<p>Tool wear prediction is a critical issue in intelligent manufacturing and condition monitoring, where the key challenge lies in accurately characterizing the temporal degradation process of cutting tools. Existing deep learning approaches mainly rely on end-to-end feature learning and regression, achieving certain improvements in prediction accuracy. However, they often overlook the intrinsic structural organization of deep features in the feature space during degradation, which makes the models sensitive to noise disturbances and operating condition variations, thereby limiting their generalization capability. To address these limitations, this paper proposes MDSC (Model with Degenerate Structural Constraints), a structure-constrained degradation modeling method based on an implicit degenerate manifold in the feature space. From a geometric perspective, MDSC introduces a parameterized variable to describe the relative positions of samples along an implicit degradation trajectory. Sequential consistency and approximate linear structure constraints are further imposed to guide the model toward learning structurally consistent and degradation-ordered feature representations. Without relying on explicit physical models, the proposed constraints embed prior degenerate structure information into the learning process in a data-driven manner, enhancing the model’s ability to represent degradation process. Experiments on a self-collected tool wear dataset, including ablation studies and comparisons with multiple baseline models, validate the effectiveness of MDSC. Moreover, robustness evaluations are conducted by introducing random noise perturbations with varying intensities during both training and testing. The results demonstrate that, compared with conventional 1DCNNs and several representative deep learning models, MDSC achieves superior prediction accuracy, improved feature structure stability, and stronger robustness under noisy conditions.</p>