What is a good way to distinguish aircraft noise from other sounds?

Question

The FAA's NextGen routes, which were modeled, not measured, have created a bunch of aircraft noise for tens of thousands of people. I want to make it easy for those people (including me) to buy a cheap USB microphone, plug it into a laptop or PC, detect aircraft noise, correlate that with publicly available flight-path data, and file complaints with the FAA, airport(s), and legislators. With enough actual (as opposed to theoretical) data, we might be able to get them to roll back the route changes. Or not, but it's worth a try.

To do that, I need some help with the "detect aircraft noise" part of the problem. It's not enough to trigger off noise above a certain dB threshold, because lawnmowers, leaf blowers, trucks, etc. also make noise. I can weed out some over-threshold noise that doesn't last long enough or lasts too long (aircraft take 5-10 seconds to pass overhead, vs. minutes for lawnmowers and a couple of seconds for trucks), but that's a pretty crude approach. And I still need to be able to factor multiple noise sources into individual components.

Correlation would probably be easier with a pair or array of microphones, to establish noise direction that could be tied to flight paths, but that adds a lot more hardware and complexity. Ideally we'd just need one microphone.

Aircraft include airplanes (jets, turboprops, and standard propellor-driven), as well as helicopters.

I'm a good coder (Java expert, former C expert), but have zero experience with DSP or training algorithms. Can you help me get started?

Answer 1

You need to train a classifier. The way I would approach it is to allow users to submit audio recordings through your app, and indicate which sections of the recordings contain the plane. This would enable you to quickly crowd source a lot of data with which to build a model. Then you would find the STFT and volume of the signal over time. Finally, train the classifier (such as an SVM or neural network) using the amplitude spectrum and volume features. Once you have trained the classifier, you can use it to detect planes by extracting the same features on live audio (STFT isn't computationally expensive, so it should be feasible to do on the phone).

Here's a backgrounder on Audio signal classification.

Answer 2

Some suggested that you go for some machine learning madness.

An other solution is to see if you can extract narrow band noise from the motors of aircrafts / road traffic and see if you can discriminate them based only on that. Additionaly, even if the narrow band noise are in the same band, you could use doppler to discriminate them (For example, if your microphone is close to road traffic, the doppler shift from it should be much more sharp than the doppler from aircrafts).

But before talking to much about how you could discriminate aircraft noise from road traffic, set up some microphone, gather the data and perform a time-frequency analysis to see how your signals look like.

Answer 3

Your objective is to measure Sound Pressure Level (SPL) at some position on the ground and compare it against safe levels.

There already exist instruments that are doing exactly that and are called Sound Level Meters. These instruments are not cheap because of all the hard work that is required to produce a valid measurement. Their functionality cannot be simply replaced by a "cheap USB microphone". For example, to derive SPL from the microphone's output, you need to have the microphone calibrated (not a cheap process) or use a measurement microphone (not a cheap microphone). To further ensure the validity of the measurement, you need to make sure that amplification and conversion to digital form has been linear and remains linear throughout the frequency spectrum of interest irrespectively of changing factors such as temperature, voltage, etc (not cheap hardware).

What you could do to get some indication of the problem in a relatively simple way is this:

1. Use a Sound Level Meter with data logging capability.
2. Set the meter up in the desired area
3. Examine the recording of the sound level meter and look for regions that are over safe levels. Hopefully, these are not going to be many.
4. Cross reference the timings that safe levels are exceeded with aircraft arrivals / departures (obviously, adjusted for the location of the mic)

In this way, you have the opportunity to estimate a background SPL (in between aircraft noise) that residents are constantly exposed to and the additional contribution to it by aircraft. By "cross correlating" aircraft timings with SPL measurements, you are doing away with all the problems of trying to distinguish the type of aircraft engine that contributes to the sound level. SPL peaks should correlate more with the aircraft schedule and less with the occasional leaf blower.

To further rule out false positives, that is occasions where the SPL exceeds safety levels because of any reason other than an aircraft, you could set up a cheap sound recorder too and automate the whole process by using the SPL curve to automatically preview the sound recording for what caused the peak at that time.

However, be prepared that the position of the sound level meter could be challenged and the "simplest" way out of this is to obtain many measurements at different positions to reduce its effect.

Hope this helps.

Potentially Useful Links:

Noise and Emissions

Aircraft Noise Issues