Compression at the Edge for Energy- and Bandwidth-Efficient Industrial Audio Analysis
Yinze Li, Zhaoyi Liu, Stefan Lommaert, Friedrich Wolf, David C. Newton, Sam Michiels, Danny Hughes · Procedia Computer Science · 2025
The rapid growth of sensor-based monitoring in Industry 4.0 has led to a substantial increase in the volume of data generated by industrial systems. This includes various data types, among which acoustic data is especially common due to its effectiveness in capturing machine behaviour and operational conditions. However, transmitting and storing acoustic data under resource-constrained conditions presents significant challenges. Audio compression provides a practical solution to reduce data volumes in low-power and low-bandwidth environments. This paper investigates the effectiveness of audio compression techniques in industrial anomalous sound detection and sound event segmentation. For anomalous sound detection, we achieve model performance within 3% of the baseline, saving 82% bandwidth. For sound event segmentation, we achieve model performance within 0.2% of the baseline, saving 91% of bandwidth. Furthermore, the evaluation on the hardware platform shows the power consumption can be reduced by 95% by using an LTE-M instead of a normal LTE modem, which is only possible using edge compression.