Toward Chiller Condenser Fouling Factor Prediction using Spiking EdgeAI

In this study, a recurrent spiking neural network (RSNN) has been implemented on an accelerator deployed on a reconfigurable circuit (FPGA) to predict the Condenser Fouling Factor (CFF) in a chiller as a predictive maintenance (PdM) measure. This system operates through a low-power device utilising edge computing, geared towards AIoT/EdgeAI applications. It can accurately identify CFF levels that exceed 25% based on several pressure and temperature sensor data distributed in the chiller, achieving an accuracy greater than 90% with an architecture of 256 parallel and recurrent neurons in the hidden layer. The model was trained using a proprietary dataset that recorded sensor states during controlled experiments where the condenser was manually obstructed at varying coverage percentages. The primary advantage of employing RSNN techniques lies in their dual capability: first, they are designed to detect temporal signal patterns, and second, their trainability allows for adaptation to various applications across different contexts. The training for the particular accelerator used in this work is done on the FPGA, not requiring power-hungry machines.