High-accuracy health prediction of sensor systems using improved relevant vector-machine ensemble regression

Abstract

Abstract Sensor systems have been used widely in many fields. However, sensors are prone to faults, which greatly reduce the performance of the trained pattern-recognition model. To improve the reliability and stability of the sensor system, it is essential to apply prognostics and health management to the sensor system. A novel health-prediction model of the sensor system is established based on the unascertained deep soft sensor (UDSS) and relevant vector-machine ensemble (RVME). The first step in health prediction is to extract the performance variables. Based on unascertained mathematics and the deep belief network, a novel UDSS is proposed to extract the performance variables, which are called the health reliability degree (HRD). The HRD is applied as the input of health prediction. The second step is to establish an appropriate predictor. Bagging is used as the framework, the relevant vector machine is used as the weak learner, and RVME is utilized to structure continuous single-step or multiple-step health predictions. To verify the effectiveness, the proposed method is applied to a gas-sensor system. An experimental gas-monitoring system is designed and developed to obtain sufficient experimental data. The simulation result demonstrates that compared to other methods, the proposed method has a lower average relative error of 0.60%.