Russian Federation
A critical analysis of modern methods of energy-efficient machine learning for embedded acoustic-vibration signal processing systems is presented. Hardware-algorithmic architectures based on 32-bit ARM Cortex-M microcontrollers (STM32U5, STM32H7, nRF52840), neural network quantization and pruning methods, and software tools TensorFlow Lite for Microcontrollers and CMSIS-NN are examined. Practical implementations demonstrating energy consumption of 50-150 mJ per inference cycle with classification accuracy >95% on ESC-10/ESC-50 datasets are analyzed. Limitations of current approaches are identified and promising research directions in the field of neuromorphic computing and edge distributed learning are defined.
tiny machine learning, embedded systems, energy efficiency, neural network quantization, acoustic signal processing, STM32U5, TensorFlow Lite, CMSIS-NN
1. Issledovanie sposobov zaschity informacii ot utechki po akusticheskomu i vibroakusticheskomu kanalam / V. K. Zol'nikov, S. A. Sazonova, A. I. Zarevich, S. S. Bashun // Modelirovanie sistem i processov. – 2025. – T. 18, № 2. – S. 27-40. – DOIhttps://doi.org/10.12737/2219-0767-2025-18-2-27-40. – EDN GBWORY.
2. Primenenie neyronnyh setey dlya optimizacii energopotrebleniya SBIS / A. V. Achkasov, A. S. Yagodkin, F. V. Makarenko, N. Yu. Zalenskaya // Modelirovanie sistem i processov. – 2025. – T. 18, № 1. – S. 7-16. – DOIhttps://doi.org/10.12737/2219-0767-2025-7-16. – EDN MQXMHM.
3. Sazonova, S. A. Issledovanie i razrabotka modeley proektirovaniya mikroshem cifrovoy obrabotki informacii / S. A. Sazonova, K. V. Zol'nikov, V. F. Asminin // Modelirovanie sistem i processov. – 2025. – T. 18, № 3. – S. 89-99. – DOIhttps://doi.org/10.12737/2219-0767-2025-18-3-89-99. – EDN LGIGCI.
