Feature Design for Predictive Maintenance - MATLAB
Video Player is loading.
Current Time 0:00
Duration 3:03
Loaded: 5.40%
Stream Type LIVE
Remaining Time 3:03
 
1x
  • Chapters
  • descriptions off, selected
  • en (Main), selected
    Video length is 3:03

    Feature Design for Predictive Maintenance

    Use the Diagnostic Feature Designer app to extract, explore, and rank features for training predictive maintenance algorithms. Extract features from a broad set of domains such as the time domain and frequency domain, time series models, and rotating machinery.

    Published: 30 Sep 2022

    To build a predictive maintenance algorithm, you need data. You probably have a lot of it. But it's hard to tell the difference between healthy and faulty operation. So you decide to try machine learning. In order to train your machine learning model, you need to extract the right features from your data. But feature engineering is difficult. How do you know which features are best?

    The Diagnostic Feature Designer app in Predictive Maintenance Toolbox makes it easy to explore your sensor data, apply data processing functions, interactively extract a variety of features, like time and frequency domain features, and identify the best features, all without writing any code.

    To get started quickly, you can use auto features to automatically extract a large set of features from your entire data set all at once. The features are ranked, so you can quickly see which features best differentiate between the different fault classes. But you're an engineer. You know something about what parameters are important and which ones might indicate a problem with your machine. So you might want to choose the features you generate.

    In the Diagnostic Feature Designer, you can select and generate simple or advanced time domain features, such as basic statistics, like mean and standard deviation, features specific to stationary time series signals, specialized features for rotating machinery, and advanced non-linear features. If you have high frequency data, you may want to explore features in the frequency domain. You can generate a power spectrum from your signal data, then extract frequency domain features.

    You can even extract application-specific features for bearings or gear meshes. If you need custom features for your project, you can use those too. You can create and apply any MATLAB function to compute custom features in both the time and frequency domains. Once you've generated all of these features, how do you know which ones are best? You can rank all of the features together using unsupervised, supervised, or prognostic ranking methods to see which ones are the best candidates for training your machine learning model.

    You can export them directly to the Classification Learner app to train and compare the performance of a variety of models. Or you can generate a MATLAB function to recreate the features programmatically on new data. There are lots of examples in the Predictive Maintenance Toolbox documentation to help you get started. To learn more about the Diagnostic Feature Designer and Predictive Maintenance Toolbox, visit the link below.

    View more related videos