Artificial neural network-based online stroke detection for CO2 linear refrigeration compressors

Abstract

CO₂ linear compressors are critical for sustainable and energy-efficient refrigeration systems due to the eco-friendly properties of CO₂. However, the unique characteristics of CO₂ compressors introduce significant challenges in piston stroke control. The large pressure difference between suction and discharge conditions requires high operating currents to overcome gas forces, resulting in substantial piston offsets. These offsets interact with nonlinear parameter variations, elevating the risk of resonant frequency shifts and potential valve collisions. Accurate piston stroke measurement is essential to address these issues, but traditional methods relying on displacement sensors are costly. This study presents an innovative artificial neural network (ANN) method for sensorless piston stroke measurement in CO₂ linear compressors. The proposed model requires only six inputs: voltage, current, frequency, active power, suction pressure, and discharge pressure. Optimized ANN parameters enable high prediction accuracy, with an average R² of 0.955, RMSE of 0.206, and an average error of 2.24 % on the testing set. Furthermore, a simple stroke adjustment method based on the ANN model is proposed, allowing for effective stroke control and natural frequency calculation.