With the increasing severity of food safety problems worldwide, rapid prediction of potential toxins in foods has become critical. Traditional prediction methods, such as chemical analysis and bioassay, can provide accurate results, but they are time-consuming, costly and complicated to operate, making it difficult to meet the demand for large-scale screening. In recent years, machine learning technology, with its powerful data processing capability and pattern recognition advantages, has shown a broad application prospect in the field of food biotoxin prediction. This paper first discusses the importance of biotoxin prediction in the field of food safety. Then, the basic theory, key algorithms and models of machine learning are introduced in detail, its application in biotoxin prediction is discussed, and the practical effects of different algorithms and models are analyzed. To address the problems of machine learning in biotoxin prediction, model optimization and improvement strategies are discussed, including feature selection, hyperparameter tuning, and integrated learning. The potential challenges facing the application of machine learning in this field, such as data availability, model generalization ability, and the complexity of cross-disciplinary cooperation, are pointed out, and potential future research directions are also proposed. In the future, with the continuous progress of machine learning and the gradual expansion of food biotoxin datasets, it is expected that its application in the field of food biotoxin prediction will be further developed to provide strong support for environmental protection and human health.

Liquid smoke processed by thermal pyrolysis at 450 ℃ under nitrogen protection was prepared as a flavoring agent from peach wood. Its effect on the flavor characteristics and consumer acceptability of sirloin steak was investigated. Various liquid smoke concentrations and marination times were applied to the smoking process. Sensory descriptors were summarized by focus group tests and illustrated by a word cloud method. Flavor evaluation was carried out using check-all-that-apply questions and principal component analysis (PCA). Results showed that the liquid smoke significantly improved the sensory descriptors of steak, including mild smokey, barbequed, cooked beef-like and meaty flavors, dark color, chewy and juicy texture. Correlation analysis showed that both the overall preference and odor preference were positively correlated with the flavor preference. Chemical analysis showed that the liquid smoke at low concentrations used in this study did not have significant effect on the moisture content, fat content or pH of steak.

The application of deep learning in food safety detection and risk early warning is becoming more and more extensive, thus providing new opportunities for improving food safety, quality control and authenticity identification. This paper first introduces the basic concept of deep learning and its current development in the field of food safety, and discusses the application of convolutional neural network (CNN), recursive neural network (RNN), transformer architecture and graph neural network (GNN) in food safety detection and risk prediction. Although deep learning performs well in improving the efficiency and accuracy of food safety detection, its practical application still faces challenges such as poor data quality, weak privacy protection capacity and lack of model interpretability. Next, this paper analyzes potential risks that could be brought about by these problems and proposes possible solutions such as promoting data standardization, strengthening privacy protection, and promoting the formulation of policies regarding artificial intelligence. In the future, the combination of deep learning with the Internet of Things (IoT) and blockchain technology and further development of generative artificial intelligence will promote the digital transformation of the food industry and enable the whole-process traceability of food safety monitoring.

To address the problems of the traditional international standard methods for milk powder quality detection such as subjectivity and lag, this study proposed a rapid method for the detection of the dispersibility and bulk density of milk powder based on residual network (ResNet). The dataset used in this study included 499 particle distribution images taken for 10 groups of instant whole milk powder samples under a 10 × optical microscope. Initially, these sample groups were tested for dispersibility and bulk density using the international standard methods, and classified into different levels of dispersibility and bulk density based on the test results. Subsequently, these microscopic images were used to train the ResNet to facilitate effective classification of different samples. Ultimately, the classification results were used to predict the dispersibility, loose density, and tapped density of instant whole milk powder. Additionally, this study compared the predictive performance of different deep learning models, including ResNet, EfficientNetV2, and Swin Transformer. The results indicated that the deep learning model based on ResNet 152 exhibited the best performance in predicting the dispersibility, loose density, and tapped density of instant whole milk powder, with accuracy rates of 97.50%, 98.75%, and 95.00%, respectively for the test set. The exceptional performance of these deep learning models in milk powder quality detection not only proves that this method can predict the dispersibility and bulk density of milk powder in real time and accurately, but also provides a new technological approach for online quality detection of milk powder.