Acoustic Signal Processing and Deep Learning Nexus for Brushless Direct Current Motor Status Classification

Abstract

Brushless Direct Current (BLDC) motors are ubiquitously deployed in various industries that underline the present industrial revolution. Owing to their versatility and relevance in the manufacturing industry, predictive maintenance of these motors calls for critical concern, over implications such as operational downtime and maintenance cost. This paper presents the development of a BLDC motor classification framework leveraging acoustic signal processing and deep learning. The IDMT-ISA-ELECTRIC-ENGINE dataset for the 42BLF01 Brushless Direct Current (BLDC) motor classification under artificially induced states, was imported for model development. This audio dataset contains three classes, namely, ‘engine1_good', 'engine2_broken' and 'engine3_heavyload'. Audio processing was carried out on the dataset, to convert the audio signals into mel spectrograms, via the Short-time Fast Fourier and Normalization techniques. Three recurrent deep learning models (Gated Recurrent Unit (GRU), Long Short-Term Memory (LSTM) and an LSTM-GRU hybrid), three transformer models with positional encodings (pe) (rotational_pe, sinusodal_pe and relative_pe) and a convolutional neural network (CNN) model (cnn2d_model) were implemented to fit the processed data. Model training, validation and inference were carried out and it was discovered that although all the models performed excellently, the best performing models were sinusodal_pe and relative_pe on the basis of Loss score. Furthermore, a BLDC motor status classifier application was developed using Streamlit. Developed recurrent models along with an ensemble of the models were deployed in the application and could classify BLDC motor operational status excellently