General Question about Predictive Maintenance

[drieder]

Hello Everyone,

after following this guide (the example where the failure of a motor shall be predicted) I am wondering about the process about predictive maintenance in general.

In this example it is somehow a fact, that if the grease of the motor is low it has a high chance of failure. I guess this information could have been retrieved with the Analytics Methods aswell (eventhough it is not explained how they found out this specific corelation in the first place).

So as far as I understand, the predictive model generated here can tell me whether or nor the grease is low for a given set of parameters and their values. 

I am wondering where exactly is the "predictive" aspect? Like it is now, it could only tell me in Real time that "now the grease is low" but isn't that already to late if I wanted to prevented this state in the first place? 

Please explain to me how I can use this data to actually prevent this states (to know that this state will happen soon in the future) instead of pointing out that this state "probably occurs right now". Or maybe just point out if there is something wrong with my understanding in general.

Thank you very much,

Dominik

[mgoel]

@drieder

As per the given example the predictive aspect is based on using Analytics on the different vibration sensors you can predict time to fail. Some machines ave hundreds of sensors and humans have difficulty in understanding in all the different way these sensors can cause time to fail.

It will be good to use ThingWorx analytics for you and see all the aspects using in analytics builder like Model creation, Signals and scoring. you can use trail version to see how a predictive model works with different goals. How a signals tells us accurate result which we want. Different Algorithms of analytics works behind all these.

Do let me know in case of any question.

Regards-Mohit

[DmitryTsarev]

@drieder

Dominik, my post is not a direct response to your question, but more like a generalization / reflecting on it, so sorry in advance if something written below is obvious / not of interest.

Thingworx Analytics could be useful not only for predicting events (it's still an important part of it, though), but also for providing diagnostic analytics. While it's sometimes thought that diagnostics analytics can be done without any sophisticated software, in many cases, especially for hardware with dozens and even hundreds of sensors, it can't.

In the guide you mentioned, Analytics "predicts" the low grease state (i.e. failure) in the sense that it provides you with the likely combination of parameters that leads to the failure so you could pro-actively avoid such state by, say, changing operational modes of the equipment.

Of course, this information could be retrieved with analytics methods, without using Thingworx - after all, all Analytics does, is apply known ML algorithms (currently TW Analytics supports 6 of them - Decision Tree / Random Forest, Linear and Logistic Regression, Gradient Boost and Neural Net), but one of the key features / advantages of Thingworx Analytics is that you can quickly get some useful insights without hiring a data scientist and spending time to implement ML methods with other tools.

In some cases, even if Analytics (or any other tool) was able to predict some event, it might be not of much use, since there would be not enough time to react.
Generally predicting makes sense only if there is enough time to react to prevent some undesirable event (or to enjoy the pleasure of anticipating a desired event).

[drieder]

Hello,

first I want to thank you for your detailed insight, it made the general approach clear to me. So in the case of the sensors, the developers could have used the results of Analytics to raise an alert whenever one of the sensors is nearby a critical value within a threshold of +- x in order to prevent it right?

Another thing I came accross is this webinar called Make your Data Talk! Using Visualizations as an Interface for Intelligence. At 32:46 you see a matrix with the current risk of failure for a specific component (low, medium, high etc).

Do you think they made this categorization based on the marked value on the picture? Is it the probability of the goal being true?

Thank you for your help.

Regards,

Dominik

[cmorfin · Accepted solution]

Hi Dominik

Regarding this video, I think at 34:10 they give hint on how they built the categories, and yes it based on the model output .

Regarding your initial question about prediction.
We have had some good answers earlier but I wanted to point one thing.

We can have 2 types of datasets with ThingWorx Analytics: standard and time series.
In a standard dataset, as used in the guide you followed,  each record is independent from the others and the scoring is based on the value of each features (sensor values for example) at a specific time. As @DmitryTsarev mentioned this is useful if you have enough time to react. However this is very important when you have a very large amount of parameters (feaures / sensor values) to  take into account.

In a time series dataset, the records depends on the features but also on the previous records . The scoring process in this case will predict the future value.
Those datasets are usually much smaller than the standard one, as they have a much more limited number of features. However with this type of dataset, if you take a sensor measure every 5 min, to predict a goal value. the score will predict the goal value for now + 5 min. See The Help Center on Time Series Prediction.

Hope this helps

Kind regards

Christophe

[drieder]

Hello Christophe,

I already wondered what the prediction with time series is about, so now it makes perfectly sense.

Regards,

Dominik