Thermal AI Models: Detection & Temperature Forecasting

This repository hosts official production checkpoints and research models for the AI-Thermal-detect edge monitoring system.

Repository Contents

models/
  yolo_uniform.onnx              # Ultralytics YOLOv8n exported to ONNX FP32 [1, 5, 8400]
  yolo_uniform.pt                # PyTorch checkpoint (names: {0: 'uniform'})
  yolov8n.engine                 # TensorRT FP16 compiled engine
forecasting/
  linear_reg.pkl                 # Multi-target Autoregressive Linear Regression (11 estimators, serving)
  scaler_X.pkl                   # Feature MinMaxScaler (55 inputs)
  scaler_y.pkl                   # Target MinMaxScaler (11 outputs)
  mtkan_model.keras              # Experimental Temporal KAN model (Keras 3)
  mtkan_weights.weights.h5       # MTKAN model weights
  config.json                    # Channel definitions, window size (5), horizon (5)
  metrics_test.csv               # Empirical test split metrics for forecasting
scripts/
  infer_yolo.py                  # ONNX & PyTorch YOLO inference script
  forecast_temperature.py        # Linear Regression & MTKAN inference script
metadata/
  model_manifest.json            # Exact input/output contracts, shapes, byte sizes
  checksums.sha256               # Cryptographic SHA-256 digests

1. Object Detection: Thermal Uniform / PPE Monitoring

Architecture: Ultralytics YOLOv8n
Class Names: {0: "uniform"} (Extracted directly from checkpoint metadata)
Input Tensor: images shape [1, 3, 640, 640], dtype float32, normalized to range [0.0, 1.0]
Output Tensor: output0 shape [1, 5, 8400], where each candidate contains [cx, cy, w, h, uniform_conf]
Hardware Target:
- CPU/Server: yolo_uniform.onnx via ONNX Runtime
- NVIDIA Jetson Edge: yolov8n.engine compiled via TensorRT FP16

Validation Note: No external mAP claim is asserted beyond the explicit checkpoint metadata and verifiable tensor contracts.

2. Multi-Target Temperature Forecasting

Production Serving Model: Autoregressive Linear Regression

File: forecasting/linear_reg.pkl
Type: 11 independent linear regression estimators trained on autoregressive historical features.
Input Window: 5 minutes historical window across 11 measurement points ($5 \times 11 = 55$ features), scaled via scaler_X.pkl.
Forecast Horizon: 5 minutes ahead for all 11 monitored targets, inverse-scaled via scaler_y.pkl.
Execution Latency: sub-millisecond CPU forward pass ($< 0.2\text{ ms}$).

Experimental Research Model: MTKAN (Temporal Kolmogorov-Arnold Network)

File: forecasting/mtkan_model.keras
Architecture: 2x TKAN recurrent layers + Dense layers (implemented in Keras 3.14).
Role: Research benchmark exploring non-linear spline activation forecasting on STL-decomposed thermal series.

Empirical Test Set Metrics (from `metrics_test.csv`)

Evaluation metrics across 6 designated test channels:

Target ID	MAE (°C)	RMSE (°C)	R² Score
ID_1	0.478	0.588	0.969
ID_2	0.400	0.502	0.979
ID_3	0.508	0.649	0.955
ID_4	0.560	0.668	0.920
ID_5	0.435	0.554	0.970
ID_6	0.377	0.471	0.962

3. How to Download and Verify

from huggingface_hub import snapshot_download

# Download all verified artifacts into a local directory
local_dir = snapshot_download(
    repo_id="neitpm/ai-thermal-detect-models",
    local_dir="models_from_hub",
    repo_type="model"
)

Run inference with the included scripts:

# YOLO Detection via ONNX Runtime
python scripts/infer_yolo.py --model models/yolo_uniform.onnx

# 5-Minute Ahead Temperature Forecasting
python scripts/forecast_temperature.py --dir forecasting/ --engine linear

Downloads last month: 15