Approaches and lessons learned

[ricardoevvargas]

Do you have an idea on how to explore the 3W's features and would you like to ask our community for opinions?

This is our dedicated space for this type of interaction. For each idea, leave a comment for others to reply. It is hoped that this discussion will help community members achieve their goals more efficiently by re-using approaches and lessons learned by other members.

Thanks for the contribution!

[utkuaslan85]

First of all, I am pleased to thank you for this work. It is almost impossible to find publicly available dataset related with oil & gas for freelance researchers or students to study on. This work may be really helpful for some like me.
I am a MSc Subsea Engineering student at Robert Gordon University (Online-distance learning). And I am also an R&D Specialist on AI. I am planning to work on 3W dataset in my MSc dissertation which I will start this semester. I am thinking to use GANs to create synthetic data to increase data and check whether this will increase performance of ML algorithms I will try in the study.
Can you please share your kind opinion about my proposal? If you think this is not a good idea, I would be happy to hear your advice if you want to share.

[cjmunaro]

Your proposal is very attractive, since most anomalies 5,6,7,8 have less than 8 instances (times series) of real data. Care must be taken, however, due to the diversity of behaviors of data in different time series. I suggest to work at instance level ot at least grouping instances with similar behavior. As an example, I attach a figure showing 5 variables for anomaly number 8 for the three available instances. You can see the behavior of the variables can be very different for the same anomaly in different instances. Data was normalized and time-shifted to produce this figure.
image

[utkuaslan85]

Thank you very much for the advice and the figure. I started to work in detail to find best approach to deal such data distribution.

[ricardoevvargas]

Hi, @utkuaslan85. Glad you liked the 3W and that it will be useful for your work. I agree with @cjmunaro: your proposal is really interesting. It is expected that some datasets (including the 3W dataset) have event types that are rare, critical and difficult to model with a phenomenological approach. In these scenarios, we generate simulated and hand-drawn instances. Verifying the hypothesis that a GAN is able to improve the training process of a ML algorithm to the point of increasing its performance with a test set is a very useful research. Depending on the result of this research, we will have one more approach to help in the detection/classification of real events.