Transfer Learning with Domain Adaptation for Unlabelled Sensor Faulty Data Classification

In the data analytics based IoT era, reliable sensor data is essential for maintaining the integrity and effectiveness of IoT applications. The reliability of sensor data can be compromised due to faults that that occur in the sensors and this faulty sensor signals can lead to performance degradation in IoT enabled systems. This study investigates a transfer learning method augmented by a domain adaptation technique for sensor fault classification. The domain adaptation technique allows the classifier model to accurately identify two types of sensor faults, drift and bias fault, in a target sensor node with an unlabeled dataset. We trained an ANN classifier using a semi-supervised approach, leveraging labeled samples from the source sensor node alongside unlabeled data from the target sensor. In the current work, both the source and target sensors are of the same type but are located in different locations. Results from two experiments demonstrate that our approach achieves approximately 97% accuracy when the target sensor is indoor and the source sensor is outdoor. When both sensors are located outdoors, the accuracy exceeds 99%. Other classification metrics also remain consistent. This approach not only maximizes the use of available data but also offers a robust solution for scenarios where obtaining labeled data is difficult or missing.