I have a 40k Hz time-series data of vibration, which is used to predict equipment failure. The goal here is to make a system that predicts it automatically. I am thinking of a couple of ways but not sure what the best way is to frame the problem from machine learning perspective. I'll explain how an analyst usually does it manually:

Given a 1D vibration data, we want to break it down into smaller pieces (1 second in below figure, sometimes more). And then, using Fast Fourier Transform (FFT), we can bring this chunk into frequency domain and do some analysis there. A couple of things that are analyzed include: amplitude of spikes, how often spikes appear, vibration values near frequency of interest, etc. Then we can decide whether it is anomalous or not. Finally, we do it repeatedly on different chunk of data and bring it to the final conclusion.

How do we formulate this problem using machine/deep learning? Do we need to still do FFT or can we do it with raw time-series data? What are the features and target variable like? I'd really appreciate any ideas as to how to tackle this problem.


FFT Process


1 Answer 1


One way to approach anomaly detection is by learning a threshold. Learning a threshold is more straightforward after Fast Fourier Transform (FFT). For example, define a threshold as any value that is twice the median.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.