A combined use of wavelet packet decomposition and convolutional neural network for robust bearing fault diagnosis

Rotating machines play a vital role in industrial systems, where unplanned downtime can lead to significant financial and operational losses. Rolling bearings, essential for supporting rotational motion, are prone to faults such as misalignment, imbalance, and wear, making their monitoring crucial for predictive maintenance. This paper proposes a novel fault diagnosis method combining wavelet packet decomposition (WPD) with a one-dimensional convolutional neural network (1D-CNN). WPD decomposes vibration signals to extract time-frequency fault features, from which statistical indicators such as kurtosis, energy, and crest factor are calculated. A new fault-sensitive indicator, defined as the product of these metrics, selects the most informative component for fault detection. The reconstructed signal is then classified using a 1D-CNN, achieving excellent diagnostic accuracy under varying loads. Compared with Random Forest (RF) model, the proposed method better captures non-stationary features and enhances classification performance, offering a robust solution for rotating machinery fault diagnosis.