TY - JOUR

T1 - More than spectrograms

T2 - Deep representation learning for machinery fault detection

AU - Wißbrock, Peter

AU - Ren, Zhao

AU - Pelkmann, David

N1 - Publisher Copyright: © 2024 Elsevier Ltd

PY - 2024/7/20

Y1 - 2024/7/20

N2 - Machinery fault detection is a critical task for ensuring the reliability, safety, and quality of industrial systems. In recent years, anomalous sound event detection has been of rising interest to accomplish the above goals. Deep representation learning has recently emerged as a promising approach to address the challenge of developing accurate and robust fault detection systems by extracting representations from limited sensor data. In this paper, we analyze how acoustic-based deep representation learning can be applied to industrial real-world scenarios. To address the latest advances, challenges, and potential solutions in the field of industrial sound analytics, we present the approach of deep trainless data aggregation, a method that aggregates data without requiring additional training. We address applications under the limitation of data scarcity for anomalous observations and reduce the risk of scalability challenges in data collection for industries with many assets. By using high-level representations, which can be deployed to edge devices with limited computational costs, the approach applies to manifold industrial real-world applications. The proposed pipeline includes the use of three-dimensional features, image creation, and extraction of a deep representation from pretrained neural networks. Our intense benchmark shows that our solution reaches state-of-the-art performance. In particular, we highlight the importance of the three-dimensional transformations and their parameterization in this pipeline. These transformations are the basis on which many deep learning techniques for machinery fault detection are built. We demonstrate that the psychoacoustic loudness and roughness are effective alternatives to the conventional log-Mel-spectrograms. Our study can serve as a framework for further applications and developments in the field of industrial sound analytics. Further work could include the development of specialized neural networks for the application presented.

AB - Machinery fault detection is a critical task for ensuring the reliability, safety, and quality of industrial systems. In recent years, anomalous sound event detection has been of rising interest to accomplish the above goals. Deep representation learning has recently emerged as a promising approach to address the challenge of developing accurate and robust fault detection systems by extracting representations from limited sensor data. In this paper, we analyze how acoustic-based deep representation learning can be applied to industrial real-world scenarios. To address the latest advances, challenges, and potential solutions in the field of industrial sound analytics, we present the approach of deep trainless data aggregation, a method that aggregates data without requiring additional training. We address applications under the limitation of data scarcity for anomalous observations and reduce the risk of scalability challenges in data collection for industries with many assets. By using high-level representations, which can be deployed to edge devices with limited computational costs, the approach applies to manifold industrial real-world applications. The proposed pipeline includes the use of three-dimensional features, image creation, and extraction of a deep representation from pretrained neural networks. Our intense benchmark shows that our solution reaches state-of-the-art performance. In particular, we highlight the importance of the three-dimensional transformations and their parameterization in this pipeline. These transformations are the basis on which many deep learning techniques for machinery fault detection are built. We demonstrate that the psychoacoustic loudness and roughness are effective alternatives to the conventional log-Mel-spectrograms. Our study can serve as a framework for further applications and developments in the field of industrial sound analytics. Further work could include the development of specialized neural networks for the application presented.

KW - Anomalous sound event detection

KW - Condition monitoring

KW - Data scarcity

KW - Industrial sound analytics

UR - http://www.scopus.com/inward/record.url?scp=85199052844&partnerID=8YFLogxK

U2 - 10.1016/j.apacoust.2024.110178

DO - 10.1016/j.apacoust.2024.110178

M3 - Article

AN - SCOPUS:85199052844

VL - 225

JO - Applied acoustics

JF - Applied acoustics

SN - 0003-682X

M1 - 110178

ER -