The reliability of electricity supply systems is greatly dependent on the health of isolators, which are responsible for safe isolation of circuits and protection of equipment. Faulty isolators can cause catastrophic consequences such as widespread blackouts, fire hazard, and serious monetary losses, hence the need for effective and proper fault detection systems. We here propose a deep learning architecture for machine isolation of electrical isolators into normal and defective classes with the assistance of transfer learning. Five pre-trained convolutional neural networks, i.e., ResNet50, DenseNet121, MobileNetV2, EfficientNetB0, and VGG16, were fine-tuned and experimented on a handpicked two-class isolator image dataset. Results indicate that all the models yielded accuracies higher than 95 %, and VGG16 worked better with 97.92 % accuracy and F1-scores of 98 for both classes. For increased interpretability and building up operator confidence, the use of Gradient-weighted Class Activation Mapping was made useable to give a visualization of decision-making areas. This combination of high interpretability and accuracy makes sure that the suggested method is efficient but also transparent and therefore suitable for adoption in smart grid and predictive maintenance environments.