Airplane Noise Detection based on Hidden Markov Model Classification

This project will attempt to identify jet aircraft noise from a variety of environmental noise sources. In recent years, real estate values have become affected by their proximity to airports. With the increase in air travel, aircraft noise during takeoffs, landings, and fly overs have become a common event in many parts of the country. Simple noise level monitoring has been used to create noise maps in areas around airports, however the monitoring technique is unable to distinguish between various environmental noise sources. Typically, aircraft noise is many seconds in length, which will require the classification technique to understand the long temporal structure of aircraft noise in addition to the short term spectral characteristics. Hidden Markov Models (HMM) will be used in this project to provide the classification system a method to account for the rise, level, and tail off of the spectral events during an aircraft fly over. Related Research Extensive research into noise recognition has begun recently. Automatic Noise Recognition (ANR) systems have been implemented using statistical methods in order to identify various environmental sounds. Typical industry applications for these system would include: personal PDA devices, noise monitoring, animal tracking, hearing impaired assistance. Christophe Couvreur [4,5,7] is a leading researcher into environmental noise detection. His systems have employed both statistical methods, and HMMs to identify a wide variety of environmental sounds. A new breed of "intelligent" noise monitoring solutions [5] are being developed that will automatically adapt the recognition system to the environment it is placed into, and enhance the models based on the new data. Introduction In order to understand the approach needed to classify aircraft noise, it is important to first identify characteristics of the noise spectrum. Shown below in Figure 1 and Figure 2 are the spectrograms of two different jet aircraft fly over sounds. Page 1 Figure 1 Figure 2 As can be seen from the above spectrograms, the pitch of the aircraft sound changes as the plane travels overhead. Much of this change is due to the Doppler effect, which will change the frequency of the sound waves when the airplane begins to get closer to the microphone, and then continues to move away from the microphone. Interestingly however, the frequency of the strong track is different between the two noise recordings. The first recording has the pronounced track starting at 3300 Hz, while the second recording has the pronounced track at 3700 Hz. I believe that this is due to two effects. First, it is not clear as to the actual model of the plane that created these sounds. It is likely that two different airplanes are represented in the above spectrograms, and that their engines are different. Different engines are likely to have a different sound spectrum due to the variations in rotor design, and airflow patterns through the engine. The second difference is that the angle of fly over, or speed of the aircraft with respect to the microphone location can cause a frequency shift. The second plane has a higher pronounced frequency to begin with, which could represent that this plane flew directly over the microphone location, which via the Doppler effect, would push the sound waves closer together, thus providing a higher frequency track. It may have also been traveling at a faster rate of speed. The first plane might have been 1/4 mile away from the microphone source at it's closest point, so as it approached, the sound waves were not as tightly coupled together as the second plane. As can be seen, a classification system that will detect aircraft noise will need to be able to handle this variation in frequency tracks. It is clear however, that the frequency shifts are visible, but the time at which they happen will also vary based on the time it takes for the airplane to arrive at the closest distance to the microphone, at which time, the frequency shift should be occurring. The long temporal property of aircraft noise, makes it important to design a classifier that will account for this structure. A Hidden Markov Model will be used to provide this temporal structure to the classification system. Hidden Markov Models A Hidden Markov Model (HMM) consists of a series of states connected together to form a Markov Chain [2,3]. These states can be described by a emission probability which is typically a probability density function, and a transition probability matrix. For each state in a Hidden Markov Model, a choice can be made to transition to the same state, or to another state within the Markov Chain. The probability density function is typically a Gaussian Mixture Model with a set of n input vectors describing that state. The transition probability matrix will describe the likelihood of the next set of input vectors resulting in a match for the same state, or a match in another state. While HMMs can be fully interlinked, so that any state can reach any other state within the model, speech and noise recognition models typically have a evolving time property which would prevent "backwards" movement within the Markov Chain. Therefore the aircraft noise system will rely on a left-right HMM which is described in Figure 3.