Clustering and classification of the residential noise in apartment buildings based on machine learning using spectral and temporal characteristics Jeonghun Kim 1 , Chunwon Eom 2 , Jongkwan Ryu 3 Chonnam National University 77, Yongbong-ro Buk-gu, Gwangju, Rebublic of Korea

ABSTRACT In this study, clustering and classification of the residential noise in apartment buildings based on machine learning were conducted using data of spectral and temporal characteristics. The sound sources were recorded in residential spaces and the length of each sound source was edited to 5 seconds. As a result of the first spectral characteristic clustering, three spectral characteristic clusters (C1, C2, and C3) were derived for all evaluation methods (L Aeq and L Amax ) and octave bands (1/1 and 1/3 octave band) analysis. As a result of the secondary temporal characteristic clustering, when the L Aeq using 1/1, 1/3 oct bands were used, each cluster consists of 3 time-varying clusters, resulting in a total of 9 residential noise clusters (C1-1, C1-2, ··, C3-2, and C3-3). In the clustering using L Amax for1/1 and 1/3 oct bands, a total of 11 residential noise clusters were generated according to C1 (k=3), C2 (k=3), and C3 (k=5) time-varying clusters. By labeling with the residential noise cluster for each analysis method, sound source classification was performed through machine learning (SVM, k-NN, ANN), and differences between analysis methods were investigated. In all analysis methods, it was found that the ANN model (accuracy avg: 84%) classified residential noise clusters better than the SVM and k-NN models. A total of 9 residential noise clusters composed of analysis methods (L Aeq and 1/3 oct band) for labeling, and the ANN showed the highest accuracy value of 85%. 1. INTRODUCTION

The number of people affected by exposure to noise is increasing with the development of cities. An individual’s response to noise is different according to spectral and temporal characteristics even though the noise source is at the equal sound pressure level [1-2]. In Korea, complaints about floor impact sound are the most common in apartments [3], and since floor impact noise is a sound source dominated by low frequencies, it is necessary to categorize floor impact noise. To manage the noise between upper and lower units, the regulation on sound insulation performance of floor in building and sound level of actual floor impact noise in living space are used [4]. At governmental management center for floor impact noise, investigators judge the sound source type and to exceed the criteria of sound level of floor impact noise by hearing sound recorded for 24-hours. This involves a lot of time and individual subjective judgment. To solve such a problem, a simple model that uses

1 wjdgns9752@naver.com 2 ucw1472@naver.com 3 jkryu@jnu.ac.kr

machine learning to make an objective judgement with less time is needed. Many existing sound classification studies using machine learning utilize deep learning (CNN, RNN, CRNN, etc.) algorithm. However, sound classification using deep learning takes a lot of time because it converts and analyzes images or time series data. Therefore, a simple and real-time analysis of sound sources using machine learning is needed. In this study, residential noise clustering and classification were conducted and compared through an analysis method based on spectral and temporal characteristics, which is relatively easy to analyze for residential noise sources generated in apartments.

2. DATASET OF SOUND SOURCES

2.1. Residential noises Residential noise sources were selected by referring to the status of each noise complaints surveyed by the governmental management center for floor impact noise [3]. For floor impact sound and air transmission sound, sound sources were recorded in a mock-up apartment unit with area of 84 and 59 m 2 and a slab thickness of 210 m. The traffic, environmental, and construction noise sound source were recorded from the outside and the filter of sound insulation of the façade was applied [5].

Table 1. Types of residential noise sources

Noise

Concrete pumping

Adult jumping

Bath

Adult running Shower Concrete

vibrator Adult walking Washbowl Earth auger

Drainage and

facility noise

Children`s running Toilet Excavator

loading Children`s

Floor impact

Outdoor unit

noise Forklift

walking

noise

Chair scrapping

Road 40km/h Loader Remote dropping Road 60km/h Payloader Entrance close Road 80km/h Piledriver Golf ball dropping Environmental

Traffic noise

Construction

noise

Airplane noise Road roller Hammering Railway noise Stone crush Indoor close

noise

Air-compressor

Spoon dropping Asphalt finisher Truck unloading

Dog bark Batch plant Metal sound Breaker Tunnel ventilation Piano sound Bulldozer

Construction

noise

Airborne

noise

TV sound Concrete mixer Vibrating

roller

2.2. Preprocessing and analysis The length of each sound source was edited to be 5 seconds, and the sound pressure level of all sound sources was adjusted from 45 dBA, which is a indoor noise criteria for road traffic noise, to -5 to +5 dBA, 40 to 50 dBA in Leq (1 dBA interval). To extract the spectral characteristics, L Aeq and L Amax values for each 1/1 (31.5 to 2,000 Hz) and 1/3 (25 to 2,500 Hz) octave bands were derived. For temporal characteristics, L Aeq values for each 0.1s to 4.993s were derived through sound pressure level analysis for 5 seconds (time weighting: fast with 6 ms step). 3. CLUSTERING

3.1 Method Clustering is a representative unsupervised learning algorithm that divides a given data set into groups of similar data. In this study, residential noise sources were grouped by spectral and temporal characteristics through K-means clustering [6]. The number of k in k-means clustering was selected through the silhouette coefficient. First, in order to group by spectral characteristics of residential noise sources, L Aeq and L Amax values for each octave band were derived as shown in Table 2, and primary clustering was performed and compared. Thereafter, as shown in Table 3, secondary clustering was performed based on temporal characteristic values for each cluster.

Table 2. Examples of spectral characteristics features

1/1 oct band [Hz] 31.5 63 125 250 500 1k 2k

L Aeq 65.28 53.90 46.70 37.78 33.33 34.35 34.04

L Amax 69.85 59.58 51.85 37.01 28.97 30.26 28.42

1/3 oct band [Hz] 25 31.5 40 50 63 ··· 2k 2.5k

L Aeq 57.88 64.10 53.81 44.36 51.86 ··· 28.98 30.34

L Amax 59.56 64.70 60.80 49.39 57.70 ··· 22.76 25.43

Table 3. An example of temporal characteristics features

Time[s] 0.1 0.106 0.112 ··· 0.4981 0.4987 4.993

L Aeq 36.13 36.10 36.41 ··· 39.40 39.44 39.66

3.2 Result Table 4 shows the mean and standard deviation of the silhouette coefficients used to determine k of k-means clustering. In this study, the k with the highest average and the lowest standard deviation of silhouette coefficient among k-values over 2 was used. As a result of the first spectral characteristic clustering, three temporal characteristic clusters (C1, C2, and C3) were derived for all evaluation methods and octave bands. As a result of the secondary temporal characteristic clustering, when the evaluation method L Aeq and 1/1, 1/3 oct bands were used, each cluster consists of 3 time-varying clusters, resulting in a total of 9 residential noise sources (C1-1, C1-2, · ··, C3-2, and C3-3). A total of 11 residential noise clusters were generated in the clustering using L Amax and 1/1, 1/3 oct bands according to the evaluation method C1 (k=3), C2 (k=3), and C3 (k=5) time-varying clusters. Figure 1 and Table 5 show the spectral characteristic clusters (C1, C2, and C3) by the analysis using L Aeq and 1/3 oct band and the time-varying clusters of each spectral characteristic cluster (C1-1, C1-2, ..., and C3-3) in L Aeq values. Although they were clustered into the same spectral cluster, they were temporally more finely divided. Table 5 indicates the most included sound source for each cluster.

Table 4. Average and standard deviation of silhouette coefficient of each clustering for spectral and

temporal characteristics

Spectral characteristics Temporal characteristics

k AVE STEDV k AVE STEDV

C1 3 0.25 0.12

L Aeq (1/1) 3 0.30 0.02

C2 3 0.23 0.22

C3 3 0.25 0.06

C1 3 0.25 0.14

L Aeq (1/3) 3 0.31 0.05

C2 3 0.25 0.06

C3 3 0.22 0.23

C1 3 0.25 0.07

L Amax

(1/1) 3 0.26 0.04

C2 3 0.22 0.23

C3 5 0.29 0.08

C1 3 0.25 0.07

L Amax

(1/3) 3 0.27 0.05

C2 3 0.22 0.24

C3 5 0.25 0.17

Spectral characteristics Temporal characteristics

C1 C1-1 C1-2 C1-3

C2 C2-1 C2-2 C2-3

C3 C3-1 C3-2 C3-3 Figure 1. Clustering results according to spectral and temporal characteristics

(Analysis method: L Aeq and 1/3 oct band)

Table 5. The most included sound source for each cluster

Cluster

C1-1 C1-

2 C1-3 C2-1 C2-2 C2-3 C3-1 C3-2 C3-3

Noise

Chair scrapping Bath Dry cell dropping

Road 40km Hammering Entrance

Adult jumping

Adult running

Adult walking

type

close

4. Classification

4.1 Method In this study, 9 and 11 residential noise clusters were labeled to classify sound sources through machine learning (Support Vector Machine, k-Nearest Neighbor, Artificial Neural Network) [7-9], and differences between analysis methods were investigated. All data were 70% training and 30% testing, and 30% of training data was used for validation. For the k-NN and SVM, the algorithm provided by scikit-learn was used, and it was verified using k-fold (n=10). ANN constructed the model as shown in Table 6, and the epoch was repeated 100 times.

Table 6. Artificial Neural Network model structure

4.2 Result Table 7 shows the results of each machine learning model by analysis method. In all analysis methods, it was found that the ANN (accuracy avg: 84%) model classified residential noise clusters better than the SVM (accuracy avg: 80.25%) and k-NN (accuracy avg: 76.25%) models. A total of 9 residential noise clusters composed of analysis methods ( L Aeq and 1/3 oct band) for labeling, and the ANN showed the highest accuracy value of 85% among all analysis methods. When the analysis method using L Aeq and 1/1 oct band was configured, the accuracy was 83%, and when the analysis method using L Amax for 1/1 and 1/3 oct band was configured, the accuracy value was 84%.

Table 7. Sound source classification (results for each analysis method

Analysis method

Model

L Aeq (1/1)

L Aeq (1/3)

L Amax

L Amax

AVG

(1/1)

(1/3)

ANN 83% 85% 84% 84% 84%

SVM 80% 79% 80% 82% 80%

KNN 78% 77% 76% 74% 76%

5. CONCLUSIONS

In this study, residential noises were clustered and classified based on the spectral and temporal characteristics of residential noise sources generated in apartments. As a result of the first spectral characteristic clustering, three spectral characteristic clusters (C1, C2, and C3) were found for all evaluation methods and octave bands. As a result of the secondary temporal characteristic clustering, when the evaluation method of L Aeq for 1/1 and 1/3 oct bands were used, each cluster consists of 3 time-varying clusters, resulting in a total of 9 residential noise clusters (C1-1, C1-2, · ··, C3-2, and C3-3). A total of 11 residential noise clusters were generated in the clustering using L Amax and 1/1,

Tae Te aan ae rare) ‘i, oe (me) (ne, 8) oe (ome) (re, 120 oe ome) (re, 0 st (me) (ue

1/3 oct bands according to C1 (k=3), C2 (k=3), and C3 (k=5) time-varying clusters. By labeling with the type of residential noise cluster for each analysis method, sound source classification was performed through machine learning. In all analysis methods, it was found that the ANN (accuracy avg: 84%) model classified residential noise clusters better than the SVM (accuracy avg: 80.25%) and k-NN (accuracy avg: 76.25%) models. A total of 9 residential noise clusters composed of analysis methods ( L Aeq and 1/3 oct band) for labeling, and the ANN models showed the highest accuracy value of 85% among all analysis methods. In the future, the difference in psycho-acoustic indicators between spectral and temporal clusters will be investigated, and the psychological properties of each cluster will be identified through semantic differential method. 5. ACKNOWLEDGEMENTS

This work is supported by the Korean Agency for Infrastructure Technology Advancement (KAIA) grant funded by the Ministry of Land, Infrastructure and Transport (Grant 21CTAP-C163631-01).

6. REFERENCES

1. Jeon, J. Y. Subjective evaluation of floor impact noise based on the model of ACF/IACF. Journal

of sound and Vibration, 241(1) (2001) 147-155 2. Ryu, J., Sato, H., Kurakata, K., Hiramitsu, A., Tanaka, M., & Hirota, T. Relation between

annoyance and single-number quantities for rating heavy-weight floor impact sound insulation in wooden houses. The Journal of the Acoustical Society of America, 129(5) (2011) 3047-3055 3. Floor Noise Management Centre [Monthly report]. Floor Noise, Management Centre, Korea

Environment Corporation, Republic of Korea (2018). 4. Ministry of Land and Transportation and Ministry of the Environment in Korea, Regulation of

the Scope and Criteria of Noise Between Floors in an Apartment (2014). 5. Ryu, J., & Song, H. Effect of building façade on indoor transportation noise annoyance in terms

of frequency spectrum and expectation for sound insulation. Applied Acoustics, 152, 21-30. (2019). 6. Hartigan, J. A., & Wong, M. A. Algorithm AS 136: A k-means clustering algorithm. Journal of

the royal statistical society. series c (applied statistics), 28(1), (1979) 100-108. 7. Coomans, D., & Massart, D. L. Alternative k-nearest neighbour rules in supervised pattern

recognition: Part 1. k-Nearest neighbour classification by using alternative voting rules. Analytica Chimica Acta, 136 (1982) 15-27. 8. Cortes, C., & Vapnik, V. Support-vector networks. Machine learning, 20(3), (1995) 273-297. 9. Gevrey, M., Dimopoulos, I., & Lek, S. Review and comparison of methods to study the

contribution of variables in artificial neural network models. Ecological modelling, 160(3), (2003) 249-264.