Feed forward neural network application for classroom reverberation time estimation

Abstract

Acoustic problem is a main issues of the existing classroom due to lack of absorption of surface material. Thus, a feed forward neural network system (FFNN) for classroom Reverberation Time (RT) estimation computation was built. This system was developed to assist the acoustic engineer and consultant to treat and reduce this matter. Data was collected and computed using ODEON12.10 ray tracing method, resulting in a total of 600 rectangular shaped classroom models that were modeled with various length, width, height, as well as different surface material types. The system is able to estimate RT for 500Hz, 1000Hz, and 2000Hz. Using the collected data, FFNN for each frequency were trained and simulated separately (as absorption coefficients are frequency dependent) in order to find the optimum solution. The final system was validated and compared with the actual measurement value from 15 different classrooms in Universiti Tun Hussein Onn Malaysia (UTHM). The developed system show positive results with average validation accuracy of 94.35%, 95.91%, and 96.42% for 500Hz, 1000Hz, and 2000Hz respectively.