Damage identification technology of substation instrument panel for solving imbalanced classification

Abstract

Edge computing plays an active role in empowering the power industry as a key technology for establishing data-driven Internet of things (IoT) applications. Traditional defect diagnosis mainly relies on regular inspection of equipment by operation and maintenance personnel at all levels, and its accuracy relies on the human experience. In actual production, the image data of some dashboard damage types are easy to collect in large quantities, while some dashboard damage types occur less frequently and are more difficult to collect. The use of edge computing nodes allows flexible and fast collection of smart meter data and transmission of the reduced data or results to a cloud computing center. In this study, we provide a fresh balanced training approach to address the issue of learning from unbalanced data. In the equilibrium training phase, a new impact balance loss is introduced to reduce the influence of samples on the overfitting decision boundary. Experimental results show that the proposed balance loss function effectively improves the performance of various types of imbalance learning methods.

References

[1] QiuT, ChiJ, ZhouX, et al.. Edge computing in industrial internet of things: architecture, advances and challenges[J]. IEEE communications surveys & tutorials, 2020, 22(4):2462-2488

[2] ShangF, LaiJ, ChenJ, et al.. A model compression based framework for electrical equipment intelligent inspection on edge computing environment[C], 2021, New York, IEEE: 406-410

[3] ZhuX, YangJ, ZhangC, et al.. Efficient utilization of missing data in cost-sensitive learning[J]. IEEE transactions on knowledge and data engineering, 2019, 33(6):2425-2436

[4] WuJ, BravermanV, YangL. Obtaining adjustable regularization for free via iterate averaging[C], 2020, Princeton, IMLS: 10344-10354

[5] RizkH, YamaguchiH, HigashinoT, et al.. A ubiquitous and accurate floor estimation system using deep representational learning[C], 2020, New York, ACM: 540-549

[6] DengS, ZhaoH, FangW, et al.. Edge intelligence: the confluence of edge computing and artificial intelligence[J]. IEEE internet of things journal, 2020, 7(8):7457-7469

[7] ZhouZ, ChenX, LiE, et al.. Edge intelligence: paving the last mile of artificial intelligence with edge computing[J]. Proceedings of the IEEE, 2019, 107(8):1738-1762

[8] ZAIDI S S A, ANSARI M S, ASLAM A, et al. A survey of modern deep learning based object detection models[J]. Digital signal processing, 2022: 103514.

[9] LiL, DoroslovackiM, LoewM H. Approximating the gradient of cross-entropy loss function[J]. IEEE access, 2020, 8: 111626-111635

[10] ParkD Y, ChaM H, KimD, et al.. Learning student-friendly teacher networks for knowledge distillation[J]. Advances in neural information processing systems, 2021, 34: 13292-13303

[11] ChuY, YangX, LiH, et al.. Multi-level feature aggregation network for instrument identification of endoscopic images[J]. Physics in medicine & biology, 2020, 65(16):165004

[12] SingadkarG, MahajanA, ThakurM, et al.. Deep deconvolutional residual network based automatic lung nodule segmentation[J]. Journal of digital imaging, 2020, 33(3): 678-684

[13] PASZKE A, GROSS S, MASSA F, et al. Pytorch: an imperative style, high-performance deep learning library[EB/OL]. (2019-12-03) [2022-09-20]. https://arxiv.org/abs/1912.01703.

[14] LI Y, LIANG Y. Learning overparameterized neural networks via stochastic gradient descent on structured data[EB/OL]. (2019-08-01) [2022-09-20]. https://arxiv.org/abs/1808.01204v1.