CN-121074607-B - Federal learning incremental training platform and method for identifying defects of power equipment

CN121074607BCN 121074607 BCN121074607 BCN 121074607BCN-121074607-B

Abstract

The invention relates to the technical field of artificial intelligence, in particular to a federal learning incremental training platform and a federal learning incremental training method for identifying defects of power equipment, wherein the federal learning incremental training platform comprises a computing node, a central coordination system, a differential privacy protection module, a sensitive feature shielding module, an aggregation weight adjustment module and a privacy evaluation and self-correction module, wherein the computing node is responsible for collecting image data and labels of the power equipment, carrying out local training and generating model parameters; the central coordination system aggregates the model parameters of each node to update global parameters, the differential privacy protection module dynamically allocates privacy budget and adds calibration noise, the sensitive feature shielding module carries out differential protection on image data, the aggregation weight adjustment module dynamically adjusts weights according to the quality, the quantity and the privacy protection level of the node data, the privacy evaluation and self-correction module evaluates the performance and the privacy protection level of the model and automatically adjusts the parameters to form closed loop optimization, and the platform enables different power grid companies to safely cooperate with training models and effectively protects the sensitive data.

Inventors

LIANG ZHEHENG
QIAN ZHENGHAO
WANG YECHAO
ZHANG XIAOLU
ZHOU CHUN
ZHOU FANGFANG
LI KAI
QIN QIANG
LONG ZHENYUE
SHEN WUQIANG
YAO CHAOSHENG
LU CHANGCAI
ZHANG ZIYANG
ZHANG JINBO
CUI LEI
ZENG JIJUN
SHEN GUIQUAN

Assignees

广东电网有限责任公司

Dates

Publication Date: 20260508
Application Date: 20251105

Claims (7)

1. A federal learning incremental training platform for power equipment defect identification, comprising: the computing node is used for collecting image data of the power equipment, storing the image data and corresponding tag data as a training data set, and carrying out local training on the training data set to generate local model parameters; The central coordination system is in communication connection with the computing nodes and is used for receiving the local model parameters uploaded by the computing nodes, aggregating the local model parameters to update global model parameters and transmitting the global model parameters to the computing nodes; The differential privacy protection module is in communication connection with the computing node and is used for dynamically distributing privacy budgets according to the type of the power equipment and the defect sensitivity, adding calibration noise to the local model parameters uploaded by the computing node and reducing the risk of data leakage, and comprises a privacy budgeting distribution unit, a gradient deviation monitoring unit, a noise generation unit, a noise calibration unit, a disturbance execution unit and a disturbance execution unit, wherein the privacy budgeting distribution unit is used for establishing an equipment type-sensitivity mapping table, distributing basic privacy budgets according to the type of the power equipment and the defect severity, the gradient deviation monitoring unit is used for monitoring gradient deviation values in the model aggregation process in real time, triggering a privacy protection mechanism when the gradient deviation values exceed a preset threshold, the noise generation unit is used for generating initial noise according to the privacy budgeting values distributed by the privacy budgeting distribution unit, the noise calibration unit is used for analyzing gradient change trend, calculating the difference between the current gradient and the historical gradient, and dynamically adjusting noise calibration factors according to the difference degree, and the disturbance execution unit is used for adding the calibrated noise to the original gradient to form disturbance gradient; The sensitive feature shielding module is in communication connection with the computing node and is used for identifying sensitive features and key identification features in the image data and implementing a differential protection strategy on the sensitive features and the key identification features, and comprises a feature importance analysis unit, a feature sensitivity evaluation unit, a feature classification unit, a differential protection unit and a differential protection unit, wherein the feature importance analysis unit is used for carrying out importance analysis on model middle layer features and calculating the contribution degree of each feature to a prediction result; The aggregation weight adjustment module is in communication connection with the central coordination system and is used for dynamically adjusting the aggregation weight according to the data quality, the data quantity and the privacy protection level of each computing node and balancing the privacy protection and model performance; the aggregation weight adjustment module comprises a node basic weight initialization unit, a contribution degree evaluation unit, a data quality analysis unit, a weight calculation unit, a smooth transition unit and a model analysis unit, wherein the node basic weight initialization unit is used for initializing each node basic weight based on data quantity and data distribution similarity; The privacy evaluation and self-correction module is in communication connection with the differential privacy protection module, the sensitive characteristic shielding module and the aggregation weight adjustment module and is used for evaluating the performance and the privacy protection level of the model and automatically adjusting the privacy protection parameters to form a closed-loop optimization system.
2. The federal learning incremental training platform for identifying defects of power equipment according to claim 1, wherein the privacy assessment and self-correction module comprises a multi-dimensional assessment unit, a privacy loss calculation unit, a parameter sensitivity analysis unit, an automatic optimization unit and an abnormality detection unit, wherein the multi-dimensional assessment unit is used for establishing an assessment system comprising accuracy, privacy protection degree and communication efficiency indexes, the privacy loss calculation unit is used for designing a privacy loss comprehensive assessment function and balancing model performance and privacy protection degree, the parameter sensitivity analysis unit is used for analyzing influence trend of different privacy parameters on model performance, the automatic optimization unit is used for automatically adjusting privacy budget allocation strategies, noise calibration parameters and feature shielding strategies based on assessment results, and the abnormality detection unit is used for finding and processing parameter adjustment abnormality and preventing sudden degradation of system performance.
3. The federal learning incremental training platform for power plant defect identification of claim 1, wherein the computing node comprises: the image acquisition unit is used for acquiring image data of the power equipment; the label marking unit is used for marking the defects of the image data and generating label data; the data preprocessing unit is used for normalizing, enhancing and extracting features of the image data; a local training unit for training a local model based on the image data and the tag data; The parameter uploading unit is used for uploading the local model parameters obtained through training to the central coordination system; and the model updating unit is used for receiving the global model parameters issued by the central coordination system and updating the local model.
4. The federal learning incremental training platform for power plant defect identification of claim 1, wherein the central coordination system comprises: the parameter receiving unit is used for receiving the local model parameters uploaded by each computing node; The aggregation processing unit is used for carrying out weighted aggregation on the local model parameters according to the aggregation weight provided by the aggregation weight adjustment module and updating the global model parameters; The model distribution unit is used for distributing the updated global model parameters to each computing node; The node management unit is used for managing the joining, exiting and state monitoring of the computing nodes; And the incremental updating unit is used for realizing incremental updating of the model and avoiding complete retraining.
5. The federal learning incremental training platform for power plant defect identification of claim 1, further comprising: The security authentication module is in communication connection with the computing node and the central coordination system and is used for carrying out identity authentication on the computing node and ensuring the security and credibility of the nodes participating in training, and the security authentication module comprises a single-node authentication unit and a multi-node authentication unit, wherein 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 a plurality of computing nodes.
6. The federal learning incremental training platform for power plant defect identification of claim 1, further comprising: The anomaly detection module is in communication connection with the computing node and the central coordination system and is used for carrying out anomaly detection on the local model parameters uploaded by the computing node, identifying and processing malicious or anomalous model parameters, and comprises a feature extraction unit, a feature splicing unit, a feature difference calculation unit, an outlier calculation unit and an outlier judgment unit, wherein the feature extraction unit is used for carrying out feature extraction on the local model parameters, the feature splicing unit is used for splicing the extracted local features and global features, the feature difference calculation unit is used for calculating the characteristic differences after the splicing, the outlier calculation unit is used for calculating the outlier of the feature differences, and the outlier judgment unit is used for judging the outlier.
7. A federal learning incremental training method for power plant defect identification, applied to the platform of any one of claims 1 to 6, comprising the steps of: s1, acquiring image data of power equipment by a computing node, and marking the image data to obtain a training data set; S2, the differential privacy protection module dynamically allocates privacy budgets according to the type of the power equipment and the defect sensitivity; S3, a sensitive characteristic shielding module identifies sensitive characteristics and key identification characteristics in the image data, and a differential protection strategy is implemented; S4, the computing node performs local training based on the training data set to generate local model parameters; S5, adding calibration noise to the local model parameters by the differential privacy protection module to form disturbed model parameters; s6, the calculation node uploads the disturbed model parameters to a central coordination system; s7, an aggregation weight adjusting module dynamically adjusts the aggregation weight according to the data quality, the data quantity and the privacy protection level of each computing node; S8, the central coordination system carries out weighted aggregation on the disturbed model parameters uploaded by each computing node according to the aggregation weight, and updates global model parameters; s9, the central coordination system distributes the global model parameters to each computing node; s10, the privacy evaluation and self-correction module evaluates the performance and the privacy protection level of the model, automatically adjusts privacy protection parameters, and forms a closed loop optimization system.

Description

Federal learning incremental training platform and method for identifying defects of power equipment Technical Field The invention relates to the technical field of artificial intelligence, in particular to a federal learning incremental training platform and a federal learning incremental training method for identifying defects of power equipment, which are applied to intelligent identification and analysis of defects of various equipment such as transformer substation equipment, transmission lines, distribution networks and the like in a power system. Background Along with the continuous promotion of smart power grid construction, timely discovery and treatment of defects of power equipment have important significance for ensuring safe and stable operation of the power grid. Traditional power equipment defect identification mainly relies on manual inspection, and is low in efficiency and susceptible to subjective factors. In recent years, although a defect recognition technology based on deep learning has advanced to some extent, challenges in terms of data privacy protection, cross-region collaboration, model update efficiency and the like are still faced. At present, each power grid company has a large amount of power equipment defect image data, the data has higher sensitivity, and safety risks exist in direct sharing. Meanwhile, the power equipment in different areas has different characteristics, the defect types and the expression forms are also different, and how to realize cross-area collaborative learning on the premise of protecting the data privacy becomes a problem to be solved urgently. In addition, with the appearance of new defects and the update iteration of equipment, the model needs to be updated continuously to adapt to new conditions, and the traditional retraining mode is high in calculation resource consumption and low in efficiency. Federal learning, as a distributed machine learning paradigm, allows multiple parties to co-train a model without sharing raw data, providing the potential to solve the above-described problems. However, when the existing federal learning method is applied to defect identification of electric equipment, the problems of insufficient privacy protection, difficulty in balancing model performance and privacy protection, low incremental updating efficiency and the like still exist, and the actual application requirements are difficult to meet. Disclosure of Invention The invention aims to provide a federal learning incremental training platform and a federal learning incremental training method for identifying defects of power equipment, and aims to solve the problems that in the prior art, data privacy protection and model performance are difficult to balance, inter-regional collaboration efficiency is low, model updating cost is high and the like. The invention provides a federal learning incremental training platform for identifying defects of power equipment, which comprises the following steps: the computing node is used for collecting image data of the power equipment, storing the image data and corresponding tag data as a training data set, and carrying out local training on the training data set to generate local model parameters; The central coordination system is in communication connection with the computing nodes and is used for receiving the local model parameters uploaded by the computing nodes, aggregating the local model parameters to update global model parameters and transmitting the global model parameters to the computing nodes; The differential privacy protection module is in communication connection with the computing node and is used for dynamically distributing privacy budget according to the type of the power equipment and the defect sensitivity, adding calibration noise to the local model parameters uploaded by the computing node and reducing the risk of data leakage; The sensitive characteristic shielding module is in communication connection with the computing node and is used for identifying sensitive characteristics and key identification characteristics in the image data and implementing a differential protection strategy on the sensitive characteristics and the key identification characteristics; The aggregation weight adjustment module is in communication connection with the central coordination system and is used for dynamically adjusting the aggregation weight according to the data quality, the data quantity and the privacy protection level of each computing node and balancing the privacy protection and model performance; The privacy evaluation and self-correction module is in communication connection with the differential privacy protection module, the sensitive characteristic shielding module and the aggregation weight adjustment module and is used for evaluating the performance and the privacy protection level of the model and automatically adjusting the privacy protection parameters to form a closed-loop optimization sys