Multiview Scalogram Fusion With Adaptive SE-CNNs for Bearing Fault Diagnosis

Rolling-element bearings are essential components in rotating machinery, but they are vulnerable to specific defects that may lead to performance degradation and unplanned downtime. Accurate diagnosis in noisy and varying conditions remains a major challenge for predictive maintenance. This article proposes an adaptive multiview scalogram fusion framework based on convolutional neural networks (CNNs) with a group-level squeeze-and-excitation gating mechanism, which dynamically reweights multiple signal representations on a per-sample basis. Vibration signals are processed through envelope analysis, kurtogram-guided filtering, and the continuous wavelet transform to generate four complementary time–frequency views: raw, filtered, envelope, and filtered-envelope scalograms. The proposed method allows the network to emphasize informative representations while selectively suppressing noise-dominated views. Experiments on the machine failure prevention technology (MFPT) dataset demonstrate that the proposed approach achieves superior robustness relative to conventional CNNs, ResNet-18, and CNN–SVM baselines, particularly under low signal-to-noise ratios. Furthermore, validation on real industrial machine data confirms the ability of the proposed framework to produce consistent diagnostic decisions, effectively distinguishing healthy, degraded, and faulty operating states. These results indicate that adaptive multiview fusion provides a robust and scalable solution for vibration-based fault diagnosis in realistic industrial environments.