A Novel Method for Removing Baseline Drifts in Multivariate Chemical Sensor

This article proposes a sensor signal processing method for removing baseline drifts in multimodal chemical sensors. Sensitivity and baseline drift are the primary reasons for inaccurate readings from low-cost sensors. The prediction errors from the state-of-the-art multivariate regression algorithms have been high due to the drifts. The response of a sensor is modeled as an autoregressive (AR) process associated with an observation zero-mean Gaussian noise. The parameters of the state-space model are learned using the calibration data. The variance of the Gaussian noise is estimated using an adaptive online method. The AR process equation and the observation equation are used to construct the adaptive Kalman filter. Each sensor response is passed through a separate Kalman filter, and then, a regression technique is used to predict the sensor response. The proposed model is applied to a standard air quality data set to demonstrate its efficacy in removing drifts in sensors.