In the rapidly growing realm of the Internet of Things (IoT), reliance on sensor-generated data has become crucial for the operation of multiple services and systems. As essential components of these systems, wireless sensor networks (WSNs) are installed in a wide range of diverse and often harsh environments. However, these networks are highly prone to a range of faults, including software bugs, communication failures, and hardware malfunctions. Such issues can lead data to data being transmitted incorrectly, endangering the security, reliability, and economic stability of the systems they support. Addressing the challenge of sensor fault detection, we propose a novel hybrid technique to enhance the classification of sensor fault data in WSNs. Our method leverages a publicly available dataset of temperature sensor readings to generate synthetic data by using a conditional generative adversarial networks. These synthetic samples closely resemble common temperature sensor data despite the introduction of artificial sensor faults in WSNs, including hardover, drift, spike, erratic, and stuck faults. In order to capture the temporal dependencies in time-series data, we transform the sensor readings into Gramian Angular Field images (GAF), retaining the temporal structure. These GAF images are then processed using a convolutional autoencoder to extract rich feature representations, followed by a three-layer artificial neural network for the multi-class classification of sensor faults. Our proposed method not only addresses the challenges of data scarcity and imbalance but also enhances accuracy in sensor fault detection. The proposed method demonstrates high accuracy, F1-score, recall, and sensitivity, achieving 95.93%, 95.84%, 95.88%, and 95.88% respectively.
