Abstract Purpose This work addresses the challenge of rolling element bearing fault diagnosis under varying operating conditions, where reliable generalization remains a major limitation for many existing deep learning–based approaches. Most prior studies are commonly evaluated under constrained experimental settings, often assuming a single motor operating condition and limited bearing variability, which restricts their robustness and practical applicability. The objective of this study is to develop a fault diagnosis framework that improves cross-condition robustness by fusing complementary fault representations, namely Impulse-Driven Morphological Maps (IDMM) and Fractal Dynamics Maps (FDM), within a dedicated deep learning architecture termed Impulse–Fractal Dual-Map Network (IMFD-Net). Method The proposed framework constructs two complementary diagnostic representations from vibration signals. IDMM is designed to emphasize structured impulse-related motion characteristics associated with localized bearing defects, whereas FDM captures nonlinear, scale-dependent dynamic complexity through fractal analysis. These two maps are fused to provide a joint representation that integrates deterministic impulse-related information and stochastic dynamic complexity. The fused IDMM+FDM representation is then processed by the proposed IMFD-Net architecture, which is designed to efficiently learn hierarchical features from heterogeneous diagnostic maps. The framework is validated on two benchmark datasets, namely the Case Western Reserve University (CWRU) dataset and the Paderborn Bearing Dataset, under multiple motor load, operating-condition, and mixed-condition evaluation settings. Computational efficiency is also assessed by considering feature extraction and model inference jointly. Results Experimental results demonstrate that the proposed IDMM+FDM fusion and IMFD-Net framework achieve strong robustness and generalization across different datasets and operating regimes. The classification accuracies range approximately between 90% and 98% under multiple motor load, operating-condition, and mixed-condition scenarios. The fused IDMM+FDM configuration requires 55.6 ms per segment on the CWRU dataset and 80.0 ms per segment on the Paderborn dataset when feature extraction and inference are jointly considered. These results indicate that the proposed representation effectively captures both localized impulse-driven fault patterns and nonlinear fractal dynamic characteristics, leading to reliable diagnostic performance under variable working conditions. Conclusions The study demonstrates that combining Impulse-Driven Morphological Maps and Fractal Dynamics Maps provides a complementary and robust representation for bearing fault diagnosis. By integrating these heterogeneous maps within the proposed IMFD-Net architecture, the framework improves fault classification performance across different motor loads, operating conditions, and dataset regimes. The results suggest that the proposed method offers a practically applicable route for robust bearing condition monitoring, particularly in scenarios where operating conditions and bearing characteristics vary across training and testing environments.
Hazret Tekin (Thu,) studied this question.