*3.1. The DYNAMAP Project*

The DYNAMAP LIFE project proposed the implementation of a dynamic noise mapping system [29], able to determine the acoustic impact of road infrastructures in real-time, following the European Noise Directive 2002/49/EC. A Multi-Sensor Network collects the noise level measurements in two pilot areas: In the city of Milan, and on the A90 motorway around Rome.

Each of the sensor nodes has to accomplish a set of basic specifications [42] defined to satisfy DYNAMAP requirements for each monitoring station, which are the following: (i) A 40–100 dB(A) broadband linearity range; (ii) a 35–115 dB working range with acceptable Total Harmonic Distortion

(THD); (iii) a narrow-band floor noise level; and (iv) a sampling rate of 48 kHz. The project also requires the possibility of audio recording, as well as VPN and GPRS/3G/Wi-Fi connections. The precision of the sensors is a key issue for system reliability [43].

In the case of the city of Milan, the deployment of the network is shown in Figure 2, where a team of acoustic experts chose up to 24 low-cost sensor locations to collect the data to generate noise maps. All of these locations are a key issue for the study of the features of the recorded signal and, therefore, for the performance of the ANED [12]. For more details about the location of the sensors, the reader should refer to [44,45].

**Figure 2.** Map of the sensor locations in Milan. Different colors correspond to a catalog of six groups of streets [46].

Although the final goal of the DYNAMAP project is to dynamically update the noise maps in a GIS-based platform [29,47], preliminary work has been done in the sensors to evaluate the features of the raw acoustic signals collected by the network; this work enabled the recollection of up to 100 h of outdoor urban acoustic raw data in the Milan environment.

### *3.2. Description of the Recording Campaign*

Table 1 lists the sensor nodes of the WASN, indicating their identifiers, street location, and GPS coordinates within district nine of Milan, in which they were installed. A recording campaign considering both weekends and weekdays (recording 20 min each hour every day) was performed through the multi-sensor network in real-operation conditions. The final selection of files to be labelled allowed our team to collect more than 100 h of raw urban acoustic data. Previous studies [48] showed that sampling two different days could accomplish the requirements of diversity of the urban activity during weekend days and weekdays; the different traffic flow attains both RTN and also the anomalous noise events found in an urban environment, showing the relevant differences. These previous studies took into account the experience of data set designs for other acoustic event detection projects, such as those in [8,49]. Furthermore, the recordings were exhaustive, considering the entire day period of each day, for two reasons: (i) The road traffic noise profile changes between day and night, as can be observed in the two-day equivalent noise level curve in Figure 3—noise profile curves attain a certain 24 h regularity, and this property has been taken into account in the recordings; and (ii) because the

types of events occurring are substantially different between day and night, and recording with a periodic sample widens the possibility of collecting more types of noise.


**Table 1.** List of the nodes of the Wireless Acoustic Sensor Network (WASN) deployed in district nine in Milan.

**Figure 3.** Example of two-day *LAeq*<sup>300</sup> curve of sensor hb148. The vertical axis is the measured *LAeq*300, in dB.

To that effect, recordings were planned for the 24-node network during the first 20 min of each hour, for each hour of the day, during two selected days: A weekday (Tuesday, 28th of November 2017) and a weekend day (Sunday, 3rd of December 2017), in order to maximize the diversity of the recorded ANEs and with a schedule intended to be exhaustive. A total of 1116 recordings of 20 min each were gathered during this recording campaign, producing an audio database of 372 h, from which our research group managed to manually listen and label up to 100 h.
