A conditional GAN and dual-channel hybrid deep feature framework for robust sensor fault detection in WSNs

Abstract

Sensor-generated data is vital to the operation of numerous systems and services in the rapidly growing field of the Internet of Things. Wireless Sensor Networks, as an essential setup for these systems, are frequently deployed in large, diverse, and often harsh environments. However, these networks are highly vulnerable to various faults, potentially leading to improper data transmission, reliability, and financial stability of the systems. To address these challenges, we propose a hybrid model for sensor fault detection that integrates a machine learning classifier with the deep learning (DL) model, specifically VGG-16 and ResNet-50. Synthetic samples are generated using a Conditional Generative Adversarial Network and common sensor faults, such as hardover, drift, spike, erratic, and stuck fault are introduced by leveraging a publicly available temperature sensor dataset. Time-series data is transformed into Gramian Angular Field images, from which deep features are extracted using VGG-16 and ResNet-50. These extracted features are then fused to form a hybrid feature pool. Our framework effectively addresses problems related to data imbalance and enhances accuracy. The proposed model outperforms the individual feature sets, VGG-16 (89.22%) and ResNet-50 (84.21%), achieving notable accuracy of 92.55% with the fused feature set, underscoring its potential for robust sensor fault detection.