Fast-RCNN Coupled Four-Dense Layered Deep Fully Connected Neural Network-Based Insulator Chain Defect Defection

Electric power lines may be at risk of a safety hazard as a result of defective insulators. Image detection technique can significantly save maintenance costs and increase the effectiveness of Insulator Defect Detection. Nevertheless, limited precision and a lengthy detection process are drawbacks of the present insulator fault detection methods. For images with complex backgrounds, it is challenging to detect insulator faults using the conventional methods since they focus on minimal edge detection from images and classifier design. To address this issue, this paper recommends fast-RCNN coupled four-dense layered deep fully connected neural network (FR-4DCNN) that detects the insulator defects with high accuracy using the Insulator Defect Detection dataset from KAGGLE. The proposed FR-4DCNN model uses Insulator Defect Detection dataset from KAGGLE with 1800 insulator images. The insulator images are preprocessed with convolutional layers for feature map creation followed by the formation of region proposal network that detects the insulators by bounding box regressor algorithm. The classified insulator is then fed into ROI pooling coupled with four-dense hidden layered deep fully connected neural network having single input and output layer that predicts the ROI for classification of the insulator defect types for each ROI using the bounding box regressor method. The novelty of the proposed FR-4DCNN exists in the classification of the insulator defects in the form of missing plates in the hanging or attached insulator chain, broking sheds in insulators and the presence of rust in insulators. The dataset for Insulator Defect Detection was divided into training and testing data, and the training data were fitted to both the proposed FR-4DCNN model and other deep learning models in order to compare the efficiency. Results of execution indicate that the proposed FR-4DCNN model showcases the accuracy of 99.47%, precision of 99.42%, recall of 99.25%, and F1-score of 99.37% when compared with existing CNN models.