DESIGN OF A FIELD-PROGRAMMABLE GATE ARRAY PLATFORM FOR ANIMAL VOCALIZATION CLASSIFICATION UTILIZING A CONVOLUTIONAL NEURAL NETWORK

Studies such as Manteuffel et al., (2004) and Laurjis et al., (2021) have established links between animal vocalizations and their wellbeing, including emotional states and respiratory illness. Zhao et al., (2016) have proven Field-Programmable Gate Arrays (FPGAs) to have a higher computational speed than CPU based system (four times faster), and a greater energy efficiency than GPU based systems (7.5 times reduction in power consumption) for Convolutional Neural Network (CNN) applications. The purpose of this study is to investigate the ability to implement an animal vocalization classification CNN on a custom wearable, low power (up to 123 mW), FPGA system. Specifically, this study focuses on the design of the custom hardware, and the identification of vocalizations originating from animals with respiratory illness. The CNN was trained on a dataset of one-second-long recordings of healthy and sick chickens, with the goal of detecting vocalizations from sick chickens. The performance of the vocalization CNN was evaluated on both CPU based and FPGA based systems. The testing accuracy of the CNN on a CPU system was 92.4%. The accuracy on a commercially available FPGA system was 84.2%, with a recall of 96.7% and a precision of 78.7%. The custom FPGA board was unable to successfully detect any vocalizations. The power consumption of the FPGA based systems were significantly reduced from the CPU based system; ~8mW for the FPGA systems, compared to ~95W for the CPU system (Intel, 2018).