🧰 Library for generating synthetic time series data with anomalies for testing anomaly detection algorithms.

🚧 Work in progress (Anomalies should be as good - bad? - as possible)

Entry point: main.py

• Usage Examples
  • Combination Strategies
• Base Architecture
  • BaseGenerator
• Noise-Based Generators
  • NormalGenerator
  • UniformGenerator
  • LaplaceGenerator
  • ExponentialGenerator
  • GammaGenerator
  • PoissonGenerator
  • PinkNoiseGenerator
• Pattern-Based Generators
  • CostantGenerator
  • DriftGenerator
  • SinusoidGenerator
• Utility Generators
  • MaskGenerator

import numpy as np
from generators import NormalGenerator, DriftGenerator

# Generate time series shape
shape = (1000, 3)  # 1000 time steps, 3 variates

# Example 1: Add Gaussian noise
normal_gen = NormalGenerator(
    shape=shape, 
    combine_mode="add", 
    combine_domain="time",
    mean=0.0, 
    std=0.5
)
base_ts = np.zeros(shape)
noisy_ts = normal_gen.generate_and_combine(base_ts)


# Example 2: Add drift
drift_gen = DriftGenerator(
    shape=shape,
    combine_mode="add",
    combine_domain="time",
    drift_type="linear",
    drift_rate=0.01
)
final_ts = drift_gen.generate_and_combine(spiked_ts)

• Time domain addition: Direct addition of anomaly to signal
• Time domain multiplication: Scaling of signal by anomaly factor
• Frequency domain addition: Add anomaly in frequency space (good for spectral anomalies)
• Frequency domain multiplication: Multiply in frequency space (good for filtering effects)

The abstract base class that all generators inherit from. It provides:

• Shape management: Define time series dimensions (sequence length, number of variates)
• Combine operations: Merge generated data with existing time series
• Domain control: Apply operations in time or frequency domain
• Mode control: Add or multiply generated data

Common parameters for all generators:

• shape: Tuple defining (sequence_length, number_of_variates)
• ts: Existing time series to clone shape from
• combine_domain: "time" or "frequency" domain for combination
• combine_mode: "add" or "mul" for combination operation

Generates Gaussian (normal) distributed noise.

Use case: White noise, natural measurement errors, random fluctuations

Generates uniformly distributed noise where all values within a range are equally likely.

Use case: Random noise with bounded range, testing robustness to uniform perturbations

Generates Laplace (double exponential) distributed noise with heavier tails than normal distribution.

Use case: Noise with occasional large deviations, outlier-prone data

Generates exponentially distributed noise (always positive values).

Use case: Waiting times, inter-arrival times, lifetime data anomalies

Generates gamma distributed noise with flexible shape (always positive values).

Use case: Positive-valued anomalies, skewed distributions, duration modeling

Generates Poisson distributed noise, suitable for count data.

Use case: Count anomalies, event frequency changes, discrete value noise

Generates 1/f^α noise (colored noise) with power spectral density inversely proportional to frequency.

Use case: Natural phenomena noise, long-range correlations, realistic background noise

Note: Uses FFT method to generate frequency-domain colored noise

Generates constant-value anomalies at random locations with configurable length.

Use case: Sensor freezing, stuck values, plateau anomalies

Generates gradual drift/trend patterns over time.

Use case: Sensor degradation, gradual system changes, trend anomalies

Generates sinusoidal patterns with configurable frequency, amplitude, and phase.

Use case: Periodic anomalies, oscillatory patterns, seasonal effects

Generates boolean masks for selective anomaly application.

Use case: Control which time points and variates receive anomalies

Note: Returns boolean mask instead of numeric values

© 2025 G.S.