CWT on each frequency one at a time

I have a very large matrix (512x1500000) and I want to run CWT on each row. However, since CWT creates a new dimension with its frequency bands, I am running into memory issues. I do not need to store all frequency bands; I just need to run the analysis on each frequency band individually. So, I am wondering if there is a way to run CWT on each frequency band one at a time in a for loop.

[Umar]

To address this issue, you can indeed run CWT on each frequency band one at a time in a for loop. This approach allows you to process the data sequentially, focusing on one frequency band at a time without the need to store all frequency bands simultaneously in memory. By implementing a for loop to iterate through each frequency band, you can apply the CWT analysis on a per-band basis, effectively managing memory usage and avoiding overload. This method enables you to conduct the analysis efficiently and effectively while optimizing memory utilization. Furthermore, utilizing this iterative approach in a for loop allows you to customize the processing of each frequency band, enabling more targeted and specific analysis tailored to your requirements. By executing CWT on individual frequency bands sequentially, you can achieve your desired outcome without encountering memory constraints. In conclusion, running CWT on each frequency band one at a time in a for loop is a practical solution to address your memory issues and perform the analysis efficiently. This method offers flexibility, control, and optimized memory management, allowing you to focus on analyzing specific frequency bands without overwhelming system resources. Consider implementing this approach to streamline your CWT analysis process effectively.

[ryoroy]

According to the documentation for CWT, we can specificy ts or fs. Should I call whatever frequency band I want to here to get the CWT for one frequency band? And what frequency band should I specificy? Is it the central frequency of our wavelet?

Thank you!

[Mathieu NOE]

I wonder if you could also do a n th octave analysis , so passing each data (channel) inside bandpass filter (with a for loop to switch to the next band) , then you could buffer the output of the filter and take the rms for a given time period and you would end up with a kind of n th octave spectrogram

you can then fine tune your frequency and time resoltion by playing with n (the 1/n octave band) , the buffer size and overlap.

just my 2 cents

Answer 1

[Swastik Sarkar]

Hi @ryoroy,

I attempted to replicate your issue by creating a random matrix “data” of dimension 512x1500000 and running the Continuous Wavelet Transform (CWT) on each row. Here's the sample code I used:

numRows = 512; 
numCols = 1500000; 
data = randn(numRows, numCols); 
for i=1:numRows 
    disp(i) 
    cwt(data(i,:)) 
end 

I observed that the memory usage fluctuated between two values, likely due to memory allocation and deallocation during each loop iteration. Initially, the MATLAB process consumed 18GB, which increased to 27GB during the CWT operation on a row.

To optimize the process, I modified the MATLAB code to perform CWT not only row by row but also for each frequency band. This approach reduced the RAM consumption to a constant 11GB, although it increased the computation time.

numRows = 512; 
numCols = 150000; 
data = rand(numRows, numCols); 
scales = 1:10; % Adjust Scales accordingly 
cwtResults = cell(numRows, 1); 

for i = 1:numRows 
    cwtRow = cell(1, numCols); 
    for j = 1:numCols 
        cwtRow{j} = cwt(data(i, j), scales, ‘morl’); 
    end 
    cwtResults{i} = cwtRow; 
end 

To address your query about performing CWT for a single frequency band, you can use the following approach:

desiredFrequency = 10; % Desired frequency in Hz 
desiredTimeScale = 1 / (desiredFrequency * waveletScale); 
cwtResult = cwt(data, desiredTimeScale, waveletAlgo); 

I hope you find a similar outcome, and this helps you.