Author :
Abstract
Bearing fault detection plays a crucial role in ensuring the reliability, safety, and operational efficiency of rotating machinery used in automotive and industrial applications. Early and accurate identification of bearing defects can significantly reduce unexpected downtime, maintenance costs, and potential system failures. Among the numerous diagnostic approaches, vibration signal analysis has emerged as one of the most effective and widely adopted techniques for fault diagnosis, owing to its sensitivity to subtle changes in mechanical behavior.
In recent years, deep learning models—particularly Convolutional Neural Networks (CNNs)—have demonstrated remarkable capability in automatically learning complex patterns from raw sensor data. This study conducts a comparative evaluation of two CNN-based methods: a Short-Time Fourier Transform (STFT)-driven two-dimensional (2D) CNN and a raw vibration signal-driven one-dimensional (1D) CNN. The dataset includes vibration signals representing different bearing health states, namely healthy, inner race fault, and outer race fault conditions. The STFT-based 2D CNN performs time–frequency analysis to extract spectral features, while the 1D CNN learns discriminative features directly from time-domain data. Experimental findings reveal that the 1D CNN achieves higher classification accuracy with reduced preprocessing effort, whereas the STFT-based method provides more interpretable visual representations. The comparative results offer practical insights into the trade-offs between interpretability and accuracy, guiding the selection of appropriate models for real-world bearing fault diagnosis systems.
Keywords
Abstract
Bearing fault detection plays a crucial role in ensuring the reliability, safety, and operational efficiency of rotating machinery used in automotive and industrial applications. Early and accurate identification of bearing defects can significantly reduce unexpected downtime, maintenance costs, and potential system failures. Among the numerous diagnostic approaches, vibration signal analysis has emerged as one of the most effective and widely adopted techniques for fault diagnosis, owing to its sensitivity to subtle changes in mechanical behavior.
In recent years, deep learning models—particularly Convolutional Neural Networks (CNNs)—have demonstrated remarkable capability in automatically learning complex patterns from raw sensor data. This study conducts a comparative evaluation of two CNN-based methods: a Short-Time Fourier Transform (STFT)-driven two-dimensional (2D) CNN and a raw vibration signal-driven one-dimensional (1D) CNN. The dataset includes vibration signals representing different bearing health states, namely healthy, inner race fault, and outer race fault conditions. The STFT-based 2D CNN performs time–frequency analysis to extract spectral features, while the 1D CNN learns discriminative features directly from time-domain data. Experimental findings reveal that the 1D CNN achieves higher classification accuracy with reduced preprocessing effort, whereas the STFT-based method provides more interpretable visual representations. The comparative results offer practical insights into the trade-offs between interpretability and accuracy, guiding the selection of appropriate models for real-world bearing fault diagnosis systems.
