To address the issues of poor robustness and weak generalisation in deep subspace network-based micro-expression recognition, this paper proposes a novel method that integrates nonlinear deep subspace learning with optical flow computation. The method employs kernel transformation to comprehensively extract emotional features from micro-expressions while simultaneously utilizing optical flow characteristcs to capture subtle motion dynamics, thereby enhancing recognition robustness. Experimental validation is performed on 4 widely adopted spontaneous micro-expression datasets (SMIC, SAMM, CASME and CASME Ⅱ) as well as a composite dataset 3DB-combined samples. Results demonstrate that the proposed method outperforms existing deep learning algorithms, including MACNN and Micro-Attention, achieving a recognition accuracy of 0.834 6 on the composite dataset. Furthermore, after adding 10%, 20%, 30%, and 40% random noise blocks to the SMIC dataset, the method consistently maintains superior unweighted F₁ scores compared to other algorithms. These findings substantiate its effectiveness and robustness in real-world micro-expression recognition scenarios.