Daro77/univariate-time-series-with-missingness-and-reconstructed

Experimental Datasets - Time Series Inpainting

This directory contains all datasets used in the time series inpainting experiments, including original data, corrupted versions, reconstructed results, and intermediate image representations.

---

📁 Directory Structure

data/
├── README.md                      # This file
│
├── 0_source_data/                 # Original univariate industrial time series (7 files, 8.5 MB)
│   ├── boiler_outlet_temp_univ.csv
│   ├── pump_sensor_28_univ.csv
│   ├── vibration_sensor_S1.csv
│   ├── water_level_sensors_2010_L300.csv
│   ├── water_level_sensors_2010_L308.csv
│   └── water_level_sensors_2010_L311.csv
│
├── 1_missing_data/                # Time series with injected missing values (721 files, 964 MB)
│   ├── [dataset]_[mechanism]_[rate]_[iter].csv
│   └── ...
│
├── 2_fixed_data/                  # Reconstructed time series (16,201 files, 25 GB)
│   ├── [dataset]_[mechanism]_[rate]_[iter]_[method].csv
│   └── ...
│
└── images_inpainting/             # Image representations (16,324 files, 4 GB)
    ├── 0_original_images/         # Original GAF/MTF/RP/SPEC images
    ├── 1_missing_images/          # Corrupted images
    ├── 2_fixed_images/            # Reconstructed images
    └── 3_difference_images/       # Visualization of differences

---

📊 Dataset Statistics

Directory	Files	Size	Description
0_source_data	7	8.5 MB	Original industrial time series
1_missing_data	721	964 MB	Corrupted versions (systematic missingness)
2_fixed_data	16,201	25 GB	Reconstructed by 31 methods
images_inpainting	16,324	4 GB	Image transformations (GAF/MTF/RP/SPEC)
Total	33,253	~30 GB	Complete experimental dataset

---

📈 0_source_data/

Description

Original, complete industrial time series datasets without any missing values. These serve as ground truth for all experiments.

Contents

File	Source	Points	Sampling	Description
boiler_outlet_temp_univ.csv	Industrial boiler	~10,000	1 min	Temperature sensor readings
pump_sensor_28_univ.csv	Industrial pump	~10,000	1 min	Pressure/flow measurements
vibration_sensor_S1.csv	Vibration monitor	~10,000	High freq	Mechanical vibration data
water_level_sensors_2010_L300.csv	Water monitoring	~8,760	1 hour	Water level station L300
water_level_sensors_2010_L308.csv	Water monitoring	~8,760	1 hour	Water level station L308
water_level_sensors_2010_L311.csv	Water monitoring	~8,760	1 hour	Water level station L311

Format

All files are CSV with two columns:

timestamp,value
2023-01-01 00:00:00,42.5
2023-01-01 00:01:00,42.8
...

Archive

# Compressed archive
data_0_source.tar.gz (1.8 MB)

# Extract
tar -xzf ../data_0_source.tar.gz

---

🕳️ 1_missing_data/

Description

Time series with systematically injected missing values. Each file corresponds to one experimental configuration with controlled missingness.

Naming Convention

[dataset]_[mechanism]_[rate]_[iteration].csv

Examples:

• boiler_MCAR_2p_1.csv - Boiler, MCAR mechanism, 2% missing, iteration 1
• pump_MAR_5p_3.csv - Pump, MAR mechanism, 5% missing, iteration 3
• vibration_MNAR_10p_7.csv - Vibration, MNAR mechanism, 10% missing, iteration 7

Experimental Design

Parameter	Values	Count
Datasets	3 (boiler, pump, vibration)	3
Mechanisms	MCAR, MAR, MNAR	3
Missing Rates	2%, 5%, 10%	3
Iterations	1-10	10
Total Configs	3 × 3 × 3 × 10	270 per dataset

Note: Water level datasets (L300, L308, L311) were excluded from main experiments.

Missingness Mechanisms

MCAR (Missing Completely At Random):

• Random points removed
• No systematic pattern
• Simulates: random sensor failures

MAR (Missing At Random):

• Probability depends on observed values
• Conditional missingness
• Simulates: threshold-based dropout

MNAR (Missing Not At Random):

• Probability depends on missing values themselves
• Systematic bias
• Simulates: sensor saturation, range limitations

Format

Same as source data, but with NaN values:

timestamp,value
2023-01-01 00:00:00,42.5
2023-01-01 00:01:00,NaN
2023-01-01 00:02:00,42.8
...

Archive

# Compressed archive
data_1_missing.tar.gz (192 MB)

# Extract
tar -xzf ../data_1_missing.tar.gz

---

🔧 2_fixed_data/

Description

Time series reconstructed using 31 different methods. This is the main experimental output containing all reconstruction results.

Naming Convention

[dataset]_[mechanism]_[rate]_[iteration]_[method].csv

Examples:

• boiler_MCAR_2p_1_imputemean.csv - Mean imputation
• pump_MAR_5p_3_interpolatecubic.csv - Cubic interpolation
• vibration_MNAR_10p_7_gafunet.csv - GAF + U-Net inpainting
• boiler_MCAR_2p_1_rpsd2all4.csv - RP + Stable Diffusion 2

Reconstruction Methods (31 total)

Classical Methods (15)

Statistical Imputation (3):

• imputemean - Mean of observed values
• imputemedian - Median of observed values
• imputemode - Mode of observed values

Directional Fill (2):

• imputeffill - Forward fill (propagate last valid)
• imputebfill - Backward fill (propagate next valid)

Interpolation (9):

• interpolatenearest - Nearest neighbor
• interpolatelinear - Linear interpolation
• interpolateindex - Index-based
• interpolatequadratic - Quadratic polynomial
• interpolatecubic - Cubic spline
• interpolatepolynomial - High-order polynomial
• interpolatepchip - Piecewise Cubic Hermite
• interpolateakima - Akima spline
• interpolatespline - Smoothing spline

Machine Learning (1):

• knn - K-Nearest Neighbors
• sarimax - SARIMAX time series model

Image-based Methods (16)

U-Net Models (4):

• gafunet - GAF + U-Net
• mtfunet - MTF + U-Net
• rpunet - RP + U-Net
• specunet - Spectrogram + U-Net

Stable Diffusion 2 - Fine-tuned (4):

• gafsd2all4 - GAF + SD2 (trained on all 4 types)
• mtfsd2all4 - MTF + SD2 (trained on all 4 types)
• rpsd2all4 - RP + SD2 (trained on all 4 types)
• specsd2all4 - SPEC + SD2 (trained on all 4 types)

Pipeline: Image-based Methods

Time Series → Image Transform → Inpaint → Inverse Transform → Time Series
    (CSV)    → (GAF/MTF/RP/SPEC) → (U-Net/SD2) → (Inverse) → (CSV)

Format

Same as source data, with reconstructed values:

timestamp,value
2023-01-01 00:00:00,42.5
2023-01-01 00:01:00,42.7    # Reconstructed (was NaN)
2023-01-01 00:02:00,42.8
...

Statistics

• Total files: 16,201
• Per dataset: ~5,400 files
• Per config: 31 methods
• Size: 25 GB total (1.5 MB per file average)

Archive

# Compressed archive
data_2_fixed.tar.gz (5.3 GB)

# Extract
tar -xzf ../data_2_fixed.tar.gz

---

🎨 images_inpainting/

Description

Image representations of time series at various stages of the inpainting pipeline. Used for visualization and debugging.

Subdirectories

0_original_images/

Complete time series converted to images (ground truth).

Format: [dataset]_[config]_[type].png

1_missing_images/

Corrupted time series as images (input to inpainting models).

Format: [dataset]_[config]_[type]_missing.png

2_fixed_images/

Inpainted images (output from U-Net/SD2).

Format: [dataset]_[config]_[method]_[type].png

3_difference_images/

Visual differences between original and reconstructed.

Format: [dataset]_[config]_[method]_[type]_diff.png

Image Types

Type	Full Name	Size	Description
gaf	Gramian Angular Field	64×64	Polar encoding of temporal correlations
mtf	Markov Transition Field	64×64	State transition probabilities
rp	Recurrence Plot	64×64	Phase space recurrence patterns
spec	Spectrogram	64×64	Time-frequency representation

Statistics

• Total images: 16,324
• Original: ~4,000 images
• Missing: ~4,000 images
• Fixed: ~6,000 images
• Differences: ~2,324 images
• Size: ~4 GB total

Archive

# Compressed archive
data_images.tar.gz (4.0 GB)

# Extract
tar -xzf ../data_images.tar.gz

---

🔬 Experimental Workflow

Complete Pipeline

1. SOURCE DATA (0_source_data/)
   ↓ Inject missing values (MCAR/MAR/MNAR)
   
2. MISSING DATA (1_missing_data/)
   ↓ Convert to images (optional for image methods)
   
3. IMAGES - MISSING (images_inpainting/1_missing_images/)
   ↓ Apply reconstruction (31 methods)
   
4A. IMAGES - FIXED (images_inpainting/2_fixed_images/)
    ↓ Inverse transform to time series
    
4B. FIXED DATA (2_fixed_data/)
    ↓ Compute metrics (MAPE, MAE, RMSE)
    
5. RESULTS (results/quick_experiment/df_final_*.csv)

Files Generated Per Configuration

For each experimental configuration (e.g., boiler_MCAR_2p_1):

1. 1 missing file → 1_missing_data/boiler_MCAR_2p_1.csv
2. 4 original images → images_inpainting/0_original_images/boiler_MCAR_2p_1_[gaf|mtf|rp|spec].png
3. 4 missing images → images_inpainting/1_missing_images/...
4. 31 reconstructed files → 2_fixed_data/boiler_MCAR_2p_1_[method].csv
5. 16 reconstructed images → images_inpainting/2_fixed_images/... (4 methods × 4 types)

Total per config: ~50 files

---

📊 Data Generation Scripts

Source Data

Original data was collected from industrial systems. No generation script needed.

Missing Data

# Generated by iterative_experiment.py
python iterative_experiment.py --generate-missing-only

Reconstructed Data

# Generated by main experiment
python iterative_experiment.py

Images

# Generated automatically during image-based reconstruction
# See: ts_image_inpainting.py

---

💾 Storage & Archives

Individual Archives

Archive	Original Size	Compressed Size	Compression Ratio
data_0_source.tar.gz	8.5 MB	1.8 MB	4.7:1
data_1_missing.tar.gz	964 MB	192 MB	5.0:1
data_2_fixed.tar.gz	25 GB	5.3 GB	4.7:1
data_images.tar.gz	4 GB	4.0 GB	1.0:1 (PNG)
Total	~30 GB	~10.5 GB	2.9:1

Extracting Archives

# Extract all
tar -xzf data_0_source.tar.gz
tar -xzf data_1_missing.tar.gz
tar -xzf data_2_fixed.tar.gz
tar -xzf data_images.tar.gz

# Extract to specific directory
tar -xzf data_2_fixed.tar.gz -C /path/to/destination/

Creating Archives (for backup)

# Compress individual directories
tar -czf data_0_source.tar.gz data/0_source_data/
tar -czf data_1_missing.tar.gz data/1_missing_data/
tar -czf data_2_fixed.tar.gz data/2_fixed_data/
tar -czf data_images.tar.gz data/images_inpainting/

# Compress everything
tar -czf data_complete.tar.gz data/

---

🔍 Data Access Examples

Loading Source Data

import pandas as pd

# Load original time series
df = pd.read_csv('data/0_source_data/boiler_outlet_temp_univ.csv', 
                 index_col='timestamp', parse_dates=True)

print(f"Length: {len(df)}")
print(f"Missing: {df['value'].isna().sum()}")

Loading Corrupted Data

# Load corrupted version
df_missing = pd.read_csv('data/1_missing_data/boiler_MCAR_2p_1.csv',
                         index_col='timestamp', parse_dates=True)

# Count missing values
n_missing = df_missing['value'].isna().sum()
missing_rate = n_missing / len(df_missing) * 100

print(f"Missing: {n_missing} ({missing_rate:.1f}%)")

Loading Reconstructed Data

# Load reconstruction
df_fixed = pd.read_csv('data/2_fixed_data/boiler_MCAR_2p_1_gafunet.csv',
                       index_col='timestamp', parse_dates=True)

# Compare with original
df_original = pd.read_csv('data/0_source_data/boiler_outlet_temp_univ.csv',
                          index_col='timestamp', parse_dates=True)

# Compute error on missing regions only
mask = df_missing['value'].isna()
errors = abs(df_original.loc[mask, 'value'] - df_fixed.loc[mask, 'value'])
mae = errors.mean()

print(f"MAE on missing regions: {mae:.4f}")

Loading Images

from PIL import Image
import numpy as np

# Load original image
img = Image.open('data/images_inpainting/0_original_images/boiler_MCAR_2p_1_gaf.png')
img_array = np.array(img)

print(f"Shape: {img_array.shape}")
print(f"Type: {img_array.dtype}")

---

📈 Dataset Characteristics

Time Series Properties

Dataset	Length	Min	Max	Mean	Std	Trend	Seasonality
Boiler	~10k	25.3	89.7	58.2	12.4	Stable	Weak
Pump	~10k	1.2	98.5	45.6	28.7	None	Strong
Vibration	~10k	0.001	2.456	0.542	0.389	Increasing	None
Water L300	~8.7k	2.1	5.8	3.4	0.8	Seasonal	Strong
Water L308	~8.7k	1.8	6.2	3.2	0.9	Seasonal	Strong
Water L311	~8.7k	2.4	5.5	3.6	0.7	Seasonal	Strong

Missing Data Distribution

By Mechanism:

• MCAR: 33.3% of configs (90 per dataset)
• MAR: 33.3% of configs (90 per dataset)
• MNAR: 33.3% of configs (90 per dataset)

By Rate:

• 2%: 33.3% of configs (90 per dataset)
• 5%: 33.3% of configs (90 per dataset)
• 10%: 33.3% of configs (90 per dataset)

By Dataset:

• Boiler: 270 configs
• Pump: 270 configs
• Vibration: 270 configs
• Total: 810 configs

---

🔧 Maintenance & Cleanup

Checking Data Integrity

# Count files in each directory
echo "Source: $(ls data/0_source_data/*.csv 2>/dev/null | wc -l)"
echo "Missing: $(ls data/1_missing_data/*.csv 2>/dev/null | wc -l)"
echo "Fixed: $(ls data/2_fixed_data/*.csv 2>/dev/null | wc -l)"

# Check for empty files
find data/ -type f -empty

Disk Space Management

# Check space usage
du -sh data/*/

# Remove intermediate images (if needed)
rm -rf data/images_inpainting/3_difference_images/

# Keep only essential data
# Warning: This removes all but source and final results
rm -rf data/1_missing_data/
rm -rf data/images_inpainting/

Regenerating Data

# Regenerate missing data
python iterative_experiment.py --force-regenerate-missing

# Regenerate specific method
python iterative_experiment.py --methods gafunet --force-reprocess

---

📚 Related Files & Documentation

Scripts

• iterative_experiment.py - Main experiment runner (generates 1_missing_data/, 2_fixed_data/)
• ts_image_inpainting.py - Image transformation functions
• calculate_differences.py - Computes metrics from reconstructed data
• generate_training_dataset.py - Generates synthetic training data (separate from this)

Results

• results/quick_experiment/df_final_*.csv - Aggregated metrics
• results/quick_experiment/final_results.json - Complete experimental results

Documentation

• EXPERIMENT_1_DESCRIPTION.md - Experiment methodology
• README.md (project root) - Main project documentation

---

⚠️ Important Notes

Data Provenance

• Source data: Real industrial sensors (anonymized)
• Missing data: Synthetically generated with controlled mechanisms
• Reconstructions: Generated by automated pipeline
• Images: Deterministic transformations from time series

Reproducibility

All data can be regenerated from source:

1. Source data (0_source_data/) - original, no changes
2. Missing data (1_missing_data/) - regenerate with same seeds
3. Fixed data (2_fixed_data/) - regenerate by running experiments
4. Images (images_inpainting/) - regenerate during reconstruction

Data Quality

Validated:

• ✅ All files have correct format
• ✅ No corrupted CSV files
• ✅ Missing rates match specifications
• ✅ All methods completed successfully

Known Issues:

• Some image-based methods may fail on extreme cases (handled gracefully)
• Large file count may cause filesystem issues on some systems

---

🔒 Data Usage & Citation

Usage Guidelines

This dataset is for research purposes as part of the time series inpainting project.

Allowed:

• Analysis and visualization
• Method comparison
• Result reproduction
• Academic publication

Restrictions:

• Do not redistribute source data without permission
• Cite original data sources when publishing
• Acknowledge reconstruction methods used

Citation

If you use this dataset in your research:

@misc{timeseries_inpainting_data_2025,
  title={Time Series Inpainting Experimental Dataset},
  author={Dariusz Kobiela and Jarosław Kobiela and Adam Kurowski and Agnieszka Landowska},
  year={2025},
  note={Dataset containing 270 experimental configurations across 3 industrial time series, 
        reconstructed using 31 methods}
}

---

📧 Support

Questions?

1. Check file naming conventions above
2. Review experimental workflow
3. Consult related scripts documentation
4. Check project main README

Issues?

• Missing files: Regenerate using experiment scripts
• Corrupted data: Re-run specific configurations
• Disk space: Use archives, remove intermediate data
• Performance: Use archived versions when possible

---

Dataset Version: 1.0
Last Updated: 1.12.2025
Total Size: ~30 GB (uncompressed), ~10.5 GB (compressed)
Status: Complete ✅