Applying Edge AI to DC Arc Fault Detection (Part 3): Compiling and Deploying Models on MCUs (Download)
The first two installments of this article series discussed the challenges of managing DC arc faults, and how edge AI-enabled systems can help catch faults early. The training process as described in Part 2 included collecting data, picking a convolutional-neural-network (CNN) architecture, and using tools to create a model that can tell the difference between a real arc and normal switching noise.
The next challenge in the development process is the most challenging: Taking the model and preparing it to fit on a resource-constrained MCU with kilobytes of RAM. Additional requirements include programming the model to run in under a millisecond and not having it interfere with the system’s main control loop, which keeps the system stable.
