簡易檢索 / 詳目顯示

回結果列表
研究生： 柯蘇菲
Chladek, Sophie-Marie
論文名稱： Comparative Evaluation of Sensor- and Image-Based Learning Pipelines for Photovoltaic Shading Root Cause Classification
Comparative Evaluation of Sensor- and Image-Based Learning Pipelines for Photovoltaic Shading Root Cause Classification
指導教授： 謝昱銘
Hsieh, Yu-Ming
共同指導: 鄭芳田
Cheng, Fan-Tien
學位類別： 碩士
Master
系所名稱： 智慧半導體及永續製造學院 - 半導體製程學位學程
Program on Semiconductor Manufacturing Technology
論文出版年： 2026
畢業學年度： 114
語文別： 英文
論文頁數： 119
外文關鍵詞： Photovoltaic Monitoring, Root-Cause Classification, Time-Series Classification, Computer Vision, Image Segmentation, Multimodal Fault Diagnosis, Smart PV Maintenance
相關次數： 點閱：13  下載：0 
分享至:
查詢本校圖書館目錄 查詢臺灣博碩士論文知識加值系統 勘誤回報

    • Photovoltaic systems are increasingly deployed as key components of sustainable energy infrastructure, creating a growing demand for deployable intelligent monitoring solutions capable of detecting performance degradation and identifying the underlying external root causes of power loss. Conventional electrical monitoring is effective in indicating abnormal behaviour; however, its ability to distinguish between visually different shading scenarios remains limited due to the indirect nature of temporal sensor signals. To address this challenge, this study proposes and evaluates a deployment-oriented multimodal deep learning framework for photovoltaic shading root-cause classification by integrating electrical sensor-based time-series learning with image-based visual diagnostics. The developed methodology includes: (1) a multivariate sensor classification pipeline using LSTM, TCN, and Transformer architectures, (2) an image-based computer vision pipeline using YOLOv8 object detection, RT-DETR transformer detection, and YOLOv8 segmentation for spatial shading analysis, and (3) a hybrid deployment strategy combining continuous sensor screening with targeted image-based fault verification. Experimental validation was performed on a custom laboratory dataset consisting of synchronized photovoltaic sensor measurements and annotated image samples under leaf, dirt, and geometric shadow conditions. The results demonstrate that sensor-based time-series models can identify abnormal PV behaviour to a moderate extent but remain limited for fine-grained root-cause classification, while image-based models provide direct spatial information and stronger task-specific performance. Based on these findings, the study provides a practical deployment reference for future intelligent photovoltaic maintenance systems.

    SUMMARY 3 1 Introduction 10 1.1 Research Background 10 1.2 Research Motivation 10 1.3 Research Objective 12 1.4 Research Process 14 1.5 Thesis Organization 16 2 Literature Review and Core Technologies 18 2.1 Literature Review 18 2.1.1 The Importance of Solar PV Energy 18 2.1.2 Failure Categories 18 2.1.3 Shading and Soiling in PV Systems 18 2.1.3.1 Dirt and Soiling 18 2.1.3.2 Shading Theory and Classification 19 2.1.4 The Importance of PV Monitoring 20 2.1.5 Methodologies for PV Fault Detection 20 2.1.5.1 Traditional and Sensor-Based Pipelines 20 2.1.5.2 Image-Based Pipelines and Feature Extraction 21 2.2 Core Technologies 22 2.2.1 Data Augmentation via SMOTE-Tomek and VAE-Based Methods 22 2.2.1.1 Variational Autoencoder (VAE) 23 2.2.1.2 SMOTE and SMOTE-Tomek 23 2.2.2 LSTM 24 2.2.3 TCN 26 2.2.4 Transformer 27 2.2.5 Computer Vision 29 2.2.5.1 Object Detection - Yolo 30 2.2.5.1.1 Backbone and Feature Extraction 30 2.2.5.1.2 Anchor-Free and Decoupled Head Mechanism 31 2.2.5.1.3 Loss Formulation and Detection Metrics 31 2.2.5.2 Transformer Detection - RT-DETR 32 2.2.5.3 Segmentation YoloSeg 34 3 Evaluation Framework and Methodology 37 3.1 Evaluation Framework 37 3.2 Workflow Design and Pipeline Structure 38 3.3 Evaluation Methodology 39 3.3.1 Evaluation Metrics 40 3.4 Framework Alignment with Research Questions 41 4 Experimental Set up and Data Acquisition 42 4.1 Functional Requirements Analysis 42 4.2 Laboratory Configuration 43 4.3 Sensor Data Acquisition 44 4.4 Experimental Test Matrix 46 4.5 Data Acquisition Methodology 48 4.6 Image Data Acquisition 49 4.7 Annotation Process 49 5 Sensor-Based Pipeline 52 5.1 Scope and Structure 52 5.2 Feature Correlation Analysis 53 5.3 Data Preprocessing 55 5.4 Data Mode Implementation 58 5.4.1 Standard Date Mode (Real Data only) 58 5.4.1.1 Dataset Construction 58 5.4.1.2 Feature Scaling and Integrity Preservation 59 5.4.1.3 Handling of Class Imbalance 60 5.4.2 VAE-SMOTE Augmented Data Mode 60 5.4.2.1 VAE-Based Latent Representation Learning 61 5.4.2.2 SMOTE in Latent Space 61 5.4.2.3 Decoding of Synthetic Samples 62 5.4.2.4 Dataset Augmentation and Storage 62 5.4.3 Summary 63 5.5 Model Architectures 63 5.5.1 Model Implementation 64 5.5.2 LSTM Architecture Implementation 64 5.5.2.1 LSTM - Role in This Work 65 5.5.3 Transformer Architecture Implementation 66 5.5.3.1 Transformer - Role in this Work 66 5.5.4 TCN Architecture Implementation 67 5.5.4.1 TCN - Role in this work 68 5.6 Training Integration 68 5.7 Methodological Summary 69 6 Image-Based Pipeline 70 6.1 Scope and Structure 70 6.2 Data Preprocessing 70 6.3 Model Implementation 72 6.3.1 Dataset Preparation and Integration 72 6.3.2 Unified Training Workflow 73 6.3.3 Implementation of Image-Based Model Architectures 73 6.3.4 YOLO Object Detection 73 6.3.5 RT-DETR (Transformer-Based Detection) 74 6.3.6 YOLO Segmentation (YOLO-seg) 74 6.4 Model Evaluation 75 6.5 Summary 75 7 Results and Evaluation 76 7.1 Objective 76 7.2 Sensor-based Pipeline Results 76 7.2.1 LSTM 77 7.2.1.1 LSTM Standard Mode 77 7.2.1.2 LSTM VAE-SMOTE 77 7.2.2 Temporal Convolutional Networks (TCN) 78 7.2.2.1 TCN Standard Mode 78 7.2.2.2 TCN VAE-SMOTE 78 7.2.3 Transformer 79 7.2.3.1 Transformer Standard Mode 79 7.2.3.2 Transformer VAE-SMOTE 79 7.3 Image-based Pipeline Results 79 7.3.1 YOLO 80 7.3.1.1 Overall Model Performance 80 7.3.1.2 Precision-Recall Characteristics 81 7.3.1.3 Confusion Matrix 83 7.3.1.4 Training Behavior and Model Convergence 85 7.3.1.5 Qualitative Evaluation 86 7.3.2 RT-DETR 87 7.3.2.1 Overall Model Performance 88 7.3.2.2 Precision-Recall Characteristics 89 7.3.2.3 Confusion Matrix 91 7.3.2.4 Training Behavior and Model Convergence 92 7.3.2.5 Qualitative Evaluation 94 7.3.3 YOLO-Seg 95 7.3.3.1 Overall Model Performance 95 7.3.3.2 Precision-Recall Characteristics 97 7.3.3.3 Confusion Matrix 99 7.3.3.4 Training Behavior and Model Convergence 100 7.3.3.5 Qualitative Evaluation 102 7.4 Performance Evaluation and Comparative Analysis 103 7.4.1 Sensor-Based Pipeline 103 7.4.2 Image-Based Pipeline 105 7.5 Cross-Pipeline Comparison 108 7.5.1 Limits of Sensor-Based Pipeline 108 7.5.2 Advantage of Image-Based Pipeline 108 7.6 Practical Framework and Deployment Strategy 109 7.6.1 Stage 1: Continuous Electrical Filtering 110 7.6.2 Stage 2: On-Demand Visual Diagnosis 110 8 Conclusion 113 8.1 Conclusion 113 8.2 Future Work 115 REFERENCES 117

    [1] E. Ortega et al., “New monitoring method to characterize individual modules in large photovoltaic systems”, Solar Energy, vol. 193, p. 906–914, 2019.
    [2] Bamisile, O., Acen, C., Cai, D., Huang, Q., and Staffell, “The environmental factors affecting solar photovoltaic output,” Renewable and Sustainable Energy Reviews, vol. 208, 2025.
    [3] Katleho Masita et al., “Deep learning in defects detection of PV models: A review,” Solar Energy Advances, vol. 5, 2025.
    [4] Mouadd Bougoffa, Samir Benmoussa et al., “Hybrid deep learning for fault diagnosis in PV systems,” machines, vol. 13, no. 378, 2025.
    [5] W. G. Kais Abdulmawjood, “IECON 2024 - 50th annual conference of the IEEE Industrial Electronics Society,” in Comparative Analysis of Machine Learning Algorithms for Identifying Partial Shading Conditions on PV Array, 2024.
    [6] B. Z. and X. Y. J. Wang, “39th Chinese Control Conference (CCC),” in PV Abnormal Shading Detection Based on Convolutional Neural Network, 2020.
    [7] J. A. Arenas-Prado et al., “Automated detection of shading faults in photovoltaic modules using convolutional neural networks and I-V curves,” Processes, vol. 13, no. 2999, 2025.
    [8] F. Aziz et al., “A novel convolutional neural network-based approach for fault classification in Photovoltaic Arrays,” IEEE Power & Energy Soeciety Section (IEEE Access) , vol. 8, pp. 41889-41904, 2020.
    [9] Mohamed M. Badr et al., “Fault identification of photovoltaic array based on machine learning classifiers,” IEEE Power & Energy Society Section (IEEE Access), vol. 9, pp. 159113-159132, 2021.
    [10] C. Saiprakash et al., “Improved fault detection and classification in PV arrays using stockwell transform and data mining techniques,” Results in Engineering, vol. 23, 2024.
    [11] K.I. Baradieh et al., “A Study on the impact of different PV model parameters and various DC faults on the characteristics and ferformance of the Photovoltaic Arrays,” Inventions, vol. 9, no. 93, 2024.
    [12] G. Nassreddine et al., “Fault detection and classification for photovoltaic panel system using machine learning techniques,” Applied AI Letters, no. vol. 6, 2025.
    [13] R.E.Figueredo et al., “Using Computer Vision with Edge Machine Learning to Recognize Dirt on Photovoltaic Modules,” Journal of Intelligent System and Telecommunications, vol. 1, no. 2, pp. 140-150, 2025.
    [14] H. Z. Y. Wang et al., “Vision-based fault detection for photovoltaic modules using deep learning,” Applied Energy, vol. 238, p. 2019, 162–172.
    [15] Ahmed Faris Amiri et al., “Fault detection and diagnosis of a photovoltaic system based on deep learning using the combination of a CNN and Bi-GRU,” sustainability, vol. 16, no. 1012, 2024.
    [16] Mohand Djeziri et al., “Detection and identification of degradation root causes in a photovoltaic cell based on physical modeling and deep learning,” applied sciences, vol. 15, no. 7684, pp. 1-14, 2025.
    [17] S. Oussama et al., “2nd International Conference of Advanced Technology in Electronic and Electrical Engineering (ICATEEE),” in Real-Time Fault Detection and Classification in PV Systems Through Classical Thresholding Techniques, 2025.
    [18] R. Hariharan et al., “A method to detect photovoltaic array faults and partial shading in PV systems,” IEEE Journal of Photovoltaics, vol. 6, no. 5, pp. 1278-1285, 2016.
    [19] A. Mellit et al., “A machine learning and internet of things-based online fault diagnosis method for photovoltaic arrays,” Sustainability, vol. 13, no. 13203, pp. 1-14, 2021.
    [20] J. Shin et al., “Comparative analysis of data representation methods for PV anomaly detection: raw time-series data vs. diverse binning strategies,” Mathematics, vol. 14, no. 501, pp. 1-22, 2026.
    [21] J. Suh et al., “Detection of photovoltaic panel dust using drone shooting imagery and deep learning techniques,” KSMER Journal of The Korean Society of Mineral and Energy Resources Engineers, vol. 61, no. 1, pp. 23-32, 2024.
    [22] S. Boubaker et al., “Assessment of machine and deep learning approaches for fault diagnosis in photovoltaic systems using infrared thermography,” remote sensing, vol. 15, no. 1686, pp. 1-21, 2023.
    [23] F. Pamungkas et al., “A novel approach for efficient solar panel fault classification using coupled UDenseNet,” Sensors, vol. 23, no. 4918, 2023.
    [24] Vinamra Parakh et al., “2025 IEEE 5th International Conference on Sustainable Energy and Future Electric Transportation (SEFET),” in Solar PV Panels Image Classification and Hotspot Region Detection Using Machine Learning, 2025.
    [25] Stocksieker, S., Pommeret, D., and Charpentier, A, "arxiv.org," 09 December 2024. [Online]. Available: https://arxiv.org/html/2412.07039v1. [Accessed 19 April 2026].
    [26] N.V. Chawla, “Synthetic Minority Over-sampling Technique,” Journal of Artificial Intelligence Research, vol. 16, p. 321–357, 2002.
    [27] Barak Or et al., "arxiv.org," 26 December 2025. [Online]. Available: https://arxiv.org/pdf/2501.06491. [Accessed 19 April 2026].
    [28] F. Swana et al., “Tomek Link and SMOTE approaches for machine fault classification with an imbalanced dataset,” Sensors, vol.22, no.9, 2022.
    [29] Nasraldeen Alnor Adam Khleel et al., “A novel approach for software defect prediction using CNN and GRU based on SMOTE Tomek method,” Journal of Intelligent Information Systems, vol. 60, pp. 673-707, 2023.
    [30] S. Hochreiter and J. Schmidhuber, “Long Short-Term Memory,” Neural Computation, vol. 9, no. 8, p. 1735–1780, 1997.
    [31] PyTorch, “PyTorch, LSTM,” [Online]. Available: https://docs.pytorch.org/docs/2.12/generated/torch.nn.LSTM.html. [Accessed 05 April 2026].
    [32] Shaojie Bai et al., “An empirical evaluation of generic convolutional and recurrent networks for sequence modeling,” 4 March 2018. [Online]. Available: https://arxiv.org/abs/1803.01271. [Accessed 24 April 2026].
    [33] Ashish Vaswani et al., “31st Conference on Neural Information Processing Systems (NIPS 2017),” in Attention Is All You Need, Long Beach, CA, USA, 2017.
    [34] Joseph Redmon et al., “You Only Look Once: Unified, Real-Time Object Detection, ” 9 April 2016. [Online]. Available: https://arxiv.org/abs/1506.02640. [Accessed 5 April 2026].
    [35] J. Terven. and D. Cordova-Esparza, "ResearchGate," April 2023. [Online]. Available: https://www.researchgate.net/publication/369760111_A_Comprehensive_Review_of_YOLO_From_YOLOv1_to_YOLOv8_and_Beyond. [Accessed 5 April 2026].
    [36] Ultralytics, “Ultralytics - Yolov8, ” 2023. [Online]. Available: https://docs.ultralytics.com/models/yolov8/. [Accessed 5 April 2026].
    [37] Yian Zhao et al., "arxiv.org," 3 April 2024. [Online]. Available: https://arxiv.org/abs/2304.08069. [Accessed 5 April 2026].
    [38] Kaiming He et al., “Proc. IEEE Int. Conf. Computer Vision (ICCV),” 24 January 2018. [Online]. Available: https://arxiv.org/abs/1703.06870. [Accessed 05 April 2026].
    [39] Ultralytics, “Ultralytics - Yolov8-vs-Yolov9,” 2024. [Online]. Available: https://docs.ultralytics.com/compare/yolov8-vs-yolov9/. [Accessed 5 April 2026].
    [40] Zheng Ge et al., "arxiv.org," 6 August 2021. [Online]. Available: https://arxiv.org/abs/2107.08430. [Accessed 5 April 2026].
    [41] Ultralytics, “Ultralytics - Loss,” 2023. [Online]. Available: https://docs.ultralytics.com/reference/utils/loss/. [Accessed 5 April 2026].
    [42] X. Li, "arxiv.org," 8 June 2020. [Online]. Available: https://arxiv.org/abs/2006.04388. [Accessed 5 April 2026].
    [43] Nicolas Carion et al., "arxiv.org," 26 May 2020. [Online]. Available: https://arxiv.org/abs/2005.12872. [Accessed 5 April 2026].
    [44] Alexey Dosovitskiy et al., "arxiv.org," 3 June 2021. [Online]. Available: https://arxiv.org/abs/2010.11929. [Accessed 5 April 2026].
    [45] Yian Zhao, "arxiv.org," 17 April 2023. [Online]. Available: https://arxiv.org/abs/2304.08069. [Accessed 19 April 2026].
    [46] Xizhou Zhu, "arxiv.org," 18 March 2021. [Online]. Available: https://arxiv.org/abs/2010.04159. [Accessed 6 April 2026].
    [47] D. Bolya et al., “IEEE/CVF Int. Conf. Computer Vision (ICCV),” in YOLACT: Real-time Instance Segmentation, 2019.
    [48] Z. Sun et al., “YOLOv8E: an efficient YOLOv8 method for instance segmentation of individual tree crowns in Wellington City, New Zealand,” Journal of the Royal Society of New Zealand, vol. 55, no. 6, pp. 1854-1879, 2024.
    [49] G. Lapalme et al., “A systematic analysis of performance measures for classification tasks,” Information Processing & Management, vol. 45, no. 4, p. 427–437, 2009.
    [50] Tsung-Yi Lin et al., "arxiv.org," 21 February 2015. [Online]. Available: https://arxiv.org/abs/1405.0312. [Accessed 20 April 2026].
    [51] J. Long et al., “Proc. IEEE Conf. Computer Vision and Pattern Recognition (CVPR),” in Fully Convolutional Networks for Semantic Segmentation, 2015.
    [52] L. d. O. Santos et al., “Photovoltaic power estimation and forecast models integrating physics and machine learning: A review on hybrid techniques,” Solar Energy, vol. 275, no. 113044, pp. 1-16, 2024.
    [53] Z. Yang et al., "https://arxiv.org/," 18 December 2025. [Online]. Available: https://doi.org/10.48550/arXiv.2512.17107. [Accessed 7 April 2026].
    [54] Q. Tang et al., “A Comparative Review on Multi-Modal Sensors Fusion Based on Deep Learning,” Signal Processing, vol. 213, 2023.
    [55] “Dual‐transformer deep learning framework for seasonal forecasting of Great Lakes water levels - Scientific Figure on ResearchGate,” November 2024. [Online]. Available: https://www.researchgate.net/figure/The-structure-of-a-Long-Short-Term-Memory-LSTM-memory-cell-The-diagram-illustrates-the_fig2_391871569. [Accessed 24 April 2026].

    QR CODE