Abstract:
With recent advancements in AI algorithms, data processing for AI applications is primarily done in a cloud-based data center with deep-learning models that require massive computing capacity. The decision-making process happening at a centralized level in these cloud-based AI has caused massive traffic at the Cloud resulting in scalability issues, excessive power consumption, connectivity, and latency. By bringing the AI computation as close as possible to the sensor (Edge-AI), data bandwidth can be minimized, system scalability and responsiveness can be improved, and real-time requirements can be fulfilled. With Edge-AI, intelligence can be distributed, and inference can now happen at the edge, which diminishes the measure of network traffic streaming back to the Cloud with the reaction time for IoT gadgets being cut to a minimum as well as management decisions will be available on-premise, which is near the devices which can bring many advantages. With the help of STM32CubeAI (solution of ST Microelectronics for AI at the edge), we can implement Edge AI on an STM32-bit microcontroller. In our research work, we have developed a deep learning model for Direction-of-Arrival (DOA) estimation of signals having varying SNRs. The DOA estimation is widely used in 5G wireless communication. Using STM32CubeAI, the pre-trained neural network for DOA estimation has been mapped to STM32 Microcontroller & it is able to classify the DOA (ranging from 0 to 345 degrees) of a signal having SNR in the range of 0 – 20 dB with greater than 90% accuracy. We are also developing an improved version of python-based GUI, which can interact with the microcontroller & show the results of classification, including accuracy for generated test signals in front of the user's screen for the demo purpose of this application.
