Desain Rangkaian Terintegrasi Spiking Neural Network Rendah Daya Untuk Pengenalan Aktivitas Manusia Berbasis Photopletysmogram

Ilmi, Ahmad Jabar (2025) Desain Rangkaian Terintegrasi Spiking Neural Network Rendah Daya Untuk Pengenalan Aktivitas Manusia Berbasis Photopletysmogram. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5022211038-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (8MB) | Request a copy

Abstract

Human Activity Recognition (HAR) berbasis Artificial Neural Network (ANN) pada sinyal Photoplethysmogram (PPG) telah berhasil diterapkan dan menghasilkan performa yang menjanjikan. Namun, tantangan utama dalam pengembangan HAR berbasis PPG adalah efisiensi daya dan kinerja inferensi, khususnya pada perangkat wearable yang memiliki keterbatasan sumberdaya. Tugasakhir ini mengusulkanimplementasi Spiking Neural Network (SNN) berbasis neuron Leaky Integrate-and-Fire (LIF) pada Field Programmable Gate Array (FPGA) sebagai solusi alternatif. Metode yang digunakan meliputi preprocessing PPG, pelatihan model SNN offline, Desain Register Transfer Languange (RTL) arsitektur model SNN, serta implementasi model SNN pada PFGA. Hasil eksperimen menunjukkan proses preprocessing motion artifact cancelation (MAC) pada sinyal PPG signifikan dalam meningkatkan akurasi klasifikasi dari 34,9 % menjadi 82 % pada kelas aktivitas rest, walk, dan run. Rancangan arsitektur RTL SNN dirancang secara modular sehingga parameter neuron, jumlah neuron, serta jumlah layer dapat disesuaikan. Implementasi model SNN pada FPGA memeroleh akurasi sebesar 81,60 % pada kuantisasi bobot fixed-point Q2.6 (berbeda 1,07 % dari pelatihan offline). Latensi inferensi rata-rata pada implementasi FPGA adalah 15,608 ms dengan durasi komputasi inti SNN sebesar 0,086 ms serta konsumsi energi rata-rata yang diperoleh sebesar 0,462 mJ per inferensi.
========================================================================================================================================
Human Activity Recognition (HAR) using Photoplethysmogram (PPG) signals has been implemented with Artificial Neural Network (ANN) approaches and has shown promising performance. However, the main challenges in developing PPG-based HAR are power efficiency and inference performance, especially on wearable devices with limited resources. This final project proposes the implementation of a Spiking Neural Network (SNN) based on Leaky Integrate-and-Fire (LIF) neurons on Field Programmable Gate Array (FPGA) as an alternative solution. The methods used include PPG preprocessing, offline SNN model training, Register Transfer Level (RTL) architecture design of the SNN model, and implementation of the SNN model on FPGA. The experimental results show that the motion artifact cancelation (MAC) preprocessing process on the PPG signal significantly improves classification accuracy from 34,9 % to 82% for activity classes of rest, walk, and run. The SNN RTL architecture design is modularly designed so that neuron parameters, number of neurons, and number of layers can be adjusted. The implementation of the SNN model on FPGA achieves an accuracy of 81,60% with Q2.6 fixed-point weight quantization (differing by 1,07% from offline training). The average inference latency on the FPGA implementation is 15,608 ms with a core SNN computation duration of 0,086 ms and an average energy consumption of 0,462 mJ per inference.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Spiking Neural Network, Human Activity Recognition, Rendah Daya, Photoplethysmogram, FPGA, Spiking Neural Network, Human Activity Recognition, Low Power, Photoplethysmogram, FPGA.
Subjects:	T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK7882.P3 Pattern recognition systems T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK7895.G36 Field programmable gate arrays--Design and construction.
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Electrical Engineering > 20201-(S1) Undergraduate Thesis
Depositing User:	Ahmad Jabar Ilmi
Date Deposited:	27 Jul 2025 14:10
Last Modified:	29 Jul 2025 06:48
URI:	http://repository.its.ac.id/id/eprint/122256

Actions (login required)

View Item