Multi-task deep learning meets hyperspectral imaging: A unified modeling framework for WEEE plastic identification and flame retardant quantification

Rapid plastic identification and additive/contaminant quantitation are critical for efficient recycling of waste electrical and electronic equipment (WEEE). Conventional single-task approaches treat these objectives in isolation, limiting their performance when analyzing high-dimensional spectral data. This study proposes a framework that integrates a multitask convolutional neural network, shared and task-specific learning, and cross-task fusion and attention to simultaneously classify polymer substrates, identify flame retardant types, and quantify flame retardant loading based on hyperspectral imaging data. The dataset comprised laboratory-prepared samples of three common WEEE plastics (acrylonitrile butadiene styrene, high-impact polystyrene, and polypropylene) compounded with four representative flame retardants (ammonium polyphosphate, aluminum trihydrate, decabromodiphenyl ether, and decabromodiphenyl ethane) across graded loadings (1–30 %). The proposed model outperformed single-task baseline models on near-infrared hyperspectral imaging data, improving the classification accuracy of polymer substrates and flame retardant type by 7–10 % and increasing the R 2 score for flame retardant loading regression by 0.04–0.071. Comparable performance gains were achieved using the proposed model on mid-wave infrared data. The findings of the study provide a solid theoretical foundation and technical support for intelligent recognition and efficient sorting of WEEE plastics. Notably, the framework may be adopted in other two-dimensional spectral sensing applications, including solid waste sorting, pollutant monitoring, and broader intelligent environmental sensing.