基于监督抽检数据的肉类食品安全风险分析及预测

为通过数据挖掘预测食品安全风险、隐患和趋势,进行预警和快速反应,基于2015—2017年国家肉类食品监督抽检的18 378 批次样品数据,分析我国肉与肉制品主要安全现状与风险种类,并基于检测指标及属性运用BP（back propagation）神经网络方法构建以抽样省份、产品类型、产地、生产日期、年份、是否大型企业6大属性指标为输入层、包含2 个隐藏层、以是否合格为输出层的肉类食品安全神经网络预测模型。结果表明：经数据准备、模型生成、数据训练和验证及参数优化,得到的3 层BP神经网络预警模型总体百分比矫正为96.2%;对于合格样本,判定正确的概率为96.5%,错判概率为3.5%,预测模型具有较好的参考和应用价值。基于BP神经网络的食品安全预警方法能够对输入样本进行有效预测,为食品安全风险研判和风险预警提供技术支撑。     

A model for reduced fat food product development success

The majority of new food products fail following launch. The aim of this research has therefore been to model the food product development process and to determine the relative importance of factors comprising the model in terms of product outcome. Data was gathered by qualitative interview. Food product developers (n=47) from food companies based in England and Northern Ireland (n=27) were required to describe their experiences and provide insight into problems encountered in the invention, development, launch and marketing of new standard and reduced fat food products. Interviews were recorded, transcribed and content analysed thematically. Data from 127 product case histories were binary coded, screened and modelled through a series of logistic regression analyses against product outcome (success/failure) as defined by the company concerned. The resulting model implies that actions taken during the concept phase of the product development process and the nature of expertise employed during the process are particularly important for product success. It is recommended that appropriate expertise is enlisted at product conception and that a food technologist is employed to ensure that the texture of the product is acceptable. Whilst it is important to liaise with retailers throughout the process, manufacturers should formulate their own product recipes and avoid following trends in food and ‘me too’ product development.     

长短期记忆神经网络在肉制品中铅含量风险预警的应用

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。     

Artificial neural network approach to simultaneously predict shelf life of two varieties of packaged rice snacks

Actual storage shelf life test by storing a packaged product under typical storage conditions is costly and time consuming. A new approach using an artificial neural network (ANN) algorithm for shelf life prediction of two varieties of moisture-sensitive rice snacks packaged in polyethylene and polypropylene bags and stored at various storage conditions was established. The ANN used to predict the shelf life was based on multilayer perceptron with back propagation algorithm. The ANN algorithm employed the data of product characteristics, package properties and storage conditions. The neural network comprised an input, one hidden and one output layers. The network was trained using Bayesian regularisation. The performance of ANN was measured using regression coefficient ( R ² = 0.23–0.28) and root mean square error (RMSE = 0.96–0.99). The ANN-predicted shelf lives agreed very well with actual shelf life data. ANN could be used as an alternative method for shelf life prediction of moisture-sensitive food products as well as product/package optimisation.     

Prediction of freezing time for food products using a neural network

An artificial neural network (ANN) was developed to predict the freezing time of food products of any shape. The Pham model was used to generate freezing time data and to train ANN based on Wardnets. The product thickness (a), width (b), length (c), convective heat transfer coefficient (h_c), thermal conductivity of frozen product (k), product density (ρ), specific heat of unfrozen product (C_pu), moisture content of the product (m), initial product temperature (Ti), and ambient temperature (T_∞) were taken as input variables of the ANN to predict freezing time. The effects of the number of hidden layer nodes, learning rate, momentum on prediction accuracy were analyzed. The performance of the ANN was checked using experimental data. Predicted freezing time using the ANN was proved a simple, convenient and accurate method. Selection of hidden nodes, learning rate and momentum were important to ANN predictions.     

基于电子鼻的鸡蛋新鲜度检测

鸡蛋新鲜度检测十分重要,为实现无损检测鸡蛋新鲜度,该文利用电子鼻技术,通过挥发物的检测来尝试对20℃、70%RH贮藏条件下的鸡蛋新鲜度进行预测。并测量鸡蛋的理化指标(哈夫单位和蛋黄指数)作为新鲜度的衡量标准。通过线性判别分析对储藏不同天数的鸡蛋进行分类分析,发现线性判别分析能较好地区分不同储藏天数的鸡蛋,判别函数的总贡献率为75.70%;利用多元线性回归和BP神经网络分析法建立电子鼻响应信号和鸡蛋理化指标之间的关系模型,所建多元线性回归模型的相关系数达0.84以上,相对误差在8.00%左右;所建BP神经网络模型的相关系数达0.84以上,相对误差在9.00%左右。说明电子鼻技术对鸡蛋新鲜度具有一定的预测能力,该研究可为鸡蛋新鲜度的无损检测提供参考。     

山药片阶段降湿促干特性及多物理场耦合模型

目的:探明热风干燥过程中阶段降湿调控策略对农产品物料干燥特性的影响。方法:研究了60℃干燥条件下,恒湿干燥(相对湿度15%,25%,35%,45%)和阶段降湿干燥(第一阶段相对湿度45%保持10,20,30,60 min,第二阶段相对湿度15%)对山药片的干燥特性的影响,构建热质传递多物理场耦合模型进行验证,并测定干燥后产品的复水比和微观结构变化。结果:①恒湿干燥时,干燥速率随相对湿度的增加而降低;但阶段降湿干燥(50%相对湿度保持15 min而后降为20%)的干燥时间比恒湿干燥(相对湿度20%)的缩短了14.3%。②山药片干燥速率总体呈先升速后降速的趋势,水分有效扩散系数D_eff变化范围为7.16×10^-10~2.34×10^-9 m²/s;多物理场耦合模型拟合结果表明,相对湿度越高,物料升温速率越快,内部水分向外扩散迁移速度也越快。③复水比随着相对湿度的升高呈先升高后降低趋势;阶段降湿(相对湿度50%保持15 min而后降为20%)干燥时,山药片内部呈蜂窝状的多孔结构,有利于内部水分向外迁移,此条件下复水比最高为6.85±0.05;而恒湿干燥(相对湿度为20%)时,山药片内部水分扩散通道发生收缩坍塌、结壳堵塞,复水率较低。结论:热风干燥时,阶段降湿调控策略可显著缩短干燥时间,并有效改善物料微观结构,提高产品复水率;构建的多物理场耦合模型能够准确模拟阶段降湿干燥过程中山药片内部的热质传递过程。     

Evaluation of additional cooking procedures to achieve lethality microbiological performance standards for large, intact meat products

The U.S. Department of Agriculture Food Safety and Inspection Service (USDA-FSIS) has a specific lethality performance standard for ready-to-eat products. To assist meat processing establishments in meeting the performance standard, USDA-FSIS developed Appendix A, which provides guidelines for cooking temperatures, times, and relative humidity. This project determined whether the USDA-FSIS performance standards for lethality were met when using parameters other than those identified in Appendix A to cook large hams and beef inside rounds. The effects of alternative lethality parameters on the reduction of Salmonella Typhimurium and coliforms and on the toxin production of Staphylococcus aureus were evaluated. Large (9- to 12-kg) cured bone-in hams (n = 80) and large (8- to 13-kg) uncured beef inside rounds (n = 80) were used in this study. The products were subjected to 1 of 10 treatments defined by combinations of final internal product temperatures (48.9, 54.4, 60.0, 65.6, or 71.1°C) and batch oven relative humidities (50 or 90 % ). For all treatments, at least a 6.5-log reduction in Salmonella Typhimurium was achieved. The coliform counts were also substantially reduced for both hams and rounds. Across all treatments for both products, S. aureus toxin production was not detected. The relative humidity did not alter the lethality effectiveness for any of the treatments. The final internal temperatures and relative humidity combinations used in this project achieved the lethality performance standard established by USDA-FSIS for fully cooked, ready-to-eat products.     

Artificial intelligence-based identification of butter variations as a model study for detecting food adulteration

The demand for high-quality food products is increasing globally at unprecedented rates in response to growing health concerns and consumer awareness about healthy food options. Yet, the tools for determining food quality remain restricted to well-equipped laboratories, not readily accessible to consumers. Unfortunately, the current inspection mechanisms are limited and cannot keep track of all the products continuously, which exposes weakness in the system towards adulteration, falsification, and mislabeling products. Consumers rely on manufacturer labeling alone, with no convenient and user-friendly tool to confirm quality, especially for organic products. The advancement of Artificial Intelligence (AI) provides an opportunity for these tools to be developed. In this study, we demonstrate that simple sound vibrations traversing the food products can be used in conjunction with deep learning models to verify high quality products with no additives, as well as organic food products. Our neural network models, namely Parallel CNN-RNN and CRNN, achieve high accuracy on the defined classification tasks. To our knowledge, this is the first report of an AI-based tool utilizing simple sound vibrations to identify adulteration in food products.     

COMPUTER SIMULATION OF STORAGE LIFE OF FOODS UNDERGOING SPOILAGE BY TWO INTERACTING MECHANISMS

A mathematical model was developed to describe the deterioration of potato chips, a model of dry food products in flexible packages. In this product, deterioration occurred by two mechanisms acting simultaneously with interaction between them. The mechanisms were: oxidation by atmospheric oxygen and textural changes due to water absorption. We found that the extent of oxidation and the equilibrium relative humidity increase due to water absorption could be predicted as a function of time. The technique can be applied to investigate the storage life of products for different package configurations, initial conditions and environmental conditions.     

ACOUSTIC WAVE ANALYSIS FOR FOOD CRISPNESS EVALUATION

Sound signals for five snack food products at two moisture levels were recorded digitally and product crispness was evaluated by a trained sensory panel. Sound signal features were extracted by analyzing signal-value and power-value dependencies. Principal component regression and neural network techniques were used to determine the usefulness of the sound signal features as predictors of sensory crispness. In a validation test, a trained neural network model predicted sensory crispness from sound signal features to an R²-value of 0.89. The results show the effectiveness of the techniques employed to extract and use sound signal features for crispness evaluation.     

Prediction of temperature and moisture content of frankfurters during thermal processing using neural network

An artificial neural network (ANN) was developed to predict temperature and moisture content of frankfurters during smokehouse cooking. Fat protein ratio (FP), initial moisture content, initial temperature, radius of frankfurter, ambient temperature, relative humidity and process time were input variables. Temperature at the frankfurter centre, average temperature of the frankfurter and average moisture content (d.b) of the frankfurter were outputs. Network training data were obtained from validated heat and mass transfer models simulating temperature and moisture profiles of a frankfurter. Backpropagation method was used for ANN training. Selection of hidden nodes, learning rate, momentum and range of input variables were important to ANN prediction. The FP was not an important factor in predictions.     

EFFECTS OF OVEN HUMIDITY ON FOODS BAKED IN GAS CONVECTION OVENS1

Bread, cakes, and cookies were baked in direct‐ and indirect‐gas fired, foodservice style convection ovens. Oven humidity was varied by including different numbers of perlite and water dummy loads during baking. Oven, product, and dummy temperatures were continuously recorded during baking, as was oven humidity. The properties of the finished baked foods were measured. Increased oven humidity resulted in products with higher yield, greater volume or spread, lighter color, and reduced firmness.     

腐乳白点控制生产工艺参数的研究

腐乳白点常见于成品腐乳,无食用安全问题,但严重影响产品外观和销售。该研究以白腐乳为研究材料,通过单因素试验及正交试验考察腐乳生产过程中发花温度、发花相对湿度、发花时间、摆胚间距、毛霉接种量对腐乳白点及氨基酸态氮含量的影响。结果表明,腐乳最佳生产工艺参数为：摆坯间距2.0 cm、毛霉接种量（显微镜单视野下孢子数）5个、发花温度28 ℃、发花相对湿度90%、发花时间38 h。在此优化工艺条件下,腐乳氨基酸态氮含量为1.09 g/100 g,未见腐乳白点产生。     

NaCl content and temperature effects on moisture diffusivity in the Gluteus medius muscle of pork ham

Dry-cured meat products are salted with different amounts of NaCl and then dried under different conditions (temperature and relative humidity) in order to stabilize the product. The effective moisture diffusivity values (D(e)) in Gluteus medius muscles, unsalted or soaked in different salting solutions (0.02, 0.05 and 0.08 kg NaCl/kg H(2)O), and dried at different temperatures (5, 13, 19 and 26?°C) and 80% of air relative humidity were determined. Water diffusion was only allowed parallel to meat fibre direction. The D(e) values were determined by applying Fick's second law. As temperature increased or NaCl content decreased, the D(e) increased. At low temperatures, the D(e) was less dependent on NaCl content than at higher ones. The energy of activation, obtained from the Arrhenius equation, is lower for unsalted compared with salted samples. Therefore, the effective moisture diffusivity of salted meat products is more sensitive to temperature variations than that of unsalted meat.     

Edible films and coatings to prevent the detrimental effect of oxygen on food quality: Possibilities and limitations

Oxygen has a deleterious effect on the quality of a wide variety of food products. The application of edible films and coatings to food products represents a new approach to solve this problem. Edible films and coatings can include antioxidant agents in their formulation and at the same time, they represent a barrier to oxygen, which results in a better preservation of quality. The water activity of the product, as well as the ambient relative humidity, determines the antioxidant effect of films and coatings. This paper reviews the latest studies dealing with the effectiveness and application of antioxidant films and coatings.     

THERMAL PROCESS CALCULATION USING ARTIFICIAL NEURAL NETWORKS AND OTHER TRADITIONAL METHODS

This work aimed to develop a method of thermal process calculation of canned food based on an artificial neural network and to compare it to the traditional Ball and Stumbo methods. A back‐propagation through time network was used as a model of heat conduction occurring in canned food, in order to predict the cold point temperature as a function of the autoclave and the initial product temperature. For training the network, a dataset of the cold point and retort temperature along time was obtained by processing a paste‐based product in autoclave. The network architecture (5‐8‐9‐1) was selected, presenting an excellent generalization capacity, with a mean relative error of 2.2% between the calculated F‐values of the observed and predicted time–temperature data. The Ball and Stumbo methods presented higher errors as the retort temperature was not constant. Neural network showed a great potential for online process calculation by having advantages such as precision and simplicity.     

Simultaneous Determination of Carmoisine and Ponceau 4R

A new, simple, sensitive, and inexpensive method for simultaneous determination of binary mixtures of food colorants, carmoisine and ponceau 4R, which are species with highly overlapped spectra, by using H-point standard addition method (HPSAM) is proposed. Absorbances at the two pair of wavelengths, 460 and 549 nm or 490 and 541 nm, were monitored while adding standard solutions of carmoisine or ponceau 4R, respectively. The accuracy and reproducibility of the determination method for various known amounts of carmoisine and ponceau 4R in their binary mixtures were evaluated. This method was satisfactorily applied for the determination of carmoisine and ponceau 4R in synthetic binary mixtures and several commercial products. The total relative standard error for applying the HPSAM to five synthetic samples was 3.26% or 4.96% with carmoisine or ponceau 4R addition, respectively.     

Numbers and strains of lactobacilli in some probiotic products

The numbers and types of Lactobacillus strains in probiotic feed or food products for sale in Europe were compared with the information stated on the product labels. The labels of all 10 products gave information on the genus and species of Lactobacillus, and seven gave information on the numbers of a specified strain(s). Lactobacilli were recovered using a recently validated method, developed as part of the European Community Project SMT4 CT98-2235 for the official control of probiotic microorganisms used as feed additives. Polymerase chain reaction (PCR)-based methods were used to assess the accuracy of labelling with regard to genus and species, and pulsed-field gel electrophoresis (PFGE) was used to identify strains. Five products were mislabelled with respect to the numbers and three with respect to species of lactobacilli. In four cases, the specified strains were not detected. Four fermented milks sold under three trademarks contained the same strain, which was named differently on each label. As safety and functionality of probiotics are strain dependent, these results demonstrated the need to control lactobacilli present in commercially probiotic human food products, not only at the species but also at the strain level, to ensure their quality and protect the consumer. The usefulness of the official methods developed for animal feed additives to identify lactobacilli in probiotic foods for humans was demonstrated.     

Delicious words – Assessing the impact of short storytelling messages on consumer preferences for variations of a new processed meat product

Continuous food product improvements require communication that induces consumers to test the new products. The paper presents an experiment that investigated the characteristics of consumers who reacted positively to variations of a new meat product, and explores the effect of short storytelling messages on choice. 1087 Danish consumers were presented with binary choice sets consisting of conventional and new variations of a cubed ham product; half of the respondents in addition saw short storytelling messages. It was found that while some consumers – especially those characterized by values related to stimulation seeking and hedonism and with a tendency towards variety seeking – reacted positively to the new product, others were less interested. The introduction of elements of storytelling, however, had a favorable effect on choice for the segment previously less interested. Thus, the results suggest that consumers who otherwise were not open for the new food product became more positive when storytelling was introduced.