Implementation of machine learning algorithm in embedded devices

This paper describes the usage of neural networks in microcontrollers for deployment in embedded devices. The issue is focused on the design of a suitable neural network, its optimization and deployment in a 32-bit microcontroller with regards to the limiting factors of the chosen microcontroller. The introductory part of the article is a description of the used technology and hardware on which the solution will be implemented. Accelerometer motion recognition was chosen as a practical application. The proposed solution recognizes 6 basic movements, respectively movement in three axes. Tensorflow and Keras frameworks were used to design and implement a neural network. The created neural network model was after optimization implemented in the firmware of the STM32L4x microcontroller. The proposed solution implements automatic motion detection and its subsequent classification. The proposed principle can be applied to a group of sensors connected to the available interfaces of the microcontroller. Application with an accelerometer can be used to detect specific vibrations, application with MEMS microphones can be used to detect specific sound patterns that indicate a possible fault condition of the monitored device in industry.