updateDetector
Update settings of a trained deep learning anomaly detector and recompute detection threshold
Since R2025a
Syntax
Description
Add-On Required: This feature requires the Time Series Anomaly Detection for MATLAB add-on.
Set New Threshold Manually
detector = updateDetector(detector,ThresholdMethod="manual",Threshold=Threshold)Threshold that you
specify. Use this syntax when you want to adjust the threshold directly rather than
recompute it from input time series data.
For examples that you can open and experiment with, see Use UpdateDetector to Tune Anomaly Detector Performance Without Retraining
This function requires Deep Learning Toolbox™.
Recompute Threshold Using New Data
detector = updateDetector(detector,data)detector using the
data in data and the current settings of the detector model, as long
as detector.Threshold is not set to
"manual".
Use this syntax when you want to compute the threshold using a different set of data
than the original training data. The data data that you use must
consist only of only normal data, that is, that contains no known anomalies or anomalous
data.
Set Additional Options
detector = updateDetector(___,Name=Value)Name=Value arguments. These
options set the corresponding properties of detector, include
thresholding and execution-related properties. The options you can set depend on the
detector model you are updating.
Note that, when you use updateDetector, there are no default
values for model settings, as the model is already configured. You must explicitly specify
any settings, such as the threshold method and the corresponding threshold parameter, that
differ from the initial detector settings.
For example, suppose you first create a time series isolated forest detector model
using the "kSigma" threshold method, which has an associated default
threshold parameter of 3. After viewing the results, you decide to try
the "ContaminationFraction" method instead, using a threshold parameter
value that identifies the top 5% of the anomaly scores as anomalous. To do so, use the
following command. detector =
updateDetector(detector,data,ThresholdMethod="contaminationFraction",ThresholdParameter=0.05)
Examples
Suppose that you have a trained TCN detector model, and you want to try to improve performance by updating the threshold with a specific value that you specify.
The current detection performance and the anomaly scores are shown in the following figures.
In the top plot, the anomaly appears as vertical spikes against a smooth sinusoidal signal. Red spikes indicate detections. Blue spikes indicate undetected anomalies. In the bottom plot, the horizontal threshold line is higher than some of the anomaly spikes, causing the detector to miss the lower-score anomalies.
Change the threshold to 2.0 by using the following commands.
detector_tcn = updateDetector(detector_tcn,ThresholdMethod="manual",Threshold=2.0);
results = detect(detector_tcn, sineWaveAbnormal);
plot(detector_tcn,sineWaveAbnormal{3})Assess the resulting detection and anomaly score plots.
With the new threshold, the detector recognizes all the spikes as anomalies.
The USAD detector model has tuning parameters that you can adjust after training.
The current detection performance using default settings is shown in the following figure. The detector identifies only a fraction of the anomaly spikes.
The USAD detector includes two sensitivity
coefficients—Alpha and Aeta—which must sum
to 1. The default values are 0.7 and 0.3. Change these to 0.9 and
0.1.
detector_US = updateDetector(detector_US,sineWaveNormal,ThresholdMethod="kSigma",... ThresholdParameter=3,Alpha=0.9,Beta=0.1); results = detect(detector_US, sineWaveAbnormal); plot(detector_US,sineWaveAbnormal{3})
Assess the resulting detection and anomaly score plots.
With the new threshold, the detector recognizes all the spikes as anomalies.
