电工技术学报  2024, Vol. 39 Issue (17): 5434-5449    DOI: 10.19595/j.cnki.1000-6753.tces.231243
电力系统与综合能源 |
面向电力巡检图像目标检测的联邦学习激励机制
仲林林1, 刘柯妤2
1.东南大学电气工程学院 南京 210096;
2.东南大学-蒙纳士大学苏州联合研究生院(东南大学) 苏州 215123
Federated-Learning Incentive Mechanism for Object Detection in Power Inspection Images
Zhong Linlin1, Liu Keyu2
1. School of Electrical Engineering Southeast University Nanjing 210096 China;
2. SEU-Monash Joint Graduate School (Suzhou) Southeast University Suzhou 215123 China
全文: PDF (6085 KB)   HTML
输出: BibTeX | EndNote (RIS)      
摘要 基于深度学习的目标检测算法能够高效处理电力巡检图像,及时发现故障隐患。然而,由于数据整合困难及数据隐私保护等原因,单个电力公司或第三方机构可能不足以训练出高性能模型。为解决这一问题并激励更多参与方加入面向电力巡检图像目标检测的联邦学习,该文构建了基于模型公平和基于收益公平的电力巡检图像目标检测联邦激励机制。基于模型公平的激励机制适用于所有参与方都是数据拥有方的情况,通过贡献评估分配不同性能的模型;基于收益公平的激励机制针对同时存在数据拥有方和数据需求方的模式,数据拥有方获得相应的收益,而数据需求方获得高性能模型。实验结果显示,在这两种激励机制中,公平性相关系数分别可达到0.96和1。这表明所提出的激励机制可有效地提升公平性,并能够激励更多的参与方加入到面向电力巡检图像目标检测的联邦学习中。
服务
把本文推荐给朋友
加入我的书架
加入引用管理器
E-mail Alert
RSS
作者相关文章
仲林林
刘柯妤
关键词 电力巡检目标检测联邦学习激励机制公平性    
Abstract:In recent years, the automated inspection approaches have been widely used in power system, which is safer and more efficient than traditional manual inspection. In order to effectively process the large number of power inspection images generated by automated inspection machines, such as unmanned aerial vehicles (UAVs), the deep learning-based object detection technology is commonly used to timely detect potential faults. However, in practical applications, power lines are widely distributed, and the collected inspection images are not easily integrated. In addition, due to data privacy and restrictions from relevant laws and regulations, these inspection images cannot be shared with or provided to third-party organizations. The data exist as “data islands” scattered among various power companies. In this situation, a single power company has a limited amount and variety of inspection image data, making it difficult to obtain a robust object detection model with good generalization performance through traditional data-driven deep learning algorithms.
Federated learning (FL), as an emerging distributed technology, can be efficiently applied to visual tasks of power inspection by constructing a high-performance model through collaboration among participants while protecting data privacy. In traditional federated learning, the final distributed model to each participant is a global model with the same performance without considering the contribution differences by different participants, which means different participants receive the same model. However, in real scenarios, the quantity, quality, and cost of power inspection image data provided by different participants are really different, and their contributions to the training of global model are also different. If the participants in federal learning cannot receive fair and reasonable returns, it will cause some participants to be unwilling to participate in the federal training of object detection model for processing power inspection images, and cannot guarantee the number of participants in federal learning, resulting in the inability to continue federal training or poor training performance.
To address these challenges and incentivize more participants to join federated learning for object detection of power inspection images, this paper proposes two federated incentive mechanisms based on model fairness and revenue fairness, respectively. The model fairness incentive mechanism is designed for the scenarios where all participants are data owners, and it involves evaluating contributions to obtain models with varying performance levels. In this incentive mechanism, we use an evaluation strategy based on marginal benefit, and compare the contribution measurement methods based on unilateral training, dynamic accuracy, and impact function. The revenue fairness incentive mechanism targets the scenarios involving both data owners and data demanders, wherein data owners receive corresponding benefits, and data demanders obtain high-performance models. In this incentive mechanism, the fairness of the federated learning is modeled by evaluating contributions, expected losses, and expected temporal losses.
The above two incentive mechanisms are validated by the experiments on the homemade UAV power tower inspection image dataset. The three kinds of data heterogeneity are considered including feature skew, quantity skew and mixed skew. The results demonstrate that the fairness correlation coefficients in the two incentive mechanisms can reach 0.96 and 1, respectively. This indicates that the proposed incentive mechanisms can effectively enhance fairness and motivate more participants to engage in federated learning for object detection of power inspection images. Also, it is noted that although we use power object detection as the main application scenario in this work, the proposed incentive mechanism can be applied to various power vision tasks including image classification, object detection, and anomaly detection, and has a certain degree of generality.
Key wordsPower inspection    object detection    federated learning    incentive mechanism    fairness   
收稿日期: 2023-08-01     
PACS: TM11  
基金资助:国家自然科学基金(92066106)、江苏省科协青年科技人才托举工程(2021031)和东南大学“至善青年学者”支持计划(中央高校基本科研业务费)(2242022R40022)资助项目
通讯作者: 仲林林 男,1990年生,副研究员,博士生导师,研究方向为高电压技术、放电等离子体技术、人工智能技术。E-mail:linlin@seu.edu.cn   
作者简介: 刘柯妤 女,1998年生,硕士研究生,研究方向为联邦学习与目标检测在电气领域的应用。E-mail:lky@seu.edu.cn
引用本文:   
仲林林, 刘柯妤. 面向电力巡检图像目标检测的联邦学习激励机制[J]. 电工技术学报, 2024, 39(17): 5434-5449. Zhong Linlin, Liu Keyu. Federated-Learning Incentive Mechanism for Object Detection in Power Inspection Images. Transactions of China Electrotechnical Society, 2024, 39(17): 5434-5449.
链接本文:  
https://dgjsxb.ces-transaction.com/CN/10.19595/j.cnki.1000-6753.tces.231243          https://dgjsxb.ces-transaction.com/CN/Y2024/V39/I17/5434