Title Low-power wearable system for asthmatic wheeze detection Title (croatian) Nosivi sustav za detekciju asmatskih fićuka s malom potrošnjom energije Author Dinko Oletić Mentor Vedran Bilas (mentor)Committee member Vedran Bilas (član povjerenstva)Granter University of Zagreb Defense date and country 2016, Croatia Scientific / art field, TECHNICAL SCIENCES Universal decimal classificationUDC ) 621.3 - Electrical engineering Abstract Area of this research is minimization of energy consumption of electronic systems for long-term monitoring of common symptom of chronic asthma - asthmatic wheezing. The proposed electronic sensor system consists of a wearable sensor measuring the mechanical vibrations on skin surface originating from respiration, and a wireless mobile device (smartphone). The system conducts automatic recognition of respiratory sounds. The essential design requirement is energy consumption at which the sensor system is able to quantify symptoms. In this context, research compared two system architectures: with asthmatic wheeze detection on the wearable sensor, and the architecture in which information capture on occurrence of wheeze is distributed between a sensor and a mobile device. Firstly, the problem of power-efficient asthmatic wheeze detection on-board wearable sensor was addressed. Building up on the literature state-of-the-art, four algorithms based on features drawn from STFT time-frequency decomposition have been compared on an low-power audio processing DSP. Best trade-off between classification performance and execution speed was obtained by STFT spectral crest shapes tracking algorithm yielding 87.51% sensitivity (SE), 93.42% specificity (SP), and 92.53% accuracy (ACC) at less than 3% of DSP’s active time. Secondly, application of compressed sensing (CS) was evaluated in a distributed sensing system as a mean of simultaneous reduction of power consumption of sensor, and the data-rate reduction in wireless communication. A prototype was demonstrated, implementing a combination of a time-domain subsampling CS encoder on the sensor, and the real-time DFT/DCT spectra CS reconstruction OMP algorithm on smartphone. Compression ratios of 4x to 5.33x were proved feasible. A robust algorithm for wheeze detection from CS reconstructed spectra, based on HMM, was proposed. It enabled for up to SE of 89.99%, SP of 94.03%, and 93.45% ACC at the CS compression rate of 4x, at the processing cost comparable to spectral crest shapes tracking algorithm. Finally, system-level power analysis confirmed that lowest system-level power, estimated from 328 to 428 uW, may be achieved in architecture with wheeze detection on-board wearable sensor. However, in the distributed architecture, CS enables for the lowest sensor power (216 - 357 uW), while simultaneously keeping the total system power (775 - 2605 uW) lower or equal to streaming of the uncompressed signal.
Abstract (croatian) Područje ovog istraživanja je minimizacija potrošnje energije elektroničkih sustava za dugotrajno praćenje učestalog simptoma kronične astme - astmatskih fićuka. Predloženi elektronički senzorski sustav sastoji se od nosivog senzora koji mjeri vibracije na površini kože koje potječu od disanja, i bežičnog mobilnog uređaja (pametnog telefona). Sustav automatski raspoznaje zvukove disanja. Esencijalni konstrukcijski zahtjev odnosi se na količinu energije potrebne sustavu za kvantifikaciju simptoma. U tom kontekstu, istraživanje uspoređuje dvije arhitekture sustava: s detekcijom astmatskih fićuka na senzoru i arhitekturu u kojoj je dohvat informacije o pojavi fićuka raspodijeljen između senzora i mobilnog uređaja. Najprije je obrađen problem energetski učinkovite detekcije fićuka na procesoru nosivog senzora. Krećući od postojećeg stanja tehnike opisanog u literaturi, četiri algoritma temeljena na značajkama izvađenim iz kratkotrajne Fouriereove transformacije (engl. STFT) su uspoređena na audio procesoru za digitalnu obradu signala niske potriošnje. Najbolji odnos performansi klasifikacije i brzine izvođenja je dobiven algoritmom za praćenje spektralnih vrhova STFT-a dajući osjetljivost od 87,51%, specifičnost od 93,42% i točnost od 92,53%, pri manje od 3% aktivnog vremena procesora. Nadalje, primjena sažetog otipkavanja (engl. CS) je evaluirana u raspodijeljenom senzorskom sustavu kao način simultanog smanjenja potrošnje nosivog senzora te za smanjenje količine podatkovnog prometa u bežičnoj komunikaciji. Demonstriran je prototip koji implementira kombinaciju enkodera za sažeto otipkavanje s podotipkavanjem u vremenskoj domeni na senzoru i OMP algoritam implementiran na pametnom telefonu koji rekonstruira DFT/DCT spektar sažeto otipkanog signala. Pokazano je da se mogu ostvariti omjeri kompresije od 4x do 5,33x puta. Predložen je robustan algoritam detekcije fićuka iz rekonstruiranog spektra sažeto otipkanog signala, baziran na skrivenom Markovljevom modelu (engl. HMM). Algoritam omogućava osjetljivost do 89,99%, specifičnost do 94,03% i točnost do 93,45% pri omjeru kompresije 4x, uz trošak procesorske obrade sumjerljiv algoritmu za praćenje spektralnih vrhova u STFT-u. Konačno, analiza potrošnje na razini sustava potvrdila je da je najmanja srednja potrošnja, estimirana između 328 i 428 uW, ostvariva u arhitekturi s detekcijom fićuka na nosivom senzoru. Ipak, u raspodijeljenoj arhitekturi, sažeto otipkavanje omogućava najmanju potrošnju nosivog senzora (216 - 357 uW), uz simultano održavanje ukupne potrošnje senzorskog sustava (775 - 2605 uW) nižom ili podjednakom slučaju kontinuiranog prijenosa nekomprimiranog signala.
Keywords
biomedical sensors
wireless sensor networks
electronic auscultation
asthmatic wheeze detection
short-term Fourier transform (STFT)
compressed sensing (CS)
hidden Markov model (HMM)
low-power electronic design
Keywords (croatian)
biomedicinski senzori
bežične senzorske mreže
elektronička auskultacija
detekcija astmatskih fićuka
kratkotrajna Fourierova transformacija
sažeto otipkavanje
skriveni Markovljev model
elektronički dizajn uz nisku potrošnju energije
Language english URN:NBN urn:nbn:hr:168:945084 Study programme Title: Electrical Engineering and Computing Type of resource Text Extent 165 str. ; 30 cm. File origin Born digital Access conditions Closed access Terms of use Created on 2019-04-10 15:10:46
