基于知识蒸馏的红外光伏组件缺陷检测模型

王银; 张杰; 谢刚; 赵志诚; 胡啸; 吴晓晖

doi:10.19912/j.0254-0096.tynxb.2024-0402

PDF(5521 KB)

太阳能学报 ›› 2025, Vol. 46 ›› Issue (7) : 653-662. DOI: 10.19912/j.0254-0096.tynxb.2024-0402

第二十七届中国科协年会学术论文

基于知识蒸馏的红外光伏组件缺陷检测模型

王银^1,2, 张杰^1,2, 谢刚^1,2, 赵志诚^1,2, 胡啸^1,2, 吴晓晖³

作者信息 +

DEFECT DETECTION MODEL OF INFRARED PHOTOVOLTAIC MODULE BASED ON KNOWLEDGE DISTILLATION

Wang Yin^1,2, Zhang Jie^1,2, Xie Gang^1,2, Zhao Zhicheng^1,2, Hu Xiao^1,2, Wu Xiaohui³

Author information +

文章历史 +

摘要

针对目标检测模型的参数量和计算复杂度的不断增加导致红外光伏组件缺陷检测模型难以部署到边缘设备的问题,提出一种基于模型压缩的红外光伏组件缺陷检测算法T-DINO。以ResNet-101为教师网络、ResNet-18为学生网络,提出一种动态的自适应蒸馏方法,在基于特征蒸馏中利用二者的注意力权重差异进行高效的知识转移,在基于输出响应（logit）蒸馏中也将其作为指导知识对学生网络进行蒸馏,最后在损耗极小精度的情况下大大降低模型复杂度和参数量;同时提出融合模块CSF Block对局部特征和全局特征进行建模,提高检测精度。在自主构建的红外光伏组件故障数据集上进行实验,相比基线网络DINO（ResNet-101）模型参数量减少77.3%,计算复杂度降低69.3%,AP50提高5.2%。仿真实验结果表明：压缩后的模型适合部署在边缘设备,可满足实际红外光伏组件缺陷检测要求。

Abstract

Aiming at the problem that the infrared photovoltaic module defect detection model is difficult to deploy to edge devices due to the increasing number of parameters and computational complexity of object detection model, a defect detection algorithm T-DINO （Tiny DINO） based on model compression is proposed. Using ResNet-101 as the teacher network and ResNet-18 as the student network, a dynamic adaptive distillation method is proposed. The difference in attention weights between the two is used for efficient knowledge transfer in feature-based distillation, and it is also used as guiding knowledge for distillation of the student network in output response （logit） distillation. Consequently, the complexity of the model and the number of parameters are greatly reduced with minimal accuracy loss. At the same time, the fusion module CSF Block （Conv-Self Attention Block） is proposed to model local features and global features to improve the detection accuracy. Experiments on the self-constructed infrared PV module fault dataset showed a 77.3% reduction in the number of parameters, a 69.3% reduction in computational complexity and a 5.2% increase in AP50 compared to the baseline network DINO （ResNet-101） model. The simulation results show that the compressed model is suitable for deployment in edge equipment and can meet the requirements of actual infrared photovoltaic module defect detection.

导出引用

王银, 张杰, 谢刚, 赵志诚, 胡啸, 吴晓晖. 基于知识蒸馏的红外光伏组件缺陷检测模型[J]. 太阳能学报. 2025, 46(7): 653-662 https://doi.org/10.19912/j.0254-0096.tynxb.2024-0402

Wang Yin, Zhang Jie, Xie Gang, Zhao Zhicheng, Hu Xiao, Wu Xiaohui. DEFECT DETECTION MODEL OF INFRARED PHOTOVOLTAIC MODULE BASED ON KNOWLEDGE DISTILLATION[J]. Acta Energiae Solaris Sinica. 2025, 46(7): 653-662 https://doi.org/10.19912/j.0254-0096.tynxb.2024-0402

中图分类号： TP391.4

参考文献

[1] GIELEN D, BOSHELL F, SAYGIN D, et al.The role of renewable energy in the global energy transformation[J]. Energy strategy reviews, 2019, 24: 38-50.
[2] HANSEN K, BREYER C, LUND H.Status and perspectives on 100% renewable energy systems[J]. Energy, 2019, 175: 471-480.
[3] 代云中, 唐浩, 刘铭煊, 等. 一种Z源光伏并网逆变器及其共模漏电流抑制[J]. 太阳能学报, 2024, 45(3): 517-522.
DAI Y Z, TANG H, LIU M X, et al.Z-source photovoltaic grid-connected inverter with common-mode leakage current inhibition[J]. Acta energiae solaris sinica, 2024, 45(3): 517-522.
[4] 金伟勇, 卢丽娜, 赖欢欢, 等. 基于功率特征的K-ISSA-LSTM光伏功率预测[J]. 太阳能学报, 2024, 45(2): 429-434.
JIN W Y, LU L N, LAI H H, et al.K-ISSA-LSTM photovoltaic power prediction based on power characteristic[J]. Acta energiae solaris sinica, 2024, 45(2): 429-434.
[5] 周颖, 叶红, 王彤, 等. 基于多尺度CNN的光伏组件缺陷识别[J]. 太阳能学报, 2022, 43(2): 211-216.
ZHOU Y, YE H, WANG T, et al.Photovoltaic module defect identification based on mulit-scale convolution neural network[J]. Acta energiae solaris sinica, 2022, 43(2): 211-216.
[6] GOODFELLOW I J, POUGET-ABADIE J, MIRZA M, et al.Generative adversarial networks[C]//Proceedings of the 27th International Conference on Neural Information Processing Systems. Montreal, Canada, 2014(2): 2672-2680.
[7] GUO S Q, WANG Z H, LOU Y, et al.Detection method of photovoltaic panel defect based on improved mask R-CNN[J]. Journal of Internet technology, 2022, 23(2): 397-406.
[8] HE K M, GKIOXARI G, DOLLAR P, et al.Mask R-CNN[J]. IEEE transactions on pattern analysis and machine intelligence, 2020, 42(2): 386-397.
[9] PRABHAKARAN S, ANNIE UTHRA R, PREETHAROSELYN J.Deep learning-based model for defect detection and localization on photovoltaic panels[J]. Computer systems science and engineering, 2023, 44(3): 2683-2700.
[10] CAO Y K, PANG D D, ZHAO Q C, et al.Improved YOLOv8-GD deep learning model for defect detection in electroluminescence images of solar photovoltaic modules[J]. Engineering applications of artificial intelligence, 2024, 131: 107866.
[11] LIU Z, LI J G, SHEN Z Q, et al.Learning efficient convolutional networks through network slimming[C]//2017 IEEE International Conference on Computer Vision (ICCV). Venice, Italy, 2017: 2755-2763.
[12] HINTON G, VINYALS O, DEAN J.Distilling the knowledge in a neural network[J]. Computer science, 2015, 14(7), 38-39.
[13] HAN S, MAO H Z, DALLY W J.Deep compression: compressing deep neural networks with pruning, trained quantization and huffman coding[J]. Neurocomputing, 2023(520): 152-170.
[14] BA L J, CARUANA R.Do deep nets really need to be deep?[C]//Proceedings of the 28th International Conference on Neural Information Processing Systems-Volume 2. Montreal, Canada, 2014: 2654-2662.
[15] ROMERO A, BALLAS N, KAHOU S E, et al.FitNets: hints for thin deep nets[C]//International Conference on Learning Representations (ICLR), 2015:1-10.
[16] BELAGIANNIS V, FARSHAD A, GALASSO F.Adversarial network compression[C]//European Conference on Computer Vision, 2018: 431-449.
[17] CHEN S X, ZHANG C J, DONG M.Coupled end-to-end transfer learning with generalized fisher information[C]//2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Salt Lake City, UT, USA, 2018: 4329-4338.
[18] RAJASEGARAN J, KHAN S, HAYAT M, et al.Self-supervised knowledge distillation for few-shot learning[J]. Computer vision and pattern recognition, 2020: 1-11.
[19] VASWANI, SHAZEER, PARMAR, et al. Attention is all you need[C]//Advances in neural information processing systems, 2017: 5998-6008.
[20] ZAGORUYKO S, KOMODAKIS N.Paying more attention to attention: improving the performance of convolutional neural networks via attention transfer[C]//International Conference on Learning Representations, 2017:1-10.
[21] ZHU X Z, SU W J, LU L W, et al.Deformable detr: Deformable transformers for end-to-end object detection[C]//The Ninth International Conference on Learning Representations, 2021:1-14.
[22] Liu S., Li F., ZHANG H,et al.DINO: DETR with improved DeNoising anchor boxes for end-to-end object detection[C]//Proceedings of the International Conference on Learning Representations, 2023:1-17.
[23] LIN T Y, GOYAL P, GIRSHICK R, et al.Focal loss for dense object detection[J]. IEEE transactions on pattern analysis and machine intelligence, 2020, 42(2): 318-327.
[24] CHO Y J.Weighted Intersection over Union (wIoU): A New Evaluation Metric for Image Segmentation[J]. Pattern recongnition letters, 2024(185):101-107.
[25] LIU Z, LIN Y T, CAO Y, et al.Swin transformer: hierarchical vision transformer using shifted windows[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision , 2021:1-19.
[26] CARION N, MASSA F, SYNNAEVE G, et al.End-to-end object detection with transformers[C]//European Conference on Computer Vision, 2020: 213-229.
[27] LIU S L, LI F, ZHANG H, et al.DAB-DETR: dynamic anchor boxes are better queries for DETR[C]//International Conference on Learning Representations (ICLR), 2022: 1-19.
[28] LI F, ZHANG H, LIU S L, et al.DN-DETR: Accelerate DETR Training by Introducing Query DeNoising[J]. IEEE transactions on pattern analysis and machine intelligence, 2024, 46(4): 2239-2251.
[29] AHN S, HU S X, DAMIANOU A, et al.Variational information distillation for knowledge transfer[C]//2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Long Beach, CA, USA, 2019: 9155-9163.
[30] CHEN D F, MEI J P, ZHANG Y, et al.Cross-layer distillation with semantic calibration[J]. Proceedings of the AAAI conference on artificial intelligence, 2021, 35(8): 7028-7036.