SAS AIoT in Oil & Gas Refinarie

[Abhishek_Kapure]

Dear All, Hope you are good & healthy. I am working in one of SAS AIoT platform in Oil & Gas refineries project. In the given project I have created predictive models like Logistic regression, Decision tree for fault prediction using SAS VDMML. Now, I am trying to predict Remaining Useful Life of Asset or Equipment but not getting from where to start. So can anyone guide or give any inputs on how we can proceed on this topic ahead. Although anyone can suggest some another models (Alternate approach) to predict fault & remaining useful life or some other scenarios on IoT Data. All suggestion is always welcome. Thanks!

[sbxkoenk]

Hello,

This is very interesting.

In 2015 I have worked almost a year on a project to predict "Remaining Useful Life of Asset / Equipment".
And it's indeed a bit different than modelling / predicting when a piece of equipment breaks down. Before breaking down the piece is not behaving properly for months with very bad production quality as a result. You want to avoid that obviously.

I will dig into my old documentation later this weekend to give you more input, but I still remember the type of survival analysis I used to model this :
It was a discrete-time logistic-hazard model using PROC LOGISTIC (SAS/STAT).

And PROC LIFEREG (SAS/STAT) can also be used to model Right-Censored Failure Time Data.

Yes, you will have to deal with censoring indeed. That's an additional challenge in this type of modelling.

"talk" to you later,
Koen

[sbxkoenk]

Hello again,

I shortly have dived into my old projects. 
I could find a lot on that particular 2015 project (remaining useful life [RUL] prediction), but unfortunately NOT the great .pdf papers I had on this.

Let me know if I can help further.
The project I did in 2015 was also with big data (sensor data) and in the Cloud (AWS).

The model that I used was a

→ Multinomial discrete-time logistic hazard regression.

This model allows for:

• Time-dependent covariates: inputs whose value change over the tenure (the time the asset has been or is at risk for the event(s))
• Time-varying effects: interactions between tenure (time) and a covariate

This model supports right-censoring and left truncation.

This model allows for multiple outcomes. Multiple reasons for failure or the event can be modeled as competing risks. The event or failure types need to be mutually exclusive and exhaustive.

Good luck,
Koen

[PriyaSharma]

Hi Abhishek, 

As an alternate approach to Survival Analysis you can try methodologies for condition based monitoring. Using SVDD or MWPCA you can detect anomalies or degradation in assets that can lead to faults.  

Here are few examples: 

1. Support Vector Data Description (SVDD)
   1. Analysis of Aircraft Engine Degradation
   2. Anomaly Detection in Air Handling Units
2. Moving Window PCA (MWPCA) or Subspace Tracking (SST)
   1. Anomaly Detection in Floodlights for Smart Campus
   2. Detecting Degradation in Wind Turbines

Please reach out if you need more details. 

Thanks,

Priya

[sgowardipe · Accepted solution]

AIoT new Tech in Oil and Gas