本文基于2015~2018年国家肉制品监督抽检的5万批次检测数据，使用长短期记忆神经网络模型建立肉制品中铅含量的风险预警模型。参照国家食品检验标准并结合专家打分，构建肉制品中铅含量的6个食品安全风险等级；运用Softmax和汉宁窗对风险等级数据进行预处理，使用Tensorflow建立三层长短期记忆神经网络的时间序列风险预警模型，通过500轮模型训练实现对肉制品中铅含量的风险趋势预测。结果表明，长短期记忆神经网络对肉制品中铅含量的预测有较高的准确率，31个省份的平均误差为0.27，同实际检测风险基本匹配；模型稳定重现性较好，运行十次的平均误差为0.27。此模型可以实现对全国不同地域铅含量的趋势预警，为日常监督抽检和食品安全风险预警提供技术支撑。

In this paper, the detection data was derived from 50,000 batches of meat products sampling inspection by the China Food and Drug Administration, from 2015 to 2018. A Long Short-Term Memory (LSTM) neural network model for lead content in meat products was constructed to analyze the current status of meat product safety and the types of risks. Six levels of food safety risk in lead were established by combining the national food safety standards as a reference with experts scoring. Then the food safety risk data were processed by softmax and hanning window. The Tensorflow was applied to establish a three-layer LSTM neural network of time series risk early warning model through 500 rounds of model training. The results showed that the LTSM had a high accuracy for risk warning in lead content, the data from 31 provinces basically matched the actual risk value, with an average error of 0.27. The model was a stable and reproducible method for forecasting as well, with the average error of ten runs is 0.27, which could provide a new method to implement the trend early-warning in different regions of the country and provide a technical support for government and related authorities for daily supervision.

国家重点研发计划项目（2018YFC1603602）