A Distributed Fault Diagnosis Method for Oil-Immersed Transformers Based on Data Privacy Protection
By employing federated learning and differential privacy processing, a distributed fault diagnosis model for transformers is constructed, which solves the data silo problem in transformer fault diagnosis, achieves collaborative training without data sharing and privacy protection, and improves the accuracy and security of diagnosis.
Patent Information
- Application Number
- CN202210751099.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-06-28
- Publication Date
- 2025-10-28
- Estimated Expiration
- 2042-06-28
AI Technical Summary
In existing technologies, data silos exist during transformer fault diagnosis, making data sharing difficult and user data privacy difficult to protect during machine learning diagnosis.
A distributed fault diagnosis method based on federated learning is adopted. A local convolutional neural network model is built at each transformer site, and the model parameters are aggregated using the FedAvg algorithm. At the same time, differential privacy processing is performed on the central server, and Gaussian noise is added to protect data privacy.
It enables collaborative training without sharing data, protects user information, has good generalization ability, solves the data silo problem, and avoids privacy leaks of model parameters.
Smart Images
Figure CN115239989B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of transformer fault diagnosis technology, specifically relating to a distributed fault diagnosis method for oil-immersed transformers based on data privacy protection. Background Technology
[0002] In recent years, with the increasing demand for electricity due to national economic development, the total number of power transformers put into operation in various countries has been increasing year by year. Therefore, accurate diagnosis of whether transformers have failed and what kind of failures have occurred has become a key research focus.
[0003] In existing technologies, many scholars have introduced artificial intelligence methods into the field of transformer fault diagnosis. By collecting a large amount of transformer operating data and using machine learning algorithms to simulate human learning methods, an ideal model for transformer fault diagnosis can be obtained after multiple training iterations. This method achieves high efficiency, digitalization, and intelligence in transformer fault diagnosis. However, as the protection of industrial data privacy has become increasingly important, data sharing between departments and industries within the power sector has encountered obstacles. The operating data of each power transformer is isolated locally, and the phenomenon of "data silos" is prevalent between data pools. This phenomenon makes it difficult to implement the aforementioned method of centralized training based on data collection. Therefore, in the process of machine learning-based transformer fault diagnosis, how to break down "data silos" to achieve multi-party collaborative training while protecting user data privacy is a key issue in overcoming the current application dilemma. Summary of the Invention
[0004] The purpose of this invention is to address the above-mentioned problems by proposing a distributed fault diagnosis method for oil-immersed transformers based on data privacy protection. This method can effectively solve the data silo problem in the transformer diagnosis process, protect the privacy of model parameters, effectively protect user information, and has good generalization ability.
[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0006] This invention proposes a distributed fault diagnosis method for oil-immersed transformers based on data privacy protection, comprising the following steps:
[0007] S1. Obtain dissolved gas data in oil from N transformer stations and store it in the corresponding local database. No data sharing is performed between the local databases.
[0008] S2. Normalize the dissolved gas data in oil from each local database, and group and label the normalized data based on the modified three-ratio method.
[0009] S3. Perform bidirectional tiling enhancement on the labeled data and convert it into a two-dimensional image accordingly;
[0010] S4. Construct a transformer fault diagnosis model based on federated learning. The transformer fault diagnosis model includes a central server and N participant clients. Each participant client corresponds to a transformer site. Each participant client is configured with a local model, which is a convolutional neural network model. The central server uses the FedAvg algorithm to aggregate the parameters of the local models.
[0011] S5. Input the two-dimensional images of each transformer station into the corresponding local model for training to obtain the local model parameters;
[0012] S6. Add Gaussian noise to the local model parameters for differential privacy processing;
[0013] S7. Upload the local model parameters after differential privacy processing to the central server for aggregation, and then send the aggregated local model parameters to the local model. Return to step S5 and continue until the preset number of training iterations are completed to obtain the optimal transformer fault diagnosis model.
[0014] Preferably, the dissolved gas data in the oil includes five gases: H2, CH4, C2H6, C2H4, and C2H2.
[0015] Preferably, the dissolved gas data in oil from each local database is normalized using the following formula:
[0016]
[0017]
[0018]
[0019]
[0020]
[0021] in, The total content of the five gases in each data set. This represents the total content of CH4, C2H6, C2H4, and C2H2 gases. and The dissolved gas contents in the oil after normalization treatment are H2, CH4, C2H6, C2H4, and C2H2, in that order.
[0022] Preferably, the convolutional neural network model is the LeNet-5 network model.
[0023] Preferably, gradient clipping is performed before adding Gaussian noise to the local model parameters for differential privacy processing.
[0024] Preferably, the Gaussian noise satisfies the following formula:
[0025]
[0026] in,
[0027]
[0028] In the formula, Δf is the sensitivity of differential privacy, δ is the relaxation factor, ε is the privacy budget, and D k Let C be the sample size of participant client k, k = 1, 2, ..., N, and C be the gradient clipping coefficient.
[0029] Preferably, during training, the update formula for each local model parameter is as follows:
[0030]
[0031] in, and F represents the weight parameters for participant k in the t-th and t+1-th iterations, respectively. k (ω k Let be the loss function for participant k. Here, R is the gradient operator, and R is the learning rate; N(0,σ) 2 ) is random noise that follows a Gaussian distribution.
[0032] Preferably, the local model parameters after differential privacy processing are uploaded to the central server for aggregation, as shown in the following formula:
[0033]
[0034] in, Let N be the weight parameter for participant k in the (t+1)th iteration, and N be the number of participant clients.
[0035] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0036] 1) This method uses a federated learning architecture to address the shortcomings of centralized fault diagnosis when it is impossible to collect a large amount of data. It breaks down the data silos in the transformer local area, realizes distributed collaborative training without data leaving the local area, and avoids privacy leakage of local model parameters.
[0037] 2) By adding Gaussian noise to the local model parameters and implementing an adaptive noise addition mechanism to account for the differences in the sample content of each participant's client, privacy leakage during the local model parameter upload process is further avoided, effectively protecting data privacy and exhibiting good generalization ability. Attached Figure Description
[0038] Figure 1This is a flowchart of the distributed fault diagnosis method for oil-immersed transformers based on data privacy protection according to the present invention;
[0039] Figure 2 This is a schematic diagram of the LeNet-5 network model of the present invention;
[0040] Figure 3 This is a structural diagram of the differential privacy federated learning method of this invention;
[0041] Figure 4 This is a graph showing the accuracy variation of different privacy budget models for this invention on a balanced dataset;
[0042] Figure 5 This is a graph showing the loss variation of different privacy budget models under a balanced dataset according to the present invention;
[0043] Figure 6 This is a graph showing the accuracy variation of different privacy budget models for this invention under imbalanced datasets;
[0044] Figure 7 This is a graph showing the variation of loss for different privacy budget models under an imbalanced dataset, based on the present invention. Detailed Implementation
[0045] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0046] It should be noted that, unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application.
[0047] like Figure 1-7 As shown, a distributed fault diagnosis method for oil-immersed transformers based on data privacy protection includes the following steps:
[0048] S1. Obtain dissolved gas data in oil from N transformer stations and store it in the corresponding local database. No data sharing is performed between the local databases.
[0049] S2. Normalize the dissolved gas data in oil from each local database, and group and label the normalized data based on the improved three-ratio method.
[0050] In one embodiment, the dissolved gas data in the oil includes five gases: H2, CH4, C2H6, C2H4, and C2H2.
[0051] In one embodiment, the dissolved gas data in oil from each local database is normalized using the following formula:
[0052]
[0053]
[0054]
[0055]
[0056]
[0057] in, The total content of the five gases in each data set. This represents the total content of CH4, C2H6, C2H4, and C2H2 gases. and The dissolved gas contents in the oil after normalization treatment are H2, CH4, C2H6, C2H4, and C2H2, in that order.
[0058] In real industrial environments, each transformer station collects operational data (i.e., dissolved gas data in oil) and stores it in a local database, with no data sharing between these databases. In this embodiment, five gases—H2, CH4, C2H6, C2H4, and C2H2—are selected for analysis to diagnose transformer faults based on operational conditions. and The dissolved gas contents in the oil, in order of normalization, are H2, CH4, C2H6, C2H4, and C2H2. It should be noted that the choice of gas type can be adjusted according to actual needs. Then, according to the modified three-ratio method recommended in my country's current DL / T722-2000 "Guideline," the above data are grouped and labeled, such as each group including the dissolved gas data of the above five oil types, to facilitate subsequent model training. This is existing technology and will not be elaborated further here.
[0059] S3. The labeled data is enhanced by bidirectional tiling and converted into a two-dimensional image.
[0060] In this process, by performing bidirectional tiling enhancement on the labeled data, the original data features are expanded into a 28×28 matrix feature, and then further converted into a corresponding two-dimensional image, giving the data spatial features, which helps to enhance the feature expression ability of the data.
[0061] S4. Construct a transformer fault diagnosis model based on federated learning. The transformer fault diagnosis model includes a central server and N participant clients. Each participant client corresponds to a transformer site. Each participant client is configured with a local model, which is a convolutional neural network model. The central server uses the FedAvg algorithm to aggregate the parameters of the local models.
[0062] In one embodiment, the convolutional neural network model is the LeNet-5 network model.
[0063] Among them, a transformer fault diagnosis model is built based on federated learning. In the transformer fault diagnosis model, each independent transformer site is equivalent to a participant. Each participant isolates its data to solve the "data silo" problem while meeting the needs of user privacy protection and data security.
[0064] The participant's client is equipped with a local model, which uses the LeNet-5 network model as the training model, or other convolutional neural network models from existing technologies. Figure 2 As shown, the LeNet-5 network model is a current technology, consisting of convolutional layers, pooling layers, fully connected layers, and a classifier. Layers 1 and 3 are convolutional layers, each using a 5×5 convolutional kernel to extract input features (two-dimensional images). Layers 2 and 4 are pooling layers, using a 2×2 window to pool and compress the convolutional image. Layer 5 uses a 5×5 convolutional kernel to process the output of the above four layers, resulting in 120 1×1 neurons. Layer 6 is a fully connected layer with 84 nodes, corresponding to a 7×12 bitmap. Layer 7 outputs the image category through a classifier.
[0065] S5. Input the two-dimensional images of each transformer station into the corresponding local model for training to obtain the local model parameters.
[0066] S6. Add Gaussian noise to the local model parameters for differential privacy processing.
[0067] In one embodiment, gradient clipping is performed before adding Gaussian noise to the local model parameters for differential privacy processing.
[0068] In one embodiment, the Gaussian noise satisfies the following formula:
[0069]
[0070] in,
[0071]
[0072] In the formula, Δf is the sensitivity of differential privacy, δ is the relaxation factor, ε is the privacy budget, and D k Let C be the sample size of participant client k, k = 1, 2, ..., N, and C be the gradient clipping coefficient.
[0073] like Figure 3 As shown, the user parameters are the local model parameters, and the global model is the transformer fault diagnosis model. To avoid the risk of privacy leaks caused by attackers stealing the local model parameters during the upload process, we consider adding noise during the local model training process to achieve differential privacy federated learning. Generally, gradient clipping should be performed before adding noise, such as binding the gradient between C and -C.
[0074] In a federated learning structure, it is difficult for the data held by each participant to satisfy the independent and identically distributed principle. Simply adding the same noise to all participants obviously cannot eliminate the impact of data imbalance on the model training effect. Therefore, considering the difference in the data sample size of each participant, an adaptive noise addition mechanism is achieved through formulas (6) and (7).
[0075] S7. Upload the local model parameters after differential privacy processing to the central server for aggregation, and then send the aggregated local model parameters to the local model. Return to step S5 and continue until the preset number of training iterations are completed to obtain the optimal transformer fault diagnosis model.
[0076] In one embodiment, the update formula for each local model parameter during training is as follows:
[0077]
[0078] in, and F represents the weight parameters for participant k in the t-th and t+1-th iterations, respectively. k (ω k Let be the loss function for participant k. Here, R is the gradient operator, and R is the learning rate; N(0,σ) 2 ) is random noise that follows a Gaussian distribution.
[0079] In one embodiment, the local model parameters after differential privacy processing are uploaded to a central server for aggregation, as shown in the following formula:
[0080]
[0081] in, Let N be the weight parameter for participant k in the (t+1)th iteration, and N be the number of participant clients.
[0082] The above process can be iterated multiple times to obtain the optimal transformer fault diagnosis model. Using the optimal transformer fault diagnosis model for fault diagnosis can result in better diagnostic results for the local data of each federated participant.
[0083] To more intuitively reflect the impact of privacy protection effectiveness and privacy budget on model accuracy, a comparative experiment was first designed using independent and identically distributed data (balanced dataset) to model transformer fault diagnosis under different privacy budgets. (Refer to...) Figure 4 The x-axis represents the number of training iterations, and the y-axis represents the model's accuracy under different privacy budgets. With increasing training iterations and adjustments to the privacy budget, the model accuracy can reach 97.64%. (Refer to...) Figure 5 The horizontal axis represents the number of training iterations, and the vertical axis represents the loss of the loss function under different privacy budgets. As the number of training iterations increases and the privacy budget is adjusted, the loss can be reduced to as low as 0.06.
[0084] Furthermore, considering that local data from participants in an industrial environment rarely meet the independent and identically distributed principle (imbalanced datasets), and generally exhibit differences in fault types and sample sizes, the data used in the above experiments was divided. Each participant was assigned two or three non-overlapping fault type data sets, and the privacy budget was appropriately adjusted for the experiment. (Refer to...) Figure 6 The x-axis represents the number of training iterations, and the y-axis represents the model's accuracy under different privacy budgets. With increasing training iterations and adjustments to the privacy budget, the model accuracy can reach 97.02%. (Refer to...) Figure 7 The horizontal axis represents the number of training iterations, and the vertical axis represents the loss of the loss function under different privacy budgets. As the number of training iterations increases and the privacy budget is adjusted, the loss can be reduced to as low as 0.09.
[0085] This method employs a federated learning architecture to address the shortcomings of centralized fault diagnosis in transformer fault diagnosis, which cannot collect large amounts of data. It breaks down the data silos in transformers, enabling distributed collaborative training where data remains local and avoiding privacy leaks of local model parameters. Furthermore, by adding Gaussian noise to the local model parameters and implementing an adaptive noise addition mechanism to account for the differences in sample content among participating clients, privacy leaks during the upload of local model parameters are further avoided, effectively protecting data privacy and demonstrating good generalization ability.
[0086] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0087] The embodiments described above are merely specific and detailed examples of the embodiments described in this application, and should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the scope of protection of this application. Therefore, the scope of protection of this patent application should be determined by the appended claims.
Claims
1. A distributed fault diagnosis method for oil-immersed transformers based on data privacy protection, characterized in that: The distributed fault diagnosis method for oil-immersed transformers based on data privacy protection includes the following steps: S1. Obtain dissolved gas data in oil from N transformer stations and store it in the corresponding local database. No data sharing is performed between the local databases. S2. Normalize the dissolved gas data in oil from each local database, and group and label the normalized data based on the modified three-ratio method. S3. Perform bidirectional tiling enhancement on the labeled data and convert it into a two-dimensional image accordingly; S4. Construct a transformer fault diagnosis model based on federated learning. The transformer fault diagnosis model includes a central server and N participant clients. Each participant client corresponds to a transformer site. Each participant client is configured with a local model, which is a convolutional neural network model. The central server uses the FedAvg algorithm to aggregate the parameters of the local model. S5. Input the two-dimensional images of each transformer station into the corresponding local model for training to obtain the local model parameters; S6. Add Gaussian noise to the local model parameters for differential privacy processing; S7. Upload the local model parameters after differential privacy processing to the central server for aggregation, and then send the aggregated local model parameters to the local model. Return to step S5 and continue until the preset number of training iterations are completed to obtain the optimal transformer fault diagnosis model.
2. The distributed fault diagnosis method for oil-immersed transformers based on data privacy protection as described in claim 1, characterized in that: The dissolved gas data in the oil includes five gases: H2, CH4, C2H6, C2H4, and C2H2.
3. The distributed fault diagnosis method for oil-immersed transformers based on data privacy protection as described in claim 2, characterized in that: The dissolved gas data in oil from each local database is normalized using the following formula: in, The total content of the five gases in each data set. This represents the total content of CH4, C2H6, C2H4, and C2H2 gases. and The dissolved gas contents in the oil after normalization treatment are H2, CH4, C2H6, C2H4, and C2H2, in that order.
4. The distributed fault diagnosis method for oil-immersed transformers based on data privacy protection as described in claim 1, characterized in that: The convolutional neural network model is the LeNet-5 network model.
5. The distributed fault diagnosis method for oil-immersed transformers based on data privacy protection as described in claim 1, characterized in that: Before adding Gaussian noise to the local model parameters for differential privacy processing, gradient clipping is also required.
6. The distributed fault diagnosis method for oil-immersed transformers based on data privacy protection as described in claim 5, characterized in that: Gaussian noise satisfies the following formula: in, In the formula, Δf is the sensitivity of differential privacy, δ is the relaxation factor, ε is the privacy budget, and D k Let C be the sample size of participant client k, k = 1, 2, ..., N, and C be the gradient clipping coefficient.
7. The distributed fault diagnosis method for oil-immersed transformers based on data privacy protection as described in claim 6, characterized in that: During training, the update formulas for the parameters of each local model are as follows: in, and F represents the weight parameters for participant k in the t-th and t+1-th iterations, respectively. k (ω k Let be the loss function for participant k. Here, R is the gradient operator, and R is the learning rate; N(0,σ) 2 ) is random noise that follows a Gaussian distribution.
8. The distributed fault diagnosis method for oil-immersed transformers based on data privacy protection as described in claim 7, characterized in that: The local model parameters, after differential privacy processing, are uploaded to the central server for aggregation, as follows: in, Let N be the weight parameter for participant k in the (t+1)th iteration, and N be the number of participant clients.