Solved

Can the signals of warnings, alarms, and failures be used to tag anomalies and/or failures in the compressor?

  • 4 October 2022
  • 9 replies
  • 150 views

Userlevel 2

There are some signals that in the names or the description have the words: alarm, warning, or fault. 


E.G. There is this alarm signal, (in red ) that I’m comparing with one of the vibration signals.  And the alarm seems to be on when the vibrations are close to 0 and the Compressor is off. I’m plotting the label as a combination of the name and the description, but the externalId of the alarm is: pi:160268

In most cases that alarm is on when the compressor is off, except in this case, Where I do not see any problem with the vibration signal:

 

There are these other two alarms, from the plots: one is on when the compressor is off and the other one when it is starting up. But I do not see any strange behavior in the signal when they occur. 

Besides the alarms, I’ve also looked into some of the fault signals but those seem to be on for very long times: 

Is there another way to use the alarm and fault signals to label anomalies or failures? I don’t know if using the vibration signals is the best reference to look for anomalies I just use them because usually are a good indicator of the health of the motor. 

 

Alternatively, I would just label the anomalies by eye, and I would like to have a reference to label them properly. E.G. from this image: 

 

I would label the values in the red box as anomalies, is that corrected or it can be a normal behavior?

 

And from this second image, I’m less confident that the values in the red boxes are anomalies and that may be just noise. Which would be the label for these spikes? 

 

 

icon

Best answer by kelvin 18 October 2022, 14:28

View original

9 replies

Userlevel 3
Badge

I love the digging and exploring in the OID data you are doing @Luis Ramon Ramirez Rodriguez. I’m going offshore to Valhall in a week, and I will note down the timestamps and have a look at what has been happening from the operators end. If there are some events that we don’t yet have in the publicdata CDF project. but right now I think the triggering of the common alarm of the switchgear system can be a lot of things. but I will speak to the guys what is triggering the common alarm on the switch gear for the 1.stg compressor. 

And I don’t think it is frequency components show in the vibration timeseries, yes it is aggregated but it is aggregated in max length peak to peak vibration of the driveshaft. would be nice with frequency components also from the vibration system in the publicdata project. I can see what we can do something about getting the frequency specturm aswell “live” into the CDF project. if its easily available from Bently Nevadas System 1 that is the 3. party SCADA system for the vibration monitoring of the compressor. 

 

 

 

Userlevel 2

Thanks @kelvin , 

 

One other question, I notice that in several cases the end time seems to be earlier than the start time, as in the bellow image: 

 

 

Which date is a good default option in such cases? 

Userlevel 3

The Events of type Workorder are what describe what to fix on a high level, and the Workitem’s are the actual steps that a work order consists of. 

Btw, you can filter both on event type, and on asset. In this screenshot I’m filtering for Workorder on the asset subtree of system 23. You can also filter for time periods, specific metadata, etc.

 

Userlevel 2

Great, thanks @kelvin 

I looked into the Events tab in fusion, does it makes sense to consider the work orders and worktask as "maintenance events"? I didn't see a specific label for "Maintenance "

 

 

Userlevel 3

@Luis Ramon Ramirez Rodriguez 

Maintenance events, work orders etc. Are commonly modeled as Events in Cognite Data Fusion. You can find the events in the Events tab in Fusion, or by using the client.events.list() method in the Python SDK :) 

Userlevel 2

Hi  @kelvin 
I have not found an alarm that clearly indicates an incident. Regarding the maintenance events, how can I look for them? is it in the same alarm signals or is there another place? 

Userlevel 3

Hey @Luis Ramon Ramirez Rodriguez 

The data you’re looking at seems to be already aggregated vibration data, meaning the FFT has already been performed and aggregated into an amplitude value for a certain frequency spectrum before coming into Cognite Data Fusion. A frequency of six samples per minute also strengthens this hypothesis, as raw signals from a vibration sensor would typically be in the order of 10s - 100s thousands per second (tens to hundreds of kilohertz).

Further, I do not know whether there are actual faults in any period of the data. Can you find any related maintenance events or alarms for the same equipment that would indicate any previous incidents? We would typically work closely with the subject matter experts (SMEs) at our customers to identify possible failures :)

Userlevel 2

Hi @kelvin , Thanks for the response.

I also looked into the vibration signals (e.g. VAL_23_YT_96134_01:Z.X.Value) I downloaded the data at the highest frequency available, about 6  samples per minute, and tried to find anomalies using FFT, but did not find any pattern in the frequencies. I'm guessing it is because of one of the following reasons: 

* The analysis I made was incomplete/incorrect 
* There is no pattern in the vibration frequencies 
* The sampling rate is not high enough to correctly map the signal's frequency spectrum.   

Do you have any comments on that? 

Userlevel 3

Hey @Luis Ramon Ramirez Rodriguez! What you’re digging into is one of the challenges we face every day; interpreting and generating insights from industrial data, without access to the “correct answer”.  Thus, I can not tell you what is the true label, simply because it does not exist. However, information about which data relates to each other and how it relates, what we call context, is represented in the Cognite Data Fusion project you’re exploring.

In industrial systems, it is not uncommon that you get alarms when shutting down the equipment, as pressures, temperatures etc. Drop outside their pre-defined boundaries.

We typically filter out the periods where the equipment is off first, either by generating a rule that detects it, or, ideally, use a status signal if it exists.

Do you @Stig Harald Gustavsen have anything to add about these data specifically?

Reply