Time Series Anomaly Detection Dataset Generator

A synthetic dataset generator for time series anomaly detection tasks, supporting both univariate and multivariate time series with configurable anomaly patterns.

Features

• Univariate & Multivariate Support: Generate single or multi-dimensional time series data
• Flexible Anomaly Injection: Control the ratio of anomalous samples in your dataset
• Customizable Sequence Length: Fixed or randomly distributed sequence lengths
• Rich Metadata: Each sample includes detailed attributes about the time series and anomalies
• Two Generation Modes:
  • Attribute-based generation (recommended)
  • Metric-based generation using synthetic.json

Installation

Requirements

numpy
tqdm
networkx  # for multivariate generation

Setup

pip install numpy tqdm networkx

Quick Start

Basic Usage

from src.generate_dataset import generate_dataset
import pickle

# Generate 100 univariate samples with anomalies
dataset = generate_dataset(
    num_samples=100,
    seq_len=1000,
    anomaly_sample_ratio=1.0,
    is_multivariate=False
)

# Save the dataset
with open('dataset.pkl', 'wb') as f:
    pickle.dump(dataset, f)

Multivariate Time Series

# Generate multivariate time series with 5 features
dataset = generate_dataset(
    num_samples=100,
    seq_len=2500,
    anomaly_sample_ratio=1.0,
    is_multivariate=True,
    num_features=5,
    use_attribute_set=True  # Recommended
)

Parameters

Parameter	Type	Default	Description
num_samples	int	1000	Number of samples to generate
seq_len	int, optional	None	Length of each time series. If None, randomly generated (100-10000)
anomaly_sample_ratio	float	0.5	Ratio of samples containing anomalies (0.0-1.0)
is_multivariate	bool	False	Whether to generate multivariate time series
num_features	int, optional	None	Number of features for multivariate data
activate_function	bool	False	Apply activation functions in multivariate generation
metrics	list, optional	None	Specific metrics for anomaly generation
use_attribute_set	bool	False	Use ALL_ATTRIBUTE_SET for generation (recommended)

Output Format

Each sample in the dataset is a dictionary containing:

{
    'normal_time_series': np.ndarray,  # Time series without anomalies
    'time_series': np.ndarray,         # Time series with anomalies injected
    'labels': np.ndarray,              # Binary labels (1 = anomaly, 0 = normal)
    'attribute': dict                  # Metadata about the time series
}

Univariate Attribute Structure

{
    'metric': str,           # Metric type
    'anomalies': list,       # List of anomaly descriptions
    'pattern': str,          # Base pattern type
    # ... other attributes
}

Multivariate Attribute Structure

{
    'attribute_list': list,  # Attributes for each feature
    'num_features': int,     # Number of features
    'is_endogenous': list,   # Which features have endogenous anomalies
    'dag': str              # Directed Acyclic Graph structure
}

Examples

Example 1: Anomaly-Only Dataset

# Generate dataset with only anomalous samples
dataset = generate_dataset(
    num_samples=500,
    seq_len=1000,
    anomaly_sample_ratio=1.0,  # 100% anomalies
    is_multivariate=False,
    use_attribute_set=True
)

Example 2: Mixed Dataset

# Generate dataset with 50% anomalous and 50% normal samples
dataset = generate_dataset(
    num_samples=1000,
    anomaly_sample_ratio=0.5,
    is_multivariate=False
)

Example 3: Variable-Length Sequences

# Generate samples with random sequence lengths (100-10000)
dataset = generate_dataset(
    num_samples=200,
    seq_len=None,  # Random length
    anomaly_sample_ratio=0.8
)

Running from Command Line

You can modify the configuration at the bottom of generate_dataset.py and run:

cd src
python generate_dataset.py

This will generate a dataset with the configured parameters and save it to data_generate/samples/.

Project Structure

TSAD_dataset_gen/
├── src/
│   ├── generate_dataset.py      # Main dataset generation script
│   ├── ts_generator.py           # Univariate time series generator
│   └── ts_multi_generator.py    # Multivariate time series generator
├── data_generate/
│   └── samples/                  # Output directory for datasets
└── README.md

Notes

• Recommended Mode: Use use_attribute_set=True for more flexible generation without JSON constraints
• Memory Usage: Large datasets with long sequences can consume significant memory
• Generation Time: Multivariate generation with many features may take longer
• Anomaly Guarantee: When anomaly_sample_ratio=1.0, the generator ensures all samples contain at least one anomaly

Acknowledgements

This project references and adapts parts of the code from ChatTS.

License

This project is provided as-is for research and educational purposes.