Analyzing the Impact of Signal Domain Representations on Bearing Fault Detection Performance

Bearing failures are a common issue in industrial settings, requiring effective detection methods to ensure reliability. Traditional approaches rely on either physics-based models or data-driven techniques, with the latter proving efficient in differentiating between healthy and faulty conditions. Deep learning has further advanced fault detection by automating feature extraction, but model performance is influenced by the signal domain used. While time and frequency domains are widely applied, angular and order domains are particularly useful under varying rotational conditions, yet their application in deep learning remains underexplored. This study evaluates the effectiveness of different signal representations (time, frequency, angular, and order) for bearing fault detection under varying rotational speeds. A multiclass classification approach is adopted to assess their performance, and an error analysis identifies factors influencing classification accuracy. Results highlight the advantages and limitations of each representation, providing insights into their suitability for fault detection and guiding future research directions.