Effectiveness of the Number of Extracted Features from MFCC on Deep Learning for Fault Classification of Industrial Machinery Based on Acoustic Signal

283 - 293

Abstract

Reliable and efficient fault classification of induction motors is a critical concern in industrial environments. This study focuses on four operational states of a three-phase induction motor: normal function, imbalance fault, and horizontal and vertical misalignment faults. We propose a comparative analysis of the impact of the number of features extracted from Mel-frequency Cepstral Coefficient (MFCC) for reliable fault diagnosis of induction motors using acoustic emission signal processing. In this work, Long Short-Term Memory (LSTM) is utilized as the fault classification algorithm, and its performance is compared with three other deep learning models: Feed-Forward Neural Network (FFNN), Convolutional Neural Network (CNN), and Recurrent Neural Network (RNN). The results demonstrate that using LSTM with 13 MFCC features achieves a validation accuracy of approximately 97.4%, with 26 MFCC features achieving 96.12%, and with 39 MFCC features achieving 96.3%.