An End-to-End Diagnostic Framework: Multichannel Vibration SPP Imaging CRHA-Net for Bearing Fault Recognition

As industrial equipment evolves toward high precision and reliability, its fault diagnosis becomes increasingly important. Traditional vibration signal-based methods remain limited in both extracting weak impact features and recognizing complex operating conditions. More efficient signal enhancement and intelligent feature learning strategies are urgently needed. To this end, this paper proposes an end-to-end diagnostic framework which improves Symmetrized Dot Pattern (SDP)—Symmetry-Preserving Projection (SPP) and Cross-attention fused Residual Hierarchical Adaptive-attention Network (CRHA-Net). Specifically, the three-channel vibration signals are first adaptively weighted and fused based on information entropy and kurtosis to fully integrate their information for highlighting these components related closely to bearing faults. Variational mode decomposition (VMD) is then used to obtain physically meaningful intrinsic mode functions (IMFs). This is combined with a fixed lag and angle magnification mechanism to generate a structurally stable "six-petal" SPP image. Moreover, the CRHA-Net consists of a CNN branch with Squeeze-and-Excitation (SE) channel attention and a hierarchical residual encoder (RHA) with head-level adaptive attention. Residual bridging injects mid-layer priors, and cross-attention fusion enables adaptive aggregation across branches and scales. Comparative experimental results demonstrate that CRHA-Net achieves higher recognition accuracy (99.81%) than representative CNN models, visual Transformer models, and CNN-Transformer hybrid architectures.