How can I design a neural network to detect and classify faults in a system?
3 views (last 30 days)
David Willingham on 7 Mar 2022
Edited: David Willingham on 8 Mar 2022
Before even attempting to design and then train a neural network model, I recommend starting with the data.
Does your simulation data capture the faults you want to detect?
Does it capture them multiple times?
Can features be added to the simulation data that makes detecting features easier? For example, do standard statistical measures such as moving averages, max peaks, range etc differ from normal sometime before a fault occurs?
Image Analyst on 8 Mar 2022
One additional thing to consider is that you have to have lots of training faults. Otherwise it won't find them. Let's say you have faults 0.1% of the time and you have a dataset that has 10 thousand samples but only 10 had faults. Well, when if you train with that dataset, and then try to use it, it might not find faults. I mean it could say everything is without fault and be 99.9% accurate just by saying all samples are perfect. So obviously you don't want it to miss valid faults so you have to have lots of them in the training set.
Also do you want it to just say the signal has a fault or has no faults? Or do you want it to find the location and duration of the fault?