ScaleMixer: Selective excitation of feature-enhanced food time series prediction model

Food safety data analysis faces complex and variable characteristics, including cyclical fluctuations, trend changes, and seasonal features, which pose great challenges for accurate forecasting, Traditional time series analysis methods adopt a single perspective, which makes it difficult to fully capture the multidimensional information in the data. Food safety data exhibit different characteristics at different time scales:micro-scale shows high-frequency fluctuations,while macro-scale shows long-term trends. In this paper, we propose the ScaleMixer to realize more accurate food safety risk assessment by dynamically fusing the time series features of different sampling scales. The model contains two core modules: a multi-scale decomposition mixing module to handle historical data and a multi-predictor integration module for future prediction. Meanwhile, Selective excitation of feature-enhanced methods is introduced to adaptively adjust the feature channel weights based on the attention mechanism to enhance the importance of key information dimensions. We apply the model to camel milk safety monitoring dataset to analyze 26 key indicators including protein, fat content and acidity. The experimental results show that outperforms existing methods in both long-and short-term prediction tasks, improves the performance by 8.71% compared to the benchmark model, and significantly improves the accuracy and efficiency of food safety risk assessment.