I am currently working on a project that requires fault detection of engines based on the available data.
The available information are parameters that were recorded during annunciation of the previous faults, historical data on the engines of that model and obviously the engine itself, and some rules that were defined in the control logic to trigger the faults.
Unfortunately, the triggering logic provides many false positives despite rigorous QA before deployment. Therefore, using an ML or AI algorithm or any other approach that could be helpful in this area, I would like to update the triggering logic that would minimize the false positives and improves accuracy and precision of detection while does not affect performance (speed) significantly.
Any help in this area is appreciated.
Hi Afshin,
Well, usually autoenconders can deal with this kind of problem.
Another strategy is to create a synthetic dataset with all your data and you put random values between 0 and 1 for each instance. Then you train a deep learning model. Usually, anomalies will result in values out of 0-1 range.
The paper entitled Ensemble and Deep Learning Applications for Showcase Fault Analysis is a good example of this application.
I hope that my answer might help you!
Best regards,
Fábio
Dear Afshin Rahimi,
Well, I am interested to exchange you some of my experiences, instead of some static statements, for example, fault detection and diagnosis (FDD) can be divided into two major categories— model-driven and data-driven.
Data-driven, you are specially talking about data, has three major distinct parts namely feature extraction (FE), feature selection (FS) and classification, in which you can keep your footprint. It can be seen these three processes are inter-linked to achieve a high accuracy.
Hope you are aware of the difference between detection and diagnosis in the context FDD. So, FE clearly presents the detection and faults visualization issues. Obviously, ML/AI is the best option for classification. Now you can explore relevant literature and AI-based state-of-the-art methods for FDD.
You may also review some of my papers. I am looking forward to seeing you. Thanks
Regards,
M M Manjurul Islam
University of Ulsan, South Korea
A review of process fault detection and diagnosis
Part III: Process history based methods
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Artificial Intelligence