Author(s):
Luis Fernández -
National Polytechnic Institute
Luis A. Pérez -
National Polytechnic Institute
Sergio S. Guerra -
National Polytechnic Institute
Noise levels from heavy automobile and aircraft traffic during morning and evening rush hours in Mexico City could potentially cause hearing loss. The Committee of Aerial Transport has proposed a new aircraft classification system in which the aircraft operator would pay a fee based on noise production, not weight or aircraft type. As a result, we put together a novel computational model that not only measures noise production, but also identifies an aircraft based on the noise it generates.
Each node in our wireless monitoring system contains a half-inch prepolarized microphone, a data acquisition card that measures noise level, an industrial computer, and a wireless Internet connection using Wi-Fi or 3G. Each microphone is in a weatherproof case mounted 4 meters above the ground. The node measures noise levels every 30 seconds and streams the data back to the control center every 5 minutes.
Figure 1. Distributed Wireless Monitoring System Diagram
Identifying an aircraft based on the noise spectrum characteristics it generates is complicated in the real world because background noise, the weather, takeoff speed, and the aircraft’s load can interfere with analysis. Recently, measurement equipment that uses neural networks to identify noise has appeared on the market, but it can only distinguish between jet aircrafts, propeller aircrafts, helicopters, and background noise. We decided to create a computational model to measure and interpret noise. Using only the noise created during the 24 seconds after takeoff, our system can correctly identify the aircraft.
System Development
Using a wireless topology reduces costs and provides flexibility in setting up a monitoring system. Each monitoring node is based on a headless industrial PC running Windows XP with a Wi-Fi adapter and an NI USB-9234. During the hardware evaluation process, we decided to use NI products because of their high measurement quality, ruggedness, and reliability compared to lower cost sound-level meters.
Even though each node is connected to the city’s electrical system, we use an uninterruptible power supply to prevent data loss. The node measures noise levels every 30 seconds and the system can collect data locally for up to 14 days. The government plans to use our data to identify the times and locations in Mexico City with the highest noise level, create noise maps, and implement regulatory actions to control the noise and promote a healthier environment for its citizens.
Our system can record traditional metrics used for road traffic noise, such as continuous equivalent sound level (Leq), and it can also record fractional octave analysis and measure prominent tones. In addition, the system can transfer WAV files to a central server to study transient signals that may trigger alarms, which helps identify isolated sound sources that interfere with accurate measurements.
Figure 2. Noise Patterns From Two Weeks in the City Square México City.
We originally planned to use the public Wi-Fi that the government installed in 2008, but some nodes had to be converted to a slower 3G system provided by a wireless carrier. Although we can use speech and data services simultaneously on the 3G network, it has significantly slower data transfer rates.
Communicating via TCP/IP
Our control center has a static IP address. Each node has a dynamic address assigned by a DHCP server. To simplify, the control center is similar to a server, and the nodes are similar to a client. The nodes attempt to open a TCP connection, and if the control center receives this connection request and the node’s validation key, it accepts the connection.
Figure 3. Control Center Central Server Interface
Figure 4. Signal Analysis of Audio That Exceeds Thresholds
Figure 5. Noise Level, Time, and Date, and Amplitude Measured in dBA
Figure 6. 3D Noise Map Displaying Noise Level, Time, Date, and Amplitude
Below is the system block diagram we created for pattern generation and recognition. We considered takeoff noise a nonstationary transient signal because it starts and ends at a zero level and has a finite duration. As Figure 8 shows, most of the signal’s energy is below 2 kHz. In this case, we notice the background noise more strongly at the ends of the signal because the aircraft-generated noise masks it in the middle portion.
Figure 7. System Block Diagram for Pttern Generation and Recognition
Figure 8. Typical Noise Signal and Spectrum of a Boeng 747 Taking Off
For all aircraft noise, we observed the typical form of the amplitude spectrum from 0 to 5,000 Hz. We chose to use a sampling frequency of 11,025 Hz in order to reduce the number of samples taken in 24 seconds to 264,600 samples. In other aircraft noise analyses, the recommended sampling frequency is 25 kS/s and D-, C- and A-weighting filters.
Reducing Spectral Resolution
We decided to reduce the spectral resolution because the amplitude spectrum has 132,300 harmonics, which would result in very complex processing. In addition, we were only interested in gathering data about the spectral form.
We present the following hypotheses:
- Any method to reduce spectral resolution introduces a tolerance in the initial and final times within the measurement interval of aircraft noise. For example, a feedforward neural network is trained with one noise pattern, which was acquired from zero seconds from the aircraft takeoff until 24 seconds later. In run time, if the aircraft takeoff noise is acquired from 5 seconds until 24 seconds, this 5 second time displacement will have little effect on the spectral form if its spectral resolution has been reduced.
- A median filter (moving average filter) creates a typical form of the aircraft’s takeoff noise spectrums.
- The decimation of average spectrum, with a rate X, conserves the spectral form of an aircraft’s takeoff noises.
Conclusion
Our 10-node system successfully measures the noise produced during takeoff by airplanes at the International Airport of Mexico. The system makes many different types of spectral analyses and obtains the most-used statistical indicators for noise measurement, expressed in dB(A) or dB(C). We can store the data collected by our system and later come back to perform more in-depth analysis. We can determine potential health risks from this noise, and gain an idea of how noise levels fluctuate throughout the day.
In the future, we would like to measure the differences of the yield when applying this technique after segmenting the original signal. We also plan to test new parameters to create the neural network.
Author Information:
Luis Fernández
National Polytechnic Institute
Mexico City
Mexico
Print
PDF
