Data Collection System and Stream Learning Model for ADL Classification

📌 Overview

This repository contains the complete implementation of the thesis project:

"Data Collection System for Training a Stream Learning Model for Activities of Daily Living (ADL) Classification"

The system was designed to support real-time classification of human activities using inertial data collected from smartphones (accelerometer and gyroscope).

The project integrates:

• Offline supervised training
• Online incremental (stream) learning
• Selective label request policy
• Stability and reaction analysis over time

The architecture follows a client–server model with persistent storage and incremental adaptation.

---

🎯 Objective

To design and implement a client-server architecture capable of:

• Collecting inertial sensor data from smartphones
• Segmenting data into temporal windows
• Training supervised models (SVM, Random Forest)
• Deploying an incremental Stream Learning model
• Performing real-time ADL classification
• Requesting labels selectively based on model uncertainty
• Evaluating stability and reaction performance over time

---

🏗 System Architecture

The system is divided into two major phases:

---

🔹 Offline Phase (Local Data Collection & Model Selection)

In the offline phase:

• Smartphones collect raw IMU signals.
• Data are stored locally and exported as CSV files.
• The dataset is cleaned and filtered.
• Signals are segmented into time windows (20Hz / 50Hz).
• Class balancing techniques are applied (Random Oversampling / SMOTE).
• Supervised models (SVM, Random Forest) are trained and compared.
• The best configuration is selected and exported as the baseline model.

Offline Architecture

---

🔹 Online Phase (Streaming & Incremental Learning)

In the online phase:

• The client app streams windowed data (features + metadata).
• The server receives and validates windows.
• A classification module generates predictions + confidence.
• A policy engine decides whether to request a label.
• Labeled windows are aligned and stored.
• An incremental learner (HT / ARF) updates the model continuously.
• Performance is monitored using learning curves, stability and reaction metrics.

High-Level Client–Server Architecture

---

🔹 Detailed Server Modules

The server is divided into:

• Ingestion/Validation module
• Classification module
• Rule/Policy engine
• Label aligner
• Incremental trainer
• Database persistence

---

🔹 Labeling & Incremental Training Pipeline

This diagram summarizes the internal decision flow:

1. Window → Baseline classifier (SVM + Calibrator)
2. Prediction + confidence → Policy engine
3. Label request (if needed)
4. Label alignment
5. Incremental update (HT / ARF)

---

---

📊 Key Results

• Final selected offline model: SVM – Pocket position – 20Hz
• Comparative analysis against Random Forest
• Incremental evaluation using River
• Stability analysis across sessions
• Reaction time analysis after model updates
• Per-class performance metrics
• Learning curve overlays

---

⚙️ Technologies Used

• Python
• Scikit-learn
• River (Stream Learning)
• Pandas
• NumPy
• Matplotlib
• Imbalanced-learn (SMOTE)

---

🧠 Research Contributions

• Integration of offline supervised learning with online incremental adaptation
• Selective labeling strategy to reduce annotation burden
• Stability and reaction metrics for continuous evaluation
• End-to-end architecture from data collection to deployment simulation

---

👩‍🔬 Author

Paula Sofía Muñoz.
Electronic and Telecommunications Engineering
Universidad del Cauca

---

📜 License

This project is intended for academic and research purposes.