Energy-efficient time series analysis with machine learning and deep learning on embedded computing platforms

The present Ph.D. thesis presents techniques and solutions for energy-efficient time-series analysis based on automated learning executed on resource-constrained, low-power computing platforms, with an interest in both Deep Learning and traditional, non-deep Machine Learning. This dissertation spans diverse domains, from algorithmic research on the accuracy-efficiency tradeoff in processing different biosignals to applied research inspired by industrial scenarios. The unifying methodology that brings all the research questions addressed in this thesis under the same perspective is the interest in time-series analysis as a task to be performed in the presence of the resource constraints characteristic of low-power edge computing devices. This dissertation covers the three major types of automated learning tasks: binary classification, multi-class (single-label) classification, and regression. Starting from binary classification, this work presents a proximity sensor for active safety in industrial machinery and a setup for epilepsy detection from intracranial electroencephalography. Both solutions are based on a Temporal Convolutional Network (TCN) executed on an embedded MCU. Moving to multi-class (single-label) classification and regression, the research addresses hand modeling from the surface electromyographic (sEMG) signal. Starting with off-device TCNs for the recognition of discrete hand gestures, the classification setup is advanced via deployment on a multi-core MCU and heuristics for unsupervised adaptation to arm posture. Then, regression was addressed for a more versatile control of Human-Machine Interfaces (HMIs). Developing an embedded TCN accurate in hand kinematics estimation, I addressed the modeling of hand kinematics and force with event-based features, which are computationally cheaper and promising for future porting onto event-driven devices with reduced latency and energy consumption. The sEMG contributions advance the field of non-invasive intuitive wearable HMIs. This research work proves the success of the embedded approach to time-series Machine Learning, achieving SoA accuracy and efficiency and proving promising for impactful applications in the industrial, clinical, and consumer domains.