Overhead transmission line insulators are non-conductive materials that separate conductors from grounded transmission towers. Once in operation, they frequently experience environmental pollution and electrical or mechanical stress. Since adverse operational conditions can lead to insulation failure, regular inspections are essential to prevent power outages. To this end, this paper proposes a novel technique based on deep convolutional neural networks (CNNs) to classify high-voltage insulator surface conditions based on their image. Successful applications of CNNs in computer vision have led to several pretrained architectures for image classification. To use these pretrained models, a practitioner typically fine-tunes and selects one final model via a model selection stage and discards all other models. In contrast with many existing studies that use such a “winner-takes-all” approach, here, we identify the best subset of seven popular pretrained CNN architectures that are combined by soft voting to form an ensemble classifier. From a machine learning (ML) perspective, this focus is warranted because the convolutional base of each pretrained architecture operates as a feature extractor and an ensemble of them works as a combination of various feature extraction rules. Our numerical experiments demonstrate the advantage of the identified ensemble model to individual pretrained architectures.

Insulators are vital protection and isolation barriers used in power transmission systems. To prevent unexpected failures caused by severe weather conditions, it is important to develop intelligent insulator fault prognosis systems. To this end, this study has focused to examine and analyze normal (clean) and contaminated (wet) insulators through a new technique that can act as a catalyst for possible future solutions to the data-driven classification of insulator surface conditions. In particular, a novel stacked ensemble learning based on six pretrained deep convolutional neural networks (CNNs), which are used as level-0 generalizers in the stacked structure, is proposed. As the level-1 generalizer that is used in the stacked structure, a majority vote is considered among selected subsets of level-0 generalizers. In other words, training the level-1 generalizer is equivalent to selecting the best subset of pretrained CNNs for classifying a clean insulator surface from those sprayed with water droplets. Since a wetted insulator surface can enhance the electric field intensity and may lead to flashover when combined with other contaminants (dust, soil, cement, etc.), water droplets are considered a type of contamination in our study. Considering a tradeoff between performance and model complexity in training the level-1 generalizer points to two combinations of pretrained CNNs, namely, EfficientNetB2-ResNet50-Xception, and MobileNet-DenseNet121-Xception. Our empirical results show that (i) both combinations lead to better performance when compared with individual pretrained models, and (ii) the latter combination leads to a considerably lower complexity (~39% less parameters) at the expense of ~9% reduction in accuracy.

Outdoor high-voltage (HV) insulators are prone to surface contamination, increasing flashover risk and threatening power transmission systems reliability. Timely inspection is essential, but conventional visual inspections are costly, and inefficient for long transmission lines. An automated inspection using unmanned aerial vehicle images can be both efficient and low-cost. Given an appropriate training sample, a deep neural network such as YOLO (You Only Look Once) can be trained to localize insulators in diverse backgrounds and classify surface contamination. However, the wide variety of YOLO architectures proposed recently makes model selection challenging due to the accuracy-complexity trade-off. Rather than proposing a new model, this study focuses on selecting the “optimal” one. To this end: 1) a “laboratory” dataset of 15,000 insulator images with various surface contaminations is collected; 2) 21 YOLO architectures (YOLOv3–v11 and their variants) are trained, followed by a domain-specific (DS) model selection based on accuracy-complexity trade-offs; and 3) the selected DS-YOLOv11m and DS-YOLOv11n are fine-tuned on a small “industry” dataset of 356 images from a local HV substation to validate DS-specific pre-training in real-world scenarios. This work presents the first DS benchmarking of YOLOv3–v11 for insulator detection and contamination classification, resulting in a lightweight model optimized for real-time edge deployment. Evaluation of the fine-tuned models shows mAP@0.5 scores of 0.983 (DS-YOLOv11m) and 0.977 (DS-YOLOv11n). Although DS-YOLOv11n achieves slightly lower accuracy, it uses significantly fewer resources, reducing FLOPs by ~10.7 times. Finally, DS-YOLOv11n is deployed on a Raspberry Pi 5 to enable real-time contamination classification in edge-computing scenarios.

Regular monitoring of outdoor insulators is crucial to ensure the reliable functioning of the power grid. With recent progress in computer vision technologies, traditional manual and expensive visual inspections can now be replaced by automated analysis using images captured by unmanned aerial vehicles (UAVs). In such applications, a practitioner might opt to choose a state‐of‐the‐art object detection and classification deep learning architecture, including You Look Only Once (YOLO). The variety of existing YOLO architectures per se makes selecting the best application‐dependent YOLO model challenging. However, selecting the best architecture solely based on performance without considering the model complexity limits its deployment on resource‐limited embedded devices. Consequently, we conduct a rigorous, systematic model selection based on the performance–complexity trade‐off across 13 YOLO architectures to determine the most effective model for detecting common mechanical faults in insulators using images captured by UAVs. A dataset comprising 15,000 images of insulators, categorized into normal condition, bird‐pecking damage, cracks, and missing caps, has been compiled for training the models. Specifically, all considered YOLO architectures are compared using model complexity and the mAP@0.5:0.95. During the model selection stage, YOLOv8l proved to be the best model in terms of mAP@0.5:0.95, while YOLOv5n was the model of choice in terms of complexity at the expense of a slight reduction in performance. Alongside YOLOv8l and YOLOv5n, an “optimal” model (OP‐YOLO) was selected using a multicriteria decision‐making approach, balancing detection accuracy and computational efficiency. In particular, in terms of test‐set performance, YOLOv8l, YOLOv5n, and OP‐YOLO achieved 0.919, 0.901, and 0.896 mAP@0.5:0.95, respectively. Although YOLOv8l reported a higher mAP@0.5:0.95, YOLOv5n requires ∼20.9 times less memory and ∼40.2 times less floating‐point operations per second (FLOPs). Also, YOLOv5n outperforms the OP‐YOLO model, still requiring ∼12 times less memory and ∼19 times less FLOPs.