A Composite Sensor Calibration Algorithm Based on Deep Neural Network

Accurate vacuum measurement is critical in numerous fields. Aiming at the issues of nonlinear errors, temperature drift, and gas composition interference in the overlapping region of Pirani and piezoresistive composite vacuum sensors, this paper proposes an adaptive calibration algorithm based on a hysteresis-type deep neural network (H-DNN). Through analyzing the structure and principle of the composite sensor, it is found that the quadratic nonlinear characteristics of the piezoresistive sensor in the overlapping region significantly differ from the logarithmic response of the Pirani sensor. Moreover, a temperature fluctuation of 10°C can cause the error of the former to change by ±0.8% and that of the latter by ±1.2%, while gas composition changes can make the Pirani error reach up to 15%. To address these challenges, a four-level algorithm framework is designed, and a hysteresis-type deep neural network architecture for the overlapping segment is constructed. By employing a bidirectional feedback mechanism and multi-physical-field coupling modeling, environmental interference suppression and dynamic weight allocation are achieved. Experimental results show that this algorithm reduces the long-term stability error in the overlapping region from ±5.0% to ±0.8%. Full-range verification indicates that after calibration, the error in the overlapping region is ≤±1%, and the errors of the Pirani and piezoresistive sensors in the non-overlapping regions are reduced to ±3% and ±4%, respectively. This significantly enhances the reliability of composite sensors under complex working conditions and provides an efficient solution for wide-range vacuum measurement in fields such as semiconductor manufacturing and aerospace exploration.