In man-hour prediction, machining process texts with similar semantics but distinct processing logic make it difficult to effectively distinguish working hours using differentiating features. This leads to ineffective matching of man-hour features. A two-level classification method integrating multi-channel features and word-weighted clustering was proposed to address this problem. A hierarchical framework for process categories and features was constructed. At the process category level, based on BERT word embeddings, a “verb+noun” recombination channel was designed to enhance key information, and CNN, BiLSTM, and R-Transformer were fused to capture local, contextual, and global features. At the feature level, relying on the classification results at the process category level, a K-Means algorithm combined with a domain lexicon weighting mechanism was adopted to perform unsupervised clustering based on geometric features, thereby establishing a fine-grained classification system for machining process categories and geometric features. Experimental results demonstrate that the classification model for process categories outperforms the single-trained BERT model by 4.21, 4.49, and 4.40 percentage points in accuracy, recall, and F1-score, respectively. Compared with traditional K-Means, the feature-level clustering model improves the silhouette coefficient and CH score by factors of 1.02 and 4.86, respectively, and the mean relative error of final man-hour prediction decreases by 4.59 percentage points. Experimental results validate the effectiveness of the two-level framework in fine-grained classification and provide correctly matchable structured input for man-hour quota models.