Predictive Maintenance on time series without failure occurence

matteosoldini95

Hi,

I need your help to solve a problem of predictive maintenance.

I have a time series in which two machines are checked every 10 seconds by some sensors.

As in the title, I don't have a historic failure occurence for these machines, so I can't apply the model of predictive maintenance example.

Could you explain me what can I use to know if one of these machines is working badly and need maintenance or if there are some sensors with weird values?

Thank you all.

land

Ahh, good thing is: You don't need deployment then in a real factory, right? Academic results of how well that would work are sufficient?

Then it comes down to time series prediction, which is far easier than deploying it in a real world scenario.

The Series Extension of RapidMiner should be helpful there. Take a look at the Windowing operator to get multiple past time points into one example. Then you can predict the current values from them. If it works, you have a working baseline model modeling the normal behavior of your machine. If reality is multiple times far away from your model's predictions, machine starts to behave strange...

So that's the idea. But still as hard as any data science and machine learning problem.

Greetings, 

Sebastian

 

Thomas_Ott

@matteosoldini95 boy, your asking a tough question because the answer will depend on the operating characteristics of the sensors themselves, as well as the tolerances of what they're monitoring. For example, if a temperature sensor can operate at 50C and breaks down at 60C, then might want to know that. Likewise if the item that's being made ranges in temperature from 10C to 40C and any reading above or below those thresholds, the quality suffers, well you need to know that too. 

 

I would start with run some standard deviations over the time series to see where they suddenly expand or contract. That might give you a clue to get started on. 

land

Hi,

well, not an easy problem. Obviously I would recommend to get some help to get started. What you then can do, if you want to go further than the univariate approach Thomas depicted, is to build a baseline model. Hence you predict each sensor value for the next time point and compare what happened from what was expected. If the prediction is off farther from what happened than usual, you see somethings going to be out of order.

Can be hard to deploy such a setup as you need sort of real time distributed in an usually IT unfriendly environment. Are you going to deploy that at all or is it just an exercise?

 

Greetings,

 Sebastian

matteosoldini95

Hi,

my teacher asked me to try to solve this problem for my graduation thesis but it could require too much time. Anyway I'm interested in learning how to solve it, so I asked your help.

Thank you for your attention,

Matteo

 

matteosoldini95

Thank you Thomas, I'll start following the approach that you explained.

Greetings,

Matteo

land

Ahh, good thing is: You don't need deployment then in a real factory, right? Academic results of how well that would work are sufficient?

Then it comes down to time series prediction, which is far easier than deploying it in a real world scenario.

The Series Extension of RapidMiner should be helpful there. Take a look at the Windowing operator to get multiple past time points into one example. Then you can predict the current values from them. If it works, you have a working baseline model modeling the normal behavior of your machine. If reality is multiple times far away from your model's predictions, machine starts to behave strange...

So that's the idea. But still as hard as any data science and machine learning problem.

Greetings, 

Sebastian

 

matteosoldini95

Yes, I just need some academic results to understand if the behavior of my machine is normal. Maybe, if I get some good results, we will try to apply the model on other data sets to do predictive maintenance.

Thank you,

Matteo.