PT2.210 - Towards fully hardware-based neuromorphic encoding for efficient vibration signal recognition
- Event
- EUROSENSORS XXXVI
2024-09-01 - 2024-09-04
Debrecen (Hungary) - Band
- Poster
- Chapter
- PT2 - Smart Systems and Artificial Intelligence in Sensing
- Author(s)
- T. Zeffer, T. N. Török, L. Pósa, F. Braun, J. Volk - HUN-REN Centre for Energy Research, Budapest (Hungary)
- Pages
- 265 - 266
- DOI
- 10.5162/EUROSENSORSXXXVI/PT2.210
- ISBN
- 978-3-910600-03-4
- Price
- free
Abstract
Neuromorphic signal processing can enhance the efficiency of IoT sensors and support edge compu-ting solutions. However, the preprocessing of the encoded signals, to be transferred to a spiking neu-ral network, requires high computational power. In this work, we propose an energy-efficient hard-ware-based solution for the analysis of rapidly changing vibration or acoustic signals. This was in-spired by the human cochlear implant, which exploits the plasticity of the human brain to enable clear speech recognition even on a very limited number of frequency channels (16-22). Our proposed hard-ware consists of a 16-channel frequency-selective MEMS cantilever array, and a VO² memristor nanogap based oscillator for amplitude sensitive spiking signal generation. To test our solution, we used Google Command Speech benchmark database.