🧪 Sensor Signal Classification – ML for Real-World Time Series

This repository explores machine learning approaches for classifying multivariate time series data collected from real-world sensor experiments involving different chemical substances.

📌 Problem Overview

Each experiment consists of a variable-length time series with 16 sensor channels, capturing diverse signal behaviors under different environmental and chemical conditions. The task is to classify each experiment into one of 17 chemical classes.

To tackle this challenge, the project investigates multiple modeling strategies with different levels of temporal abstraction:

• 🧩 Instant Time Approach: Classify each individual time step as a standalone 1×16 feature vector.
• 📐 Time Window Approach: Use fixed-length slices of 100×16 to incorporate short-term temporal dynamics.
• 🎯 RBF Approximation: Compress entire time series into Radial Basis Function (RBF) coefficients and classify in a low-dimensional space.
• 🔁 Temporal Processing (RNN): Leverage recurrent neural architectures (GRU, LSTM) to process the entire sequence with variable length for global temporal encoding.

---

📁 Project Structure

sensor-ts-classification/
│
├── docs/
│   ├── report_ita.pdf                      # Final thesis in Italian 🇮🇹
│   └── report_eng.pdf                      # [English version of the thesis](docs/report_eng.pdf) 🇬🇧
│
├── instant_approach/                      # Models on isolated 1×16 time steps
│   ├── preprocessing.ipynb
│   ├── training.ipynb
│   ├── evaluating.ipynb
│   ├── explaining.ipynb
│   └── training_evaluating_binary.ipynb
│
├── time_windows_approach/                 # Models on 100×16 windows with attention/encoders
│   ├── preprocessing.ipynb
│   ├── training.ipynb
│   ├── evaluating.ipynb
│   ├── explaining.ipynb
│   └── training_evaluating_binary.ipynb
│
├── RBF_approach/                          # Time series compression with RBF coefficients
│   ├── preprocessing.ipynb
│   ├── training.ipynb
│   └── rbf_features.csv
│
├── RNN_approach/                          # Sequence-level classification with GRU/LSTM
│   ├── training.ipynb
│   └── utils.py
│
└── data_analysis.ipynb                     

---

🛠️ Implementation Notes

• Implemented in PyTorch Lightning for modular training, evaluation, and logging.
• Each approach directory contains independent preprocessing, training, and evaluation code.
• Data filtering excludes experiments with fewer than 200 time steps.
• Train/validation/test splits are done per-experiment to maintain time coherence and prevent data leakage.

---

📊 Modeling Summary

Approach	Input Shape	Models Used
Instant Time	1×16	MLP, Encoder MLP
Time Window	100×16	MLP, CNN, Encoder, CSE
RBF Approximation	RBF Coeffs (16×k)	MLP, Random Forest
Temporal (RNN)	Variable × 16	GRU, LSTM

---

📄 Read the Full Thesis

📘 Read the full thesis in English (PDF)