32th Congress of the International Council of the Aeronautical Sciences

08 - Manufacturing and Supply Chain Management


R. Rameshbabu¹, D. Rajaram¹, O.J. Pinon Fischer¹, T.G. Puranik¹, D.N. Mavris¹; ¹Georgia Institute of Technology, United States

Rolling bearings play a critical role in the health of rotating machinery. Consequently developing an approach that is structured, repeatable and scalable for the identification of anomalies in rolling bearings is of significant value o the manufacturing industry. Machine learning techniques have been traditionally used for the detection of anomalies in rolling bearing data; however, theirrnperformance relies extensively on feature engineering, making them highly dependent on subject matter expertise and reducing their generalizability. Traditional machine learning techniques also struggle with large amounts of data. Deep learning architectures, on the other hand, perform well with large amounts of data. More importantly, they are able to learn complex representations through successive layers of data processing units, hence circumventing the significant amount of domain expertise and time dedicated to feature engineering. While the literature on the application of deep learning for bearing anomaly detection is growing, the vast majority of studies limit themselves to a specific Deep Learning architecture and/or variant for the application of interest. When benchmarking studies are conducted, the models being evaluated are fine-tuned manually, reducing the generalizability and transferability of the approach and models to other processes, tools, or machines.rnThe present paper addresses such limitations by 1) conducting a benchmarking and assessment study of most relevant feature extraction techniques and deep learning architectures for rolling bearing anomaly detection and 2) implementing a structured approach to the selection of feature extraction techniques and deep learning architectures that is repeatable and scalable across a wide variety of architectures.

View Paper