Andreas Dickow, Klemens Graf, and Gregor Feiertag
Pressure Sensor Calibration, Temperature Control, Neural Networks, System Identification
Calibration of piezoresistive pressure sensors requires an extensive acquisition of sensor characteristics. Calibration devices have to fulfill high standards of temperature and pressure accuracy and trajectory speed. While pressure control performs quite well, state-of-the-art devices show poor temperature control performance, due to their retarding and nonlinear temperature dynamics. To overcome this problem, the standard industrial PID temperature controller can be replaced by more sophisticated control methods. The vast majority of these control methods, such as model predictive control, require a model of the device’s temperature behavior. Black-box models can deliver all necessary information for model-based control applications. This paper focuses on black-box model identification of a calibration device’s temperature behavior. A single-input-single-output (SISO) model of the nonlinear temperature dynamics is learned by a recurrent neural-network (RNN) with Long-short-term-memory (LSTM) using the resilient propagation (RPROP−) training method. The method is applied to a real calibration fixture, which uses convective cooled Peltier-elements for temperature control. The use of LSTM leads to a timeseries-reproduction error, which is more than 10 times smaller, compared to a RNN having tanh activation functions instead of LSTM.
Important Links:
Go Back
