Session: 17-01-01: Research Posters

Paper Number: 143231

143231 - Enhancing the Accuracy of Machinery Fault Diagnosis Through Fault Source Isolation of Complex Mixture of Industrial Sound Signals 

Machinery health monitoring techniques provide valuable insights into the performance and condition of machines. Among various monitoring techniques, acoustic sensor-based monitoring has become a significant area of interest for the industry due to its ability to accurately capture fault signatures, thereby improving the detection accuracies of deviations from regular operations. In the context of monitoring machine conditions through collecting acoustic sensor data, a generic complication arises in the form of a mixture of diverse sound sources combined with redundant machine-generated sounds. Such sounds complicate isolating and analyzing the desired sound signals related to the machine’s performance. The diverse nature of machine sounds, sound contamination, or complexity could lead to suboptimal fault features and yield inaccurate failure estimation. This inconsistency can potentially risk the entire decision-making process regarding machinery health, a consistent concern in the industry. It demands attention to explore new approaches that can optimize and recover actual machine sounds from a mixture of sounds with greater quality, ease of adaptation in the industry, and dependability. The significance of denoising methods and noise reduction techniques for machinery fault diagnosis applications has been extensively examined in the literature. However, the focus has been limited to recovering the actual fault sound source from a complex mixture of sounds in fault diagnosis applications. This study proposes a novel framework that enhances the accuracy of machinery fault diagnosis using audio source separation of a complex mixture of sound signals. The proposed approach employs a Deep Extractor for Music Source Separation (DEMUCS), a state-of-the-art music source separation approach for achieving high-quality industrial sound separation. It consists of an encoder/decoder comprising a convolutional encoder, a bidirectional Long short-term memory (LSTM) network, and a convolutional decoder interconnected with customized U-Net architecture. The proposed methodology consists of two steps. In the first step, fault sound isolation and recovering individual fault sounds from a complex mixture of sound signals are enabled using DEMUCS. In the second step, the isolated fault sounds are fed through a 1D-convolutional neural network (1D-CNN) classifier for adequate classification.

The DEMUCS-CNN approach was validated through a case study on bearing faults-based system fault diagnosis, four classes (“normal,” “inner,” “outer,” and “ball” faults) classification using a machine fault simulator by Spectra Quest equipped with a condenser mic for identifying multiple faults. To examine its relative performance on fault diagnosis, the DEMUCS-CNN approach was compared to the traditional blind source separation enabled 1D-CNN (BSS-CNN). The separation results of DEMUCS were evaluated based on objective measures using source-to-distortion (SDR) ratio, which demonstrated that the waveforms of the inner and ball faults were similar to the target waveforms, with SDR values of 7.80 and 4.83, respectively. However, the normal and outer fault conditions exhibited relatively lower separation results, with SDR scores of 3.20 and 2.35, respectively, and the distribution patterns had slight variations from their original signals. The second phase of the approach showed that fault classification using DEMUCS-recovered fault sounds led to enhanced fault classification overall accuracy (99%) and superior efficiency on unseen data compared to conventional BSS-recovered fault sounds classification overall accuracy (48.33%). According to the analysis of the area under curve (AUC) of receiver operating characteristics (ROC), it has been confirmed that the DEMUCS-CNN method performs better than the BSS-CNN method. The DEMUCS-CNN method has demonstrated a high AUC score of around 1.0 for all classes, while BSS-CNN achieved an average AUC score of 0.58, indicating the proposed method’s excellent performance on new data. The experimental outcomes of the study have demonstrated that the DEMUCS-CNN methodology has achieved a higher level of accuracy in prediction when encountering new or unseen data. Additionally, the comparison of the fault isolation approach by DEMUCS revealed superior fault classification accuracy, further supporting its effectiveness over the conventional BSS-CNN method. These findings suggest DEMUCS-CNN methodology has significant potential to be an effective tool for industries and researchers looking to enhance machinery fault classification and prediction capabilities.

Presenting Author: Ayantha Senanayaka Center for Advanced Vehicular Systems at Mississippi State University

Presenting Author Biography: Ayantha Senanayaka, currently working as a Postdoctoral Associate at the Center for Advanced Vehicular Systems located at Mississippi State University, completed his Ph.D. in industrial and systems engineering under the guidance of Dr. Linkan Bian. Prior to that, he earned his M.Sc. in Statistics from Mississippi State University and a bachelor's degree in Mechatronic Engineering from the University of Wolverhampton in the UK. His research interests include transfer learning, adversarial learning, ensemble learning, sensor fusion, and data-driven process monitoring on advanced manufacturing and maintenance applications.

Authors:

Ayantha Senanayaka Center for Advanced Vehicular Systems at Mississippi State University
Philku Lee Korea Automotive Technology Institute
Nayeon Lee Center for Advanced Vehicular Systems at Mississippi State University
Charles Dickerson US Army Corps of Engineers Engineer Research and Development Center
Anton Netchaev US Army Corps of Engineers Engineer Research and Development Center
Sungkwang Mun Center for Advanced Vehicular Systems at Mississippi State University