부산대학교 도서관

During transformer operations, the production of particles in transformer oil is inevitable, which might jeopardize the transformer’s ability to operate safely; therefore, identifying pollutants in transformer oil is crucial. Here, we propose a recognition technique that combines lens-free holography with deep learning based on object detection algorithms. First, we capture the impurity image of transformer oil with a lens-free device and then reconstruct the image to enhance the feature extraction efficiency of the convolutional network. We validated this concept through experiments and trained the reconstructed impurity image of transformer oil using the YOLOv5 and Faster RCNN networks. Between these two network models, the YOLOv5 model showed better overall performance. Both networks improved the recognition performance of the reconstructed images, with YOLOv5 achieving the highest detection performance, with an F 1 value exceeding 73%. This technology can successfully identify and mark various pollutants in transformer oil; moreover, the experimental device is simple and can detect a large number of holograms of transformer oil samples in a short time. The average detection time for each image is only 15 ms. This quality detection method for transformer oil samples can be extended to other small particle detection applications.