An Improved Sensor Calibration with Anomaly Detection and Removal

Sensor calibration is a widely adopted process for improving data quality of low-cost sensors. However, such a process may not address data issues caused by anomalies. Anomalies are considered as data errors that are inconsistent to the actual physical phenomena. This paper presents an improved sensor calibration, which applies a process for detection and removal of anomalies before the sensor calibration process. A Bayesian-based method is used for anomaly detection that takes advantage of cross-sensitive parameters in a sensor array. The method utilises dependencies between cross-sensitive parameters, which allows underlying physical phenomena to be modelled and anomalies to be detected. The calibration approach is based on stepwise regression, which automatically and systematically selects suitable supporting parameters for a calibration function. The evaluation for anomaly detection shows that the results are better than the state-of-the-art methods, in terms of accuracy, precision and completeness. The overall evaluation confirms that data quality can be further enhanced when anomalies are removed before the calibration.