TransCNN: Deep Hybrid Model for Effective Intermittent Fault Diagnosis in Wireless Sensor Networks

Abstract

Intermittent fault diagnosis (IFD) is a vital subject in modern communication and computing networks. Unlike permanent problems, which are simpler to discover and localize, intermittent defects appear irregularly, disappear autonomously, and frequently resemble natural network variations. This renders them more difficult to track, yet significantly disruptive. Modern deep learning methods, such as CNN (Convolutional Neural Network) and LSTMs and their variations, sometimes don’t take long-term dependencies or localized spatial anomalies into consideration. This causes a lot of false alarms and makes the accuracy go down. The lack of annotated intermittent failure datasets and their irregularity make it much harder to train and generalize models. To solve these problems and make dataset-related problems less of a problem, there is a need to create a hybrid strategy that combines Transformer-based temporal modelling with CNN-based spatial feature extraction and Conditional Tabular Generative Adversarial Networks (CTGAN)-based synthetic data augmentation. The proposed method meets the immediate need for a precise, scalable, and robust IFD solution that works in real time in mission-critical Wireless Sensor Network (WSN) applications, that cuts down on false alarms, and that makes the network more reliable. Meticulous preprocessing steps, including normalization, window sliding, and balancing, ensured that the model training received optimal input. The enhanced datasets indicate that the proposed model significantly outperforms CNN, CNN-LSTM, and Convolutional Autoencoder. The model demonstrates an accuracy of 98.3%, a precision of 98.7%, a recall of 98%, and an F1-score of 98.3%.