Ajou University repository

저전력 경량화 NNoM 프레임워크를 활용한 반려동물의 호흡수 기반 위험도 및 시계열 예측 시스템
  • 정도영
Citations

SCOPUS

0

Citation Export

DC Field Value Language
dc.contributor.advisor김영진-
dc.contributor.author정도영-
dc.date.issued2024-08-
dc.identifier.other33984-
dc.identifier.urihttps://aurora.ajou.ac.kr/handle/2018.oak/39340-
dc.description학위논문(석사)--IT융합공학과,2024. 8-
dc.description.abstractAs artificial intelligence technology develops, research on high-speed, low-power, and lightweight models in servers is steadily progressing. However, the growth of IoT technology and markets will require more data and computation, which will require specifications greater than the amount of computation and speed that servers can perform. Therefore, in this paper, to solve the problem of intensive artificial intelligence computation, we conducted research on a lightweight, low-power system using the NNoM (Neural Network on Microcontroller) framework, which is lightweight enough to be used in small networks and devices._x000D_ <br> The system proposed in this paper ensures data stability through the selection of highly accurate sensors and is designed to collect more than 10 hours of data per day through low-power Bluetooth communication. In addition, noise was removed from the collected data using windowing techniques and moving average techniques, and data normalization was performed to extract meaningful features through analysis and preprocessing of time series data. Based on the signs of heart disease and pulmonary edema in pets, the characteristics of specific symptoms were modeled, and data on asymptomatic symptoms was used based on datasets collected in real environments and previously collected datasets. Finally, we applied the Fully-Connected Deep Neural Network (FFDNN) and Long Short-Term Memory (LSTM) models to predict respiratory rate and risk, and compared the results._x000D_ <br> Through this study, we propose a real-time data processing and prediction system in small network devices. In addition, the importance of low-power and lightweight models was emphasized, and directions for future research in various application fields were suggested._x000D_ <br>| 인공지능 기술이 발전하면서 서버에서의 고속, 저전력, 경량화 모델에 관한 연구는 꾸준히 진행되고 있다. 하지만 IoT 기술과 시장의 성장은 더 많은 데이터와 연산을 요구할 것이고, 이는 서버가 수행할 수 있는 연산량과 속도보다 더 큰 사양이 필요할 것이다. 따라서 본 논문에서는 인공지능 연산이 집중되는 문제점을 해결하기 위해, 소규모 네트워크와 장치에서 사용할 수 있을 정도로 경량화된 NNoM(Neural Network on Microcontroller) 프레임워크를 활용하여 경량화, 저전력 시스템에 관한 연구를 수행하였다. _x000D_ <br> 본 논문에서 제안하는 시스템은 정확도가 높은 센서의 선정을 통해 데이터의 안정성을 확보하였고, 저전력 블루투스 통신을 통하여 하루 10시간 이상의 데이터를 수집할 수 있도록 설계하였다. 또한 수집된 데이터는 윈도우 기법과 이동 평균 기법을 통하여 노이즈 제거를 수행하였고, 데이터 정규화를 수행함으로써 시계열 데이터에 대한 분석과 전처리를 통해 유의미한 특징을 추출하였다. 애완동물의 심장병 및 폐수종에 징후를 기반으로 하여, 특정 증상에 대한 징을 모델링 하였고, 실제 환경에서 수집된 데이터셋과 기존에 수집되었던 데이터셋을 기반으로 하여 무증상에 대한 데이터로 활용하였다. 최종적으로 FFDNN(Fully-Connected Deep Neural Network)과 LSTM(Long Short-Term Memory) 모델에 적용하여 호흡수 및 위험도 예측을 수행하였고 그 결과를 비교하였다._x000D_ <br> 본 연구를 통하여 소규모 네트워크 장치에서의 실시간 데이터 처리 및 예측 시스템을 제안한다. 또한, 저전력 및 경량화 모델의 중요성을 강조하며, 향후 다양한 응용 분야에서의 연구 방향을 제시하였다._x000D_-
dc.description.tableofcontents1. 서론 1_x000D_ <br>2. 관련 연구 5_x000D_ <br> 2.1. 반려동물의 생체 신호 선정 및 모델링 5_x000D_ <br> 2.2. 인공지능 및 NNoM(Neural Network on Microcontroller) 7_x000D_ <br> 2.3. 심장병 예후 신호 예측 8_x000D_ <br> 2.4. 호흡수 측정 센서 10_x000D_ <br> 2.5. 블루투스 통신 12_x000D_ <br> 2.6. 히스테리시스와 필터링 13_x000D_ <br> 2.7. FFDNN, RNN, LSTM을 활용한 데이터 예측 17_x000D_ <br>3. 동기 및 기여점 22_x000D_ <br>4. 애완동물의 위급상황 예측을 위한 인공지능 시스템 구조 25_x000D_ <br> 4.1. 헬스케어 하드웨어 아키텍처 25_x000D_ <br> 4.2. 헬스케어 소프트웨어 아키텍처 26_x000D_ <br> 4.2.1. 계측 및 통신구조 26_x000D_ <br> 4.2.2. NNoM 레이어 구성 28_x000D_ <br>5. 실험 결과 30_x000D_ <br> 5.1. 실험 환경 30_x000D_ <br> 5.2. 제안하는 기법 및 비교 35_x000D_ <br> 5.2.1. 케이스별 입력 분류/모델링 결과 35_x000D_ <br> 5.2.2. NNoM 레이어 구성에 따른 결과 비교 37_x000D_ <br> 5.2.3. 위급상황 예측 시뮬레이션 테스트 결과 39_x000D_ <br>6. 결론 42_x000D_ <br>참고문헌 43_x000D_ <br>Abstract 48_x000D_-
dc.language.isokor-
dc.publisherThe Graduate School, Ajou University-
dc.rights아주대학교 논문은 저작권에 의해 보호받습니다.-
dc.title저전력 경량화 NNoM 프레임워크를 활용한 반려동물의 호흡수 기반 위험도 및 시계열 예측 시스템-
dc.title.alternativeRisk and Time Series Prediction System Based on Respiratory Rate of Pet-Animals using Low-Power, Lightweight NNoM Framework-
dc.typeThesis-
dc.contributor.affiliation아주대학교 대학원-
dc.contributor.alternativeNameChung Doyoung-
dc.contributor.departmentIT융합대학원 IT융합공학과-
dc.date.awarded2024-08-
dc.description.degreeMaster-
dc.identifier.urlhttps://dcoll.ajou.ac.kr/dcollection/common/orgView/000000033984-
dc.subject.keywordArtificial intelligence-
dc.subject.keywordNeural Network on Microcontrollers-
dc.subject.keywordRespiratory Rate-
dc.subject.keywordTime series prediction-
dc.subject.keyworddisease monitoring-
dc.subject.keywordpet health care-
dc.description.alternativeAbstractAs artificial intelligence technology develops, research on high-speed, low-power, and lightweight models in servers is steadily progressing. However, the growth of IoT technology and markets will require more data and computation, which will require specifications greater than the amount of computation and speed that servers can perform. Therefore, in this paper, to solve the problem of intensive artificial intelligence computation, we conducted research on a lightweight, low-power system using the NNoM (Neural Network on Microcontroller) framework, which is lightweight enough to be used in small networks and devices._x000D_ <br> The system proposed in this paper ensures data stability through the selection of highly accurate sensors and is designed to collect more than 10 hours of data per day through low-power Bluetooth communication. In addition, noise was removed from the collected data using windowing techniques and moving average techniques, and data normalization was performed to extract meaningful features through analysis and preprocessing of time series data. Based on the signs of heart disease and pulmonary edema in pets, the characteristics of specific symptoms were modeled, and data on asymptomatic symptoms was used based on datasets collected in real environments and previously collected datasets. Finally, we applied the Fully-Connected Deep Neural Network (FFDNN) and Long Short-Term Memory (LSTM) models to predict respiratory rate and risk, and compared the results._x000D_ <br> Through this study, we propose a real-time data processing and prediction system in small network devices. In addition, the importance of low-power and lightweight models was emphasized, and directions for future research in various application fields were suggested._x000D_-
Show simple item record

Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.

Total Views & Downloads

File Download

  • There are no files associated with this item.