Enabling System Support for General-Purpose Sensing in Smart Environments

The Internet-of-Things (IoT) sensing systems have the potential to revolutionize our living environments, yet their transformative potential remains largely unrealized.  Despite the rapid proliferation of IoT devices and their immense potential for a range of applications like building maintenance and healthcare monitoring, their integration into real-world environments faces significant hurdles due to practical deployment challenges and escalating privacy concerns. 

Current IoT sensing systems are typically built with monolithic, purpose-specific architectures that focus on a limited range of sensing capabilities designed for specific  applications.   This  results  in  isolated,  vendor-controlled  solutions  with  limited features to support diverse application requirements for machine learning (ML),scale, and reliability. As a result, IoT ecosystems become fragmented, which hinders both widespread adoption and long-term viability. 

To address these limitations of current IoT systems, this thesis proposes a shift to-wards general-purpose sensing systems that support current and future applications, adapt to evolving stakeholder needs, and provide robust privacy safeguards. This the-sis introduces several novel system design approaches to achieve this vision. Starting with Mites, a scalable, general-purpose sensing platform that delivers fine-grained environmental data and establishes the foundational architecture for extensible and adaptable IoT deployments across various application scenarios.  Building on this, MLIoT is presented as an end-to-end general-purpose machine learning system de-signed to transform raw sensor data into high-level inferences, supporting the entire ML lifecycle for IoT applications.  To further enhance the interpretability of these inferences, TAO, a context recognition framework, is developed to detect semantically meaningful contexts from the inference, improving understanding and usability agnostic to the underlying ML inference pipelines.  Complementing these advancements, Kirigami showcases a general-purpose edge audio speech filter that removes human speech segments while preserving other sounds,  thereby maintaining high accuracy for non-speech inferences and balancing privacy with utility.  The thesis demonstrates how comprehensive system support for general-purpose sensing facilitates various applications and meets the needs of diverse stakeholders through the real-world deployment of more than 300 multimodal sensor devices in a fully occupied, five-story university building at Carnegie Mellon University (CMU). Through these innovative system design approaches,  this thesis advocates a transformative shift towards scalable, privacy-preserving, and general-purpose IoT sensing systems, unlocking the full potential of smart environments.