In our densely populated country, a quiet living environment is a rare luxury. Consequently, noise pollution is a common problem. The good news: through the use of sensors and hybrid AI, constant monitoring and automatic recognition of different types of sounds is now a viable option. Munisense, Peutz and Embedded Acoustics are currently working together to take that technology further so that it will actually be possible to take long-term measurements and use AI to determine the type and origin of sound.
Someone is shouting, a motorcycle is driving by, and construction work is taking place nearby, complete with music blaring from a radio. In the background, there's the drone of an airplane and industrial noise. While the human ear can recognize some of these sounds, it's impossible to unravel everything based purely on hearing. Moreover, accurately pinpointing the location of all those sound sources is even more challenging.
Previously: snapshots
"The fact that sounds interfere with each other is what makes sound measurements so complex," emphasizes George Boersma, co-founder and director of Munisense. "In the Netherlands, sound maps have been used for a long time: maps showing the average annual sound levels. These are required by European law for large municipalities and these maps are made once every five years. However, annual averages do not provide sufficient insight into the actual noise levels, which can vary greatly in time, space, and type of activity. As a result, the effectiveness of policy choices remains unclear for a long time. And that, of course, doesn't help with gaining the trust of residents."
On to the continuous sound scan
"But things can be done differently," Boersma points out. "As Munisense, we wanted to work with Peutz and Embedded Acoustics and applied for and received the MIT AI grant in 2022. We used the grant to develop AI algorithms that not only help unravel what sounds our sensors pick up in an area, but through a network application can also determine fairly accurately where the various sound sources are located. So it's a continuous sound scan, where we can also locate the sound sources. And this makes it possible to create a dynamic noise map, visualizing the real-time situation. Even before residents complain about noise pollution, noise peaks can be seen on that map."
Self-contained sensors
Creating a networked infrastructure of sensors in a neighborhood and having all the captured data analyzed in real time by a hybrid AI application: this is something that was not possible five to 10 years ago. That it is now an option is largely because both sensors and AI technology are becoming more sophisticated and cheaper. "In doing so, the sensors can operate completely independently because they are able to analyze the data locally," indicates Pim Wubben, project manager at Embedded Acoustics. "That local analysis not only leads to time savings, efficiency and lower costs, but is also an important condition to ensure privacy. This is because the sounds are analyzed directly, thanks to AI in the sensors, and only information about the characteristics of the sound is transmitted, and therefore not the sound itself."
Characteristic audio profiles
But how can you differentiate sounds? "For that, we need a database of common sounds," says Robbert-Jan Dikken. He's a senior researcher in data-driven physics & AI and at Peutz, among other things, he's doing research on ways in which AI can help get a better picture of noise pollution. "Our company has been doing noise measurements for 70 years, so there are already a lot of datasets available here. And we're currently working to determine the characteristics of those audio recordings. That will provide profiles that will soon make it possible to use AI to directly determine exactly what sound is involved."
The database Dikken is talking about is also supplemented by sound samples that Munisense automatically collects and classifies at various municipalities, using hundreds of sensors. This data can then be used to train new and improved machine learning models.
Multiple sound sources, wind and reverberation
The classification of sounds based on available data is the first step. Now that this phase is nearing completion, the three collaborating parties are focusing on integrating measurements and models (noise emission mapping), with Peutz taking the lead. This presents them with the next challenge: particularly in urban environments, a combination of sounds is almost always audible. Furthermore, the range of a specific sound is influenced by wind speed and direction. The urban layout also plays a role; for instance, a high-rise building can cause sound reflections, making it difficult to pinpoint the exact location of the sound source, at least for the human ear. However, with sensors and hybrid AI, both specifying and locating sound sources become feasible. This requires intricate formulas and complex computational models.
Measuring the sound of sea vessels
Peutz has already applied the new sound measurement technology several times, primarily in construction projects. However, the largest pilot project by far is currently underway in the Port of Rotterdam, where the noise source volumes of ocean-going vessels are being determined automatically. For this, Peutz developed a hybrid AI system that focuses on the emission mapping of the noise. And Munisense's sensor technology is being used for the measurement network. The port area is a rather complex acoustic environment, however, so the sound of ocean vessels is separated from the sounds of all other activities taking place there. The results are promising and also help contribute to the development of dynamic noise maps.
Tackling road rage idiots
Plenty more practical applications are possible, though. For example, in the summer of 2023, at the request of the municipalities of Rotterdam and Amsterdam, Munisense installed noise meters at various locations in the city for real-time monitoring of traffic noise. And there was clearly a need for this, because in these municipalities, especially in the summer, there is a lot of noise pollution caused by unnecessarily speeding cars and motorcycles, popping exhausts and loud music from car radios. In some cases, the noise meters are equipped with a matrix sign alerting the "traffic anti-socials" to the noise nuisance caused and the amount of the fine for doing so.
Sound predictions
Things are currently moving fast. So fast in fact, that potential future applications are already being considered. "Perhaps we will soon be able to offer residents an instrument that not only allows them to see what is currently happening in terms of noise in their environment, but also indicates the noise levels expected in the near future. So a kind of outdoor radar, but for noise," suggests Wim van der Maarl, project leader for noise and environment at Peutz. "Of course, you can never predict every sound. But the more data we collect, the more insight we get into the main noise sources at different times. Based on that, using AI, it should certainly be possible to detect a general trend, enabling the system to detect deviations. In this way, you can ensure, for example, that residents have direct insight into what is happening in their living environment."
Quick adjustment
"But much more is possible," Van der Maarl continues. "For example, dynamic noise prediction maps can provide real-time input to digital twins for urban areas and enable data-driven decision-making. That means local administrators can work closely with the residents and users - such as businesses - of an area to come up with measures together that will contribute to a more pleasant living, working and living environment. The effect of those measures can then be monitored in real time and, where necessary, adjusted quickly. The latter can have an enormous impact, because in many cases noise pollution can be prevented if the person causing it is more aware of the nuisance that a certain activity can cause. And nuisance that can be prevented, doesn't need to be enforced."
