Research on intelligent diagnosis of deep-sea submersible bearing failures based on dual adaptive modeling

The deep-sea environment is extremely complex, and rolling bearings in submersible propulsion systems are the core moving components. Facing complex conditions and multiple faults, traditional bearing fault diagnosis methods lack adaptability—failing to dynamically adjust feature weights with changing conditions—resulting in poor model generalization, low accuracy, and weak robustness. To address this, this study fuses the Convolutional Kolmogorov–Arnold Network (CKAN) and Squeeze-and-Excitation Network (SENet) to build an end-to-end intelligent diagnosis model. The model first uses CKAN’s learnable convolutional kernels and activation functions to automatically learn optimal feature expression from data; then, via SENet’s attention mechanism, dynamically evaluates and assigns feature importance weights for each input sample to adapt to condition changes; and, finally, synergistically optimizes all links through backpropagation to achieve comprehensive adaptability and a feedback loop. Experimental analysis verified its effectiveness. Results show that with “feature form adaptation” and “feature importance adaptation,” it outperforms traditional methods in accuracy and robustness in complex variable-condition environments, promising in intelligent fault diagnosis.