ABSTRACT
Design for human-borne sensing faces a key challenge: to provide increasingly high-quality, day-by-day sensing accuracy and reporting from an energy-constrained and aggressively miniaturized computing form factor. Long-term maintenance-free operation is an another important goal for devices intended to be carried by people throughout their daily life. The human sensor form factor is driven by its energy storage requirements, hence power consumption resulting from data sensing, processing, and communication.
This thesis studies the energy costs in the full end-to-end human sensor platform, however specific attention is paid to optimizing energy use in the worn sensor device. Three computing layers comprising the human sensor platform are examined: the human sensor device, the mobile data aggregator, including smart phone and smart watch, and cloud-side data warehousing.
The heterogeneous compute and energy capacity qualities of the layers are exploited for both intra-layer and cross-layer improvements in energy efficiency. Opportunities to offload power consumption from the sensor device, thus enabling smaller battery capacity and further scaling of sensor device form factor are prioritized.
The full data handling flow, including data sensing, data cleaning, feature extraction and classification, data communications and storage, is considered, and tradeoffs between computed result accuracy and energy cost are tailored across a range of applications.
Wearable human sensor applications implemented and reported on in this thesis include mobile online gait analysis for runners, grocery store aisle localization with augmented reality driven item recommendation, and wearable in-field electroencephalographic brain sensing.
Results include improvements in energy-efficiency over the state-of-the-art, including an 11X speedup in cloud data processing, a 47% power reduction in a wearable running sensor when applying a smartphone-to-wearable collaboration, and, most significantly, a one-order-of-magnitude power reduction when applying an event-driven sparse adaptive sampling method to a wearable human running gait analysis sensor.
BACKGROUND
Wearables are a leading category in the Internet of Things. Compared with mainstream mobile phones, wearables target one order of magnitude form factor reduction, and offer the potential of providing ubiquitous, personalized services to end users. Aggressive reduction in size imposes serious limits on battery capacity. Wearables are equipped with a range of sensors, such as accelerometers and gyroscopes.
Related Work:
The concept of wearable technology is not entirely new-people started wearing electronic watches back in the 1980s. Since late 2000, the technology development and market penetration of wearables have experienced astonishing growth, fostered by several technology drivers. First, low-power semiconductor technology is the key enabler for aggressive form factor scaling. Wearable devices are powered by batteries.
A Real-world Wearable Sports Monitor. System Teardown Photo (Left), and Generalized Wearable System Archetype (Right).
Wearable Sensing System Architecture:
The teardown of a recently released real-world wearable sensing product. Worn on the wrist or ankle, the device senses and analyzes the athlete’s performance, records training histories, and provides real-time coaching. It uses a range of MEMS IMUs for high-precision motion sensing and recognition in sports, such as running, boxing, swimming, and cycling.
Sensing Power Consumption Sources:
This section studies the energy characteristics of wearable data sensing and analysis flow. The key energy contributor of each of the phases is identified, and energy optimization opportunities are explored.
LOW-POWER ALGORITHM DESIGN FOR HETEROGENEOUS CLOUD ARCHITECTURES
The mainstream hierarchical design methodology of modern VLSI CAD is to separate the design flow into a sequence of design optimization steps ranging in abstraction from system-level design exploration down to the physical design. Reusing prior work reduces the complexity of each design step, and abstraction makes early-stage design optimization feasible without being overwhelmed by the low-level design details. However, with the increasing role of physical effects such as interconnects, process variation, and power and thermal profiles in the final design’s cost, this separation between layers makes the overall CAD process more difficult and error prone.
Parallel Cross-Layer Optimization:
In this work, we deliver a parallel algorithm solution for cross-layer power optimization of unified high-level and physical synthesis using a non-deterministic transactional model. Through the cross-layer optimization framework, we globally optimize the decisions made in the individual layers to produce an IC holistically minimized for power.
Non-deterministic Transactional Algorithm for Unified High-level and Physical Synthesis.
Mapping of Unified Cross-Layer Optimization to Heterogeneous Architectures:
In this section, we discuss how the parallel cross-layer optimization technique is applied to
a heterogeneous computing system, which is composed of a multicore CPU and multiple high performance Nvidia GPUs. We will discuss in detail the appropriate programming considerations encountered and their corresponding performance tradeoffs.
Experimental Results:
In this section, we present the results of the parallel high-level and physical level synthesis.
We implemented both a sequential version and the proposed parallel version in the C++ programming language and CUDA SDK, and will compare their running time and performance. We further evaluate the proposed GPU floor planner with local and global convergence test against a traditional SA GPU floor planner.
MOBILE AND CLOUD BASED HUMAN LOCALIZATION TO SUPPORT AUGMENTED REALITY
System Overview:
Our system consists of an external image labeling service, a mobile component, and a remote cloud server component. To determine the initial location of a user in the grocery store, the mobile component sends a product snapshot to the cloud server component, which forwards that snapshot to an external image labeling service.
This external image labeling service returns the product identity to the cloud server, which then determines the current location (aisle) of the user by referring to an indoor layout of the grocery store. After determining the identity of the aisle in the grocery store, the mobile component estimates user motion, thus providing a position estimate within the aisle, as well as orientation.
System Architecture.
Experimental Results:
In-person Survey Design
In order to validate our system we collected in-person feedback from 15 users. These users provided us feedback in a couple of ways: first, they were asked to take an online survey so we could collect some basic demographics and information about their shopping needs and habits and any specific health/dietary restrictions. Second, users did an in-person survey with the researcher, after having accompanied the researcher while shopping in a grocery store for one hour and using our system on an Android phone.
WEARABLE SENSOR SIZE, WEIGHT, AND POWER (SWAP) ANALYSIS
The current state-of-the-art in physiological monitoring solutions are ill-positioned: relying
either on bio-sensors that measure intrinsically low-dimensional or sparse data (e.g. heart rate, blood pressure, body temperature) or on highly complex sensors (PET, fMRI, MEG) that are too difficult or impractical to integrate into a soldier’s operational routine.
Few systems leverage the electroencephalogram (EEG), which is the primary sensing technology for cognitive health monitoring. Furthermore, none of these solutions attack the problem from a total-system perspective, instead often concentrating on advancement of individual sensor components or improvement in a single targeted objective. For this work, we present a Soldier-borne wearable and wireless system for physiological monitoring of Soldier cognitive state, combining EEG with a small biomedical sensor
suite.
Mobile Electroencephalography:
The electroencephalogram is a non-invasive technology for sensing and recording the neural activities of the brain. Due to the low size, weight, and power (SWaP) possible with EEG, it is one of only two neural imaging techniques realistically suitable for the mobile or wearable application the other being near infrared spectroscopy (NIRS), a new technology still under active research. By contrast with NIRS, EEG enjoys a long and established history and is currently less susceptible to the environmental effects of the mobile application.
Wearable Soldier-borne Physiological Monitoring System Architecture and Associated Data Flow.
System Architecture:
The Soldier-borne mobile EEG system architecture, and is comprised of an embedded computer hardware system paired with a soldier’s military-issued–or perhaps even personal–Android smart phone. The embedded system samples, preprocesses, and transmits sensor data over a wireless Bluetooth connection to the Android device carried by the soldier.
From there, the Android device can either interact directly with the soldier–providing neurofeedback in the form of alerts to better inform or warn the soldier about his own current cognitive state, forward this information up the chain of communications to higher commanding officers and improve their decision making, or record this information for later offline mission correlation and analysis.
Block Diagram of the Main Circuit Board (Left) and Modularized Eeg Channel Amplification and Filtering Circuit Boards (Right). Starred* Outputs Are Unused.
LONG-TERM ENERGY-EFFICIENT WEARABLE GAIT ANALYSIS FOR RUNNING
Running is the number one participatory sport. It is estimated that there are over 200 million regular runners in the world. Runners have a yearly injury rate of 50%–70%. There is a consensus among physiologists that poor running form has a major impact on injury rates. Analyzing and improving running form can reduce injury rate and can also help runners to improve performance.
Gazelle System Design:
The Gazelle wearable system architecture. It consists of a (1) system-on-chip with a 16 MHz low-power ARM Cortex-M0 and BLE/ANT+ wireless interface, (2) a 9-axis high-precision, high-power MEMS IMU suite with accelerometer, gyroscope, and magnetometer, (3) a standalone ultra-low-power, low-precision accelerometer, (4) an ultra-low-power watchdog timer, (5) a system power management unit, and (6) a standard CR2032 225 mAh coincell battery.
The Gazelle Wearable Sensor and System Architecture.
CONCLUSIONS
This thesis explored opportunities to reduce energy consumption while upholding human sensor device classification and reporting accuracy. The entire end-to-end human-borne sensing platform, from the wearable sensor device, to the mobile device data aggregation layer, to the back end cloud data storage for service delivery and personalized auto-analytics has been considered for low energy consumption optimization.
Energy efficiency improvements of up to 11X were found in the cloud layer, on average 83.6% and from 73-99% found in the wearable layer, and improvements of up to 47% when introducing intelligent collaborative assistance to the wearable device from the mobile smart phone. This chapter summarizes the research contributions made in this thesis, and then explores further avenues for extending the research work that has been presented.
FUTURE RESEARCH
Moore’s law predicts an 18 month cyclical two times reduction in transistor area, which continues until today. Battery chemistry improvements drive higher charge voltage capability, increasing battery capacity and energy density. Both technology trends will naturally improve battery lifetime in wearable devices as time passes by. In the meantime, there is more that we can investigate.
Source: University of Colorado Boulder
Author: James Alexander Williamson
>> More Wireless Bluetooth Projects for Final Year Students
>> More Wireless Biomedical Projects for Final Year Students
>> More Diy Wireless Projects for Engineering Students
>> More Wireless Embedded Projects for Engineering Students