Personal tools

Low Power Neural Network For Multi Sensors Wearable Devices

From iis-projects

Revision as of 16:02, 28 July 2017 by Magnom (talk | contribs)
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)
Jump to: navigation, search
SensorsNN.jpg

Short Description

Wearable device with sensors are most commonly used for monitoring, and detection of events related to the wearer (e.g. abnormalities in physiological functions, detection of seizures and symptoms, patterns of movement, behavior etc.) and/or his surroundings (e.g. determining the context of his actions, location...).

Event sensing implemented by digital signal processing traditionally implies continuous (non-interrupted), periodic signal sampling, storing and processing, at some rate (e.g. sampling frequency) sufficient for the particular application, in order not to miss the event. However, the rate of occurrence of the event of interest is usually very low (rare). Thus, even by employing aggressive power management strategies, such as cycling to lowest-power modes between successive samples, or shortening processing time by optimizing executed algorithms, much energy is wasted. Energy inefficiency of the approach may compromise wearable sensor's battery autonomy, or the rate at which device is able to harvest the energy from the environment.

In order to mitigate the problem of the total power spent for event detection, we propose an alternative, two-stage system architecture consisting of: 1. "wake-up sensing" (WUS) circuit, and 2. main microprocessor (MCU or DSP). WUS is an ultra-low power, but always-on circuitry, continuously monitoring the sensor signal. Circuit usually operates in the analog or mixed signal domain, and provides a coarse recognition of some pattern related to occurrence of the monitored event. Upon recognition, it outputs a digital single-bit wake-up signal engaging the second stage implementing the classical signal sampling and digital signal processing for a more thorough detection. This enables for energy-hungry main microprocessor to be completely turned-off until reception of the wake-up signal.

In the course of our previous research, a laboratory-prototype of a WUS circuit was designed, aimed at the detection of audio events in urban environments (e.g. human voice, cry, train horns, police sirens, tram bells, engine noise etc.) MEMS microphone is used as an signal input. Prototype implements analog-domain spectral decomposition using multiple band-pass filtering banks. Frequency, and pass-band width of each bank (channel) are programmable. Output of each channel is fed to the detector integrating the signal energy contained in the particular spectral band over a preprogrammed time-window. Circuit elements are constructed using discrete electronic components (operational amplifiers, comparators, passives etc.). Wake-up signal is generated by comparing the temporal sequence of outputs of each channel to some pre-set classification template. Sequential template matching state-machine logics is implemented on a low-power MCU.

The idea behind this project would be bringing the laboratory prototype of a WUS circuitry one step closer to the implementation suitable for integration into the existing wearable smart-watch device.

Depending on the applicant's profile and project type, his tasks may involve some of the following:

  • lab. testing/characterization of the existing prototype: verification of the prototype's characteristics w.r. design specification (simulations), measuring power-consumption, and assessing detection performance in lab. conditions
  • implementing the template-matching wake-up logics on the low-power microcontroller
  • implementing and testing of advanced functionality: in-operation tuning of circuit's parameters by digital potentiometers
  • redesign of the signal energy detector subsystem
  • programming the circuit for specific application, field testing
  • revision of printed circuit-board design to make it suitable for integration in smart watch
  • design of template-matching wake-up logics using integrated programmable logics (e.g. PAL, CPLD)

Depending on the applicant's profile and project type, his tasks may involve some of the following:

  • lab. testing/characterization of the existing prototype: verification of the prototype's characteristics w.r. design specification (simulations), measuring power-consumption, and assessing detection performance in lab. conditions
  • programming the circuit for specific application, field testing
  • printed circuit-board design to make it suitable for bird monitoring.

Status: Available

  • Looking for Semester and Master Project Students
Supervisors: Michele Magno

Prerequisites

(not all need to be met by the single candidate)

  • experience using the laboratory instrumentation - signal generators, oscilloscopes, DAQ cards, Matlab etc.
  • analog electronics and signal conditioning with operational amplifiers: amplifiers, filters, integrators etc.
  • knowledge of microcontroller programming (C, preferably Texas Instruments MSP430)
  • basic knowledge on signal processing is a plus.
  • plus is knowledge on printed circuit board (PCB) using Altium.

Character

30% Theory
50% Implementation
20% Testing

Professor

Luca Benini

↑ top

Detailed Task Description

Goals

Practical Details

Results