The dramatic increase in silicon-based devices and social networks have caused significant network overheads. This has motivated research into embedded machine-learning applications for edge computing devices. We exploit the actual information contained in the input signal which are typically much smaller than the physical Nyquist-rate bandwidths. For example, the key signals in Wearable Body Sensor Networks (e.g. Electrocardiogram, Electroencephalogram etc.) have bandwidths of a few kHz, dynamic ranges of 40–70 dB, and structures that are sparse (i.e. compressible) in some domain. This has led to the emergence of alternative sampling techniques through Analog-to-Information Converters (AIC) suited for random and sparse datasets. A key feature of AICs, critical for achieving ultra-low power and decentralized operation, is the information bandwidth rate, much lower than the Nyquist bandwidth. Information bandwidth refers to low-rate of significant events in sparse signals whereas Nyquist bandwidth (or physical bandwidth) processes are based on the highest rate of change in the input signal.
We investigate in this project an information-aware analog compressive-sensing system towards energy-efficient operation with higher accuracy in dynamic IoT environments. Compared to digital-only implementations, the low-resolution analog feature-extractor and classifier enables low-latency and can have very attractive low-power solution.
Publications: ISCAS’18, MWSCAS’17, TCAS-1’17 (under review), BioCAS’16, JSSC’14 (pre-WSU).
Touch sensors have created a paradigm shift in the human-machine interaction in modern electronic devices. Several emerging applications require that the sensors conform to curved 3-D surfaces and provide an improved spatial resolution through miniaturized dimensions. The proliferation of sensor application also requires that the environmental impact from their manufacturing be minimized. This work demonstrated and characterized interdigitated capacitive touch sensors manufactured using an Aerosol Jet based additive technique that reduces waste and minimizes the use of harmful chemicals. The sensors are manufactured with the capacitive elements at an in-plane length scale of about 50mm by 1.5 to 5 mm, a thickness of 0.5mm, and a native capacitance of a 1 to 5 pF. The sensor capacitance variation is within 8% over multiple samples, establishing the repeatability of the Aerosol Jet method. Scanning electron microscopy and atomic force microscopy are used to characterize the sensor electrodes. 3-D electromagnetic simulations are carried out to predict the capacitance of the printed sensors and the electric field distribution.
Conventional ECG measurement techniques involves a series of steps that disincentives a patient towards getting the right medical diagnosis at the right time. These steps include: scheduling appointments, time to transport, registration, patient preparation and measurement by a trained nurse/ doctor that adds to the cost of basic healthcare and further, leading to underusing the ECG. For example, over one third of patients evaluated for angina pectoris in outpatient settings do not have an ECG recorded. Only one fourth of patients with ST-segment elevation myocardial infarction trans- ported to emergency rooms, have a pre- hospital ECG, with resulting delays in re-vascularization procedures.
In addition, errors in interpreting ECGs are common and may be increasing in frequency. Studies have reported that correct lead positions with an error less than 1 cm was achieved by trained technicians only for 50% of studied men and 20% of studied women with worst-case placement errors up to 6 cm. It has been shown that displacements of the precordial electrodes located nearby the signal source have a greater influence on the ECG signal than shifts of the distal limb electrodes.