Energy-Efficient Serialized Walsh-Hadamard Transform Based Feature-Extraction for Information-Aware Compressive Sensing

The performance of traditional compressive sensing (CS) architectures has been tempered by dynamically changing real-world data. This paper demonstrates an information-aware compressive sensing (CS) architecture for dynamic artifact detection of biophysiological signals in wearable applications. Artifacts such as long pause, baseline wandering, and saturation often corrupt recorded data due to environmental factors. In wearable applications where power conservation and ultra-low power operation are paramount, this can lead to wasted power. By combining earlier proposed CS based architectures with an efficient analog feature-extraction (FE) and digital decision making, the sampling rate of the ADC and the integration window of the multiplying DAC can be reduced in presence of artifacts to save power. As shown, this technique can reduce the system power consumption by up to 70% in the more extreme cases of signal corruption. A serialized Walsh-Hadamard Transform (WHT) used for FE is proposed that dramatically simplifies the circuit implementation while the digital classifier comprising of quadratic Support Vector Machine (SVM) classifier ensures low power operation with accurate decision outcomes.