A PREDICTION MODEL FOR THE FLOOR IMPACT SOUND USING RANDOM FOREST REGRESSION

Abstract

 In Japan, measurement of the floor impact sound using standard heavy impact sources is stated in JIS A 1418-2:2019, and measurement of the reduction level of the floor impact sound using standard heavy impact sources is stated in JIS A 1440-2:2007. In both standards specifies the tire as a standard heavy impact source (1) and the rubber ball as a standard heavy impact source (2).

 Even though both tire and rubber ball are standardized in JIS, the use of the rubber ball is limited. For example, the ball can be used in AIJES-S001-2008. But the tire can be used in Housing Performance Indication System, Promotion of Housing Quality Assurance, and the sound insulation performance standards of the Architectural Institute of Japan, etc. Considering this situation, it would be preferable to use both tire and ball than to use only one of them. However, the translation of the measurement done by the tire to the rubber ball or vice versa is investigated, but only in the low-frequency range and yet solved.

 Hence, this paper investigated the use of a machine learning algorithm as a solution to the translation between two different standard heavy impact source measurements. The Random forest regression model was investigated as a prediction approach for the floor impact sound and the floor impact sound reduction level. At present, the sample is biased toward the data of the wall-type structural test chamber at the Building Research Institute, hence, the verification is limited. However, the outcome suggests that the maximum difference between measured and predicted floor impact sound level (L_Fmax) was within 2.3 dB for tire source and 2.5 dB for rubber ball source. In terms of the floor impact sound reduction level (ΔL_Fmax), it was within 3.9 dB for tire source and 3.6 dB for rubber ball source. For both L_Fmax and ΔL_Fmax, the variance by inter quarter range shows a good agreement between measured and predicted values. Hence, the paper showed by using random forest regressor, it is possible to translate the measurement done by the tire to the rubber ball or vice versa by changing the feature of the excitation source from the tire to the rubber ball.