A TabM-CatBoost Model for Soy Sauce Fermentation Aroma Classification with Small-Sample Feature Fusion

Source abstract

Soy sauce aroma classification is challenging because labeled fermentation batches are scarce, sensor responses drift across tanks and production cycles, and volatile fingerprints contain many weak but complementary signals. This study developed a TabM-CatBoost model for classifying four aroma states: clean mellow, caramel roasted, ester fruity, and sulfur off-note. Electronic-nose descriptors, GC-MS volatile intensities, physicochemical indices, sensory attributes, and process variables were integrated for 164 fermentation batches, and label efficiency was evaluated under a five-shot protocol. TabM learned compact multi-predictor embeddings from standardized continuous variables, whereas CatBoost combined the embeddings with categorical information through ordered target statistics without validation leakage. The proposed model achieved 95.6% accuracy, 94.9% macro-F1, and 0.973 macro-AUC, while reducing the sulfur off-note false-negative rate from 10.7% for CatBoost to 3.6%. Removing GC-MS features reduced macro-F1 by 5.4 percentage points, and the model achieved 92.7% macro-F1 under 10% instrumental perturbation. Across the label regimes, macro-F1 increased from 89.8% under five-shot training to 94.9% with full labels, while expected calibration error decreased to 0.041. External-season evaluation on 31 later-cycle batches yielded 92.8% macro-F1, and retrospective thresholding retained 24 of 25 off-note alerts while reducing simulated expert review load by 46.2%. These results indicate that compact neural fusion followed by ordered boosting can support label-efficient aroma supervision in high-salt fermented-food production.