Abstract
Outdoor high voltage (HV) insulators are essential components for entire power grids. They are utilized to provide mechanical structural support and isolate the grounded towers from the live conductors. Insulators are exposed to numerous environmental pollutions such as soluble and insoluble contaminations, along with natural pollutions like dust and snow that could lead to poor performance or malfunction of such insulators. To address this issue, deep learning-based image processing tools can be used to detect pollutants efficiently and accurately from insulator surface images. This approach generally follows a common methodology: 1) collect images for various types of insulators and under different surface conditions; and 2) train a convolutional neural network (CNN) to classify images-and this training is without any consideration being made on the insulator type. Fitting one CNN for all insulator types could be problematic since the appearance of an insulator generally depends on its type and the appearance of the pollutants on its surface thereof. Subsequently, in the process of data collection, some types of insulators could be underrepresented (i.e., data being imbalanced with respect to the insulator type), and this, in turn, could lead to inaccurate classification for those insulators. In contrast with this common approach, in this study, we show that by stratifying the data based on the insulator materials type, which is generally known in advance, and then training insulator-specific CNNs for surface condition classification, one can considerably boost the classification performance. In particular, our empirical results show that the training insulator-specific CNNs lead to an 8-10%improvement in classification accuracy when compared with a single CNN that is trained for all insulator types.