⚡ ESD Risk Prediction Framework

Explainable Attention-Based Situational Awareness for Wearable IoT Devices

---

Project Structure

esd_project/
├── data/
│   └── generate_dataset.py     # 🔵 Member 1 — synthetic ESD sensor data
├── models/
│   ├── attention_model.py      # 🟣 Member 2 — multi-head attention model
│   └── train.py                # 🟣 Member 2 — training + ablation + baselines
├── xai/
│   └── explainer.py            # 🔴 Member 3 — SHAP saliency + attention viz + KG
├── dashboard/
│   └── app.py                  # 🔴 Member 3 — real-time Streamlit dashboard
├── notebooks/
│   └── ESD_Full_Pipeline.ipynb # 📓 End-to-end pipeline (all members)
├── requirements.txt
└── README.md

---

Quickstart

Option A - Local

pip install -r requirements.txt

# Step 1: Generate dataset
python data/generate_dataset.py

# Step 2: Train model
python models/train.py

# Step 3: Launch dashboard
streamlit run dashboard/app.py

The scripts resolve paths relative to the project folder, so they can be run from the repo root without editing hardcoded Colab or machine-specific paths.

Option B - Google Colab

1. Upload all .py files to /content/esd_project/ in Colab
2. Open notebooks/ESD_Full_Pipeline.ipynb
3. Runtime -> Run All

---

System Architecture

Wearable Sensors (Member 1)
  Humidity | Temperature | E-Field | Contact Voltage | Movement
       ↓
Data Pipeline
  Edge Preprocessing → Timestamped CSV Dataset
       ↓
Attention Model (Member 2)
  Input Encoder → Positional Encoding
  → Multi-Head Self-Attention (×2 layers, 4 heads)
  → Context Fusion (activity, environment, fabric embeddings)
  → Risk Classifier → Low / Medium / High
       ↓
XAI + Knowledge Graph (Member 3)
  Gradient Saliency → Feature Importance
  Attention Maps → Temporal Explanation
  KG Reasoning → Context-Aware Risk Paths
  Plain-English Explanation + Mitigation
       ↓
Real-Time Dashboard
  Live sensor streams | Risk gauge | Attention heatmap | Explanation

---

Model Architecture

Component	Detail
Input	5 numerical sensors, window=50 steps
Context	3 categorical embeddings (dim=8 each)
Encoder	Linear projection → d_model=64
Positional Encoding	Sinusoidal
Attention	2 × Multi-Head (4 heads, ff_dim=128)
Pooling	Global average over time
Fusion	Concat numerical + categorical
Classifier	64→3 softmax (Low/Medium/High)
Parameters	~45,000

---

Dataset

Property	Value
Samples	15,000 time steps
Sample Rate	50 Hz
Sensors	Humidity, Temp, E-Field, Voltage, Movement
Context	Activity, Environment, Fabric
Labels	3-class ESD risk (Low/Med/High)
Generation	Physics-inspired simulation

---

Evaluation Metrics

• Accuracy, Macro F1, AUC-ROC
• Ablation study (heads, layers, context)
• Baseline comparison (Random Forest, SVM)
• Attention faithfulness (qualitative)
• Explanation quality (coverage, consistency)

---

Team

Member	Responsibility	Thesis Chapters
Member 1	IoT sensing, ESD physics, dataset	System Architecture, Hardware
Member 2	Attention model, training, ablation	Methodology, Results
Member 3	XAI, Knowledge Graph, dashboard	Explainability, Discussion
All	Integration, intro, conclusion	Introduction, Conclusion