Sensor fault diagnosis and calibration based on voting mechanism for online application using virtual in-situ calibration and time series prediction

Sensors are essential components in building energy control systems. Sensor fault can result in inappropriate control, thereby increasing energy consumption or discomfort. This study proposes a novel method that combines virtual in-situ calibration and time series prediction (VIC-TSP) to diagnose and calibrate sensor faults for online application to guarantee data accuracy. The method is applied to an actual heating, ventilation, and air conditioning system for the real-time comparison of residuals from measurement, calibration, and prediction values. Subsequently, sensor faults are diagnosed and calibrated via a voting mechanism. The results indicate the following: (1) Faults in the measurement values are identified by discrepancies between the residuals of the measurement and calibration predictions. After determining the measurement value faults, performing virtualization can decrease residuals by more than 73.61 %. (2) Calibration and prediction value faults indicate residuals that exceed predefined thresholds. A retraining interval of one week reduces the calibration and prediction residuals by more than 81.63 % and 78.82 %, respectively. (3) The VIC-TSP method can reduce pump energy consumption by 10 % and increase the adjustment frequency to the supply fan by 9.83 times per day.