Soft-sensing for compressor test time reduction with time-delay neural networks.

This study proposes a soft-sensor-based method to significantly shorten compressor performance evaluation tests and presents the results of its industrial application over five years by a compressor manufacturer. Traditional approaches demand long testing times to reach steady-state conditions, with overall test durations frequently surpassing two hours. In this work, a soft sensor based on a time-delay neural network (TDNN) was developed to monitor steady-state conditions of key performance parameters - cooling capacity, power consumption, and coefficient of performance - and to predict their final values. A dataset of 392 compressor evaluations was used for model development, and the proposed method achieved an overall reduction in test duration of close to 50%. This is accomplished because the proposed method incorporates a soft-sensing approach, trained on historical data, facilitating early detection of steady-state conditions and accelerating testing procedures. During five years of real industrial application, with the proposed approach tested in 9184 performance evaluations, this method demonstrated a 55% improvement in total test time, with more than 95% of the tests showing prediction errors below 2%. Therefore, the proposed tools resulted in consistent time savings and increased operational efficiency during their industrial evaluation.