This fascination with MED led me to begin experimenting with different filter design optimization cost functions that were better-designed for rotating machine faults. If anyone else is exploring new deconvolution techniques - I recommend using the non-linear optimization Matlab toolkit to solve any cost functions you come up with iteratively. Using this method, you can try out any cost function you can dream up without solving any of the math. Save the math until you have good results here first. Anyways, this led to the papers below for Maximum Correlated Kurtosis Deconvolution (MCKD) and Multipoint Optimal Minimum Entropy Deconvolution (MOMEDA), as well as the convolution region fix for MED (MEDA). Generally, I feel the deconvolution methods make more sense than most other vibration fault detection methods. For example, with the Wavelet Transform then selecting a band that has maximum Kurtosis can be thought of generally as just applying a bandpass filter - it is taking a frequency range of the signal and discarding the rest of the information + spectrum. A lot of information is lost here.
Although the deconvolution techniques have come a long way in rotating machine fault detection, there is a lot more work to be done and discovered. Here are a few brainstorming bullets for inspiration:
- The deconvolution methods solve for the inverse filter yielding the fault impulses. What does this filter describe? Can you solve for the original fault path system?
- How does MOMEDA normalized MKurt change as a fault progresses?
- MOMEDA is better than MED and MCKD. Focus on this instead I think.
- Multiple accelerometers mounted on the same machine grading the same fault? I can't think how, and maybe that's fine.
- Can you separate the fault period from multiples of the fault period?
- With MOMEDA you seem to be able to accurately track the rotational speed of elements in a machine from just the acceleration. Is this valuable to any applications?
- Maybe you can pose deconvolution problems with different MOMEDA target goals, such as goals based on fault dynamics associated with rotor imbalance?
- Fault detection on a machine with changing speeds should be possible MOMEDA, something like a pathfinding approach for the impulses for this?
- A description of the MOMEDA spectrum peaks and amplitudes as a feature set for machine learning classification?
- Bearing fault detections? Inner outer race failures should work, right?
As I was getting close to graduating from my MSc, I happened to come across a job position where I was lucky enough to turn a hobby passion of mine (reverse-engineering/computer hacking/security/vulnerabilities) into a career. I worked for Symantec for a couple years fighting and researching targeted attacks, and now work as an Anti-Virus Researcher for Microsoft working on Windows Defender. I've been here since 2013, and work a lot in machine learning to fight and detect malware. I work on developing hobby reverse-engineering tools and hobby vulnerability fuzzing tools in this field on the side as well.
Feel free to reach out to me or even to meet up if you are ever passing through Vancouver Canada and up for catching me up on latest in fault detection,
Geoff McDonald, glmcdona@gmail.com
Linked In
https://github.com/glmcdona