This is going to be a bit of a long post. My questions are:
- Do the 5 signal processing steps look "correct"?
- Next steps could be to perform order analysis, and extract features from time, frequency, order and time-freq. domain.
I am trying to study vibration from 4 wind turbines. I have data from the gearbox of every wind turbine from August 2019 to January 2020. There are 400 vibration signals for every component in every turbine (one signal is sampled each day for every turbine). Each signal is sampled with a frequency of 25.6 kHz over the course of 10 seconds.
I wish to look for early fault development in the gearbox.
- For the same turbine: Compare early vibration signals (August 2018) with more recent ones (January 2020) to look for degradation. Use clustering for this task.
- Compare vibration from all 4 turbines together. (The suspicion is that gearbox 4 is more degraded).
- The vibration is measured under different operation conditions. The vibration is correlated with many parameters such as wind speed, rotational speed of the shafts and power production. This makes it more challenging to compare the results.
- We do not know for a fact that there is a fault on the turbine(s). There may not be anything wrong on any of them.
- As of right now, we do not possess any specific information/dimensions of the components inside the gearbox or bearings. This means that we e.g. are unable to calculate the theoretical bearing defect frequencies (like BPFO, BPFI etc.). With these available, we would be able to validate our eventual findings.
- Non-stationary behaviour
Signal processing overview
The first part of our research includes preprocessing the vibration signal which is measured on the high speed shaft in the gearbox. With the proposed preprocessing method, the goal is to discover faults at an early stage in the gearbox. In stage 5 of the process, an envelope power spectrum analysis is performed. This process is also known as High Frequency Resonance Technique (HFRT). This technique amplifies weak transients at higher frequencies masked by noise. More on this technique is written in this1 book.
A general illustration for fault development of bearings is shown in figure 5. Because of the sampling frequency of our data, 25.6 kHz, we will most likely not discover any fault before “Stage 2”, that is, if any fault is present. This area is depicted “Bearing Natural Resonances”.
Walkthrough of proposed signal preprocessing method
Here, the process from figure 5 will be described in more detail. The first 10-second signal from wind turbine 4 will be used as an example. The time and frequency domain plots are shown for each step. 0.2 seconds of the 10 second interval is used in this demonstration.
- High Pass filter (6000 Hz):
- Spectral Kurtosis: SK is used to find the optimal bandwidth for a bandpass filter, to preprocess the signal further and amplify the transients. This step is performed in Matlab.
The figure above shows the kurtogram which recommends a Bandwidth (BW) og Center Frequency (CF) f or a bandpass filter. The mean BW and CF across all intervals for each turbine is used to find the following parameters for the bandpass filter:
lowcut = mean_cf - mean_bw/2 highcut = mean_cf + mean_bw/2
3. EEMD (Ensemble Empirical Mode Decomposition):
Used to reduce noise by decomposing the signal.
4. Correlation Coefficient: Correlation between the original (in red) and decomposed signals (in green) is calculated. The decomposed signal with the highest correlation is selected to be used further.
5. Envelope Spectrum Analysis: Consist of two steps. The signal is : 1. Rectified: signal = signal**2 2. Low pass filtered (2000 Hz)
With this FFT, I want to perform feature extraction to be able to carry out the objectives stated in the introduction.