Hydrogen Leak Detection with XGBoost (Single-Sensor Version)

This repository contains a machine learning pipeline for predicting hydrogen system states using a single hydrogen sensor and time information.

The model predicts system safety conditions and recommended actions based on hydrogen concentration over time.

---

Problem

Hydrogen systems require reliable monitoring to detect hazardous conditions.

In this simplified setup, the system uses:

• a single hydrogen concentration sensor (ppm)
• time (seconds)

to predict key operational states:

• leak severity – hazard level of the system
• action plan – recommended response

⚠️ Note:
The previous version included leak location, but this has been removed since it cannot be reliably inferred from a single sensor.

---

Input Features

The model uses only two features:

• ppm – hydrogen concentration from a single sensor
• time (s) – elapsed time in seconds

---

Dataset

The dataset is provided as an Excel file with two sheets:

Data → sensor measurements and labels (used for training)
Des. → system component descriptions

Only the Data sheet is used during training.

Feature Transformation

The original dataset contains multiple sensor columns:

• compressor sensor (ppm)
• storage sensor (ppm)
• pipeline sensor (ppm)
• fuelcell sensor (ppm)

These are converted into a single feature:

ppm = mean(all sensor readings)

---

Methodology

1. Load data
2. Clean dataset
3. Create ppm feature
4. Select features: ["time (s)", "ppm"]
5. Encode labels
6. Train/test split
7. Train XGBoost models
8. Hyperparameter tuning
9. Evaluation
10. Save model

Targets:

• leak severity
• action plan

---

Using the Model

import joblib
import pandas as pd

bundle = joblib.load("hydrogen_leakage_prediction_model_single_sensor.pkl")

sample = pd.DataFrame([{
    "time (s)": 120,
    "ppm": 18.5
}])

predictions = {}

for target, model in bundle["models"].items():
    features = bundle["selected_feature_sets"][target]
    pred = model.predict(sample[features])[0]
    predictions[target] = bundle["label_encoders"][target].inverse_transform([pred])[0]

print(predictions)

---

Changes from Previous Version

• Removed multi-sensor inputs
• Removed leak location
• Uses only ppm + time (s)
• Simplified deployment