A Study on Training Data Mixing and Selection Methods for Detecting the Sounds of Faulty Machinery

To train a classifier using a Deep Neural Network (DNN), a substantial amount of data sets is required. However, in cases where data acquisition is challenging or the environment undergoes changes, obtaining sufficient data for training can be problematic. Data augmentation and synthesis can be used to increase the quantity of data for training. The data generated through augmentation or synthesis should closely resemble real data and accurately reflect the environments and characteristics that users aim to model. Without this resemblance, using the generated data may not yield the desired results in the actual environment. In this paper, we propose an empirical method for selecting synthetic training data that enhances the performance of a belt conveyor fault classifier model in environments where data acquisition is challenging, without compromising the existing performance of the model. Copyright © The Korean Institute of Electrical Engineers.