the last layer of the predictor after inputting sample n. j They have the same meaning. Similarly, This represents the c-th element of the softmax layer output class probability of the k-th edge predictor corresponding to sample n when the temperature T is increased. The calculation formula is the same as above. Where α...k The weights representing the dark knowledge of the k-th edge predictor corresponding to sample n are calculated from the class probabilities output by each edge predictor corresponding to sample n, similar to the weights of the feature map knowledge. Specifically, for each generated sample, each edge predictor provides a classification probability vector after passing through a softmax layer with a temperature increased by T. Extract the class probability values ​​corresponding to the correct class labels from the vector and calculate the weighting coefficients:

[0082]

[0083] in T represents the predicted probability value of the correct class label in the class probability vector output by the k-th edge predictor corresponding to sample n, where T is 5.

[0084] (4.3.3) In this embodiment, the cross-entropy loss between the output class probability of the global model and the true label is calculated as the classification loss of the global predictor. To simplify the notation, the loss function for a single sample n is used as an example:

[0085]

[0086] Where y c This represents the c-th dimension of the one-hot encoding of sample n. This sample label is the pre-defined fault label concatenated with the generated sample from the cloud-generated data; the superscript h indicates that this probability vector is the output of a traditional softmax layer. This indicates that sample n corresponds to the c-th dimension of the output class probability of the traditional softmax layer of the cloud predictor. The traditional softmax layer refers to the layer that has not undergone temperature increase operation.

[0087] (4.3.4) Finally, the second loss function is calculated, the gradient with respect to the global model parameters is obtained, and backpropagation is performed. The second loss function is as follows:

[0088]

[0089] Here, μ and ν are both hyperparameters used to adjust the ratio between different loss functions, and both are 1 here. N represents the total number of cloud-sampled samples mentioned above, and n represents the nth sample among them.

[0090] (4.4) Deploy the global conditional generative adversarial network trained in step (4.3) to each edge and update the conditional generative adversarial network of each edge.

[0091] (4.5) Iterate through the steps of edge conditional generative adversarial network training, uploading the edge conditional generative adversarial network, global conditional generative adversarial network training, and deploying the cloud conditional generative adversarial network to the edge until the model converges or reaches the preset iteration rounds R. The trained edge feature extractor and edge predictor are the fault diagnosis models for the corresponding edges. Here, the preset iteration rounds are used to stop the model training, with R set to 100 rounds.

[0092] Furthermore, in the above steps, the learning rate used for updating the parameters of both the edge local model and the global model is a dynamic learning rate, as shown in the following formula:

[0093] lr f =lr p =10 -3 ·γ r

[0094] lr g =10 -2 ·γ r

[0095] lr cloud =10 -3 ·γ r

[0096] lr f ,lr p The learning rate represents the learning rate of the edge feature extractor and the edge predictor; lr g The learning rate of the edge feature generator; lr cloud This represents the learning rate used to update the global conditional generative adversarial network during knowledge fusion. In the three formulas above, γ is the same hyperparameter used to make the learning rate decrease as the number of iterations increases; here, γ = 0.99 is taken. r is the current iteration number, starting from 0 (0 ≤ r < R).

[0097] Seven cases from three endpoints were selected to verify the effectiveness of the method. The fault categories in the training data of the three endpoints for each case are shown in Table 1. It should be noted that the test categories for each endpoint in each case are all fault categories. The fault categories, in order, are: abnormal temperature in the exhaust transition section of the gas turbine, abnormal lubricating oil pressure in the steam turbine lubricating oil cooler, abnormal natural gas temperature before combustion in the gas turbine fuel pre-module, abnormal pressure drop of circulating water in the condenser cooling pipe, abnormal cooling air pressure in the gas turbine compressor, abnormal acceleration in the gas turbine combustion chamber, abnormal natural gas flow in the gas turbine fuel pre-module, abnormal temperature difference between hot and cold hydrogen in the generator hydrogen cooler, and large vibration of the gas turbine support bearing.

[0098] Table 1. Fault category classification in training data from different edges

[0099]

[0100] Table 2 shows the comparison results between the proposed method and some classic heterogeneous federated learning methods. "Seen" represents the average test results of the three edges for known fault categories in the training set, and "Not Seen" represents the average test results of the three edges for unseen fault categories in the training set. The score represents the geometric mean of seen and unseen results. As can be seen from the table, the proposed method performs best in the vast majority of cases, achieving the highest scores in cases A, B, D, E, and G, showing a significant improvement over the other four methods. Furthermore, in all cases, the proposed method achieves an accuracy rate of over 95% for known fault categories, indicating that the model can effectively identify known faults at each edge. Moreover, in all cases, the proposed method achieves an accuracy rate of around 80% for unseen fault categories, significantly higher than other methods, demonstrating that the proposed method can effectively aggregate knowledge from the edge models to assist each edge in identifying unseen fault categories.

[0101] Table 2 Comparison results between the method of this invention and the classical heterogeneous federated learning method

[0102]

[0103] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A federated knowledge fusion and fault diagnosis method for gas turbine power generation equipment, characterized in that, Specifically: Real-time acquisition of process variable data of gas turbine power generation equipment operation; The process variable data is input into the trained fault diagnosis model to obtain the fault diagnosis results; The trained fault diagnosis model is obtained based on training in a cloud-edge collaborative scenario, which includes... There are one edge and one cloud, where each edge corresponds to one power plant; the training method is as follows: Each edge constructs a training dataset based on the fault data of the local gas turbine power generation equipment; each sample in the training dataset contains a segment of process variable data and the corresponding fault label; Each edge constructs an edge-based conditional generative adversarial network (GAN), and the cloud constructs a global conditional GAN. Each edge-based GAN includes an edge feature extractor, an edge feature generator, and an edge predictor. The edge feature extractor extracts true fault features based on the process variable data of the input samples. The edge feature generator maps the concatenated features constructed based on the input samples into generated fault features. The concatenated features are a one-dimensional vector randomly sampled from a standard normal distribution and concatenated with the true fault labels in the input samples. The edge predictor outputs corresponding fault diagnosis results and authenticity judgment results for the true and generated fault features. The global conditional GAN ​​includes a global feature extractor, a global feature generator, and a global predictor. Each edge inputs each sample from the training dataset into its corresponding edge conditional generative adversarial network (GAN) to train by minimizing the first loss function. During training, each edge uploads its GAN to the cloud at fixed intervals. The cloud performs parameter averaging on the edge GAN's model parameters to obtain the initial model parameters of the global GAN. The cloud data is then input into both the global GAN ​​and each edge GAN, and the global GAN ​​is trained by minimizing the second loss function. The trained global GAN ​​is then redistributed to each edge. Upon receiving the global GAN, each edge uses its global GAN's model parameters as the basis for the final parameters of its edge GAN. The training continues from the starting point; this process is repeated until the preset number of rounds is reached, completing the training of the edge-conditional generative adversarial network and the global conditional generative adversarial network for each edge. The trained edge feature extractor and edge predictor are the fault diagnosis models for the corresponding edges. The cloud data is a concatenation of several one-dimensional vectors sampled from a standard normal distribution with the same number of preset fault labels. The first loss function includes the classification loss of the edge predictor and the adversarial loss between the edge feature generator and the edge predictor regarding feature authenticity. The second loss function includes the alignment loss between the feature map knowledge of the global feature generator and the feature map knowledge of each edge feature generator, the alignment loss between the output class probabilities of the global predictor and each edge predictor, and the classification loss of the global predictor.

2. The method according to claim 1, characterized in that, The classification loss of the edge predictor is: ； In the formula, Indicates the first The mapping function output from the second layer of the edge predictor outputs the fault diagnosis result. Indicates edge The A true fault characteristic, Indicates edge The One generated fault feature, Indicates edge The Fault labels representing real fault characteristics, Indicates edge The A fault label that generates fault characteristics; Indicates edge Model parameters of the edge predictor Indicates edge The total sample size.

3. The method according to claim 1, characterized in that, The adversarial loss between the edge feature generator and the edge predictor regarding feature authenticity is expressed as follows: ； In the formula, This represents the mapping function output by the first layer of the edge predictor, which outputs the truth value judgment result; Indicates edge The set of real fault characteristics Indicates edge The generation of fault feature set, Indicates edge The set of fault labels, This represents the expectation of the loss function on the true fault characteristics. This represents the expectation of the loss function on the generated fault features. Indicates edge The model parameters of the edge feature generator, Indicates edge The model parameters of the edge predictor are given by G, where G is the mapping function of the edge feature generator.

4. The method according to claim 1, characterized in that, The first loss function also includes regularization loss, expressed as follows: ； These represent the feature extractor, feature generator, and predictor, respectively. These represent the model parameters for the edge feature extractor, edge feature generator, and edge predictor at each edge. This represents the model parameters of the global conditional generative adversarial network distributed from the cloud in the previous round, and ||*|| represents the regularization function.

5. The method according to claim 1, characterized in that, The loss for each sample in the alignment loss of the output class probabilities of the global predictor and each edge predictor is expressed as: ； ； ； in It is a sample The corresponding number The output class probability of each edge predictor Indicates the total number of all fault types, indicated by superscript. Indicates that the temperature has been increased The softmax layer, Indicates that the cloud predictor is raising the temperature. The softmax layer outputs the class probability of the first One element, Indicates sample The corresponding number The edge predictor increases temperature The softmax layer outputs the class probability of the first One element, Indicates sample The corresponding number The weights of the dark knowledge of each edge predictor are determined by the samples. The class probabilities are calculated from the outputs of the corresponding edge predictors.

6. The method according to claim 1, characterized in that, The classification loss of the global predictor is the cross-entropy loss.

7. The method according to claim 1, characterized in that, The error for each sample in the error between the feature map knowledge of the global feature generator and the feature maps of each edge feature generator is expressed as: ； in Indicates sample The corresponding global feature generator outputs feature maps from each layer of the network, totaling... layer; Indicates sample The corresponding number The feature map output by each layer of the network in the edge feature generator; Indicates sample The corresponding number The weights of the feature map knowledge of each edge feature generator are determined by the samples. The class probabilities output by the corresponding edge predictors are calculated. This represents the regularization function.

8. The method according to claim 1, characterized in that, The learning rate used for parameter updates during the training of both the global conditional generative adversarial network and the edge conditional generative adversarial network is a dynamic learning rate, as shown in the following formula: ； ； ； in This represents the learning rate of the edge feature extractor and the edge predictor; This represents the learning rate of the edge feature generator; This represents the learning rate of the global conditional generative adversarial network. This is a hyperparameter used to make the learning rate decrease continuously with each iteration. This represents the current iteration round number.

Images