vaelstmAD

Create anomaly detector model that combines variational autoencoder (VAE) and long short-term memory (LSTM) networks to detect anomalies in time series

Since R2025a

collapse all in page

Syntax

detector = vaelstmAD(numChannels)
detector = vaelstmAD(numChannels,Name=Value)

Description

detector = vaelstmAD(numChannels) creates a VaelstmDetector model with numChannels channels for each time series input to the detector.

After you create the detector model, you can train, test, and modify it to obtain the level of performance you require. For more information about the anomaly detector workflow, see Detecting Anomalies in Time Series Using Deep Learning Detector Models.

detector = vaelstmAD(numChannels,Name=Value) sets additional options using one or more name-value arguments.

For example, detector = vaelstmAD(3,Normalization="range") creates a detector model for data containing three input channels and with a data normalization method of "range", which, by default, rescales the data range to [0 1].

Input Arguments

collapse all

Input

Number of input channels in each time series, specified as a positive integer. All time series inputs must have the same number of channels.

Name-Value Arguments

expand all

Specify optional pairs of arguments as Name1=Value1,...,NameN=ValueN, where Name is the argument name and Value is the corresponding value. Name-value arguments must appear after other arguments, but the order of the pairs does not matter.

Example: detector = vaelstmAD(3,Normalization="range") creates a detector model for data containing three input channels and with a data normalization method of "range", which, by default, rescales the data range to [0 1].

Window

expand all

Observation Window Length for each time series segment, specified as a positive integer. The detector uses this value to divide each input time series into segments.

Training stride length of the sliding window in the training stage, specified as a positive integer. TrainingStride controls the number of overlapping samples. If you do not specify TrainingStride, the software sets the stride length to the value of 1 to create nonoverlapping windows.

Detection window length of each time series segment, specified as a positive integer that is smaller than ObservationWindowLength. This value determines the length of the prediction segment that the model uses for detection.

Detection stride length of sliding window in detection stage, specified as a positive integer. DetectionStride controls the number of overlapping samples. If you do not specify DetectionStride, the software sets the stride length to the value of DetectionWindowLength to create nonoverlapping windows.

Threshold

expand all

Method for computing the detection threshold, specified as one of these:

  • "kSigma" — Standard deviation of the normalized anomaly scores. The parameter k determines the threshold within the standard deviation levels that identifies an anomaly. The value of k is specified by ThresholdParameter.

  • "contaminationFraction" — Percentage of anomalies within a specified fraction of windows, measured over the entire training set. The fraction value is specified by ThresholdParameter.

  • "max" — Maximum window loss measured over the entire training data set and multiplied by ThresholdParameter.

  • "mean" — Mean window loss measured over the entire training data set and multiplied by ThresholdParameter.

  • "median" — Median window loss measured over the entire training data set and multiplied by ThresholdParameter.

  • "manual" — Manual detection threshold value based on Threshold.

  • "customFunction" — Custom detection threshold method based on ThresholdFunction.

If you specify ThresholdMethod, you can also specify ThresholdParameter, Threshold, or ThresholdParameter. The available threshold parameter depends on the specified detection method.

Parameter for determining the detection threshold, specified as a numeric scalar.

The way you specify ThresholdParameter depends on the specified value for ThresholdMethod. This list describes the specification of ThresholdParameter for each possible value of ThresholdMethod.

  • "kSigma" — Specify ThresholdParameter as a positive numeric scalar. If you do not specify ThresholdParameter, the detector sets the threshold to 3.

  • "contaminationFraction"— Specify ThresholdParameter as a as a nonnegative scalar less than 0.5. For example, if you specify "contaminationFraction" as 0.05, then the threshold is set to identify the top 5% of the anomaly scores as anomalous. If you do not specify ThresholdParameter, the detector sets the threshold to 0.01.

  • "max", "mean", or "median" — Specify ThresholdParameter as a positive numeric scalar. If you do not specify ThresholdParameter, the detector sets the threshold to 1.

  • "customFunction" or "manual" — ThresholdParameter does not apply.

Detection threshold, specified as a positive scalar. The source of the Threshold value depends on the setting of ThresholdMethod.

  • If ThresholdMethod is "manual", you set the value.

  • If ThresholdMethod is "customFunction", the function you specify in ThresholdFunction computes the value.

  • For other values of ThresholdMethod, specify ThresholdParameter as the input to the specified method. The software uses this method to compute the threshold value.

Function for computing a custom detection threshold, specified as a function handle. This argument applies only when ThresholdMethod is specified as "customFunction".

  • The function must have two inputs:

    • The first input is a vector of scalar window anomaly scores.

    • The second input is a vector representing all point-level anomalies.

    For example, suppose that the value of detectionWindowLength is 10, the value of detectionStride is set to be nonoverlapping, and the time series length is 10001. Then the first input vector has a length of 1000 and the second input vector has a length of 10000.

  • The function must return a positive scalar corresponding to the detection threshold.

Model

expand all

Number of convolutional layers in the downsampling section of the model, specified as a positive integer.

Filter size of each convolutional layer in the VAE network, specified as a positive integer or integer vector. The value of NumDownsampleLayers determines how many convolutional layers the network contains.

  • If you specify FilterSize as a scalar, the size of each filter is the same in all layers.

  • If you specify FilterSize as a vector, the size of each filter in the ith vector element is equal to the value ith vector element. The length of the vector must be equal to the number of layers you specified in NumDownsampleLayers.

Number of filters in each convolutional layer for the VAE network, specified as a positive integer.

Dimensionality of the compressed representation of the input signal by the VAE model, specified as a positive integer. This value impacts the ability of the VAE model to capture the most important features when reconstructing the input data.

Number of hidden units in the LSTM layers, specified as an integer row vector. The length of this vector determines the number of LSTM layers in the LSTM model.

Dropout probability used to avoid overfitting, specified as a nonnegative numeric scalar less than 1. All convolution layers share the same dropout probability.

Normalization

expand all

Normalization technique for training and testing, specified as "zscore", "range", or "off".

Output Arguments

collapse all

Anomaly detector model, returned as a VaelstmDetector object.

Version History

Introduced in R2025a

See Also

| | |

Topics