Optimized Low-Power Mixed-Signal Sensor Systems: Improved Cross-Domain Approach to Maintain Classification Accuracy

Optimizing the mixed-signal Systems-on-Chips (SoCs) is a difficult task, particularly while including analog developing blocks and Machine Learning (ML)-based algorithms that make overall performance hard to predict. These designs for minimizing energy consumption increases the lifetime from harvested energy from the environment. In this manuscript, a method is developed for optimizing difficult mixed-signal sensor system that includes three phases. Initially, Pareto-optimal local performance of individual circuits is modelled, developing the macro-model of mixed-signal system performance involving local nonidealities and identifying optimum group of system-level parameters utilizing numerical optimization. The method is employed for Electrocardiogram (ECG) sensor system with integrated arrhythmia classification with minimizing system energy consumption and constraints on inference accuracy. Various numerical optimization algorithms are compared to evaluate its performance for minimizing task are Bayesian Optimization and Gradient-based optimization. The experimental results obtained optimization in reasonable execution time of 9.5 h while compared to existing algorithms on complex mixed-signal system.