融合多尺度空洞卷积与Kansformer的跨域睡眠分期研究OA

To address the issue of insufficient generalization performance in cross-dataset applications of existing sleep staging models,this paper proposes a hybrid encoding architecture that combines multi-scale dilated convolutions with Kansformer.First,a dual-branch multi-scale feature extraction structure that combines 1D convolution and dilated convo-lution is designed,effectively enhancing the representation capability of multi-scale local features.Second,this paper con-structs a Transformer encoder based on a fusion attention mechanism and the KAN network,which realizes global depen-dency modeling in temporal and channel dimensions,and improves the representation of complex nonlinear features.Finally,this paper adopts a training strategy that combines signal reconstruction and multi-level feature alignment to synchronously achieve distribution alignment across datasets during model optimization,thereby constructing a feature representation space with cross-domain invariance.Rotational target domain validation experiments are conducted on five public datasets with diverse population characteristics.The results show that the model achieves an average accuracy of 75.31%and an average macro F1 score of 70.52%across the five cross-domain experiments.Compared with the existing SleepDG model,the average accuracy and macro F1 score improve by 0.28 and 0.88 percentage points,respectively.