⚙️ MetroPT Train Compressor - Predictive Maintenance Using Unsupervised Learning

📌 Project Overview

This project develops an advanced predictive maintenance (PdM) and anomaly detection pipeline for a critical rolling stock subsystem: the Metro Train Air Compressor (PT3 Dataset). The goal is to move away from traditional static thresholds and utilize Machine Learning to capture multivariate anomalies, providing early warnings for critical failure events like Air Leaks before they cause severe transit delays.

---

📊 Dataset & Temporal Splitting

To simulate a real-world industrial deployment scenario and prevent data leakage:

• Training Data (Normal Baseline): February & March (The compressor operated under completely clean, failure-free conditions).
• Testing Data (Evaluation Phase): April (Contains the first high-stress Air Leak catastrophic failure event documented on April 18th).

---

📊 Exploratory Data Analysis (EDA)

Understanding the interconnected thermodynamic and electrical relationships between the compressor sensors is vital. Below is the statistical cross-correlation mapping computed from the baseline telemetry via sns.heatmap:

---

🧠 Engineering Insights: How the Model Works Mechanically

Traditional maintenance triggers an alert only when a single sensor exceeds a specific safety limit. This project utilizes One-Class SVM (RBF Kernel) to evaluate the multivariate relationships across the entire compressor system simultaneously.

Mechanically, it learns the normal "Baseline Cluster" where:

• High Motor Current must co-occur with a proportional rise in TP2 (Secondary Pressure).
• Active COMP signals align with structured duty cycles.

When an Air Leak develops, this physics-based correlation breaks down—pressure drops sharply while the motor current spikes or continues operating indefinitely to compensate for the loss. In the multi-dimensional feature space, these broken correlations force the data points to drift outside the learned normal boundary, triggering an immediate predictive anomaly alert.

---

📈 Model Performance & Evolution

1. Baseline Model: Isolation Forest

Initially, Isolation Forest was deployed as a baseline. While computationally fast, it struggled to capture the gradual, micro-leak degradation phase, yielding an unacceptable Recall of only 2% on the failure event. The confusion matrix below highlights how the baseline model missed 8,458 true anomaly slices (False Negatives):

2. Final Champion Model: One-Class SVM

By pivoting to One-Class SVM and applying smart 10% stratified sampling to overcome the high computational complexity ($O(N^2)$) on massive high-frequency data, the system achieved a massive performance leap. As shown in the confusion matrix below, the model successfully captured 8,567 out of 8,657 true anomaly operations, reducing False Negatives drastically:

📊 Comparative Performance Matrix

Metric	Baseline (Isolation Forest)	Final Model (One-Class SVM)
Failure Recall (Detection Rate)	2%	99% 🚀
Precision	8%	37% 📈
F1-Score	0.04	0.54 🔥

---

🛠️ Tech Stack Used

• Data Ingestion & Handling: Pandas, NumPy
• Data Visualization: Matplotlib, Seaborn
• Machine Learning Framework: Scikit-Learn (OneClassSVM, IsolationForest, StandardScaler)