Innovative Research
Project's Preliminary Research
March 22, 2022
This section includes our full project proposal and preliminary research completed to get under way. This provides a background and fundamental results for the Mammal Identifier System.
Project Proposal
For our main project idea, we decided to pursue underwater communication of mammals. This topic was selected in the hopes of expanding our knowledge and understanding of audio processing. By completing this project we will be able to fine tune our MATLAB skills and increase experience using different toolboxes the software has to offer. We are looking forward to implementing different filters and determining which filter type provides the optimal design to differentiate between the voices of different underwater mammals.
Initially, our project idea aimed to compare underwater and in-air audio communication. Now, the focus has shifted to solely underwater audio analysis. This goal will allow for utilization of the Fourier Domain Technique and application of a variety of filters to aid in production of an audible mammal voice signal. This result will surface after removing the additional noise from an underwater environment, which includes echos, boat motors, and other external factors.
Our goal for this project is to create an alert system that can notify large ships when their vessel is approaching too closely to a marine animal. Cargo ships and other large vessels have propellers that can kill and damage marine life if they get too close. This alert system can notify ships to avoid areas with known marine life or stop the propellers in an emergency.
Given that we are focusing on marine mammal noise, it is still a key idea to understand the sounds produced by ships and submarines as those noises will need to be accounted for in the alert system. It is also important to consider the frequencies in underwater audio processing and which ranges account will aid in noise reduction.
The alert system as a whole will be collecting various types of audible data which include ship noise and underwater echos. Identifying the frequency range of these machines as well as the operating frequency of underwater audio will be the starting point to eliminating noise in our audio signals.
Initial Tasks and Progress
In order to standardize and benchmark our initial data, it has been separated into two categorical sets: open water and species-specific. The open water set of data is aimed to provide “real-world” data with recordings that are more likely to be in a non-controlled environment. This data will be used as our test data for the machine learning algorithm for classifying the different types of mammals. This implementation will occur later in the training and testing portion of the classifier. Regarding the series specific data, this will serve as the baseline for designing the optimal filter for the data. Since each audio signal is for one specific species, it will act as the monitored data when training the classifier algorithm.
The species-specific data set contains some background noise such as water splashing, ship noise (propellers), and other undistinguishable noise. The plan of eliminating most of the external noise is to review spectrograms and apply bandpass filters to insignificant frequencies. For our initial trial with this, the MATLAB results appear to be promising with the spectrogram figures shown below.
Another noise reduction technique is CEEMDAN-MMSVC-LMSAF, which is laid out in the article, “A novel noise reduction technique for underwater acoustic signals”, written by Yu-xing Li. This article can be found under the resources page.In this article, Li provides a three step approach to eliminating noise from underwater signals. MATLAB provides built in functions that will allow the math described to be solved and implemented. It is thought that the application of these techniques will create a clear audio sample which will increase the accuracy of the mammal voice recognition.
In conclusion, we have two methods to eliminate noise from underwater audio recordings: bandpass/bandstop filtering and the CEEMDAN-MMSVC-LMSAF method laid out in the paper above. We plan to try both methods and compare the results.
Spectrogram Figures
This section displays the Spectrogram figures that correspond to an unfiltered and filter audio signal. In this example, the sound recording for an Atlantic Spotted Dolphin produces sounds in the 2000-14000 Hz range. Knowing this, a bandpass filter was applied to eliminate noise specifically around 0 HZ and 16000 Hz. When playing the two different sound recordings, the filtered one clearly displays less noise especially regarding water splashing. The first figure represents the unfiltered signal, where the second image demonstrates the signal with the applied filter.
