UsadDetector
Description
Add-On Required: This feature requires the Time Series Anomaly Detection for MATLAB add-on.
The UsadDetector object uses a deep learning network
architecture with dual encoders to implement a detector model capable of being trained to
detect anomalies in time series data using only normal data. You create this object with the
usAD function.
This detector requires Deep Learning Toolbox™.
By comparing reconstructed values with observed data within a detection window, the USAD detector identifies anomalies as significant deviations from expected patterns. You can control the sensitivity of the detector by modifying a set of threshold properties. You can also adjust two sensitivity parameters associated with the dual encoders.
Creating a UsadDetector object is the first step in a design workflow
that includes creating, training, testing, assessing, and, if necessary, modifying the
detector. For information on the workflow for developing a Predictive Maintenance Toolbox™ anomaly detector, see Detecting Anomalies in Time Series.
This anomaly detector model was inspired by the architecture proposed in the paper in [1].
For more information on the functions this workflow uses, see Object Functions.
Creation
Create a UsadDetector object by using the usAD
function.
Properties
Object Functions
train | Train deep learning anomaly detector and obtain detection threshold |
detect | Detect anomalies in time series using trained deep learning detector model |
plot | Plot detected anomalies and anomaly scores generated from deep learning anomaly detectors |
plotHistogram | Plot histogram of anomaly scores and detection threshold for trained deep learning anomaly detector |
updateDetector | Update settings of a trained deep learning anomaly detector and recompute detection threshold |
References
[1] Audibert, Julien, Pietro Michiardi, Frédéric Guyard, Sébastien Marti, and Maria A. Zuluaga. “USAD: UnSupervised Anomaly Detection on Multivariate Time Series.” In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, 3395–3404. Virtual Event CA USA: ACM, 2020. https://doi.org/10.1145/3394486.3403392.
