DEVELOPMENT OF NEURO-CONTROLLER BASED ON STM 32

Abstract

A step-by-step approach for the development of neuro-controller system is considered in this work. In order to utilize STM 32 microcontroller as a hardware solution in the meaning of edge computing we propose to deploy a deep learning model for the tasks of prediction and anomaly detection. The data which are used for the training phase as well as for the inference consist of the smart home indoor temperature time series with the time related patterns. The developed deep learning LSTM model automatically discovers time dependencies during the training and uses the time series features for walk-forward validation. The data preparation stage followed by the re-definition of time series forecasting to supervised learning problem is presented. To calculate a set of the optimal LSTM hyper-parameters the grid search algorithm that implies a step-wise full training neural network process has been applied. Trained network was deployed and tested on STM 32 microcontroller with installed X-CUBE-AI expansion package. The obtained validation and test results show a possibility of using such approach for the development and improvement of the LSTM architectures for edge and mist computing applications.

Key words: edge computing, deep learning, LSTM, smart home, STM 32.