Embedding Sequence Model in STM32 Based Neuro-Controller

The problem of sensor data modeling, short and long term forecasting as well as the use of edge devices to operate and provide useful inferences is increasingly becoming a vital factor in modern embedded AI solutions. The paper represents the pipeline development of the sensor data processing, designing the sequence model with subsequent deployment on a S$T$M32F407VG microcontroller. To implement the intended study, we have considered an annual smart home temperature data which allowed us to conduct numerical experiments with the developed sequence models. In order to proceed with LSTM and GRU cells, which were chosen as a basis of the main sequence models, the data preparation and transformation process was carefully described. We opted to produce multistep generalization model using the popular deep learning approaches. Since STM32 is a memory limited microcontroller and can not hold large models, it is crucial to optimize the best model configurations - hyper-parameters set and architecture. Addressing this problem, we applied and investigated the tuning routine via genetic optimization metaheuristic. The best configured and validated model has been chosen as a candidate to be deployed on S$T$M32F407VG. The process of model conversion, serialization, deployment with X-CUBE-AI extension pack and inference step is presented and discussed. The obtained results and conclusions can be practically useful for researchers and embedded engineers who work with AI and microcontrollers of STM32 family.