Algorithms for investigation

[Ahmed Sagarwala]

I'm at the early stages of researching disaggregation for my masters thesis. My goal is to test several algorithms used for disaggregation and reveal their accuracy. If you have any tips on algorithms you would like test, please post them here.

So far, I'm looking into time and frequency domain algorithms such as MFCC. Most of the research I'm looking into relates to sound processing, but I'm open to other suggestions.

[Oliver Parson]

We actually had a paper accepted in JAIR recently about detecting the call of a species of Cicada from audio recordings. I know it's not quite energy disaggregation, but the model might also be interesting to you.

[Ahmed Sagarwala]

Was the recording by Jim Grant the only signature used for comparison? The reason I ask is if there are variations in calls made by insects of the same location. Any attempt to compare calls from other cicada relatives around the world?

I'm wondering if this model would be useful for devices made by a manufacturer that experience degradation over time. I'm still trying to understand the mathematics behind how the algorithm works. I might have some questions in the near future.

It's a great paper. Thanks for sharing!

[Davide Zilli]

No, the recording by Jim Grant was the first one used as it's the only one we have from the New Forest. However, we have cicada recordings from other countries in Europe and we use those to learn the parameters for the HMM (note that these parameters are hand-picked, so 'learnt' manually if you wish. There is no automated algorithm used for learning).

MFCCs are good and quite strong, but be aware that the mel spectrum was invented to mimic human hearing. If you have no reason to prefer low-frequencies as the mel spectrum does, it makes sense to use linear coefficients (as opposed to mel-scale ones).

Hope this helps,
Davide

[arfa...@gmail.com]

I'm working on energy disaggregation for my research, and as I am on early stage of my thesis. I want to use clustering or svm in matlab or similar algo. Can you please send me a sample code in matlab for this?
I found data on kaggle and they have used KNN, but when I run that, the data loads, but KNN don't run.
Can anyone help pls.

Regards