Here we look at the difference between predictive and preventive maintenance and use an industrial case study to illustrate the point.
Predictive maintenance uses the condition of equipment to determine when to do maintenance. Preventive maintenance uses statistics alone.
For example, a car manufacturer might say to change the oil every 16,000 KM. A boat engine manufacturer might say do that after X number of hours. Those numbers are based on statistical models developed by the engineers who build and designed those machines. Those models relate wear to time and cumulative force.
Predictive maintenance tosses out the time-based or usage-based approach to doing maintenance and instead looks at some metric that can be measured, like vibration, that correlates with wear.
Below we look at an example of this from the steel manufacturing industry. In that case, a transducer is fit onto the housing of a shaft to measure vibration. When vibration goes outside norms this indicates that the bearings need replacement.
Using the Fourier Transformation to do Industrial Predictive Maintenance
In a PhD thesis, “Application of Predictive Maintenance to Industry Including Cepstrum Analysis of a Gearbox,” Matthew Aladaesay looks at PM practices at several New Zealand manufacturers. The goal is to predict when a furnace used to make steel will need to have its bearings changed.
Workers know that the bearing is failing because it begins to leak oil. The goal is to change bearings before they get to that point. If they wear out then the machine has to be taken apart and the teeth on the gears have to be replaced. That is costly downtime. So a project was undertaken to monitor the vibration of the machine.
As the bearings wear out, the drive shaft increases vibration. The approach is to fit a transducer on the bearing housing. This converts mechanical energy to electrical, which is something that can be measured.
This type of vibration analysis is a common technique used in predictive maintenance models. It takes the premise that vibration increases as bearings wear out and then plots the observed vibration over time. Converting that time series to a frequency distribution is called spectrum analysis.
The model uses what is called Fourier Transformation (FT) to calculate the amplitude and frequency of sine waves produced by the spectrum analysis. Amplitude means how high the curve is and frequency means how far apart these peaks in the curve are. It is something easily read from a graph.
Below is a graph from that plant showing the frequency spectrum. This is a graphical display of the FT. One can easily see the amplitude increasing as well as the frequency. In this plot, the bearing is shown to be failing.
The frequency distribution fits into a standard or other statistical distribution. This lets the model flag outliers and data points outside expected norms, thus showing when the bearings need replacing.
So, we have shown how sensor input can be used to predict when a component is going to fail. This lets workers do maintenance before the machine breaks down, which would require a costly repair. This type of data and analysis fits perfecting into the IoT, where a sensor—in this example a transducer—can feed data into a model and transfer that to the cloud where it can flag this situation. Big data and the wide availability of machine language programming frameworks make this kind of spectrum analysis and Fourier transformations relatively easy to do, so than one does not need their own PhD statistician.
Because Industry 4.0 emerged and IoT has gained a wide following, a common practice today is that the manufacturer would operate the PM data model and gather the data in their cloud where they monitor these machines and then alert their customer when they need maintenance. That is especially true in very expensive machines.