What are some of the advanced feature engineering techniques that you use for predictive maintenance?

One item we have implemented where I work is to indirectly measure the density of material that is cut by a dicer. The dicer feeds a loss in weight feeder. As the dice blades get worn, the material is not cut as finely and the material density becomes less. The feeder compensates by feeding material at a faster speed. When the feed ratio (material fed/feeder speed) drops below a certain threshold, I programmed the PLC that has the calculation to send an email to our planning department to schedule a blade replacement.

The first step is distinguishing between historical time and waveform analysis. Historical time analysis looks at global values over set periods (seconds, minutes, hours, days, or weeks) to establish alert and alarm thresholds. Conversely, waveform time analysis, especially with vibration signals, involves more intricate processes like frequency analysis, pattern identification of beating, and impacts at specific frequencies or modulations. Early in my career as a vibration analyst, it seemed daunting to identify faults from waveforms; however, I quickly learned that bearing failures and gear faults are more discernibly analyzed through waveforms than spectral analysis.

Reflecting on the graph analysis of wind turbines, a successful monitoring strategy I've use involved graphically correlating vibration levels with the power output of the generator. Whenever possible, we should correlate any operational parameter with the machine performance data. This practice helps create 'normal curves' which are invaluable for monitoring and predictive maintenance. Such correlations not only enhance the reliability of condition monitoring systems but also aid in proactive maintenance planning.

At Power-MI, we've taken a comprehensive approach to catalog all types of assets, inspection techniques, detectable failures, and industry sectors. This structured data approach helps maintain a detailed record of failures, facilitating statistical analyses to spot recurring issues or patterns. Additionally, processing maintenance analysts' recommendations (free text) through automatic text analysis and categorization tools have proven extremely beneficial.

After a period of time , you already know to change the blades. You don't need the computer to tell you. It should be a part of the maintenance program on a routine basis.

From a feature selection perspective, it is recommended not to overwhelm the predictive models with too many features. Techniques such as correlation analysis, information gain and lasso regression are enough.

Well, I am not an expert and have no experience on these tools, but what I've seen is this: -TSFRESH helps automatically pull out information from time series data, such as tracking how machine temperatures change over time. - NetworkX examines connections within a factory's network to pinpoint crucial points for monitoring. -NLTK analyzes written maintenance logs to find common issues. This why a client asked us to export our Power-MI failures records. -Scikit-learn selects the most important features, like vibration levels in machines, to focus on. Featuretools simplifies linking data from different sources, such as performance logs across various machines. Feature engineering tools help make sense of complex data.

A common misconception I’ve noticed among predictive maintenance engineers is the overemphasis on real-time data analysis. Predictive maintenance typically benefits more from timely (not necessarily real-time) data collection, which is sufficient for most machinery since equipment degradation occurs over time, providing ample warning before a potential failure. This has exceptions where protection systems should be used.