Establishment of machine learning-based predictive models for food spoilage risk in chilled meat under aerobic and vacuum storage, and for flavor classification of dry-aged beef
學生姓名:
黃鉉喅
指導教授:
凌明沛
學 期:
113下
摘 要:
Spoilage bacteria are the main cause of food spoilage, which is a common issue in daily life. Fresh meat is highly perishable and typically stored under refrigeration using either aerobic or vacuum packaging. This includes aged beef, which has gained attention for its quality and flavor. This study aims to develop traditional and machine learning (ML) models to predict bacterial concentration under various storage conditions and integrate them into a quantitative microbial spoilage risk assessment (QMSRA) framework. A flavor classification model for aged beef is also built to predict flavor and umami scores under different aging conditions. Data on chilled meat stored under aerobic and vacuum conditions were collected from literature, focusing on dominant spoilage bacteria—Pseudomonas spp. for aerobic and lactic acid bacteria for vacuum storage. Flavor data for aged beef were also compiled. After preprocessing, bacterial growth models were developed using traditional equations and ML algorithms. Data were split into 80% training and 20% validation sets, and evaluated using 10-fold cross validation and performance metrics. Spoilage risk models were constructed by linking predicted bacterial concentrations with spoilage rejection probabilities derived from literature. ML models showed the best performance during training, with Gaussian Process Regression (GPR) and Random Forest Regression (RFR). However, in validation, traditional models performed, achieving up to 91.43% accuracy, likely due to limited data and high experimental variability, which may have led to ML overfitting or poor generalization. Due to limited data, the flavor classification model remained basic. This study using the best-performing model for vacuum storage, a Monte Carlo simulation (10000 trials) predicted spoilage progression—showing spoilage beginning on day 6 and affecting about half the products by day 8, indicating the need for storage strategy adjustments. Future work will focus on expanding the dataset to improve ML performance and refine both the spoilage risk prediction and aged beef flavor classification.
黃鉉喅.pdf