Skip to content

Latest commit

 

History

History
280 lines (198 loc) · 13.8 KB

base-oscillations.md

File metadata and controls

280 lines (198 loc) · 13.8 KB
Raw

Base Oscillations

The generator comes with the following base oscillations in gutenTAG/base_oscillations:

  • sine
  • cosine
  • square
  • random-walk
  • cylinder-bell-funnel
  • ecg
  • polynomial
  • random-mode-jump
  • formula
  • sawtooth
  • dirichlet
  • mls
  • custom-input

Base oscillations can have an underlying trend. This trend can be any of the above base oscillations.

Using the variance property, you can add noise to the base oscillation. The general kind of base oscillation is always the same for all channels of a time series. However, noise and other random parameters differ between channels.

Common Parameters

Parameters for a new time series.

Name Type Description
name String Optional name for timeseries
length Int Length of entire time series
channels Int Number of dimensions
semi-supervised Bool Whether a train file without anomalies should be generated
supervised Bool Whether the train file should contain labels

These parameters can be set for all base oscillations.

Name Type Description
variance Float Noise factor dependent on amplitude
trend Object Defines another base oscillation as trend that gets added to its parent object. Can be recursively used!
offset Float Gets added to the generated time series

Sine

sine

Parameters

Name Type Description
frequency Float Number of sine waves per 100 points
amplitude Float +/- deviation from 0
freq-mod Float Factor (of base frequency) of the frequency modulation that changes the amplitude of the sine wave over time. The carrier wave always has an amplitude of 1.

Cosine

cosine

Parameters

Name Type Description
frequency Float Number of cosine waves per 100 points
amplitude Float +/- deviation from 0
freq-mod Float Factor (of base frequency) of the frequency modulation that changes the amplitude of the wave over time. The carrier wave always has an amplitude of 1.

Square

square

Parameters

Name Type Description
frequency Float Number of square waves per 100 points
amplitude Float +/- deviation from 0
freq-mod Float Factor (of base frequency) of the frequency modulation that changes the amplitude of the wave over time. The carrier wave always has an amplitude of 1.
duty Float Load percentage (time spent in high state compared to low state).

Random Walk

rw

Parameters

Name Type Description
amplitude Float +/- deviation from 0
smoothing Float 0.01 Smoothing factor for convolution dependent on length

Cylinder Bell Funnel

cbf

Parameters

Name Type Description
avg-pattern-length Int Average length of pattern in time series
amplitude Float Average amplitude of pattern in time series
variance-pattern-length Float Variance of pattern length in time series
variance-amplitude Float Variance of amplitude of pattern in time series

ECG

ecg

Parameters

Name Type Description
frequency Float Number of hear beats per 100 points

Polynomial

polynomial

Parameters

Name Type Description
polynomial List[Float] Please see numpy documentation

Random Mode Jump

random-mode-jump

Parameters

Name Type Description
frequency Float Number of jumps in Time Series
channel_diff Float Value difference of absolute mode values between channels
channel_offset Float Value offset from 0 in both directions
random_seed Int Random seed to have the similar channels

Sawtooth

sawtooth

Parameters

Name Type Description
frequency Float Number of waves per 100 points
amplitude Float +/- deviation from 0
freq-mod Float Factor (of base frequency) of the frequency modulation that changes the amplitude of the wave over time. The carrier wave always has an amplitude of 1.
width Float >0.5 rising, <0.5 falling; sets proportional width of rising slope.

Dirichlet

dirichlet

Parameters

Name Type Description
frequency Float Number of waves per 100 points
amplitude Float +/- deviation from 0
periodicity Float Sets the periodicity of the dirichlet wave (higher numbers have a higher complexity), but the high-level period length is not changed. However, even periodicity numbers double the period length.

Maximum Length Sequence (MLS)

mls

Parameters

Name Type Description
amplitude Float +/- deviation from 0
complexity Float The number of bits used to generate the sequence. This controls the length of the repeating sequence and its complexity.
smoothing Float Smoothing factor for convolutional smoothing of the generated bit sequence (highly recommended). Default is 0.01.

Custom Input (custom-input)

Enables the user to add synthetic labeled anomalies to a time series file.

The current requirements on the data are:

  • Column headers
  • Always select the column for test data creation
  • Optionally select column for training data creation
  • No anomalies (and no labels) present
  • The input file path(s) may be specified relative to the working directory

Parameters

Name Type Description
input-timeseries-path-test String location of an input time-series csv file relative to the working directory used to make test set
input-timeseries-path-train String location of an input time-series csv file relative to the working directory used to make train set
use-column-test Int, String columns from test input file to use
use-column-train Int, String columns from train input file to use

Formula

Parameters

Name Type Description
base Int, Object Base time series to operate on.
operation Object Operation done on base object.
aggregation Object Aggregation done on base object.

The formula base oscillation is a simple math engine that can perform primitive math operations on multiple already generated channels.

Syntax

Formula Object

The FormulaObj is the base to every formula. It consists of three fields: base, operation, and aggregation. The base defines the first operand. It can either be a channel id (int) or another FormulaObj. The base can be the only field in a FormulaObj and then simply copies the channel with the right id (base:int). It can also have a field operation or aggregation. However, it cannot have both at the same time.

# FormulaObj
base: Union[int, FormulaObj]
operation: Optional[Operation]
aggregation: Optional[Aggregation]

Operation

The Operation is the vectorized math operation that is performed on the base. It consists of two required fields: kind, and operand. The kind field is an enum that defines which math operator to perform. The operand field is the operand that is calculated to the before-defined base. The operand can be a scalar value or another FormulaObj.

# Operation
kind: Enum[+, -, *, /, **]
operand: Union[float, FormulaObj]

Aggregation

The alternative to the Operation is the Aggregation field. This field defines an aggregation operation on the base. Its result is either a scalar or an array. It consists of two fields: kind, and axis. The kind field defines the kind of aggregation performed on a numpy array. The optional axis field defines the axis of the array the aggregation is performed on. If no axis is defined, the aggregation will return a scalar.

# Aggregation
kind: Enum[sum, min, max, median, std, var]
axis: Optional[int]

Example

The following formula:

import numpy as np
a = np.arange(10)
b = np.ones(10)
channels = [a, b]

# formula
(channels[0] + channels[0]) * channels[1].sum()

can be defined as the following FormulaObj:

formula:
  base:
    base: 0           # (channels[0]
    operation:
      kind: "+"       # +
      operand:
        base: 0       # channels[0])
  operation:
    kind: "*"         # *
    operand:
      base: 1         # channels[1]
      aggregation:
        kind: "sum"   # .sum()