Steam trap opening sound recognition

Abstract

This study aims to evaluate simple recurrent neural network (simple RNN), long short-term memory (LSTM) and gate recurrent unit (GRU) for steam trap opening sound classification. This study conducts three experiments using different activation functions: ReLU, tanh and sigmoid functions at the output layer. Five feature extraction methods are employed including zero-crossing rate, spectral centroid, Mel-frequency cepstral coefficient, spectral rolloff, and short-term Fourier transform. We collected 24, 512 real-world audio files of steam trap operations, dividing into three subsets for evaluating the models. The results showed that an LSTM model employing sigmoid function yielded the highest precision, recall, specificity and F1 score of 75.00%, 66.67%, 99.97% and 70.59%, respectively. In addition, the LSTM model also outperformed the GRU and the Simple RNN.