Confused by envelope detection and windowing+overlapping

[mrkbrnlmr]

TL;DR

Should both envelope detection and windowing+overlapping be used before FFT?

Summary: Should both envelope detection and windowing+overlapping be used before FFT?

Summary: Should both envelope detection and windowing+overlapping be used before FFT?

Hi everyone!

I'm somewhat a newbie to signal analysis, but I've also noticed that there seem to be many different solutions to a particular probem and no one correct answer.

This is specifically related to bearing fault analysis, but I guess the principals remain the same regardless of the application.

The dilemma I’m having, is that, on one hand, envelope detection is needed in order for FFT to detect the bearing fault, but on the other hand, FFT is first required so that the signal can be bandpass filtered around the bearing fault and then returned to the time domain so that envelope detection can be applied. Is it not a "catch 22" situation?

So what is the right order for FFT, bandpass filtering and enveloping?
Also, should windowing+overlapping be performed together with enveloping and if so, which comes first?

Finally, if you can provide an academic reference for your answer, I'd be most thankful.

In any case, a huge thank you in advance!

 

[jrmichler]

You start by learning as much as possible about what you are looking for. For example, rolling element bearings have a number of known frequencies:

Outer race ball pass frequency
Inner race ball pass frequency
Ball rolling frequency - inner and outer race
Since those are impacts, there may be harmonics.

Be careful with bandpass filtering by FFT, a simple brick wall filter has infinite ringing. You are better off to look at frequencies per revolution. Study the engineering guides from bearing companies like SKF. I see that SKF has courses on the subject: https://www.skf.com/us/services/customer-training/classroom/WE204.html.

Noise and vibration signals have a huge noise component. Do not expect a simple answer, especially in a complex machine. Start with a simple system - a single electric motor or pump.

 

[mrkbrnlmr]

Thanks for the reply jrmichler. However, I was already familiar with all that.

The frequency that I want to extract is in deed the resonance frequency. I have read numerous prodecural papers on the topic, but none go into such detail and there is also conflicting information.

So I'm still looking for an answer

 

[Baluncore]

The frequency that I want to extract is in deed the resonance frequency.

What is "resonant"? Do you actually mean repetition?
What do you mean by "envelope detection"?
I would perform power spectrum accumulation to raise the bearing spectrum out of the noise.

 

[mrkbrnlmr]

Hi Baluncore,
and thanks for your reply. By resonant frequency I mean the structural resonance that is excited when a ball passes over a defect in the bearing.

This article essentially contains my question: https://link.springer.com/article/10.1007/s11633-014-0862-x

The first flow chart shows that band pass filtering must be done before enveloping. Enveloping is then followed by the FFT to obtain the frequency spectrum.

However, the reason for enveloping is to accentuate the fault frequencies, which are otherwise hidden under noise. On the other hand, we need to first bandpass filter before we can apply the envelope, but since we have not yeat applied the envleope, the fault frequencies are not accentuated, hence how can we filter - a "catch 22" situation in my opinion.

Any thoughts?

 

[Baluncore]

However, the reason for enveloping is to accentuate the fault frequencies, which are otherwise hidden under noise.

The envelope of the resonance has much lower bandwidth than the input signal. The envelope preserves the "click" rate as the spall impacts opposed rolling surfaces.

The initial band-pass filter is not implemented by an FFT. It is a simpler filter that only passes the high frequency structural resonance components excited by the fault impulses. That front-end BPF could be implemented by the transducer coupling circuit, followed by a digital filter after the A-D converter. In the paper it says; “Here, an 80-order finite impulse response (FIR) bandpass filter with 1000Hz pass-band is employed”. That can be done sequentially and continuously as the input data is converted.

 

[mrkbrnlmr]

So the FIR filter can be implemented without going into the frequency doamin - ok. I already have the data, so I will just apply the FIR to the data. Thanks for the clarification!