Building a real-time fatigue detection system for STM32H7 deployment.<\/p>\n
Constraints:<\/p>\n
- \n
- 512KB RAM<\/li>\n
- <100ms inference<\/li>\n
- preprocessing on laptop<\/li>\n
- inference on-device only<\/li>\n<\/ul>\n
Dataset:
~1M rows from asynchronous wearable sensors.<\/p>\n\n\n
\n \nSensor<\/th>\n Native Frequency<\/th>\n Notes<\/th>\n<\/tr>\n<\/thead>\n \n ACC<\/td>\n 32 Hz<\/td>\n wrist accelerometer<\/td>\n<\/tr>\n \n EDA<\/td>\n 4 Hz<\/td>\n electrodermal activity<\/td>\n<\/tr>\n \n Temp<\/td>\n 4 Hz<\/td>\n skin temperature<\/td>\n<\/tr>\n \n HR<\/td>\n 1 Hz<\/td>\n heart rate<\/td>\n<\/tr>\n \n Breathing<\/td>\n 1 Hz<\/td>\n respiration<\/td>\n<\/tr>\n \n IBI<\/td>\n ~0.59 Hz irregular<\/td>\n inter-beat interval<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n Labels:<\/p>\n
- \n
- fatigue<\/li>\n
- activity<\/li>\n
- baseline<\/li>\n<\/ul>\n
Current preprocessing strategy:
Resample everything to 4Hz.<\/p>\n\n\n
\n \nSignal<\/th>\n Strategy<\/th>\n<\/tr>\n<\/thead>\n \n ACC 32\u21924Hz<\/td>\n mean over 8 samples<\/td>\n<\/tr>\n \n EDA\/Temp<\/td>\n native 4Hz<\/td>\n<\/tr>\n \n HR 1\u21924Hz<\/td>\n linear interpolation<\/td>\n<\/tr>\n \n Breathing 1\u21924Hz<\/td>\n linear interpolation<\/td>\n<\/tr>\n \n IBI ~0.59\u21924Hz<\/td>\n forward-fill<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n Result:
~1M rows \u2192 ~85k synchronized rows.<\/p>\nCurrent doubts:<\/p>\n
- \n
- ACC to 4Hz: Using only the mean feels too lossy. Should I also include:<\/li>\n<\/ol>\n
- \n
- std<\/li>\n
- max\/min<\/li>\n
- magnitude<\/li>\n
- energy<\/li>\n<\/ul>\n
per 250ms window?<\/p>\n
- \n
- IBI: Forward-fill feels mathematically dirty for HRV-related information. Would it be better to:<\/li>\n<\/ol>\n
- \n
- keep IBI irregular<\/li>\n
- compute RMSSD\/SDNN at native timing<\/li>\n
- feed only HRV features downstream?<\/li>\n<\/ul>\n
- \n
- HR\/Breathing: Does interpolating 1Hz \u2192 4Hz introduce fake temporal resolution? Would keeping them at 1Hz be cleaner?<\/li>\n<\/ol>\n
Considering switching to a multi-rate pipeline:<\/p>\n
\n\n
\n \nSignal Group<\/th>\n Frequency<\/th>\n<\/tr>\n<\/thead>\n \n ACC<\/td>\n 8 Hz<\/td>\n<\/tr>\n \n EDA\/Temp<\/td>\n 4 Hz<\/td>\n<\/tr>\n \n HR\/IBI\/Breathing<\/td>\n 1 Hz<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n Question:
For embedded ML \/ TinyML deployment, is multi-rate worth the added pipeline complexity, or is synchronized 4Hz generally the better engineering tradeoff?<\/p>\nWould appreciate advice from anyone working with:<\/p>\n
- \n
- wearable signals<\/li>\n
- HRV<\/li>\n
- TinyML<\/li>\n
- embedded inference<\/li>\n
- multimodal physiological data<\/li>\n<\/ul><\/div>\n
submitted by \/u\/Aziz_2002 <\/a>
[link]<\/a><\/span> [comments]<\/a><\/span><\/p>
- HR\/Breathing: Does interpolating 1Hz \u2192 4Hz introduce fake temporal resolution? Would keeping them at 1Hz be cleaner?<\/li>\n<\/ol>\n
- IBI: Forward-fill feels mathematically dirty for HRV-related information. Would it be better to:<\/li>\n<\/ol>\n
- ACC to 4Hz: Using only the mean feels too lossy. Should I also include:<\/li>\n<\/ol>\n
