🏭 Fault Detection Ensemble Pipeline

An Advanced Machine Learning Architecture for High-Precision Sub-System Fault Diagnostics

📌 Project Overview

The Fault Detection Ensemble Pipeline is a robust, state-of-the-art machine learning solution engineered for the IEEE SB, GEHU ML Challenge. The core objective of this project is to determine the precise operational status of an embedded device—classifying it as either Normal (0) or Faulty (1)—by analyzing 47 anonymized numeric features representing real-time internal state, performance metrics, and environmental interactions.

By leveraging advanced dimensionality reduction, row-wise statistical profiling, and a multi-algorithmic ensemble architecture, this pipeline achieves a highly stable Out-of-Fold (OOF) Accuracy and F1 Score of ~0.985.

---

🧠 Methodology & Architecture

Our approach moves beyond basic classification by dynamically generating a holistic "behavioral profile" for each operational cycle, feeding this enriched data into a triad of specialized Gradient Boosted Models.

1. Dynamic Feature Engineering

Instead of relying solely on individual sensor readings, the system engineers deep, localized context for every single operational event:

• Row-Wise Statistical Signatures: For every row (device cycle), we compute the mean, std, min, max, skew, and kurtosis across all 47 features. This mathematically captures the overall variation, asymmetry, and extremes of the hardware's operating state in that exact moment.
• Dimensionality Reduction (PCA & ICA): Complex failure modes often manifest as subtle anomalies spanning multiple sensors simultaneously. To detect this, we apply Principal Component Analysis (PCA) and Independent Component Analysis (ICA) to the scaled feature set. We extract 10 highly compressed, orthogonal meta-features that isolate global variance patterns and independent noise signals that standard decision trees would struggle to split on individually.

2. The 5-Fold Stratified Ensemble

Single models are prone to high variance and overfitting, especially given the class imbalance present in tabular sensor data (60% Normal / 40% Faulty). To maximize generalization on unseen test data, we deployed an ensemble:

• Stratified K-Fold Cross Validation (K=5): The training data is split into 5 distinct folds, ensuring the ratio of Normal to Faulty devices remains perfectly consistent across every split. The models are trained on 4 folds and validated on the 5th, guaranteeing that out-of-fold metrics are rigorous and reliable.
• Tri-Algorithmic Blending: Three distinct architectures are trained in parallel to capture different patterns in the structured data:
  1. XGBoost (Extreme Gradient Boosting): Tuned with severe Depth regularization (max_depth=6), Subsampling (0.8), and scale_pos_weight to aggressively penalize misclassified minority faults.
  2. LightGBM: Chosen for its rapid, leaf-wise growth strategy, handling the dense matrix of statistical and PCA features highly efficiently. Weighted using internal class_weight='balanced'.
  3. CatBoost: Applied for its exceptional handling of oblivious trees and automated regularization. It acts as the stabilizing force within the ensemble.
• Soft Voting Mechanism: The final operational status prediction is derived not by taking a hard majority vote, but by averaging the probability logits of all three models.

---

📂 Repository Structure

File	Description
solution.py	The complete ML pipeline encompassing preprocessing, PCA/ICA extraction, model training, K-Fold loops, and inference logic.
readme.txt	Official dataset instructions and problem statement as provided by the ML Challenge.
FINAL.csv	The generated prediction file is formatted precisely for the competition evaluation (ID, CLASS).
.gitignore	Excludes extremely large TRAIN.csv/TEST.csv dataset blobs from version control.

Note: Due to size constraints, the raw training and testing datasets (TRAIN.csv, TEST.csv) are not included in this repository. They must be placed in the project root to run the pipeline.

---

🚀 Execution & Usage

To reproduce the ~0.985 F1-Score results, ensure your environment has the following modern data science libraries:

pip install pandas numpy scikit-learn xgboost lightgbm catboost

Once the TRAIN.csv and TEST.csv files are in your directory, execute the autonomous pipeline:

python solution.py

The script will sequentially output the training progress for each fold, compute the final ensemble validation metrics, and automatically generate the compliant FINAL.csv output.