A Novel Dual-Channel Feature Fusion Network for Rolling Bearing Fault Diagnosis Under Variable-Speed Conditions Enhanced by Improved Polar Lights Optimizer-Guided VMD

In industrial fault diagnosis, conventional methods are often hindered by the inadequate extraction of time-domain characteristics and inefficient cross-modal interactions, further exacerbated by the high costs of parameter tuning. Addressing these limitations, a multi-modal diagnostic framework that integrates parameter optimization and deep learning is proposed. First, we propose an Improved Polar Lights Optimizer (IPLO). By fusing chaotic initialization with a dynamic differential evolution strategy, IPLO adaptively locks onto the optimal Variational Mode Decomposition (VMD) parameters, guided by a multi-objective fitness function balancing signal energy, kurtosis, and spectral entropy. Second, distinct from traditional time-frequency analysis prone to frequency drifts, the framework introduces a structure-centric Relative Angle Matrix (RAM) to map one-dimensional signals into topologically stable 2D representations, which are then processed by a Swin Transformer backbone. Simultaneously, a parallel channel leverages a CNN-Transformer architecture to capture temporal dynamics from the optimized Intrinsic Mode Functions (IMFs), enhanced by an adaptive attention mechanism. The proposed method employs a bidirectional cross-attention strategy to bridge the semantic gap between temporal and spatial modalities. Extensive experiments on a rotor test rig and the XJTU-SY dataset demonstrate that the proposed method achieves an average accuracy of 98.9% and exhibits superior generalization under complex, non-stationary regimes compared to state-of-the-art baselines.