Incipient fault diagnosis of cable joints with data augmentation and neural network models

This paper proposes a data-driven method for diagnosing incipient faults in cable joints using a fully connected neural network. The method utilizes feature gas concentrations to identify fault types. A data augmentation technique based on the mean shift algorithm was employed to generate simulated data, enhancing the training dataset. The fully connected network, trained with this augmented data, achieved fault recognition accuracies of 87.5% and 82.5% on simulated and experimental baseline data, respectively. Comparative analysis with KNN and SVM models demonstrates the superior performance and generalization ability of the proposed approach, showcasing its potential in improving early fault diagnosis of cable joints.