This example shows how to process your data in the app in preparation for feature extraction. If you want to follow along with the steps interactively, use the data you imported in Import and Visualize Ensemble Data in Diagnostic Feature Designer. Use Open Session to reload your session data using the file name you provided.
If you have no session data, execute the steps for loading and importing data in Import and Visualize Ensemble Data in Diagnostic Feature Designer.
A key step in predictive maintenance algorithm development is identifying condition indicators. Condition indicators are features in your system data whose behavior changes in a predictable way as the system degrades. A condition indicator can be any feature that is useful for distinguishing normal from faulty operation or for predicting remaining useful life. A useful feature clusters similar system status together and sets different status apart.
Diagnostic Feature Designer lets you design features that provide these diagnostics.
For some features, you can generate features directly using signals you imported.
For other features, you must perform additional signal processing, such as filtering and averaging, to have meaningful results.
The processing you perform depends both on the computational requirements of the feature and the characteristics of your systems and your system data. This example shows how to:
Process your data in preparation for feature extraction
Generate various types of feature
Interpret the effectiveness of your features in histograms
The data for this system represents a transmission system with rotating parts. The variables include tachometer outputs that precisely mark the completion of each shaft revolution. The data, therefore, is an ideal candidate for time-synchronous averaging.
Time-synchronous averaging (TSA) is a common technique for analyzing data from rotating machinery. TSA averages rotation by rotation, and filters out any disturbances or noise that is not coherent with the rotation.
TSA is useful for isolating fault signatures that repeat each rotation, such as perturbations from gear-tooth defects. Features generated from a TSA signal rather than the original vibration signal provide clearer differentiation for rotational fault conditions. This advantage applies true even for features that are not specifically for rotating machinery.
To compute the TSA of the vibration data:
Select Filtering & Averaging > Time-Synchronous Signal Averaging
In Signal, specify
In Tacho Information, select Tacho
signal, and specify
Tacho/Data. Clear Compute nominal
Click OK to start the TSA computation for each of the 16 members of the ensemble. A progress bar shows the status while the computation progresses.
When the computation concludes:
The app adds a new signal variable
The signal trace plots
Vibration_tsa. The time axis
of this trace is less than the four seconds long. The original vibration
data was 30 seconds long. The shorter timespan reflects the duration of
a single rotation for each member.
The member shaft rates diverge. This divergence is evident in the increasing misalignment of the peaks during the rotation, and the fact that the member traces stop at different times.
The TSA signal gives you enough information to start generating time-domain features, but you must provide a spectrum to explore spectral features. To generate a power spectrum, select Spectral Estimation > Power Spectrum.
From Signal, select your TSA signal.
The power spectrum processing results in a new variable,
Vibration_tsa_ps. The associated icon represents a linear
model, consistent with the AR model default in the power spectrum dialog box.
A plot of the spectra appears in the plot area. As with Signal Trace, a Power Spectrum tab provides options for plotting. These options are similar to Signal Trace. There is no Panner option because Panner works only with time and not frequency.
Generate features based on general data statistics, using the TSA signal as your source. Select Time Domain Features > Signal Features.
Change Signal to
Vibration_tsa/Data. The default is for all the
features to be selected. Clear the selections for Shape
Factor and Signal Processing Metrics.
For every selected feature, the app computes a value for each ensemble member and displays the results in a histogram. Each histogram contains bins containing the number of feature values which fall within the bin range. The Histogram tab displays parameters that determine the content and resolution of the histograms.
The histogram groups, or color codes, the data according to the condition
faultCode in Group By.
Blue data is healthy and orange data is faulty, as indicated by the legend
(color coding may appear different in your session). For feature values where
the healthy and faulty labels overlap, the color appears brown due to the
overlap between blue and orange.
You can get a rough idea of which features are effective by assessing which
ones clearly segregate blue data from orange data.
CrestFactor appear effective. There are only small areas
of overlap. Conversely,
Kurtosis have large amounts of overlap. These features
appear ineffective for this data and this condition variable.
By default, the app plots the histograms for all the features in the feature table. You can focus on a subset of histograms by using Select Features. Use Select Features to limit the histogram plots to the first four in the feature table.
The histogram view now includes only the features you selected.
Control the appearance of the histograms using the parameters in the
Histogram tab, which activates when you generate the
CrestFactor feature appears to separate
healthy and unhealthy data almost completely. Investigate whether this result is
sensitive to resolution. In the Histogram tab, the
auto setting of bin width results in a resolution
of 0.1 for
CrestFactor. Enter a bin width 0.05, and click
At this resolution, both
ImpulseIndicator appear to completely separate healthy
from faulty data.
ClearanceFactor still has some mixed data,
but to a lesser degree than with the larger bin width.
Kurtosis had a smaller bin width of 0.002 with the
auto bin width setting. Changing the bin width to
0.05 results in a single bin that contains all the
Histograms visualize the ability of features to separate healthy from
unhealthy data. You can also get a numerical assessment using Group
Distance. The group distance represents the separation between
the healthy and unhealthy data distributions. Click Group
Distance. In the dialog box, select
CrestFactor in Show grouping for
The group distance, represented by the KS Statistic, is 1. This probability value represents complete separation.
Kurtosis histogram showed substantial intermixing.
The KS Statistic in this case is 0.5, reflecting the intermixing in the histogram.
Restore Bin Width to
Since you have rotating machinery, compute rotating machinery features. Select Time-Domain Features > Rotating machinery Features. In the rotating machinery dialog box, specify your TSA signal to analyze, and select the TSA signal metrics.
Other feature choices in the dialog box use the filtered TSA difference and regular signals as a source. This example does not use the difference and regular signal-based features because computation for those signals assumes common shaft speed.
The app automatically adds the new features to the feature table
and the Select Features list, and plots the new histograms
at the top of the histogram display.
Kurtosis histograms are essentially the same whether they
were computed as signal features or rotating machinery features, since both
computations used the TSA signal as a source.
Compute spectral features from the power spectrum you generated earlier. Click
Spectral Features. In Spectrum,
Set the frequency band. The power spectrum x scale changes automatically from log to linear when you open the spectral features dialog box. When you move the frequency slider, the plot shades the region that the slider setting covers. To capture the power spectrum peaks efficiently, limit the frequency range to 10 Hz.
The histograms show substantial intermixing of healthy and unhealthy data in one or more of the bins for all three features.
You now have a diverse set of features.
Save your session data. You need this data to run the Rank and Export Features in Diagnostic Feature Designer example.
The next step is to rank those features to determine which ones provide the best indication of system condition. For more information, see Rank and Export Features in Diagnostic Feature Designer.