The algorithms used by the current implementation of the acoustic goniometer have proven adequate to the task of event detections and DOA calculations. However, more research would be needed to determine whether these algorithms represent the best solution or simply a functional implementation. While determining the absolute best algorithms in general to use for acoustic goniometry may not be possible due to
dependencies upon application (varying noise sources, cross-sensitivities between event signatures, etc.), some algorithms may prove more adept in a wider range of
circumstances than others. The current acoustic goniometer design should be used to research and implement event detection and time delay estimation algorithms more thoroughly to determine the best implementation for a general purpose system (i.e., highest accuracy for the widest range of applications). This research would take a considerable investment of time and a deployment of acoustic goniometers to measure a significant number of phenomena. A project as staggering as this may require the cooperation of multiple agencies, but the determination of the best set of algorithms (or even more information on performance of the various combinations) would be invaluable to the further reduction of barriers to research in this field.
Event detection algorithms have been explored the least in the current research and would be the area of algorithm research, which would benefit the most from further experimentation. The current method for event detection employed by the goniometer prototype uses a simple threshold technique coupled with a predetermined data window
to detect an event. This simple method was easy to implement and requires few system resources, but it suffers from a few significant drawbacks. A purely threshold based algorithm can not differentiate between an event of interest and other phenomena (e.g., acoustic/electrical noise). Further, determination of the threshold can be problematic if the sound produced by the event of interest experiences drastic attenuation over distances or objects along the path of the sound create reflected versions of the signal leading to multiple detected events for a single phenomenon. Such problems have been mitigated in the current research by limiting signals of interest to the infrasound domain (little
attenuation over large distances) and employing a windowing technique around a detected event in order to ignore reflections of the original sound. However, these solutions still do not address the issue of event differentiation (e.g., distinguish between an avalanche and a gun shot). As an alternative, fingerprinting and correlation could be explored instead of thresholding in order to make the process selective to events of interest. By first characterizing the acoustic fingerprint of an event and using the fingerprint as part of a correlation filter, events occurring in the same frequency range could be differentiated without putting a significant strain on system resources. The drawback to this method would be the difficulty in event characterization and the challenge of keeping the goniometer simple to use. The threshold method can be implemented such that a researcher could adjust the thresholds and window length to meet the needs of their research. A fingerprinting algorithm, however, would require them to somehow store the fingerprint of their event for use in the correlation filter. Another option worth further exploration is the use of the Fisher Statistic approach (see Chapter 3.5.3). Due to the computational complexity of this method, either the system
would need to be modified to use a more capable processor (e.g., the newest generation of the upgraded part from ST, SOC, DSP, etc.), or the firmware would need to be adjusted to use a lower sampling rate (limiting the viable frequency range of the
goniometer) and to move from ChibiOS to a simpler (less resource hungry) cooperative scheduling algorithm. Still, other options for detecting events may exist which could be simple for any given researcher to use and solve the problem of event differentiation. Further research is needed to flesh these out and determine the best solution.
The other area of algorithm research that requires further exploration is the time delay calculation algorithm (or correlation algorithm). The current implementation of the acoustic goniometer uses a threshold to detect an event in the signal from each sensor followed by a standard correlation operation between each signal within an event window to determine the time delay for the event arriving at each sensor. Regardless of whether this method continues to be employed or the fingerprinting algorithm discussed in the preceding paragraph is employed instead, finding the best correlation method is an important aspect to furthering this research. Part of this question has been addressed in the current research (see the comparison of various algorithms in Chapter 4.2). However, the standard correlation approach selected as the best was determined under limited testing conditions. The raw data showed little evidence of unwanted noise especially in the frequency range of interest. All unwanted frequencies were easily removed with a simple digital low pass filter and a wind screen placed over the microphone. Under these conditions, the standard correlation method significantly outperformed the method of applying correlation in the frequency domain. However, according to [13], correlation in the frequency domain provides better performance in the presence of noise. Further
testing in noisier conditions could help to validate this claim and could make frequency domain correlation a more appealing option. Finally, the technique implemented for testing feature selection and comparison to determine time delays was extremely simple. If a more complex set of features was used (e.g., slope, second derivative of the signal, etc.), the results of such an approach might be improved significantly enough to make feature comparison a contender. However, using overly complex features could make the algorithm become so complex as to make it the worst choice for an embedded system with limited resources. Further testing is required to determine the viability of this method as well as its ranking among the other correlation algorithms. The acoustic goniometer developed as part of this dissertation would serve as the perfect platform for continuing this research and answering these questions.
7.2 Hardware Research
The hardware developed for the acoustic goniometer research presented in this dissertation has been proven adequate to the task of determining DOA and has been shown to be adaptable to fit multiple research needs. The design is flexible and can monitor any number of phenomena if the system is properly deployed with its gain and SD card parameter adjusted to meet the needs of the given project. Furthermore, the antenna can be easily reconfigured into any desired geometry by simply building a new antenna structure and modifying the dimension matrix in the SD card. However, as with any research project, the current implementation of the acoustic goniometer leaves room for improvement in multiple areas. The following paragraphs detail some ideas that would enhance the performance of the system and/or further lower barriers to research in the field of acoustic goniometry.