Bringing AI to Sensors – Simulation of Hardware-Aware AI Models

Future smart sensors will incorporate AI algorithms that can perform increasingly complex tasks. The movement from AI algorithms from powerful servers to limited devices, however, requires the development of customized hardware and efficient AI networks. A promising approach is the development of analog neuromorphic hardware, which can be both powerful and energy efficient. Contrary to exactly reproduceable software implementations, the usage of analog hardware implies a couple of effects like quantization, noise, saturation, and offset effects. This paper shows how such effects can be taken into account during simulation and later on for optimization of hardware/AI co-designed systems. Results of experiments using small networks for time-series vibration data analysis from an industrial application are presented.