Prognostic maintenance has been a goal in many industries for many years. Now it’s on the horizon.
Prognostic maintenance (also known as condition-based maintenance, predictive maintenance, or simply prognostics1) is the ability to know the condition of equipment, and to plan and perform maintenance accordingly before a critical failure. There are two schools within prognostic maintenance: data-driven prognostics and model-based prognostics.
Data-driven prognostics require massive sensor data on the condition of the system at hand (the ‘cause’ side, measuring potential causes of performance change) and on the performance of the system (the ‘effect’ side). This data is continuously collected and run through machine learning (ML) algorithms in hopes of detecting correlations and using these to make prognoses of future failures.
Model-based prognostics involve physical failure models that are combined with suitable sensor readings to perform failure prognoses. This is based on tried-and-true technology (e.g. stress-strain models used in design of structures), but in recent years the massive deployment of sensors, and the related data collection and processing has made this technology much more useful for prognostics.
Data-driven prognostics have some different capabilities than model-based prognostics and are therefore suitable for different purposes. Data-driven prognostics are, in principle, simple to set up based on a generic approach. However, there are few critical failures that ML algorithms can learn from, and without good data and examples to learn from, the prognoses of the ML algorithms may not be very accurate. Model-based prognostics have advantages as well. Although they are more effort to set up, and are more specific to a particular piece of equipment, they have the potential for immediate success. They can lead to better prognoses, with less data, than ‘black box’ data-driven prognostics. As a result, industry should continue to strive for obtaining good failure models for all critical equipment. Data-driven methods should be used as an aid in constructing and verifying failure models.What lies ahead?
We expect to see prognostic maintenance begin to spread near 2025. This is when we expect the beginnings of a wide rollout in the industry, with at least some working systems on large assets. Early versions are already in use today on simple systems, but the tools and techniques available are not yet reliable, nor do they work for complex machines and systems. Ultimately, successful and reliable prognostic maintenance will require us to produce good models and connect them to data, ML algorithms, and data-driven prognostics. Prognostic maintenance can be said to have arrived when it becomes a dependable standard in the industry, greatly reducing the need for corrective maintenance. Still, not every failure will be predictable.Contributors
Main author: Siegfried Eisinger
Contributor: Sture Angelsen
Editor: Thomas Fries
- K. Goebel et al. Prognostics: The Science of Making Predictions. CreateSpace Independent Publishing Platform. 2017.