A sensor-less stroke detection technique for linear refrigeration compressors using artificial neural network

Linear compressors are very attractive for domestic refrigeration due to elimination of crank mechanism, high efficiency and compactness compared with conventional compressors. The significance of stroke control in a linear compressor not only lies in avoiding the piston collision into the cylinder head but also enabling cooling capacity modulation. Predicting piston stroke without a displacement sensor reduces the cost and facilitates the stroke control especially in miniature linear compressor where there is very limited space for installing sensors. This paper reports an artificial neural network (ANN) based stroke detection approach that can be used in linear compressors and any other linear (free-piston) machines. Experimental tests were conducted in a novel linear compressor driven refrigeration system to sample and record voltage, current and displacement. Fast Fourier transform (FFT) analysis was performed on current and voltage signals to extract harmonic terms as inputs of the neural network model to predict the stroke. Six cases with different numbers of harmonic term for current and voltage were compared. Both the mean squared errors and correlation coefficients are significantly improved with the increase of harmonic terms from one to three. However, small difference is indicated between the cases with three and six terms. Best percentage error distribution was achieved in the case with six harmonic terms with the majority of percentage errors falling within ±0.7% and a maximum percentage error of 3.5%. It can be concluded that the present ANN based stroke prediction approach can be effectively adopted for linear compressors without expensive displacement sensors. This is a key step towards the commercialization of linear refrigeration compressor technologies.