Using Long Short-Term Memory (LSTM) networks with the toy model concept for compressible pulsatile flow metering

The metering of periodically oscillating pulsating flow has traditionally relied on eliminating pulsations, which is not possible for many systems such as internal combustion engines. Recent advances in the Deep Learning Suite of tools allow the extraction of useful information from complex signals acquired by inexpensive sensors. In this work, Long Short-Term Memory (LSTM) networks have been proposed to meter highly compressible pulsatile flow by learning the relationship between the average flow rate and the temporal patterns of standard orifice measurements. The model was built and evaluated with separate training and testing datasets that had different pulsation frequencies and waveforms. The waveforms were complex combinations of multiple harmonics, and several operating conditions involved reverse flow. LSTM predictions were improved by using the recently proposed 'Toy Model' concept that uses imprecise/incomplete physics to reduce dimensionality and add robustness. The effect of LSTM network architecture was explored. The use of LSTM networks for periodic time-series features that can be independently varied is a unique aspect of the work; associated advantages and problems have been discussed.