TSAG - Time Series Anomaly Generator

A time series anomaly generation tool for quickly generating anomalies in time series machine learning problems.

Applications in Anomaly Detection

Model Evaluation

• The lack of labelled data severely hinders progress in anomaly detection development due to unreliable performance evaluation.
• TSAG generates different classes of anomalies to evaluate classification performance on different anomaly types.

Class Imbalance

• Many machine learning techniques require much anomalous training data to be sufficiently performant in anomaly classification.
• TSAG generates different types of anomalies, enabling Root Cause Analysis of detected anomalies.

Installation

pip install tsag

Usage

Anomaly Template

# Create a reference template time series
n = 10
template = timeseries[:n]

Single Tranformation

from tsag import PointAnomaly

# Generate point anomaly
point_anomaly = PointAnomaly(template)
point_anomaly.plot()

# Insert generated anomaly into time series data
augmented_timeseries = point_anomaly.insert(timeseries, index=None)

Multiple Transformations

from tsag import FrequencyShiftAnomaly, AmplitudeShiftAnomaly, RangeShiftAnomaly, CompoundAnomaly

args = [
    # [Generator, {Arguments}],
    [FrequencyShiftAnomaly, {'ratio': 1/3}],
    [AmplitudeShiftAnomaly, {'ratio': 1/3}],
    [RangeShiftAnomaly, {'ratio': 1/2}],
]

# Generate compound anomaly
compound_anomaly = CompoundAnomaly(template, *args)
compound_anomaly.plot()

# Insert generated anomaly into time series data
augmented_timeseries = compound_anomaly.insert(template, index=None)