Applying Data Science Group Project – PHMAP 2023 Challenge

Welcome to our Applying Data Science group project repository!

This project explores the development of predictive algorithms for system health and performance monitoring using multivariate time-series data from the PHMAP 2023 Data Challenge.

We compare deep learning and traditional machine learning approaches—including CNN, LSTM, SVM, XGBoost, and hybrid models—across a sequence of diagnostic tasks. The objective is to assess the effectiveness, limitations, and real-world applicability of each model for fault detection and classification in complex systems.

---

Repository Structure

• dataset/
  Raw and preprocessed train/test data

• EDA&DataProcessing/
  Exploratory analysis and feature engineering

• models/
  All model implementations (CNN, LSTM, SVM, XGB)
  Task 5 is a regression task, so it is handled separately in a subfolder due to its distinct model type

---

Project Outline

• Multistage Downstream Tasks

  • Task 1: Normal/Abnormal classification
  • Task 2: Fault type classification (bubble, valve, unknown)
  • Task 3: Bubble source localization
  • Task 4: Faulty SV identification
  • Task 5: SV opening ratio prediction (regression)

• Modeling

  • CNN for capturing local patterns from raw pressure sequences
  • CNN + LSTM for learning temporal dynamics and long-term dependencies
  • XGBoost & SVM using handcrafted statistical features for robust tabular modeling

• Data Processing

  • Feature extraction (mean, std, min, max for each sensor)
  • Label extraction

---

Tools & Libraries

• Programming: Python (Jupyter notebooks)
• ML/DL Frameworks: PyTorch, scikit-learn, XGBoost..
• Visualization: Seaborn, Matplotlib

---

Acknowledgement

Dataset provided by the PHMAP 2023 Data Challenge.
This project is part of the Applying Data Science course at the University of Manchester.

---