Classification of noisy vehicles from unsupervised measurements Bert Peeters 1 M+P Wolfskamerweg 47, Vught The Netherlands Ard Kuijpers 2 M+P Wolfskamerweg 47, Vught The Netherlands

ABSTRACT The NEMO-project (https://nemo-cities.eu/) aims to identify noisy and polluting road and rail vehicles, using remote sensing technology. Noise levels from individual road vehicles are measured from the roadside, in normal traffic. Road authorities may use these data to enforce noise limits, to limit access to Low Emission Zones or to influence driving behavior. Whether a vehicle is a 'high noise emitter' is a complex question, as the noise level depends on vehicle type and condition, driving style, weather and location-specific characteristics. From a legal perspective, the question may be answered in relation to type approval noise limits, or in relation to local noise disturbance regulations. Within NEMO, a classification model is developed from a large dataset of unsupervised pass-by noise measurements, from different locations. The model labels noisy vehicles based on the noise measurements, technical vehicle data, driving conditions, and external factors. Several modeling and machine learning techniques were evaluated, to find the most accurate solution. This paper presents the results, and it looks forward to how the technological solution could be applied to enforce regulations, leading to a reduction of traffic noise annoyance. 1. INTRODUCTION

The fight against traffic noise is mainly focused on the year average noise level ( L den , L night ). This should remain a continuous effort or will even have to be intensified in order to meet the ambitions of the European Commission’s Zero Pollution Action Plan [1]. Additionally, there is increasing attention towards the handful of road users that produce excessive noise. Several municipalities in Europe have indicated that they want to employ noise measurements to keep noisy show-off cars and motorcycles with illegal exhausts out of the city center and away from scenic country routes. Technological development of such ‘noise radars’ is done by several projects and institutions around Europe. In the Netherlands, for example, the Minister of Justice answered in 2021 to questions from parliament that such innovative solutions are wanted for better enforcement, but that reliable

1 bertpeeters@mp.nl 2 ardkuijpers@mp.nl

Jai. inter noise 21-24 AUGUST SCOTTISH EVENT CAMPUS ? O ? . GLASGOW

equipment is not yet available. Besides technological development, legislation also needs to be adapted or developed, with regards to in-traffic noise enforcement as well as with regards to privacy regulations. The Horizon2020 project NEMO ( Noise and Emissions Monitoring and Radical Mitigation ) aims to develop roadside measurement equipment, or remote sensing devices (RSDs), that can measure noise levels and air pollutant emissions by individual road and rail vehicles in normal traffic. This paper focuses on noise levels from road vehicles, although the technological challenges to detect individual noisy wagons within trains are similar. More information about the rail noise or the air emissions measurement systems can be found through the NEMO project website ( https://nemo-cities.eu ) . This paper explains how, from the measurement data, vehicles can be classified as being too noisy, so-called ‘high emitters’ (HE). It describes the development of this classification model based on an extensive dataset measured with the NEMO noise-RSD. The acoustic measurement system itself and the algorithms developed to correctly separate noise levels from individual vehicles are described in a separate INTER-NOISE 2022 paper [2]. 2. LEGAL OPTIONS

It is technologically possible to measure and classify noisy vehicles in traffic. But enforcement of too noisy vehicles by the road authorities or the policy, based on the roadside measurement information, also requires a legal ground. There are several options for this:

1. Vehicles and their exhausts must comply with EU-defined noise limits through their type

approval. For type approval testing, a specific measurement method is defined by UNECE Regulation 51 [3]. The method consists of several different measurements under specific driving conditions, which cannot be reproduced by a single roadside measurement. However, a vehicle must comply during its entire lifetime and “shall not deviate from the test result in a significant manner […] under typical on-road driving conditions” [4]. Modifications to the vehicle such as a replacement exhaust system or an engine software update are only allowed if these are approved by a designated authority. 2. Vehicles must also comply to certain limits upon roadside inspection, as sometimes performed

by the police. For that purpose, another measurement method is defined in the type approval regulations [3]. This requires a noise measurement at 0.5 m distance from the exhaust of a stationary vehicle, with the engine revved up to a designated engine speed in neutral gear. These conditions also cannot be reproduced in normal traffic, although a noise-RSD system could be used as a pre-selection device, to efficiently and objectively select which vehicles should be stopped for further roadside inspection. The EU Directive 2014/47/EU on technical roadside inspection currently only allows ‘subjective evaluation’ as a means to perform this pre-selection, and not yet remote sensing as is the case for air pollutant emissions. 3. National or local regulations may state that it is not allowed to make ‘ unnecessary noise’ .

There are examples (e.g. NL) of such regulations applying specifically for vehicles, which applies to e.g. unnecessary car horn usage but also to noisy exhausts or noisy driving behavior. There are also examples of more general regulations against ‘unnecessary noise’ which could be from dogs or bars and restaurants, but which can also apply to the ‘Easy Rider’ down the street. 4. More and more European cities have implemented Low Emission Zones (LEZs) with

particular environmental regulations for entering traffic. A maximum noise limit could be made part of the LEZ admission criteria, if in-situ measurements would be used to verify these. An unsupervised measurement and classification system such as the N-RSD being developed in NEMO, could be employed for these various options. This may then be combined with in-situ measurements of air pollutants, using the exhaust emission RSD that is also being developed.

3. MEASUREMENTS AND DATA

3.1. Measurement setup The measurement setup is described in more detail in [2], but it is summarized here. The measurement set-up as used for test measurements in Rotterdam is shown in Figure 1. The system contains: • several microphones, to measure the noise levels, but also for source localization: the time delay between microphone pairs in horizontal and vertical direction indicate in what direction and at what distance the dominant noise source is located. From the acoustic signal, using frequency analysis, the engine speed of the vehicle is also estimated; • a speed radar that measures the vehicle speed over a certain distance, multiple times per second. This provides not only the speed of the vehicle, but also its acceleration or deceleration; • a camera with automatic license plate recognition (ALPR). The license plate is not used for enforcement or to gather information about the vehicle owner or driver, but only to get anonymous technical information about the vehicle from the national vehicle registry, e.g. the vehicle type (car, truck, motorcycle), its mass and engine capacity, its fuel type, etc.; • local weather information was gathered with sensors measuring wind direction and speed, precipitation and air temperature. Intelligent algorithms are used to assess if an individual vehicle could be measured correctly and separately from other vehicles, and to correct the noise level for the influence of other nearby vehicles. These algorithms are based on the microphone information as well as on the speed radar measurement. Finally, the noise data are synchronized with the license plate using the timestamp, and the vehicle information is automatically requested from the national registry. All data is sent to a cloud-based processing software (‘datahub’) where each pass-by is classified as ‘high emitter’, or not.

Figure 1: The NEMO noise-RSD measurement setup in Rotterdam, October 2021

3.2. Data description and selection Measurement data for the classification model were collected in Rotterdam, near an 2x2 road with a posted speed of 80 km/h (G.K. van Hogendorpweg), ca. 50 m behind the traffic lights at an intersection, see Figure 2. We chose a location behind the traffic lights as this provides both through traffic at constant speed and accelerating vehicles, from the main road as well as from the side road with traffic turning onto the main road. Measurements have been done from the period between 6 October 2021 and 9 March 2022, in which ca. 260.000 pass-bys were registered. That is, individual vehicle pass-bys that could be well separated from the traffic stream, and for which a license plate was also recorded. The main measurement microphone was placed at 4 m from the center of the rightmost driving lane, at 1.5 m height, with other microphones above and to the side.

Figure 2: Rotterdam test location: an 80 km/h road 50 m behind the traffic lights, capturing vehicles

from the main road (1) as well as vehicles from the side road (2) Not all recorded pass-bys can be used for training the classification model. Several (quality) selection criteria are applied to filter out unusable data: • less than 0.1 mm rain in the last hour; • wind speeds below 5 m/s; • a minimum noise peak prominence of 6 dB: the maximum noise level at the main microphone must be at least 6 dB higher than the noise level just before and after the peak, as required also by ISO11819-1 [6]; • an approximately constant acceleration (may also be zero or negative); • a distance from the main microphone of 5 meters or less, so only traffic on the rightmost driving lane was included. After this quality selection, ca. 54.000 (21%) of the original pass-by events remain. For every vehicle pass-by, we have the following data available: • noise levels: the maximum noise level ( L Amax ) with its 1/3-octave band spectrum and the sound exposure level ( L AE ); • driving conditions: speed, acceleration and engine speed, with 95% confidence bounds; • technical vehicle information, of which the most important features are:

- vehicle category in terms of the UNECE class definition, e.g. passenger car (M1), light duty vehicles < 3.5 tons (N1), heavy motorcycles (L3), etc. - the number of engine cylinders, which is needed to calculate the engine speed; - the engine capacity and maximum rated power; - the vehicle mass (unladen); - the fuel type, e.g. diesel, petrol, LPG/CNG, electric or hybrid. Motorcycles have license plates on the rear, while for trucks with trailers the license plate of the towing vehicle is on the front. As we had only one ALPR camera available, we had to make a choice. As we expected more high emitters in the motorcycle category, we chose to record the rear license plates. Trucks with trailers can still be recognized as such, as trailer license plates in NL all start with ‘O’, but we could not differentiate between different subcategories of trucks. The maximum noise level ( L A,max ) is plotted as a function of speed and acceleration in Figure 3 for all vehicles. Of these 53.955 vehicles, ca. 80% are passenger cars (80%), 10% are light duty vehicles (N1) and the remainder are a mixture of passenger and freight vehicles. Motorcycles, unfortunately, were only scarcely measured (≈ 150 in total), most likely due to the winter season and the fact that these often drive on the left side of the road, at least directly after passing the traffic lights.

Figure 3: Measurement data from Rotterdam, Oct 2021 – March 2022 after filtering: L Amax noise

levels plotted against vehicle speed, with colors indicating the vehicle acceleration

3.3. Privacy and GDPR The NEMO project and the N-RSD measurements must comply with the European data protection regulations (GDPR). In good agreement with the privacy experts from the City of Rotterdam as well as the NEMO privacy officer, we took the following measures: • Camera images are analyzed on-the-fly by the ANPR device using text recognition; the original video images are not stored, to prevent vehicle drivers or other persons to be recognizable. • Audio recordings are also directly analyzed in terms of various values (dB levels, spectra, prominence, etc.), and the original recording is automatically deleted. This way, no sensitive audio data (e.g. speech) could be accidentally recorded. In the future, a speech recognition algorithm may be integrated to eliminate this risk, which would allow the vehicle audio to be stored.

Laymaz |AB(A)} 2 So. ey 8 65: Selected passbys: noise level vs. speed 700 0.5 05 1 acceleration 2

• The license plate is automatically sent to the vehicle registry server which returns the vehicle information. Information about the owner or the driver of the vehicle is not available and not needed. The license plate itself is pseudonymized : using a hashing algorithm, it is encoded into a pseudocode that cannot be reversed into the original license plate text and then the original license plate text is deleted. The hashing algorithm does however return the same pseudocode if the same license plate is offered, so we are able to recognize if a particular vehicle has been measured already before. A privacy statement is available (in Dutch) online and an information sign is placed near the measurement location to inform drivers and the public and provide contact information. As data is stored anonymously, GDPR articles that require us to provide people with access or deletion of their data do not apply, which the GDPR legislation allows for research purposes. 4. CLASSIFICATION MODEL

4.1. Goal Based on the measurement data, every vehicle is classified automatically as a ‘high emitter’ or ‘normal emitter’. A clear definition of ‘high emitters’ is currently not available. Depending on the eventual legal implementation (see section 2), a high emitter could be defined as a vehicle that makes more noise than it is allowed to make (e.g. significantly above the type approval limit), more than it should make (e.g. aggressive driving) or more than what is expected (e.g. more than other, similar vehicles).

If the noise levels are compared to absolute type approval limits, the differences in test conditions need to be taken into account, such as different pavement, different microphone positions, and different meteorological conditions. This requires additional measurements or modelling to estimate the systematic influence of these factors and/or a certain margin to account for random errors. This is explained in section 4.4. As an alternative, the threshold for ‘high emitters’ can be defined locally, based on the actual measurement data. Using a certain initial training period for the classification model, ‘high emitters’ could then be defined as those vehicles that make significantly more noise relative to other, comparable vehicles, e.g. the top 1% noisiest passenger cars, or all cars > 10 dB louder than the average car. This is explained in section 4.3.

4.2. Driving conditions For the classification of high emitters, it is important to know the driving conditions during the pass- by, i.e. speed, acceleration and engine speed. Vehicles and tyres at higher driving speeds make more noise, and vehicles that accelerate faster or in low gear also make more noise. Although this may be unwanted, aggressive or ‘sporty’ driving is not forbidden, as long as it is below the posted maximum speed and it is not dangerous. It is important to distinguish between noisy drivers (e.g. fast acceleration in low gear) and noisy vehicles (e.g. with illegal or modified exhaust systems, broken parts or other modifications). Besides relevant for legal implementation, this distinction is also important for noise abatement measures. Noisy driving may be prevented by measures that influence traffic flow dynamics, or behavioral measures, such as informing the driver and raising awareness. Within NEMO, technology to communicate with the driver or with the vehicle dashboard is being developed to this end. To combat noisy vehicles, the vehicle should be sent to the workshop or the owner should be obliged to fix it and have it reapproved. And finally, if we want to link the classification to the type approval testing, we need to know the vehicle and engine speed, and correct for these. In our classification model, we therefore include not only vehicle information and noise levels, but also the driving conditions.

4.3. Classification relative to other vehicles For the relative classification, the goal is to identify which vehicles make significantly more noise than other, ‘normal’ vehicles in the same category, under the same driving conditions. For the

‘normal’ vehicles, we try to establish a reference noise level L Amax,ref as a function of the vehicle speed v in km/h and acceleration a in m/s 2 , by fitting the function of equation (1):

𝐿 ஺௠௔௫,௥௘௙ (𝑣, 𝑎) = 𝑐 ଴ + 𝑐 ௩ ∙𝑣+ 𝑐 ௔ ∙𝑎 (1) The coefficients c 0 , c v and c a are determined for each separate vehicle category (M1, N1, etc.). The engine speed is not yet included in this analysis. Different linear regression methods, including robust methods, have been compared to reduce the influence of outliers. A fundamental issue is that our dataset contains both normal and high emitters, and that our data are unlabeled, i.e. we do not know which datapoints are the high emitters. We can assume, however, that the normal vehicles will have the lowest noise levels, as these pass-bys will be dominated by tyre/road noise, which is then the lower boundary of the dataset. Following this assumption, we fit equation (1) not by taking the least squared error around the average, but by using a quantile regression approach to find the 20- percentile. For this, we used the QuantReg implementation found in the Python statsmodels package [7], wrapped in a scikit-learn [8] estimator class to enable cross-validation and other functionalities. The resulting reference function for M1 vehicles (passenger cars) is given in Figure 4. The blue band indicates the L Amax,ref reference level as a function of speed, with the width of the band indicating the influence of vehicle acceleration, from ca. -2 to +3 m/s 2 .

Figure 4: Measurement data from Rotterdam for M1 vehicles as L Amax vs. vehicle speed; the blue band indicates the reference level L Amax,ref as a function of speed, with the band width indicating the

influence of vehicle acceleration (min – max) With the fitted function (1) representing the reference noise level for any ‘normal’ vehicle under specific driving conditions, a high emitter is then defined as a vehicle pass-by for which the actual measured L Amax level is significantly higher than the reference level. Figure 5 shows the difference between the measured L Amax and the reference L Amax,ref for that vehicle category at the same speed and acceleration. The figure shows that: • ca. 90% of the vehicles are within a range of 8 dB(A); these are considered ‘normal emitters’. The peak of this group is slightly above 0 dB(A) as we use the 20-percentile as a reference value; • the loudest 1% is 10 dB(A) louder than the average normal emitter; these may be considered ‘high emitters’.

LaAymaz |AB(A)] 8 g 8 a 70 LAmax for vehicle category M1 ‘measurement reference 100

Further examination of the data shows that the vast majority of the high emitters are passenger cars (M1), with a few light duty vehicles (N1). Motorcycles have been measured only scarcely and are not classified. For enforcement, a single measurement may not be accurate enough to sanction the vehicle owner. However, the dataset contains several ‘repeating offenders’: vehicles that have been classified as high emitters on more than one occasion. What we heard from Rotterdam City and what the Dutch Minister wrote to parliament is that it is mainly a small group of offenders that cause these excessive noise events. Therefore, it is helpful to be able to register multiple measurements of the same high emitting vehicle, thereby increasing accuracy of the classification. To give an impression, Table 1 shows a list of the individual vehicles that are in the top 1% more than once. At least on this location, it does not seem to be only high-end sports cars that cause the most nuisance.

Figure 5: Difference between measured L Amax and reference L Amax,ref for each vehicle pass-by. Left :

difference vs. vehicle speed, with red dots indicating the highest 1% (‘high emitters’) and orange

indicating the highest 2 – 5% (‘medium emitters’). Right: statistical distribution of difference

values, with 95- and 99-percentile values indicated in the legend.

Table 1: List of vehicles identified as top 1% loudest vehicles more than once

Model Make # High emitter

Model Make # High emitter

classifications

classifications

Ford Fiesta 2 Volkswagen Crafter 2

Seat Ibiza ST 2 Opel Zafira Tourer 2

Honda Jazz 2 Jeep Compass 2

Nissan Juke 2 BMW X1 SDRIVE18D

2

Renault Kangoo 2 Ford Mustang EcoBoost Coupe

3

Seat Arosa 2

measured - reference La, max [4B(A)] 20 15 10 measured LAmax - calculated La. max 10 dB(A) KDE - 95% pct = +7.9 dB 99% pet = +12.1 dB lm histogram 0.05 0.10 0.15 0.20 speed [km/h]

4.4. Classification related to type approval The type approval regulations defined in R51 [3] are based on a combination of different vehicle tests on a test track under controlled conditions: one noise measurement at constant speed (cruise-by) and a second measurement under full acceleration conditions (wide open throttle). In later amendments to the R51, the Additional Sound Emission Provisions were added, which are currently being adapted to include Real Driving conditions (RD-ASEP). These provisions widen the scope of the measurement method to cover a much wider range of driving conditions. The method we followed is described in detail in a recent proposal [9]. The document describes a model that calculates the expected L Amax sound level for a particular vehicle, based on the vehicle and engine speed. The model, however, still needs the noise results from a cruise-by ( L CRS ) and a from full acceleration test ( L ACC ) as input, as well as measured reference values for the engine speed under accelerating conditions. We have used this method as a Sound Expectation Model (SEM) that calculates a more sophisticated reference noise level than the previous method (section 4.3). For the L CRS and L ACC we have calculated median values from the measurement data, by taking selections of only constant speed measurements and high acceleration measurements, respectively. For the engine speed, we calculated a fixed average ratio between these two driving conditions. All such reference values are calculated for subsets of the data, separating between vehicle categories, fuel type and other parameters such as the power-to-mass ratio. Using these assumptions, we were able to calculate a reference value for every vehicle using the SEM, and get a better indication of which pass-bys in our dataset are actually high emitters. This indication is then used as a means to label our data and train a machine learning (ML) model. The results of this approach are currently work-in-progress and will be available soon. The model will be applied to new measurements that will be performed in Q2 and Q3 of 2022.

5. CONCLUSIONS AND NEXT STEPS

In this paper, we have shown how high noise emitting vehicles can be separated from normal vehicles, using data from an unsupervised, roadside measurement system as developed in NEMO. Such an RSD device for noise will fulfil an important role in future enforcement of excessively noisy vehicles; we have identified several legal options that could be employed for this goal. Furthermore, the NEMO measurement and communication system will help to positively influence driver behavior. We have explained why it is important to including information on driving conditions. Currently, it seems that detailed technical vehicle information is not needed per se, other than the vehicle category and the number of engine cylinders (to calculate the engine speed). Applying a relative classification model to the Rotterdam dataset showed that after correcting for driving behavior, the 1% noisiest vehicles are at least 10 dB louder than the average normal vehicle within the same category. Adding the influence of driving conditions will enlarge this difference further. Currently, a more advanced classification model is being developed that makes use of European type approval regulations, which will strengthen its credibility and representativity. However, some assumptions still have to be made in order to classify vehicles based on a single in-traffic roadside measurement under arbitrary driving conditions. With that, the NEMO project also aims to formulate recommendations to adapt or expand EU vehicle noise regulations, to enable such remote sensing devices to be used for in-traffic enforcement of high noise emitters. Conversations with the European Commission on such adaptations will follow later this year, and recommendations will be part of the final project Deliverables. These will be publicly available at the end of the NEMO project in May 2023.

6. ACKNOWLEDGEMENTS

The NEMO project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No. 860441. The work described in this paper was made possible with the help of our NEMO Work Package 3 partners, Müller-BBM (DE) and SINTEF (NO). More information about the project can be found on https://nemo-cities.eu/ . 6. REFERENCES

1. European Commission, Communication from the Commission to the European Parliament, the

Council, the European Economic and Social Committee and the Committee of the Regions . Pathway to a Healthy Planet for All – EU Action Plan: ‘Towards Zero Pollution for Air, Water and Soil’ , COM/2021/400 final (2021) 2. Kirchoff N, Männel M, Ertsey-Bayer M. Advancements in autonomous detection of high noise

emitters in road traffic Proceedings of INTER-NOISE 2022 , Glasgow (2022) 3. Regulation No 51 of the Economic Commission for Europe of the United Nations (UNECE) —

Uniform provisions concerning the approval of motor vehicles having at least four wheels with regard to their sound emissions [2018/798], including amendments up to Supplement 2 to the 03 series of amendments (2018) 4. Regulation (EU) No 540/2014 of the European Parliament and of the Council of 16 April 2014

on the sound level of motor vehicles and of replacement silencing systems, and amending Directive 2007/46/EC and repealing Directive 70/157/EEC (2014) 5. Directive 2014/47/EU of the European Parliament and of the Council of 3 April 2014 on the

technical roadside inspection of the roadworthiness of commercial vehicles circulating in the Union and repealing Directive 2000/30/EC (2014) 6. ISO 11819-1:1997. Acoustics — Measurement of the influence of road surfaces on traffic noise

— Part 1: Statistical Pass-By method (1997) 7. https://www.statsmodels.org/dev/examples/notebooks/gen er a te d/ q ua n ti l e_r e gr e ss i on.html 8. https://scikit-learn.org/ 9. Economic Commission for Europe of the United Nations (UNECE), Proposal for Supplement 9

to 03 series of amendments to UN Regulation No. 51, submitted by the Informal Working Group on Additional Sound Emission Provisions to the 75 th session of the Working Party on Noise and Tyres, Geneva, February 2022