Federal learning incremental training platform and method for power equipment defect identification

Abstract

The present application relates to the technical field of artificial intelligence, in particular to a federal learning incremental training platform and method for power equipment defect identification, comprising a computing node, a central coordination system, a differential privacy protection module, a sensitive feature shielding module, an aggregated weight adjustment module, and a privacy evaluation and self-correction module, the computing node is responsible for collecting power equipment image data and labels, performing local training and generating model parameters; the central coordination system aggregates node model parameters to update global parameters; the differential privacy protection module dynamically allocates privacy budgets and adds calibration noise, and the sensitive feature shielding module differentially protects image data; the aggregated weight adjustment module dynamically adjusts the weight according to the node data quality, quantity and privacy protection level; the privacy evaluation and self-correction module evaluates the model performance and privacy protection level, and automatically adjusts the parameters to form a closed-loop optimization; the platform enables different power grid companies to safely collaborate in model training and effectively protect sensitive data security.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a federated learning incremental training platform and method for identifying defects in power equipment. It is applied to the intelligent identification and analysis of defects in various equipment such as substation equipment, transmission lines, and distribution networks in power systems. Background Technology

[0002] With the continuous advancement of smart grid construction, the timely detection and handling of power equipment defects are of great significance to ensuring the safe and stable operation of the power grid. Traditional power equipment defect identification mainly relies on manual inspection, which is inefficient and easily affected by subjective factors. In recent years, although deep learning-based defect identification technology has made some progress, it still faces challenges in areas such as data privacy protection, cross-regional collaboration, and model update efficiency.

[0003] Currently, power grid companies possess a large amount of image data on power equipment defects. This data is highly sensitive, and direct sharing poses security risks. Furthermore, power equipment in different regions has different characteristics, and the types and manifestations of defects also vary. How to achieve cross-regional collaborative learning while protecting data privacy has become an urgent problem to be solved. In addition, with the emergence of new defects and the upgrading of equipment, models need to be constantly updated to adapt to new situations. Traditional retraining methods are computationally expensive and inefficient.

[0004] Federated learning, as a distributed machine learning paradigm, allows multiple participants to jointly train a model without sharing the original data, offering a potential solution to the aforementioned problems. However, existing federated learning methods, when applied to power equipment defect identification, still suffer from insufficient privacy protection, difficulty in balancing model performance and privacy protection, and low efficiency in incremental updates, making it difficult to meet practical application needs. Summary of the Invention

[0005] The purpose of this invention is to provide a federated learning incremental training platform and method for power equipment defect identification, aiming to solve the problems of difficulty in balancing data privacy protection and model performance, low efficiency of cross-regional collaboration, and high model update costs in the existing technology.

[0006] This invention proposes a federated learning incremental training platform for power equipment defect identification, comprising:

[0007] A computing node is used to collect image data of power equipment, save the image data and corresponding label data as a training dataset, perform local training on the training dataset, and generate local model parameters.

[0008] A central coordination system, which is communicatively connected to the computing nodes, is used to receive the local model parameters uploaded by the computing nodes, aggregate the local model parameters to update the global model parameters, and distribute the global model parameters to the computing nodes.

[0009] The differential privacy protection module is communicatively connected to the computing node and is used to dynamically allocate a privacy budget based on the type of power equipment and the sensitivity to defects, and to add calibration noise to the local model parameters uploaded by the computing node to reduce the risk of data leakage.

[0010] A sensitive feature masking module, which is communicatively connected to the computing node, is used to identify sensitive features and key identification features in the image data and implement differentiated protection strategies for the sensitive features and key identification features.

[0011] The aggregation weight adjustment module is connected to the central coordination system and is used to dynamically adjust the aggregation weight according to the data quality, data volume and privacy protection level of each computing node to balance privacy protection and model performance.

[0012] The privacy assessment and self-correction module is communicatively connected to the differential privacy protection module, the sensitive feature masking module, and the aggregate weight adjustment module. It is used to evaluate the model performance and privacy protection level, automatically adjust the privacy protection parameters, and form a closed-loop optimization system.

[0013] Preferably, the differential privacy protection module includes:

[0014] The privacy budget allocation unit is used to establish a device type-sensitivity mapping table and allocate basic privacy budget values ​​based on the type of power equipment and the severity of defects.

[0015] The gradient deviation monitoring unit is used to monitor the gradient deviation value in real time during the model aggregation process. When the gradient deviation value exceeds a preset threshold, a privacy protection mechanism is triggered.

[0016] A noise generation unit is used to generate initial noise based on the privacy budget value allocated by the privacy budget allocation unit;

[0017] The noise calibration unit is used to analyze the gradient change trend, calculate the difference between the current gradient and the historical gradient, and dynamically adjust the noise calibration factor according to the degree of difference.

[0018] The perturbation execution unit is used to add calibrated noise to the original gradient to form a perturbation gradient.

[0019] Preferably, the sensitive feature shielding module includes:

[0020] The feature importance analysis unit is used to perform importance analysis on the intermediate layer features of the model and calculate the contribution of each feature to the prediction result.

[0021] The feature sensitivity assessment unit is used to establish feature sensitivity assessment standards, taking into account both the value of feature recognition and the risk of privacy leakage.

[0022] The feature classification unit is used to classify features into three categories: high sensitivity, medium sensitivity, and low sensitivity.

[0023] The differential protection unit is used to apply strong noise protection to high-sensitivity features, moderate noise protection to medium-sensitivity features, and minimal noise protection to low-sensitivity but critical features.

[0024] Preferably, the aggregation weight adjustment module includes:

[0025] The node basic weight initialization unit is used to initialize the basic weights of each node based on the amount of data and the similarity of data distribution.

[0026] The contribution evaluation unit is used to establish a historical contribution evaluation system for nodes and record the effectiveness of each node's model update.

[0027] The data quality analysis unit is used to analyze the relationship between the node's privacy protection level and its data quality.

[0028] The weight calculation unit is used to dynamically calculate new aggregate weight coefficients by combining node contribution and data quality factors.

[0029] The smooth transition unit is used to achieve a smooth transition of weights and prevent sudden changes in weights from causing model instability.

[0030] Preferably, the privacy assessment and self-correction module includes:

[0031] A multi-dimensional evaluation unit is used to establish an evaluation system that includes indicators such as accuracy, privacy protection, and communication efficiency.

[0032] The privacy loss calculation unit is used to design a comprehensive privacy loss evaluation function to balance model performance and the degree of privacy protection.

[0033] The parameter sensitivity analysis unit is used to analyze the impact trends of different privacy parameters on model performance.

[0034] An automatic tuning unit is used to automatically adjust the privacy budget allocation strategy, noise calibration parameters, and feature masking strategy based on the evaluation results;

[0035] The anomaly detection unit is used to detect and handle parameter adjustment anomalies to prevent sudden degradation of system performance.

[0036] Preferably, the computing node includes:

[0037] Image acquisition unit, used to acquire image data of power equipment;

[0038] A labeling unit is used to annotate defects in the image data and generate label data.

[0039] The data preprocessing unit is used to normalize, enhance, and extract features from the image data;

[0040] A local training unit is used to train a local model based on the image data and the label data;

[0041] The parameter uploading unit is used to upload the trained local model parameters to the central coordination system.

[0042] The model update unit is used to receive the global model parameters issued by the central coordination system and update the local model.

[0043] Preferably, the central coordination system includes:

[0044] A parameter receiving unit is used to receive the local model parameters uploaded by each of the computing nodes;

[0045] An aggregation processing unit is used to perform weighted aggregation on the local model parameters according to the aggregation weights provided by the aggregation weight adjustment module, and update the global model parameters.

[0046] The model distribution unit is used to distribute the updated global model parameters to each computing node;

[0047] The node management unit is used to manage the addition, removal, and status monitoring of computing nodes;

[0048] The incremental update unit is used to implement incremental updates of the model, avoiding complete retraining.

[0049] As a preferred option, it also includes:

[0050] The security authentication module is communicatively connected to the computing node and the central coordination system, and is used to authenticate the identity of the computing node to ensure that the nodes participating in the training are secure and trustworthy.

[0051] The security authentication module includes a single-node authentication unit and a multi-node authentication unit. The single-node authentication unit is used for authentication between a single computing node and the central coordination system, and the multi-node authentication unit is used for mutual authentication between multiple computing nodes.

[0052] As a preferred option, it also includes:

[0053] An anomaly detection module, which is communicatively connected to the computing node and the central coordination system, is used to perform anomaly detection on the local model parameters uploaded by the computing node, and to identify and process malicious or abnormal model parameters.

[0054] The anomaly detection module includes a feature extraction unit, a feature concatenation unit, a feature difference calculation unit, an outlier calculation unit, and an outlier determination unit. The feature extraction unit is used to extract features from the local model parameters. The feature concatenation unit is used to concatenate the extracted local features and global features. The feature difference calculation unit is used to calculate the feature differences after concatenation. The outlier calculation unit is used to calculate the outliers of the feature differences. The outlier determination unit is used to determine the outliers.

[0055] A federated learning incremental training method for power equipment defect identification, applied to any of the platforms described herein, includes the following steps:

[0056] S1, the computing node collects image data of the power equipment, annotates the image data, and obtains a training dataset;

[0057] S2, the differential privacy protection module dynamically allocates a privacy budget based on the type of power equipment and defect sensitivity;

[0058] S3, the sensitive feature shielding module identifies sensitive features and key identification features in the image data and implements a differentiated protection strategy;

[0059] S4, the computing node performs local training based on the training dataset to generate local model parameters;

[0060] S5, the differential privacy protection module adds calibration noise to the local model parameters to form perturbed model parameters;

[0061] S6, the computing node uploads the perturbed model parameters to the central coordination system;

[0062] S7, the aggregation weight adjustment module dynamically adjusts the aggregation weight based on the data quality, data volume and privacy protection level of each computing node;

[0063] S8, the central coordination system performs weighted aggregation of the perturbed model parameters uploaded by each computing node according to the aggregation weight, and updates the global model parameters;

[0064] S9, The central coordination system distributes the global model parameters to each computing node;

[0065] S10, the privacy assessment and self-correction module evaluates the model performance and privacy protection level, automatically adjusts privacy protection parameters, and forms a closed-loop optimization system.

[0066] This invention achieves the goal of maintaining high model performance while protecting data privacy by designing differential privacy protection mechanisms, sensitive feature masking strategies, adaptive adjustment of aggregation weights, and self-correction of privacy assessment techniques. The main advantages of this invention are:

[0067] 1. Significantly improved data privacy protection, reduced the risk of information leakage, enabled different power grid companies to safely conduct collaborative model training, and effectively protected the security of sensitive data.

[0068] 2. By employing differentiated feature protection strategies and adaptive weight adjustment, a high defect identification accuracy is maintained while ensuring privacy protection, thus solving the problem of balancing privacy protection and model performance in traditional methods.

[0069] 3. The incremental learning framework significantly improves model update efficiency, reduces computational resource consumption, and supports rapid adaptation to new defects and equipment updates.

[0070] 4. A closed-loop optimization system was designed, which can automatically evaluate model performance and privacy protection level, dynamically adjust relevant parameters, reduce the need for manual intervention, and improve the system's autonomous operation capability.

[0071] 5. A complete cross-regional collaboration architecture has been built, breaking down data silos and improving the intelligent operation and maintenance level of the entire power industry, resulting in significant economic and social benefits. Attached Figure Description

[0072] Figure 1 A schematic diagram of the overall architecture of the federated learning incremental training platform for power equipment defect identification provided in an embodiment of the present invention;

[0073] Figure 2 This is a schematic diagram of the differential privacy protection module provided in an embodiment of the present invention;

[0074] Figure 3 This is a schematic diagram of the sensitive feature shielding module provided in an embodiment of the present invention;

[0075] Figure 4 This is a schematic diagram of the aggregate weight adjustment module provided in an embodiment of the present invention;

[0076] Figure 5 This is a schematic diagram of the privacy assessment and self-correction module provided in an embodiment of the present invention;

[0077] Figure 6 A flowchart of a federated learning incremental training method for power equipment defect identification provided in an embodiment of the present invention. Detailed Implementation

[0078] Please refer to the attached document. Figure 1-6The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention, and should not be used to limit the scope of protection of the present invention.

[0079] like Figure 1 As shown, the present invention provides a federated learning incremental training platform for power equipment defect identification, including: computing node 1, central coordination system 2, differential privacy protection module 3, sensitive feature masking module 4, aggregate weight adjustment module 5, and privacy assessment and self-correction module 6.

[0080] Computation node 1 is used to collect image data of power equipment, save the image data and corresponding label data as a training dataset, and perform local training on the training dataset to generate local model parameters. In one embodiment of the present invention, computing node 1 can be a server device deployed in various provincial power grid companies. Each computing node 1 is responsible for collecting, storing, and processing image data of power equipment defects in its area. Power equipment may include substation equipment, transmission lines, distribution network equipment, etc. Image data is collected by shooting equipment carried by inspection personnel or by fixedly installed monitoring equipment. Each image is accompanied by one or more labels to identify the type and location of defects in the image. Preferably, computing node 1 uses a ResNet-34 model to preprocess and extract features from the collected image data. The ResNet-34 model includes a 49-layer network structure, which can effectively extract feature information of power equipment defects. In addition, computing node 1 also enhances the images through adaptive histogram equalization and weighted randomization to improve image contrast, which is beneficial for the extraction of defect features.

[0081] The central coordination system 2 is communicatively connected to the computing node 1, and is used to receive the local model parameters uploaded by the computing node 1, aggregate the local model parameters to update the global model parameters, and distribute the global model parameters to the computing node 1. The central coordination system 2 acts as the network hub of federated learning, responsible for coordinating model updates and parameter synchronization among the computing nodes 1. In actual deployment, the central coordination system 2 can be located in the State Grid Central Data Center, connected to each computing node 1 through a dedicated power communication network. When a new computing node 1 joins, the central coordination system 2 initializes the new computing node 1 according to the set initialization protocol and evaluates its trained model. If the evaluation result meets the requirements, the computing node 1 is registered to the federated learning network architecture; if the evaluation result does not meet the requirements, it returns to the initialization phase.

[0082] The differential privacy protection module 3 is communicatively connected to the computing node 1 and is used to dynamically allocate a privacy budget based on the type of power equipment and the sensitivity of defects. It adds calibration noise to the local model parameters uploaded by the computing node 1 to reduce the risk of data leakage. The differential privacy protection mechanism is one of the core innovations of this invention. It protects sensitive information in the model parameters by adding carefully calibrated noise while minimizing the impact on model performance. In this invention, the differential privacy protection module 3 dynamically allocates the privacy budget based on the importance of the power equipment and the sensitivity of defects, providing a higher level of protection for important equipment and sensitive defects.

[0083] The sensitive feature masking module 4 is communicatively connected to the computing node 1 and is used to identify sensitive features and key recognition features in the image data, implementing differentiated protection strategies for these features. Sensitive feature masking is a further enhancement of differential privacy protection; by implementing different levels of protection for different features, it ensures both data privacy and security while maintaining the model's recognition performance. In practical applications, some features in power equipment defect images may contain sensitive information, such as equipment model and location information, requiring stronger protection; while other features crucial for defect recognition need to retain their information integrity to ensure the model's recognition accuracy.

[0084] The aggregation weight adjustment module 5 is communicatively connected to the central coordination system 2 and is used to dynamically adjust the aggregation weights based on the data quality, data volume, and privacy protection level of each computing node 1, balancing privacy protection and model performance. In traditional federated learning, the aggregation weights of each node are usually fixed or calculated based on a simple proportion of data volume, which cannot adapt to the differences in data quality and privacy protection needs of different nodes. This invention achieves a better global model by dynamically adjusting the aggregation weights, allocating greater influence to nodes with high data quality, while considering the impact of privacy protection levels.

[0085] The privacy assessment and self-correction module 6 is communicatively connected to the differential privacy protection module 3, the sensitive feature masking module 4, and the aggregate weight adjustment module 5. It is used to evaluate model performance and privacy protection levels, automatically adjust privacy protection parameters, and form a closed-loop optimization system. This module generates a system status report by comprehensively evaluating multiple dimensions of the model, such as accuracy, privacy protection level, and communication efficiency, and automatically adjusts the parameter settings of each module accordingly, achieving system self-optimization. This significantly reduces the need for manual intervention and improves the system's autonomous operation capability and adaptability.

[0086] like Figure 2 As shown, the differential privacy protection module 3 includes: a privacy budget allocation unit 31, a gradient deviation monitoring unit 32, a noise generation unit 33, a noise calibration unit 34, and a disturbance execution unit 35.

[0087] The privacy budget allocation unit 31 is used to establish an equipment type-sensitivity mapping table and allocate a basic privacy budget value according to the type of power equipment and the severity of the defect. In this embodiment, the equipment type-sensitivity mapping table includes the following fields: equipment ID, equipment type, equipment level, basic privacy budget value, and dynamic adjustment coefficient. Power equipment is divided into three levels according to importance: ultra-high voltage equipment (level 1), high voltage equipment (level 2), and medium and low voltage equipment (level 3). The severity of defects is divided into three categories: urgent, important, and general. Through these two dimensions, a nine-square privacy budget matrix is ​​formed, and different privacy budgets are allocated to different equipment and different defect types. Preferably, the basic privacy budget value is set to 0.1 (highest protection level) for ultra-high voltage equipment, 0.3 for high voltage equipment, and 0.5 for medium and low voltage equipment. For defect severity, the adjustment coefficient is 0.8 for urgent defects (enhanced protection), 1.0 for important defects, and 1.2 for general defects.

[0088] The gradient deviation monitoring unit 32 is used to monitor the gradient deviation value during the model aggregation process in real time. When the gradient deviation value exceeds a preset threshold, a privacy protection mechanism is triggered. (Gradient deviation during model aggregation...) () is an important indicator for measuring the magnitude of model updates. When the deviation exceeds a threshold... When this occurs, it indicates a significant model update, which may contain sensitive information, necessitating enhanced privacy protection. In this embodiment, the threshold... The value was set to 0.1, which was determined through a large number of experiments. This value can effectively identify gradient updates that need to be protected while ensuring the model's convergence efficiency.

[0089] The noise generation unit 33 generates initial noise based on the privacy budget value allocated by the privacy budget allocation unit 31. In this embodiment, a Laplace noise mechanism is used to generate the initial noise. The noise magnitude is inversely proportional to the privacy budget; that is, the smaller the privacy budget value, the larger the added noise and the higher the protection strength. The mathematical expression for noise generation is:

[0090] ,

[0091] Where Noise is the added noise value. For privacy budget values, This indicates that the mean is 0 and the scale parameter is... The Laplace distribution of . The probability density function of this distribution is . ,in Let x represent the absolute value of x, and exp represent the natural exponential function. For example, consider the oil temperature anomaly in an ultra-high voltage transformer (a highly sensitive defect). The generated noise will have a larger variance, providing stronger privacy protection; while for minor corrosion of low-voltage power distribution lines (low-sensitivity defects). The noise variance is small, which has little impact on model performance.

[0092] The noise calibration unit 34 is used to analyze gradient change trends, calculate the difference between the current gradient and historical gradients, and dynamically adjust the noise calibration factor based on the degree of difference. Simply adding noise can significantly affect model performance; therefore, noise calibration is necessary. Calibration factor The calculation formula is:

[0093] ,

[0094] in, This is the noise calibration factor, with a value range of [0,1]. and Let represent the model parameters at time t and time t+1, respectively. Represents the absolute value of the change in model parameters. The calibration threshold is set to 0.2, with min and max representing the minimum and maximum values, respectively. This applies when the parameter changes are small. , When the value is close to 1, the noise impact is relatively small; when the parameter changes significantly... , Approaching zero, noise has a significant impact, thus achieving adaptive calibration. This calibration mechanism is particularly important in power equipment defect identification because the characteristics of different types of defects vary greatly. For example, for defects with obvious characteristics such as insulator breakage, the model parameters typically change significantly. The value will automatically decrease to reduce noise interference; and for defects with subtle characteristics such as early insulation degradation, the model parameters change less. The value will remain high, maintaining the noise protection strength.

[0095] The perturbation unit 35 adds calibrated noise to the original gradient to form a perturbed gradient. The formula for calculating the perturbed gradient is:

[0096] .

[0097] in, The gradient after perturbation. Represents the original gradient. Indicates the current moment. express The training dataset at each time step, express Model parameters at time 10:00 for The calibration factor at time, where Noise is the generated noise. This represents a scalar multiplication operation. The perturbed gradient is uploaded to the central coordination system 2 for updating the global model. In this way, the system can maintain the model's learning ability while protecting sensitive information. For example, when identifying contact wear in high-voltage circuit breakers, the original gradient may contain sensitive information such as the circuit breaker model. After adding calibration noise, this sensitive information is effectively masked, while retaining the key feature information of contact wear.

[0098] like Figure 3 As shown, the sensitive feature shielding module 4 includes: a feature importance analysis unit 41, a feature sensitivity evaluation unit 42, a feature classification unit 43, and a differentiation protection unit 44.

[0099] The feature importance analysis unit 41 is used to perform importance analysis on the features of the intermediate layer of the model, calculating the contribution of each feature to the prediction result. In this embodiment, the feature importance scoring method is used to calculate the contribution of each feature, and the calculation formula is as follows:

[0100] ,

[0101] in, Features Importance rating The loss function of the model. The loss function represents the feature The partial derivative (gradient) of . This indicates the absolute value operation. This represents scalar multiplication. This calculation method comprehensively considers both the influence of features on the loss function (represented by the absolute value of the gradient) and the magnitude of the features themselves (represented by the absolute value of the feature), effectively identifying features crucial to the model's prediction. In power equipment defect identification, this method can identify the most critical features for judging different types of defects. For example, for insulator flashover defects, the importance score of surface dirt distribution features is usually high, while the importance score of background information is low.

[0102] The feature sensitivity assessment unit 42 is used to establish feature sensitivity assessment standards, comprehensively considering the value of feature identification and the risk of privacy leakage. Besides the importance of the feature, it is also necessary to assess its sensitivity, i.e., the amount of privacy information that the feature may leak. In power equipment defect identification, some features may contain sensitive information such as equipment model, location information, and operating status. Feature Sensitivity Scoring Based on expert evaluation and historical data analysis, the sensitivity was categorized into three levels: high (1.0), medium (0.6), and low (0.3). For example, features containing substation location information were rated as highly sensitive. Features containing device model information were rated as medium sensitivity. Purely defective texture features were rated as low sensitivity. .

[0103] Feature classification unit 43 is used to classify features into three categories: high sensitivity, medium sensitivity, and low sensitivity. This is based on feature importance. and sensitivity Calculate the comprehensive score of features :

[0104] ,

[0105] in, Features Overall score This is a balancing factor, with a value range of [0,1], used to adjust the weights of importance and sensitivity. The initial value is set to 0.5. This represents scalar multiplication. (Based on comprehensive scoring) Features are categorized into high-sensitivity features. Medium sensitivity and low sensitivity Three categories. This classification method balances the identifying value of features with privacy risks. For example, for a feature representing transmission line tower numbers, although its importance score may not be high. However, it is highly sensitive. Overall score It is classified as a medium sensitivity feature.

[0106] The differential protection unit 44 applies strong noise protection to high-sensitivity features, moderate noise protection to medium-sensitivity features, and minimal noise protection to low-sensitivity but critical features. Specifically, the protection strategy is as follows: high-sensitivity features are protected with noise of high intensity. The noise, with a medium sensitivity feature added at an intensity of The noise, low-sensitivity feature added intensity is The noise, of which This serves as the base privacy budget value. This differentiated protection strategy protects sensitive information while preserving key identifying features, achieving a balance between privacy protection and model performance. For example, when identifying transformer oil level anomalies, the oil level gauge reading feature (high sensitivity) is protected against strong noise, the oil color change feature (medium sensitivity) is protected against medium noise, and the oil level fluctuation pattern feature (low sensitivity but crucial for identification) is protected against minimal noise to preserve its information integrity.

[0107] like Figure 4As shown, the aggregation weight adjustment module 5 includes: a node basic weight initialization unit 51, a contribution evaluation unit 52, a data quality analysis unit 53, a weight calculation unit 54, and a smooth transition unit 55.

[0108] The node basic weight initialization unit 51 is used to initialize the basic weights of each node based on the amount of data and the similarity of data distribution. In the initial stage of federated learning, the basic weights of each computing node 1 are mainly determined based on its amount of data, and the calculation formula is as follows:

[0109] ,

[0110] in, The basic weight of node i, Let N be the amount of data (number of samples) for node i, and N be the total number of nodes participating in federated learning. This represents the summation operation over all nodes (from j=1 to j=N). Furthermore, the similarity of data distributions is considered; nodes with high similarity between their data distribution and the global distribution are given appropriate weights, with similarity calculated using KL divergence. In power equipment defect identification scenarios, data distribution similarity is particularly important because equipment types and defect characteristics may differ across regions. For example, in rainy southern regions, insulator flashover and corrosion problems are more prominent; while in cold northern regions, freezing damage and metal embrittlement are more common. By appropriately adjusting the basic weights, the system can better balance the contributions of different regions.

[0111] The contribution evaluation unit 52 is used to establish a historical contribution evaluation system for nodes, recording the effectiveness of each node's model updates. A node's contribution is evaluated based on the impact of its historical model updates on the overall model performance improvement; the calculation formula is as follows:

[0112] ,

[0113] in, Let i be the contribution of node i at time t. For node i in Contribution at any moment This is a smoothing factor, with a value range of [0,1], and is set to 0.8. Let $t$ be the amount by which the model update of node $i$ improves the performance of the global model. This represents the total performance improvement of all nodes at time t. This represents scalar multiplication. This evaluation mechanism identifies nodes that continuously provide valuable updates and assigns them higher weights. In practical applications, some power grid companies may have richer and more diverse defect samples, or higher-quality annotations, and their model updates contribute more to the global model. For example, nodes with a large number of UHV equipment samples may contribute more to identifying UHV equipment defects, and the system will correspondingly increase their weight in such tasks.

[0114] The data quality analysis unit 53 is used to analyze the relationship between the node's privacy protection level and its data quality. Since privacy protection affects the quality of the model parameters provided by the node, it is necessary to analyze the relationship between the two and make adjustments. In this embodiment, data quality is evaluated through model prediction consistency, that is, calculating the consistency between the prediction results of the node's local model on the global validation set and the global model. Simultaneously, the impact of the node's privacy protection level on data quality is considered; generally, the higher the privacy protection level, the greater the impact on data quality. Data quality scoring. The calculation takes into account the combined effects of prediction consistency and privacy protection levels. In power equipment defect identification, high privacy protection may obscure the characteristics of some subtle defects, such as the texture changes of early degradation of insulation materials, and therefore needs to be considered in the weight calculation.

[0115] The weight calculation unit 54 is used to dynamically calculate new aggregate weight coefficients by combining node contribution and data quality factors. The formula for calculating the aggregate weight is:

[0116] ,

[0117] in, Let i be the aggregate weight of node i at time t. Based on the weights, As for contribution level, Score the data quality. This represents the summation operation over all nodes (from j=1 to j=N). This represents scalar multiplication. In this way, the system can assign higher weights to nodes that provide high-quality updates, while also considering the impact of privacy protection. For example, a node may have a large amount of data, but due to strong privacy protection, its data quality score is low; the system will appropriately reduce its weight to balance model performance and privacy protection.

[0118] The smooth transition unit 55 is used to achieve a smooth weight transition, preventing abrupt weight changes that could lead to model instability. To avoid drastic weight fluctuations causing model training instability, a smooth transition mechanism is introduced:

[0119] ,

[0120] in, The smoothed weights, This is the smoothing coefficient, with a value range of [0,1], and is set to 0.6. For node i in Weight of time, Let be the original calculated weight of node i at time t. This represents scalar multiplication. Furthermore, upper and lower weight constraints are set to ensure that the weight of each node is not too large or too small, preventing certain nodes from being completely excluded or dominating the entire model. Specifically, the lower weight is set to 0.01 × (1 / N), and the upper weight is set to 5 × (1 / N), where N is the total number of nodes. This smoothing mechanism is particularly important for the stability of the power equipment defect identification model, as defect identification typically requires stable and reliable model performance.

[0121] like Figure 5 As shown, the privacy assessment and self-correction module 6 includes: a multi-dimensional assessment unit 61, a privacy loss calculation unit 62, a parameter sensitivity analysis unit 63, an automatic optimization unit 64, and an anomaly detection unit 65.

[0122] The multi-dimensional evaluation unit 61 is used to establish an evaluation system that includes indicators such as accuracy, privacy protection level, and communication efficiency. In this embodiment, the evaluation indicators include: model accuracy (Acc), privacy protection level (Privacy), communication efficiency (Comm), and convergence speed (Conv). The weight of each indicator can be adjusted according to actual needs, and the default settings are: wAcc=0.4, wPrivacy=0.3, wComm=0.2, wConv=0.1. Model accuracy is measured by the recognition accuracy on the validation set; privacy protection level is measured by the reciprocal of the privacy budget value. The higher the value, the stronger the protection; communication efficiency is measured by the amount of communication in each training round; convergence speed is evaluated by the number of training rounds required to reach the target performance. A comprehensive evaluation of these indicators can fully reflect the system's operating status. In power equipment defect identification, different types of defects may have different requirements for these indicators. For example, for serious defects that endanger safety (such as abnormal transformer oil temperature), the accuracy weight may be increased to 0.6; while for defect data containing sensitive information (such as transmission line defects containing geographical location information), the privacy protection weight may be increased to 0.5.

[0123] The privacy loss calculation unit 62 is used to design a comprehensive privacy loss evaluation function to balance model performance and the degree of privacy protection. The formula for calculating the privacy loss function is as follows:

[0124] ,

[0125] in, Privacy loss represents a combined loss of privacy protection and model performance. For model performance loss, standard loss functions such as cross-entropy or mean squared error are typically used. Losses due to privacy protection Privacy budget value, and These are the weighting coefficients, initially set. , , This indicates scalar multiplication. It can be defined as:

[0126] ,

[0127] in, Here, log represents the privacy budget value, and log is the natural logarithm function. Indicates the level of privacy protection. This represents scalar multiplication. This function design ensures privacy budgeting. The smaller the size, the higher the level of protection, but the greater the potential loss, thus creating a balance. This balance is particularly important in the identification of defects in power equipment, because excessive privacy protection may prevent the accurate identification of certain subtle defects, such as tiny cracks in insulation materials or early corrosion on metal surfaces.

[0128] The parameter sensitivity analysis unit 63 is used to analyze the impact trends of different privacy parameters on model performance. The privacy budget value is adjusted using the controlled variable method. Noise calibration threshold The system analyzes parameters such as feature classification thresholds to observe their impact on model performance and establishes parameter sensitivity maps. This analysis helps understand the importance and adjustment range of each parameter, providing a basis for automatic tuning. In practical applications, the system periodically performs parameter sensitivity analysis, for example, analyzing the impact of current parameter settings on the defect identification performance of various types of power equipment after every 10 rounds of training and recording sensitivity reports. These reports show the impact curves of parameter changes on the accuracy of different types of defect identification, such as privacy budget. The changing trend of transformer oil temperature anomaly identification accuracy as the value increases from 0.1 to 0.5.

[0129] The automatic tuning unit 64 is used to automatically adjust the privacy budget allocation strategy, noise calibration parameters, and feature masking strategy based on the evaluation results. Automatic tuning employs a gradient descent approach, adjusting the gradient direction of each parameter according to the privacy loss function. For example, the adjustment:

[0130] ,

[0131] in, The adjusted privacy budget value, This is the privacy budget value before adjustment. This is the learning rate, typically ranging from [0.01, 0.1]. Let's set it to 0.05. Represents the privacy loss function relative to the privacy budget. The partial derivatives, This indicates the sign function, which returns -1 (negative gradient), 0, or 1 (positive gradient). This represents scalar multiplication. Similarly, other parameters are adjusted according to their respective gradient directions. In power equipment defect identification, this automatic tuning mechanism can adapt to the characteristics of different types of defects. For example, the system may automatically detect that for metal surface corrosion defects, a lower privacy budget (stronger protection) will significantly affect the identification accuracy, so it will appropriately increase the privacy budget value for this type of defect; while for sensitive information such as equipment serial numbers, even high-strength protection will not affect the identification of the main defect features, so a lower privacy budget value is maintained.

[0132] An anomaly detection unit 65 is used to detect and handle parameter tuning anomalies to prevent sudden performance degradation. During automatic parameter tuning, anomalies may occur that cause a sharp drop in model performance. The anomaly detection unit monitors the changing trends of key performance indicators. When an anomaly is detected (e.g., an accuracy drop exceeding 5%), it immediately rolls back to the previous stable state and reduces the learning rate for readjustment. Furthermore, safety boundaries for parameter tuning are set to prevent parameters from exceeding reasonable ranges. Specifically, privacy budgeting... The safety range is set to [0.05, 1.0], and the noise calibration threshold is... The safety range is set to [0.05, 0.5], and the safety range for the feature classification threshold is set to [0.3, 0.9]. These safety mechanisms ensure that the system does not deviate too far during automatic tuning, maintaining overall stability. In actual operation, the system saves the parameter configurations of the five most recent stable states so that it can quickly roll back in case of anomalies.

[0133] This invention also provides a federated learning incremental training method for power equipment defect identification, applied to the aforementioned platform, comprising the following steps:

[0134] S1, Computing node 1 collects image data of power equipment and annotates the image data to obtain a training dataset. In practical applications, computing nodes 1 deployed by provincial power grid companies collect images of power equipment through inspection personnel or fixed monitoring equipment. Professionals annotate the defects in the images to generate a training dataset. Preferably, the training dataset contains at least 1 million annotated images, covering various equipment types and defect categories. For example, the computing nodes of the East China Power Grid Company may contain 500,000 images, including defect images of substation equipment (transformers, circuit breakers, disconnect switches, etc.), transmission lines (towers, conductors, insulators, etc.), and distribution equipment, covering various defect types such as insulation damage, metal corrosion, and mechanical deformation.

[0135] S2, the differential privacy protection module 3 dynamically allocates privacy budgets based on the type of power equipment and the sensitivity of defects. The system first identifies the equipment type and defect category in the image, and then assigns a corresponding privacy budget value to each image or feature according to a pre-established equipment type-sensitivity mapping table. For example, abnormal oil levels in UHV transformers are classified as high sensitivity and assigned a privacy budget value of 0.1; while slight corrosion in low-voltage distribution lines is classified as low sensitivity and assigned a privacy budget value of 0.5. Through this differentiated privacy budget allocation, the system can provide different levels of protection for data of varying importance, ensuring the security of critical and sensitive data while avoiding unnecessary impacts on model performance from over-protection.

[0136] S3, the sensitive feature masking module 4 identifies sensitive and key identification features in the image data and implements a differentiated protection strategy. For each training sample, the system analyzes the importance and sensitivity of the features in the intermediate layer of the model, classifies the features, and applies protection of different intensities. For example, for highly sensitive features containing device serial numbers, location information, etc., a protection intensity of [insert intensity here] is added. Noise; for moderately sensitive characteristics such as equipment material and environmental conditions, an intensity of [value missing] is added. Noise; for features that are low-sensitivity but crucial for identification, such as defect textures and color changes, an intensity of [value missing] is added. Noise. This differentiated protection strategy demonstrates significant advantages in practical applications, especially for complex types of power equipment defects. For example, when identifying flashover of transmission line insulators, the system can retain the key identifying feature of the dirt distribution pattern while effectively shielding sensitive information such as insulator type and geographical location.

[0137] S4, Computation node 1 performs local training based on the training dataset to generate local model parameters. Each computation node 1 trains the ResNet-34 model on the local training dataset, optimizing the model parameters through forward and backward propagation to obtain the local model. The training process uses a batch size of 64, a learning rate of 0.001, and 50 training epochs. These parameters are optimal configurations verified through extensive experiments, ensuring model convergence while avoiding overfitting. In actual deployment, power grid companies of different sizes may adjust these parameters based on their computing resources. For example, large power grid companies with high-performance GPU clusters may increase the batch size to 128, while small power grid companies with limited computing resources may decrease the batch size to 32.

[0138] S5, the differential privacy protection module 3 adds calibration noise to the local model parameters, forming perturbed model parameters. After training, computing node 1 passes the local model parameters to the differential privacy protection module 3, which adds calibration noise according to the allocated privacy budget. The noise addition follows the aforementioned Laplacian mechanism and is appropriately calibrated by the noise calibration unit 34 to ensure that the impact on model performance is minimized while protecting privacy. In practical applications, different protection strategies may be adopted for different parameter layers. For example, stronger noise protection may be applied to fully connected layer parameters that contain more sensitive information; while weaker noise protection may be applied to convolutional layer parameters that mainly extract texture features, in order to maintain the model's ability to recognize defective textures.

[0139] S6, compute node 1 uploads the perturbed model parameters to the central coordination system 2. The privacy-protected model parameters are uploaded to the central coordination system 2 through a secure communication channel. In practical applications, the transmission employs an encryption protocol to ensure security during the transmission process. Specifically, the TLS 1.3 encryption protocol can be used, encrypted with a 2048-bit RSA key to ensure that the model parameters are not stolen or tampered with during transmission. Furthermore, the system also performs integrity verification on the uploaded parameters to ensure that the data is not corrupted during transmission.

[0140] S7, the aggregation weight adjustment module 5 dynamically adjusts the aggregation weights based on the data quality, data volume, and privacy protection level of each computing node 1. After the central coordination system 2 receives the model parameters uploaded by each node, the aggregation weight adjustment module 5 calculates the weight of each node. The calculation process considers factors such as the node's data volume, historical contribution, and data quality, assigning higher weights to nodes that provide high-quality updates. This dynamic weight adjustment is particularly important in power equipment defect identification because the distribution of power equipment types and defect types may vary significantly across different regions. For example, some regions may have a large number of defect samples from ultra-high voltage equipment, and should receive higher weights in defect identification tasks for this type of equipment; while other regions may have richer samples of transmission line defects, and should receive higher weights in corresponding tasks.

[0141] S8, the central coordination system 2 performs weighted aggregation of the perturbed model parameters uploaded by each computing node 1 according to the aggregation weight, and updates the global model parameters. The aggregation process uses a weighted average method, and the calculation formula is as follows:

[0142] ,

[0143] in, These are global model parameters, which are parameter vectors or matrices with the same dimensions as the model structure. Let i be the aggregation weight of node i, and its value is usually in the range [0,1] and satisfies the following conditions: , The model parameters uploaded for node i are parameter vectors or matrices of the same dimensions as the model structure, where N is the total number of nodes participating in the aggregation. This represents the summation operation over all nodes (from i=1 to i=N). This represents a multiplication operation between corresponding elements. Through this weighted aggregation method, the system can better utilize the contributions of each node to form a superior global model. In practical applications, this aggregation method can apply different weights to parameters of different layers. For example, for convolutional layer parameters used for low-level feature extraction, it may favor nodes with high data diversity; while for fully connected layer parameters used for high-level semantic understanding, it may favor nodes with high-quality annotations.

[0144] S9, the central coordination system 2 distributes the global model parameters to each computing node 1. The updated global model parameters are then distributed back to each computing node 1 through a secure channel for the next round of local training. This step completes a full iteration cycle of federated learning. In actual deployment, the distribution of the global model also employs encrypted transmission to ensure security. Furthermore, the system optimizes the distribution strategy based on the network bandwidth and computing resources of each node; for example, for nodes with limited computing resources, only a portion of the updated parameters may be distributed to reduce their computational burden.

[0145] S10, the privacy assessment and self-correction module 6, evaluates model performance and privacy protection levels, automatically adjusts privacy protection parameters, and forms a closed-loop optimization system. At the end of each iteration cycle, the system evaluates the current model's performance and privacy protection level, calculates the privacy loss function value, and adjusts the privacy budget, noise calibration parameters, etc., accordingly. This automatic adjustment mechanism allows the system to continuously optimize during operation, finding the optimal balance between privacy protection and model performance. For example, if the system finds that the current privacy protection parameters significantly reduce the accuracy of identifying a certain type of critical defect (such as abnormal transformer oil temperature), it will automatically adjust the privacy budget for that type of defect, appropriately reducing the protection strength and improving the identification accuracy; while for less critical defect types, it may maintain a higher protection strength to ensure data privacy and security.

[0146] The steps S1 to S10 described above constitute a complete federated learning iterative cycle. As training progresses, the performance of the global model continuously improves while maintaining a high level of privacy protection. Once the system converges or reaches the preset number of training rounds, the final global model will be used for practical power equipment defect identification tasks.

[0147] This invention utilizes innovative technologies such as differential privacy protection, sensitive feature masking, aggregated weight adjustment, and automatic parameter optimization to achieve the goal of cross-regional collaborative training of a high-performance power equipment defect identification model while protecting data privacy. Experimental results show that, compared with traditional methods, this invention improves defect identification accuracy by 11.7%, reduces false alarm rate by 28.5%, and increases training speed by 42% while maintaining the same level of privacy protection, significantly improving the efficiency and accuracy of power equipment defect identification.

[0148] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of protection of the present invention.

Claims

1. A federated learning incremental training platform for power equipment defect identification, characterized in that... ,include: A computing node is used to collect image data of power equipment, save the image data and corresponding label data as a training dataset, perform local training on the training dataset, and generate local model parameters. A central coordination system, communicatively connected to the computing nodes, is used to receive the local model parameters uploaded by the computing nodes, aggregate the local model parameters to update the global model parameters, and distribute the global model parameters to the computing nodes; A differential privacy protection module, communicatively connected to the computing node, is used to dynamically allocate a privacy budget based on the power equipment type and defect sensitivity, and add calibration noise to the local model parameters uploaded by the computing node to reduce the risk of data leakage. The differential privacy protection module includes: a privacy budget allocation unit, used to establish a device type-sensitivity mapping table and allocate a basic privacy budget value based on the power equipment type and defect severity; a gradient deviation monitoring unit, used to monitor the gradient deviation value during model aggregation in real time, and trigger a privacy protection mechanism when the gradient deviation value exceeds a preset threshold; a noise generation unit, used to generate initial noise based on the privacy budget value allocated by the privacy budget allocation unit; a noise calibration unit, used to analyze gradient change trends, calculate the difference between the current gradient and historical gradients, and dynamically adjust the noise calibration factor based on the degree of difference; and a perturbation execution unit, used to add the calibrated noise to the original gradient to form a perturbation gradient. A sensitive feature masking module, communicatively connected to the computing node, is used to identify sensitive features and key recognition features in the image data and implement differentiated protection strategies for these features. The sensitive feature masking module includes: a feature importance analysis unit, used to perform importance analysis on intermediate layer features of the model and calculate the contribution of each feature to the prediction result; a feature sensitivity evaluation unit, used to establish feature sensitivity evaluation criteria, comprehensively considering feature recognition value and privacy leakage risk; a feature classification unit, used to classify features into three categories: high sensitivity, medium sensitivity, and low sensitivity; and a differentiated protection unit, used to apply strong noise protection to high-sensitivity features, moderate noise protection to medium-sensitivity features, and minimum noise protection to low-sensitivity but crucial recognition features. The aggregation weight adjustment module, which communicates with the central coordination system, is used to dynamically adjust the aggregation weights based on the data quality, data volume, and privacy protection level of each computing node, balancing privacy protection and model performance. The aggregation weight adjustment module includes: a node basic weight initialization unit, used to initialize the basic weights of each node based on data volume and data distribution similarity; a contribution evaluation unit, used to establish a historical contribution evaluation system for nodes and record the effectiveness of model updates for each node; a data quality analysis unit, used to analyze the relationship between the node's privacy protection level and its data quality; a weight calculation unit, used to dynamically calculate new aggregation weight coefficients by comprehensively considering node contribution and data quality factors; and a smooth transition unit, used to achieve a smooth weight transition and prevent sudden weight changes that could lead to model instability. The privacy assessment and self-correction module is communicatively connected to the differential privacy protection module, the sensitive feature masking module, and the aggregate weight adjustment module. It is used to evaluate the model performance and privacy protection level, automatically adjust the privacy protection parameters, and form a closed-loop optimization system.

2. The federated learning incremental training platform for power equipment defect identification according to claim 1, characterized in that, The privacy assessment and self-correction module includes: a multi-dimensional assessment unit for establishing an assessment system that includes accuracy, privacy protection level, and communication efficiency indicators; a privacy loss calculation unit for designing a comprehensive privacy loss assessment function to balance model performance and privacy protection level; a parameter sensitivity analysis unit for analyzing the impact trend of different privacy parameters on model performance; an automatic tuning unit for automatically adjusting the privacy budget allocation strategy, noise calibration parameters, and feature masking strategy based on the assessment results; and an anomaly detection unit for detecting and handling parameter adjustment anomalies to prevent sudden degradation of system performance.

3. The federated learning incremental training platform for power equipment defect identification according to claim 1, characterized in that... The computing nodes include: The image acquisition unit is used to acquire image data of power equipment. The labeling unit is used to annotate defects in the image data and generate label data; The data preprocessing unit is used to normalize, enhance, and extract features from the image data; A local training unit is used to train a local model based on the image data and the label data; The parameter uploading unit is used to upload the trained local model parameters to the central coordination system. The model update unit is used to receive the global model parameters issued by the central coordination system and update the local model.

4. The federated learning incremental training platform for power equipment defect identification according to claim 1, characterized in that... The central coordination system includes: A parameter receiving unit is used to receive the local model parameters uploaded by each of the computing nodes; An aggregation processing unit is used to perform weighted aggregation on the local model parameters according to the aggregation weights provided by the aggregation weight adjustment module, and update the global model parameters. The model distribution unit is used to distribute the updated global model parameters to each of the computing nodes; The node management unit is used to manage the addition, removal, and status monitoring of computing nodes. The incremental update unit is used to implement incremental updates of the model, avoiding complete retraining.

5. The federated learning incremental training platform for power equipment defect identification according to claim 1, characterized in that... It also includes: The security authentication module is communicatively connected to the computing nodes and the central coordination system, and is used to authenticate the identity of the computing nodes to ensure that the nodes participating in the training are secure and trustworthy. The security authentication module includes a single-node authentication unit and a multi-node authentication unit. The single-node authentication unit is used for authentication between a single computing node and the central coordination system, and the multi-node authentication unit is used for mutual authentication between multiple computing nodes.

6. The federated learning incremental training platform for power equipment defect identification according to claim 1, characterized in that... It also includes: An anomaly detection module, communicatively connected to the computing node and the central coordination system, is used to perform anomaly detection on the local model parameters uploaded by the computing node, and to identify and process malicious or abnormal model parameters. The anomaly detection module includes a feature extraction unit, a feature splicing unit, a feature difference calculation unit, an outlier calculation unit, and an outlier determination unit. The feature extraction unit is used to extract features from the local model parameters. The feature splicing unit is used to splice the extracted local features and global features. The feature difference calculation unit is used to calculate the feature differences after splicing. The outlier calculation unit is used to calculate the outlier values ​​of the feature differences. The outlier determination unit is used to determine the outlier values.

7. A federated learning incremental training method for power equipment defect identification, applied to the platform described in any one of claims 1 to 6, characterized in that... This includes the following steps: S1, the computing node collects image data of the power equipment, annotates the image data, and obtains a training dataset; S2, the differential privacy protection module dynamically allocates a privacy budget based on the type of power equipment and defect sensitivity; S3, the sensitive feature shielding module identifies sensitive features and key identification features in the image data and implements a differentiated protection strategy; S4, the computing node performs local training based on the training dataset to generate local model parameters; S5, the differential privacy protection module adds calibration noise to the local model parameters to form perturbed model parameters; S6, the computing nodes upload the perturbed model parameters to the central coordination system; S7, the aggregation weight adjustment module dynamically adjusts the aggregation weight based on the data quality, data volume, and privacy protection level of each computing node; S8, the central coordination system performs weighted aggregation of the perturbed model parameters uploaded by each computing node according to the aggregation weight, and updates the global model parameters; S9, The central coordination system distributes the global model parameters to each computing node; S10, the privacy assessment and self-correction module evaluates the model performance and privacy protection level, automatically adjusts privacy protection parameters, and forms a closed-loop optimization system.

Images