🚂 Predictive Maintenance for Railway Wheel Failures

LightGBM · XGBoost · SMOTE · Sensor Data · Kaggle Competition

---

🏆 Competition Result

Rank 3 — INFORMS RAS 2025 Kaggle Competition Log Loss: 0.0283 | Best Iteration: 831

---

🧩 Problem Statement

Railway wheel failures are costly, dangerous, and difficult to predict. This project builds a multi-class predictive maintenance classifier on real-world railway sensor data to identify 5 failure types before they occur:

Class	Failure Type
0	High Flange
1	High Impact
2	Thin Flange
3	Other
4	Not Failed

Given monthly sensor readings per wheel, predict the failure type probability for each wheel.

---

⚙️ Pipeline Overview

Raw Sensor Data (20M+ readings)
        │
        ▼
Chunk-based Data Loading (memory-efficient)
        │
        ▼
Feature Engineering
  ├── Aggregated features (mean, std, lag across months)
  ├── Anomaly features (deviation from per-wheel baseline)
  └── Lag features (6-month rolling window)
        │
        ▼
Class Imbalance Handling → RUS + SMOTE
        │
        ▼
Model Benchmarking
  ├── Logistic Regression
  ├── Decision Tree
  ├── Random Forest
  ├── XGBoost
  ├── CatBoost
  └── LightGBM ✅ (Best)
        │
        ▼
LightGBM (Optuna-tuned) → Log Loss: 0.0283
        │
        ▼
Kaggle Submission → Rank 3

---

📊 Key Results

Metric	Value
Competition Rank	🥉 Rank 3 — INFORMS RAS 2025
Best Model	LightGBM
Validation Log Loss	0.0283
Best Iteration	831 (early stopping)
Dataset Size	20M+ sensor readings
Failure Classes	5

---

🔑 Key Engineering Decisions

1. Memory-Efficient Chunked Loading Sensor data exceeds RAM limits — loaded in chunks using pd.read_csv(chunksize=...), merged with failure labels per chunk.

2. Per-Wheel Anomaly Features Baseline computed from "not failed" months per wheel — deviations flagged as anomaly signals. This captures deterioration patterns invisible to standard aggregations.

3. Lag Features (6-month window) Rolling lag features capture temporal degradation trends leading up to failure.

4. LightGBM Hyperparameters

LGBMClassifier(
    objective='multiclass', num_class=5,
    n_estimators=3000, learning_rate=0.01,
    num_leaves=64, feature_fraction=0.8,
    bagging_fraction=0.8, bagging_freq=5,
    reg_alpha=1.0, reg_lambda=1.0
)

---

📁 Repository Structure

📦 Railway-Wheel-Failure-Prediction/
├── Railway_Wheel_Failure_LightGBM.ipynb   # Full pipeline: features + model + submission
├── README.md
└── LICENSE

Note: Raw sensor CSV files not included (competition data — proprietary). Notebook shows full pipeline with outputs.

---

🚀 Getting Started

pip install lightgbm xgboost catboost scikit-learn pandas numpy imbalanced-learn optuna

---

🧠 Concepts Covered

Predictive Maintenance · Multi-class Classification · Sensor Data Engineering · Anomaly Detection · Lag Features · LightGBM · Class Imbalance (SMOTE/RUS) · Hyperparameter Tuning · Kaggle Competition

---

🏅 Achievement

This solution achieved Rank 3 in the INFORMS RAS 2025 Problem Solving Competition hosted on Kaggle — a railway operations research challenge open to participants globally.

---

👤 Author

Dhruv Kumar Sahu M.Tech, Industrial & Management Engineering — IIT Kanpur GATE 2024 AIR 33 | LinkedIn · GitHub