Abstract: In recent years, food supply chain security problems have occurred frequently. In order to reduce the threat of foodborne risks, risk early warning is becoming a powerful guarantee to strengthen the food safety system. However, in the face of multi-source heterogeneous unstructured data on foods, the existing risk warning method is limited by its low early warning accuracy and high labor cost. Based on analysis of a large number of data from food sample survey and related information from 26 provinces in China, this paper establishes a risk classification and early warning model using deep belief network-multiclass fuzzy support vector machine (DBN-MFSVM). First, the numerous data from grain supply chain sample survey are embedded, coded and normalized to obtain structured food data, which are then input into the DBN model to extract high-dimensional features, and self-adaptively mine the risk change and intrinsic correlation probability among the risk factors in the supply chain. Finally, the high-dimensional features are input into the optimized MFSVM for training to realize the risk classification and early warning of the main hazards in the supply chain. The comparative experimental results show that the DBN-MFSVM model has better robustness and generalizability for grain sample survey data. Its accuracy rate is 98.44%, and the running time is 85 s. It can quickly identify the risk level and priority of hazardous materials in the food supply chain, and thereby provide a scientific basis for the regulatory authorities to develop targeted sampling strategies, establish priority regulatory areas and allocate risk monitoring resources.

Key words: food supply chain security; risk early warning; deep belief network; multiclass fuzzy support vector machine