Detecting the noise of bubbles
Navigating the recent machine learning submissions on archive and found an interesting application I wanted to note:
Fossil fuel power plants deposit excess carbon dioxide in storage sites (the process is called carbon capture and storage). This process prevents the extra carbon dioxide from entering the atmosphere and contributing to global warming and ocean acidification.
The paper focuses on the early detection of leaks in these systems. The thought being that gas leaks would create bubbles and the acoustic sensors would capture the noise from the movement of said bubbles.
The authors made their own positive set:
The leakage was simulated through the use of compressed air (from scuba dive cylinders), with flow, pressure and exit diameter orifice controlled. These con- trolled leaks were performed at predetermined distances from underwater acoustics monitoring equipment
Though I appreciate the cleverness of creating a relevant simulated environment, the algorithm really detects bubbles not necessarily “gas leaks”. Longer underwater recordings would to check for false positives would have been nice.
Note for later time serious projects:
The algorithms were standard (Random Forest and Gradient Boosted Trees).
The acoustic signal was modified by taking the power spectral density which describes the strength in a signal as a function of frequency.
HMM smoothing was used to account for the time-serious nature of the classifications.
