This is the second part of a discussion on determining the future state of a machine sometimes referred to as Machinery Health Prognostics. I was fortunate enough to get some real world experience working with incredibly smart and capable people in this area, but must admit the math can be daunting. In the last blog I promised we would get into this discussion. So just to set the basics: A prognostic determination, a future state of a machine can be provided from three classical approaches.
The first is a physics based model. Physics based models look at inputs and outputs and can correlate data to a cause and effect process that result in some sort of alarm when things go south. I like to describe this in simple terms as push on the accelerator and the vehicle moves faster, right? Obviously, this simple case does not describe all the multiple inputs such as fuel flow, fuel rail pressure, filter restriction, combustion etc. However, a complex model can consider the relationships between hundreds of inputs and outputs.
Today because of computational power that is appearing in the form of cloud computing and “Big Data Analysis” this sort of data can be processed quickly and accurately. Physics based models do have some drawbacks. They are time consuming to build and therefore simplified models can be less accurate. For especially complex or poorly understood systems they can also be difficult to build.
The next type of approach used is a data-driven one. This depends less on understanding the physics of the machine but relies on large sets of data to observe change from normal. My experience in working with this approach is that it has some limitations and can be less accurate than a physics based model. Data is sent to a trending application sometimes referred to mathematically as Change Point Detection. The software is programmed to look for subtle changes or trends and then a determination is made. Data scientist talk about “training” the model where they strive to avoid false positives or narrow the alarm bands to prevent nuisance calls. The data driven approach works when, of course, data is available and the Physics based approach is considered onerous , too expensive or the machine is complex and /or poorly understood. One problem with data-driven approaches can be lower confidence than a physical model provides.
Finally, the real world solution, is to use some elements of physical and data in a Hybrid Model and almost all successful prognostic solutions are based on this approach. Real data and a Physical model can give great insights.
Making a prognostic (Remaining Useful Life) determination then requires applying advanced math to predict a future state. This is a relatively new engineering science and has some great potential.
If you are interested in really getting down in the weeds on this stuff I would suggest you go to the PHM Society website www.phmsociety.org. You will find presentations and tutorials that describe Prognostic Machinery Health Management by leaders in the field. They also host an annual conference.
So where did we go from here? If you can understand the condition of a machine and know what is wrong (Diagnosis), what to do about it (Recommendation)and how long you have to fix it (Prognostics)-this is the brass ring for a maintenance program. I can remember when a newly trained vibration analyst would burst into my office and tell me I had a problem with a machine. This was always presented with techno talk like “You have a high amplitude at one times run speed.” I would usually ask “How long will it run” Their answer: “Don’t know it is your problem.”
No wonder the predictive technologies got such a bad rap and we did not trust anything. I do have to admit, as I mastered the predictive technologies, I did enjoy using the same jargon to mystify and confuse my production colleagues. In those days we were treated as a service organization not a partner-“We break it, you fix it.”
That about wraps up a really short discussion on Prognostics and is intended to just give an idea about the science. This is a deep and evolving engineering discipline. Many suppliers are entering the space, all of course, claiming to have it licked. My experience is “buyer beware.” It takes two very important things to be successful in this area. First is you must have data. The next is domain knowledge. Generally, suppliers of these solutions have neither data or domain knowledge readily available and still depend on you to provide both. I have also found people confuse this body of work with Machine Learning and in a future blog I will talk about that very important feedback step to the process.
That about finishes this discussion. Next blog I will talk a little about reasons organizations can’t sustain success and why we keep dropping back in advancing machine reliability based on my years in the battle.
Until next time. Good luck with the struggle.
At XRT we have been in the trenches and what I would have given to know how long a machine could run until it would fail in those bygone days.