온디바이스 NPU 기반 실시간 추론을 위한 다채널 BCG-다중 스케일 STFT 특징 융합 경량 CNN 회귀 모델
Lightweight CNN Regression Model for On-Device NPU–Based Real-Time Inference via Multi-Channel BCG and Multi-Scale STFT Feature Fusion
0
조회
0
다운로드
0
인용
0
참고문헌
키워드
초록
This study estimates heart rate (HR) from piezo-based ballistocardiography (BCG) embedded in a bed topper( or mattress topper) using a five-channel acquisition and a multi-scale STFT spectrogram(windows 128,256,512) representation (final tensor 15×64×45) with a lightweight CNN regressor, and deploys it on a Hailo-8 NPU to jointly validate latency and accuracy. Preprocessing uses 100 Hz sampling and a 0.83~2.0Hz zero-phase Butterworth bandpass filter, labels come from synchronous PPG inter-peak intervals and are z-normalized (=75, =15). The PyTorch model is converted via ONNXHAR, quantized with a calibration set, compiled to HEF, and executed on device. Data splits (80%/10%/10%) and metrics (MAE, RMSE, correlation, mean latency) follow internal scripts. Increasing the channel count reduces MAE by ~27.3% from one to five channels and yields r=0.8780, combining multi-scale STFTs reduces MAE by ~23.7% over single-scale and achieves r=0.9023. Versus an FFT peak-picking baseline, MAE drops by ~66.6% and correlation rises from r=0.4934 to 0.9023. INT8 deployment attains MAE 1.84 BPM (+7.6% vs float) with 2.91 ms inference latency. Overall, multi-channel/multi-scale fusion with a compact CNN provides robust, real-time BCG-HR estimation on edge hardware
Abstract
This study estimates heart rate (HR) from piezo-based ballistocardiography (BCG) embedded in a bed topper( or mattress topper) using a five-channel acquisition and a multi-scale STFT spectrogram(windows 128,256,512) representation (final tensor 15×64×45) with a lightweight CNN regressor, and deploys it on a Hailo-8 NPU to jointly validate latency and accuracy. Preprocessing uses 100 Hz sampling and a 0.83~2.0Hz zero-phase Butterworth bandpass filter, labels come from synchronous PPG inter-peak intervals and are z-normalized (=75, =15). The PyTorch model is converted via ONNXHAR, quantized with a calibration set, compiled to HEF, and executed on device. Data splits (80%/10%/10%) and metrics (MAE, RMSE, correlation, mean latency) follow internal scripts. Increasing the channel count reduces MAE by ~27.3% from one to five channels and yields r=0.8780, combining multi-scale STFTs reduces MAE by ~23.7% over single-scale and achieves r=0.9023. Versus an FFT peak-picking baseline, MAE drops by ~66.6% and correlation rises from r=0.4934 to 0.9023. INT8 deployment attains MAE 1.84 BPM (+7.6% vs float) with 2.91 ms inference latency. Overall, multi-channel/multi-scale fusion with a compact CNN provides robust, real-time BCG-HR estimation on edge hardware