Create README.md

# FEMTO-ST Bearing Dataset for IIoT Machine Health Monitoring

## 📋 Dataset Description

This dataset contains **vibration sensor data from industrial bearings** for machine health monitoring and predictive maintenance applications. The data is derived from the FEMTO Bearing Dataset and includes both raw vibration signals and pre-extracted features for anomaly detection and remaining useful life (RUL) prediction.

### Dataset Summary

- **Total Size:** ~5.5 GB
- **Format:** Parquet (features) + CSV (raw data)
- **Features Files:** 2.2 GB (train: 455MB, validation: 934MB, test: 800MB)
- **Raw Data:** 3.3 GB (test: 1.49GB, training: 573MB, validation: 1.33GB)
- **Sample Files:** vibration_sample.csv (1KB)
- **License:** MIT
- **Last Updated:** October 2025

## 🎯 Intended Use

### Primary Tasks
- ✅ Anomaly detection in industrial machinery
- ✅ Remaining Useful Life (RUL) prediction
- ✅ Time-series forecasting and degradation modeling
- ✅ Unsupervised learning (Isolation Forest, Autoencoders, LSTM)
- ✅ Feature engineering research
- ✅ Predictive maintenance algorithm development

### Applications
- Industrial IoT (IIoT) monitoring systems
- Predictive maintenance platforms
- Machine learning research and education
- Real-time anomaly detection pipelines
- Bearing health assessment

## 📊 Dataset Structure

### Available Files

#### Pre-extracted Features (Parquet)
| File | Size | Description |
|------|------|-------------|
| `features_train.parquet` | 455 MB | Training features |
| `features_val.parquet` | 934 MB | Validation features |
| `features_test.parquet` | 800 MB | Test features |

#### Raw Data Folders
| Folder | Size | Contents | Description |
|--------|------|----------|-------------|
| `training/` | 573 MB | Raw vibration signals | Training run-to-failure data |
| `validation/` | 1.33 GB | Raw vibration signals | Validation bearing experiments |
| `test/` | 1.49 GB | 15,687 files, 11 folders | Test bearing degradation data |
| `sim/` | - | Simulation data | Simulated bearing behavior |

#### Sample Files
| File | Size | Description |
|------|------|-------------|
| `vibration_sample.csv` | 1 KB | Quick sample for testing |

### Data Splits

| Split | Features (Parquet) | Raw Data | Total | Percentage |
|-------|-------------------|----------|-------|------------|
| **Training** | 455 MB | 573 MB | ~1 GB | ~20% |
| **Validation** | 934 MB | 1.33 GB | ~2.3 GB | ~40% |
| **Test** | 800 MB | 1.49 GB | ~2.3 GB | ~40% |

### Features Schema

Pre-extracted features include:

| Feature | Type | Description |
|---------|------|-------------|
| `timestamp` | datetime64 | Measurement timestamp |
| `bearing_id` | string | Unique bearing identifier |
| `rms` | float64 | Root Mean Square of vibration |
| `peak_to_peak` | float64 | Peak-to-peak amplitude |
| `kurtosis` | float64 | Statistical kurtosis |
| `skewness` | float64 | Statistical skewness |
| `crest_factor` | float64 | Peak amplitude / RMS ratio |
| `band_energy_low` | float64 | Low frequency band energy (0-5kHz) |
| `band_energy_mid` | float64 | Mid frequency band energy (5-10kHz) |
| `band_energy_high` | float64 | High frequency band energy (10-20kHz) |
| `fft_features` | array | FFT-derived spectral features |
| `rul` | float64 | Remaining Useful Life (hours) |
| `health_indicator` | float64 | Normalized health score (0-1) |

## 🚀 Usage

### Quick Start: Load Pre-extracted Features

```python
from datasets import load_dataset
import pandas as pd

# Load the dataset
dataset = load_dataset("Amgharr/FEMTO-ST_DATASET")

# Access splits
train_df = dataset['train'].to_pandas()
val_df = dataset['validation'].to_pandas()
test_df = dataset['test'].to_pandas()

print(f"Training samples: {len(train_df)}")
print(f"Validation samples: {len(val_df)}")
print(f"Test samples: {len(test_df)}")
```

### Load Specific Parquet Files

```python
import pandas as pd

# Load individual feature files
train_features = pd.read_parquet("hf://datasets/Amgharr/FEMTO-ST_DATASET/features_train.parquet")
val_features = pd.read_parquet("hf://datasets/Amgharr/FEMTO-ST_DATASET/features_val.parquet")
test_features = pd.read_parquet("hf://datasets/Amgharr/FEMTO-ST_DATASET/features_test.parquet")

# Quick exploration
print(train_features.head())
print(f"\nFeatures: {train_features.columns.tolist()}")
print(f"\nShape: {train_features.shape}")
```

### Load Sample Data

```python
import pandas as pd

# Quick sample for exploration
sample = pd.read_csv("hf://datasets/Amgharr/FEMTO-ST_DATASET/vibration_sample.csv")
print(sample)
```

### Example: Anomaly Detection Pipeline

```python
from datasets import load_dataset
from sklearn.ensemble import IsolationForest
from sklearn.preprocessing import StandardScaler
import numpy as np

# Load training and test data
dataset = load_dataset("Amgharr/FEMTO-ST_DATASET")
train_df = dataset['train'].to_pandas()
test_df = dataset['test'].to_pandas()

# Select numeric features
feature_cols = ['rms', 'peak_to_peak', 'kurtosis', 'skewness',
'band_energy_low', 'band_energy_mid', 'band_energy_high']

X_train = train_df[feature_cols]
X_test = test_df[feature_cols]

# Standardize
scaler = StandardScaler()
X_train_scaled = scaler.fit_transform(X_train)
X_test_scaled = scaler.transform(X_test)

# Train Isolation Forest
model = IsolationForest(
contamination=0.1,
random_state=42,
n_estimators=100,
max_samples='auto'
)

model.fit(X_train_scaled)

# Predict on test set
predictions = model.predict(X_test_scaled)
anomaly_scores = model.score_samples(X_test_scaled)

# Results
test_df['is_anomaly'] = predictions == -1
test_df['anomaly_score'] = anomaly_scores

print(f"Detected anomalies: {(predictions == -1).sum()} / {len(predictions)}")
print(f"Anomaly rate: {(predictions == -1).sum() / len(predictions) * 100:.2f}%")
```

### Example: RUL Prediction with LSTM

```python
from datasets import load_dataset
import numpy as np
import tensorflow as tf
from tensorflow import keras
from sklearn.preprocessing import MinMaxScaler

# Load data
dataset = load_dataset("Amgharr/FEMTO-ST_DATASET")
train_df = dataset['train'].to_pandas()

# Prepare sequences for LSTM
def create_sequences(data, seq_length=50):
sequences = []
labels = []
for i in range(len(data) - seq_length):
seq = data[i:i+seq_length]
label = data[i+seq_length]['rul'] # Predict RUL
sequences.append(seq)
labels.append(label)
return np.array(sequences), np.array(labels)

# Feature selection and normalization
features = ['rms', 'kurtosis', 'band_energy_low', 'band_energy_mid']
scaler = MinMaxScaler()
scaled_data = scaler.fit_transform(train_df[features])

# Create sequences
X, y = create_sequences(scaled_data, seq_length=50)

# Build LSTM model
model = keras.Sequential([
keras.layers.LSTM(64, return_sequences=True, input_shape=(50, len(features))),
keras.layers.Dropout(0.2),
keras.layers.LSTM(32),
keras.layers.Dropout(0.2),
keras.layers.Dense(16, activation='relu'),
keras.layers.Dense(1) # RUL prediction
])

model.compile(optimizer='adam', loss='mse', metrics=['mae'])
model.fit(X, y, epochs=50, batch_size=32, validation_split=0.2)
```

## 📖 Data Source

This dataset is based on the **FEMTO Bearing Dataset** from the PRONOSTIA platform:

- **Original Source:** [NASA PCoE Data Repository](https://ti.arc.nasa.gov/tech/dash/groups/pcoe/prognostic-data-repository/)
- **Institution:** FEMTO-ST Institute, Besançon, France
- **Platform:** PRONOSTIA (Prognostic and Health Management platform)
- **Experiment Type:** Accelerated degradation tests (run-to-failure)

### Preprocessing Applied
- ✅ Feature extraction from raw accelerometer signals
- ✅ Time-domain features (RMS, kurtosis, skewness, crest factor)
- ✅ Frequency-domain features (FFT, band energy)
- ✅ RUL calculation based on failure timestamps
- ✅ Health indicator computation
- ✅ Parquet format for efficient storage and loading
- ✅ Train/validation/test splits

## 🔬 Technical Details

### Original Recording Conditions
- **Sampling Rate:** 25.6 kHz
- **Recording Length:** 0.1 seconds every 10 seconds
- **Data Points per Sample:** 2,560 points
- **Channels:** 2 accelerometers (horizontal and vertical)
- **Operating Conditions:**
- Load: 4000N radial force
- Speed: 1800 rpm (30 Hz)
- **Bearing Type:** NSK 6804DD deep groove ball bearings
- **Test Type:** Run-to-failure under constant conditions

### Degradation Stages
1. **Healthy State:** Normal operation, low vibration
2. **Early Degradation:** Slight increase in high-frequency content
3. **Accelerated Wear:** Rapid growth in vibration amplitude
4. **Severe Degradation:** High RMS, kurtosis spikes
5. **Failure:** Complete bearing breakdown

## 📈 Dataset Statistics

```python
# Approximate statistics
Training:
- Size: ~1 GB
- Features: 455 MB
- Raw signals: 573 MB

Validation:
- Size: ~2.3 GB
- Features: 934 MB
- Raw signals: 1.33 GB

Test:
- Size: ~2.3 GB
- Features: 800 MB
- Raw signals: 1.49 GB (15,687 files)

Total: ~5.5 GB
```

## 🤝 Related Resources

- **GitHub Repository:** [siffror/iiot_machine_health](https://github.com/siffror/iiot_machine_health)
- **Live System:** Real-time monitoring with Azure Container Apps, InfluxDB, and Grafana
- **Notebooks:**
- `01_femto_eda.ipynb` - Exploratory Data Analysis
- `rms_analysis.ipynb` - Feature engineering and RMS analysis
- **Deployed Model:** Isolation Forest anomaly scorer running in Azure

## 📝 Citation

If you use this dataset, please cite:

```bibtex


@dataset
{amghar2025femto,
author = {Amghar},
title = {FEMTO-ST Bearing Dataset for IIoT Machine Health Monitoring},
year = {2025},
month = {10},
publisher = {Hugging Face},
version = {1.0},
url = {https://huggingface.co/datasets/Amgharr/FEMTO-ST_DATASET}
}
```

Original FEMTO dataset:
```bibtex
@inproceedings{nectoux2012pronostia,
title={PRONOSTIA: An experimental platform for bearings accelerated degradation tests},
author={Nectoux, Patrick and Gouriveau, Rafael and Medjaher, Kamal and Ramasso, Emmanuel and Zerhouni, Noureddine and Varnier, Christophe},
booktitle={IEEE International Conference on Prognostics and Health Management},
pages

README.md

@@ -0,0 +1,20 @@
+---
+license: mit
+task_categories:
+- time-series-forecasting
+language:
+- en
+tags:
+- iiot
+- predictive-maintenance
+- vibration-analysis
+- machine-health
+- femto-bearing
+- industrial-iot
+- sensor-data
+- time-series
+- bearing-degradation
+pretty_name: FEMTO-ST Bearing Dataset for IIoT Machine Health Monitoring
+size_categories:
+- 1M<n<10M
+---