A Peak-Centric Approach to Bearing Fault Diagnosis Using Progressive Moving Transform and 2D- Convolutional Neural Network

Bearing fault diagnosis is critical for predictive maintenance in industrial machinery, yet many existing data-driven methods struggle to adapt to varying operational loads and often analyze entire vibration signals, which can dilute key fault indicators. To address this, we propose a novel peak-centric approach that focuses on diagnostically rich signal regions, combining the Progressive Moving Average Transform (PMAT) with a 2D Convolutional Neural Network (CNN) for enhanced classification. Our primary contribution is a novel methodology that leverages localized peak regions for fault diagnosis, integrating the recently developed PMAT signal transformation and validating its generalization to mechanical systems to create highly discriminative 2D image representations from 1D vibration data. The method involves three key steps: extracting fixed-length signal fragments containing significant peaks, converting these fragments into 120×120 pixel images using the Left PMAT transform, and classifying the images into one of four health states using a custom 2D-CNN architecture. The model was rigorously evaluated on the CWRU dataset under a leave-one-load-out cross-validation scheme across four distinct load scenarios. It achieved exceptional performance, with macro-average F1-scores exceeding 99.83% in three of the four scenarios, specifically under loaded conditions (1-3 HP), and a top accuracy of 99.96%. A comparative analysis demonstrated that our PMAT-based method consistently outperformed a Continuous Wavelet Transform (CWT) baseline and other recent state-of-the-art models under these loaded scenarios. In conclusion, the proposed PMAT and 2D-CNN framework provides a robust and highly accurate tool for bearing fault diagnosis, successfully demonstrating PMAT's cross-domain generalization capability while establishing a competitive benchmark for future research. Future work will explore a hybrid PMAT-CWT transformation to further improve performance under zero-load conditions.