A multi-task mixture-of-experts CNN based on time-resolved LIBS for qualitative and quantitative collaborative analysis of milk powder adulteration.

To achieve high-performance qualitative and quantitative joint analysis of milk powder adulteration, a Multi-Task Mixture-of-Experts Convolutional Neural Network (MTMoE-CNN) based on time-resolved laser-induced breakdown spectroscopy (LIBS) was proposed. In this method, through the integration of a temporal convolutional network and an adaptive Inception module, both local spectral details and global temporal patterns in the time-resolved spectra were fully captured. Furthermore, by employing a multi-task learning mechanism, this method jointly modeled classification and regression tasks, enabling effective feature sharing between them. Meanwhile, a mixture-of-experts architecture was integrated into the quantitative branch, where expert sub-models handle specific subtasks and classification outcomes guide regression process. Compared to traditional methods, MTMoE-CNN achieved superior performance, with a classification accuracy of 0.950 and regression metrics of R2 = 0.864, RMSE = 2.897, and MAE = 1.766 on the test set. Finally, the ablation study, transferability evaluation, and interpretability analyses further demonstrated the robustness, scalability, and transparency of this model. Therefore, these results highlight the strong potential of MTMoE-CNN for real-world implementation in rapid food safety monitoring.