To address the issues of poor robustness and weak generalisation capabilities in deep subspace network-based micro-expression recognition, this paper proposes a novel micro-expression recognition method based on non-linear deep subspace learning and optical flow computation. The method introduces kernel transformation to comprehensively extract emotional information from micro-expressions, whilst simultaneously employing optical flow features to capture the motion dynamics of micro-expressions, thereby enhancing recognition robustness. The experiments were conducted on four widely used spontaneous micro-expression datasets, i.e., SMIC, SAMM, CASME and CASME Ⅱ, as well as the 3DB-combined composite dataset. The results indicate that the proposed method outperforms deep learning approaches such as MACNN and Micro-Attention in terms of recognition performance, achieving an accuracy of 0.8346 on the composite dataset. Furthermore, after adding random noise blocks at levels of 10%, 20%, 30%, and 40% to the SMIC, the unweighted F1 scores at various noise levels consistently surpassed those of other algorithms, thereby validating the effectiveness and robustness of the proposed method in the micro-expression recognition task.