Overhead power lines inspection and defect detection by using deep learning approaches on UAV images

The energy sector is one of the most critical industries globally in terms of sustainability and reliability. Traditional methods for grid monitoring, maintenance, and renewal decision-making have become increasingly complex. In this context, using image processing techniques to detect objects of varying sizes and automatically identify equipment that may cause faults is of critical importance. In this thesis study, a total of 588 images were used to train the YOLOV5 model during the machine learning phase, and accuracy was evaluated using 207 images during the testing phase. The applied model demonstrated better performance in detecting insulator objects by producing fewer false positives and achieving more correct detections. The mAP@0.5 threshold is 0.702 for insulator class. For broken insulators, the mean average precision (mAP) at 0.5 confidence threshold was calculated as 0.507, which is lower compared to the insulator class. In the insulator class, relatively higher recall values were observed even at high confidence thresholds, indicating that the applied model is better at recognizing insulator objects. In contrast, for the broken insulator class, higher recall values were achieved at lower confidence thresholds. However, as the confidence threshold increased, the recall dropped sharply, showing that the model struggles to distinguish broken insulators effectively. When both classes were evaluated together, the model demonstrated a strong average performance with a recall value of 0.96. This result indicates that the applied model performs better in the insulator class compared to the broken insulator class. To improve the relatively lower performance of the model in detecting broken insulators, future work could focus on increasing the dataset size and creating new classes to evaluate different types of insulators. Expanding the dataset and diversifying the insulator types for classification are expected to enhance the model's detection capabilities and overall accuracy in future studies.