Enhancing IoT Sensors Precision Through Sensor Drift Calibration With Variational Autoencoder

IoT sensors are made of physical materials, and due to natural decay in materials, sensor data drifts over time. Even though sensors are calibrated after deploying at the site, the accumulation of errors in sensor measurements due to sensor drifts renders the data progressively irrelevant, creating significant issues for end applications. In this article, we propose a software-driven drift detection and calibration framework based on probabilistic observation in latent space using variational autoencoders (VAEs). The proposed method utilizes the latent distribution of the generative model from sampled observational data, which are collected during the calibration phase of the deployed sensors. Variational inference in VAEs is employed to approximate the true posterior distribution for detecting sensor drifts, incorporating metrics such as Kullback-Leibler (KL) divergence. Additionally, reconstruction loss is utilized for calibrating the sensors. Both simulated and real-world sensor data are used to evaluate the proposed method. Experimental results demonstrate significant improvement over existing drift detection and calibration techniques.