SISA-RoutineGuard v5b — ONNX Edition

노인 일상 패턴 이상 감지 (Galaxy Watch + Jetson Orin Nano Super)
ONNX exports for fast Jetson inference (TensorRT compatible).

📦 ONNX Models (Jetson deploy)

File	Size	Stage	Input	Output
patch_encoder.onnx	0.9 MB	1 (Phone)	acc [6, 250, 3]	tokens [6, 256]
minute_encoder.onnx	192.5 MB	2	patch_tokens [B, 6, 256]	minute_embed [B, 768]
feature_adapter.onnx	2.7 MB	2	feature [B, 12]	embed [B, 768]
hourslot_encoder.onnx	570.9 MB	3	minute_embeds [B, 60, 768]	slot [B, 1024], slot_minutes [B, 60, 1024]
history_encoder.onnx	567.7 MB	4	slot_embeds [B, 90, 1024] + meta	history_embeds [B, 90, 1024]
vitalguard.onnx	48.2 MB	6	vital_features [B, 60, 5] + hrv	hr_residual_z + trend + context
output_heads.onnx	6.3 MB	—	cls_pooled [B, 1024]	anomaly + reason + confidence

QueryRefiner ONNX 는 cross-attention shape 복잡으로 미지원 (PyTorch ckpt만).

Total ONNX: ~1.4 GB

🚀 Jetson Inference Example

import onnxruntime as ort
import numpy as np

# TensorRT EP (Jetson에서 자동 가속)
providers = [
    'TensorrtExecutionProvider',  # Jetson Orin Nano TensorRT
    'CUDAExecutionProvider',       # fallback CUDA
    'CPUExecutionProvider',        # last resort
]

# 1. Phone 측 PatchEncoder (Phone ONNX)
phone_sess = ort.InferenceSession("patch_encoder.onnx", providers=['CPUExecutionProvider'])
patches = phone_sess.run(None, {"acc": acc_array})  # [6, 250, 3] → [6, 256]

# 2. Jetson Pipeline
me_sess = ort.InferenceSession("minute_encoder.onnx", providers=providers)
minute = me_sess.run(None, {"patch_tokens": patches.reshape(1, 6, 256)})[0]

hs_sess = ort.InferenceSession("hourslot_encoder.onnx", providers=providers)
slot, slot_minutes = hs_sess.run(None, {"minute_embeds": minute_batch})  # [B, 60, 768]

# ... history, query, heads chain

🎯 모델 사이즈 — 444.66M params

Component	Params	Deploy
PatchEncoder	0.22M	Phone (ONNX)
MinuteEncoder	48.09M	Jetson
FeatureAdapter	0.66M	Jetson
HourSlotEncoder	142.60M	Jetson
HistoryEncoder	141.86M	Jetson
QueryRefiner	97.65M	Jetson
VitalGuard	12.01M	Jetson
OutputHeads	1.58M	Jetson
Total	444.66M	—

📊 성능 (OOD: HAR-70+ 노인 70-95세, 학습 X)

시나리오	Score	AUC	Reason
Normal	0.0004	—	—
walk_missing	0.9995	1.000	0%
prolonged_inactivity	0.9987	1.000	5%
routine_time_shift	0.9995	1.000	100% ✓
activity_drop	0.7145	0.950	28%
Overall	—	0.987	—

🔧 Train/Val Split + Early Stopping

• Train/Val: 80/20 (3,277 / 819 samples)
• Best epoch = 1, val_loss = 0.2814 (saved)
• Early stop at epoch 13 (patience=10)
• Overfitting 방지 검증됨

📁 PyTorch checkpoints (학습 reproducibility용)

File	Size	Purpose
stage1_patch.pt	879 KB	PatchEncoder (also ONNX)
stage2_minute.pt	192 MB	MinuteEncoder (also ONNX)
stage2_adapter.pt	2.7 MB	FeatureAdapter (also ONNX)
stage3_hourslot.pt	570 MB	HourSlotEncoder (also ONNX)
stage4_history.pt	567 MB	HistoryEncoder (also ONNX)
stage4_refiner.pt	390 MB	QueryRefiner (PyTorch only)
stage5_full.pt	1.78 GB	Full model (val-split best)
stage6_vitalguard.pt	48 MB	VitalGuard (also ONNX)
normalizer.pkl	405 B	Feature normalizer

학습 데이터

CAPTURE-24 (151) + ArWISE V3 (10명/76일) + PPG-DaLiA (15) + WESAD (15) + MHEALTH (10). HAR-70+ (18명 70-95세)는 OOD 평가에만 사용.

Code

https://github.com/tlstngud/sisa-routineguard (PRESENTATION.md 포함)