Anomaly Detection: Detect anomalies in time series data using new anomaly detection models based on deep learning
In R2025a, you can now detect anomalies in time series data using four new anomaly detection models that are based on deep learning. You need only normal data to train these models. The trained models can then identify abnormal subsequences within a time series that deviate from normal behavior.
The models have a simplified interface that does not require user expertise in deep learning, but that still allows interaction with model parameters and properties for model tuning.
The four models are as follows:
For information on creating and using these models, see Detecting Anomalies in Time Series Using Deep Learning Detector Models.
Anomaly Detection: Enhancements in distance-based anomaly detection methods
The distance-based anomaly detection methods that were introduced in R2024b have been enhanced for R2025a.
In R2025a, you can now do the following:
Use a GPU to increase the processing speed for all distance functions. Previously, these algorithms could use only a CPU.
Use the STOMP algorithm (scalable time series
ordered matrix profile) algorithm for even faster results when you compute
matrixProfile with a GPU. You can still select the
STAMP algorithm (scalable time series anytime
matrix profile) if you need anytime capability, that is, you need to be able
to stop the algorithm before it completes and still obtain an acceptably
accurate solution.
Use multivariable data to compute multidimensional solutions for distanceProfile and multivariate solutions for
matrixProfile.
In addition, a new findMotif function has been added for R2025a to identify motifs.
This function complements the findDiscord function that was introduced in R2024b. Both functions
operate on matrixProfile output.
For more information, see the reference pages that are linked in the previous list.
Diagnostic Feature Designer: Support for regression workflows
In R2025a, you can now use feature ranking algorithms that are suited for regression problems when you rank your features in the app. You can also now export your features to Regression Learner to create and compare regression models. In previous releases, you could rank only by classification methods, as well as unsupervised and prognostic methods, and could export features only to Classification Learner for model development.
The app now provides two new ranking algorithms—MRMR (Minimum Redundancy Maximum Relevance algorithm) and Relieff, and implements both the classification and the regression versions of each. The app selects the version to use based on the nature of your condition variable. You can therefore use these algorithms not only for regression ranking, but also for multiclass and two-class ranking. For more information, in Diagnostic Feature Designer, under Feature Ranking Tab, see Supervised Ranking.
To export your features from the app to Regression Learner, in either
the Feature Designer tab or the Feature Ranking
tab, select Export Features to Regression
Learner from the Export menu to bring up a
new menu that lets you specify which features to export, which condition variable to
use, and whether you want to filter the features for streaming data. Then, click
Export.
For more information, see Export Features to Regression Learner.
Diagnostic Feature Designer: Additional Updates
Auto Features: In R2025a, you can now compute the core features set separately from more computationally intensive advanced features. In previous releases, the core features were grouped with more advanced features. The Standard set of features now contains primarily core features like signal features and spectral features. Features that require more complex computations such as with residual signals and detrended time series are now available with the new Advanced selection option. The Rotating Machinery feature set remains the same.
For more information on using Auto Features, see Generate Features Automatically in Diagnostic Feature Designer.
Datastore interaction: The ability to write data back to the original
fileEnsembleDatastore or
simulationEnsembleDatastore source that was available
in previous releases has been removed. In R2025a, the only option for saving
data when using a datastore as a source is to export the data to the
workspace and save the data you want from there. Note that this restriction
applies only to Diagnostic Feature Designer. The underlying
fileEnsembleDatastore and
simulationEnsembleDatastore datastores still support
writing to the source.
New Example: Accelerate Fault Diagnosis Using GPU Data Preprocessing and Deep Learning
A new example shows how to use GPU computing with Parallel Computing Toolbox™ to accelerate data preprocessing and deep learning for predictive maintenance workflows. For more information, see Accelerate Fault Diagnosis Using GPU Data Preprocessing and Deep Learning.