Prediction error-guided gated fusion of shallow and deep spatiotemporal features for flotation working condition recognition

Accurate extraction of representative features is fundamental for flotation working condition recognition, as the visual state of froth directly reflects the mineral separation efficiency and process stability. However, a considerable proportion of existing approaches rely on single-frame static features, making it difficult to capture dynamic changes and condition transitions during the flotation process. Furthermore, utilizing the complementarity between shallow features and deep features can take into account both empirical domain knowledge and data-driven knowledge. Therefore, we propose a Prediction Error-guided Gated Network (PEGNet) for the fusion of shallow and deep spatiotemporal features in flotation working condition recognition. Specifically, shallow features, including static features (such as color, morphology, and texture) and dynamic features (such as velocity and collapse rate), are extracted from image sequences via feature engineering. Meanwhile, deep spatiotemporal features of the image sequences are obtained via a cascaded architecture consisting of a residual network and a Convolutional Gated Recurrent Unit (ConvGRU) network. Finally, a generator module is designed to predict the last frame of the image sequence, and the prediction error is employed to guide a gating mechanism that adaptively adjusts fusion weights. This design emphasizes shallow features when deep features become unreliable, enhancing recognition robustness under complex and dynamic conditions. Comparative and ablation studies on a real-world flotation dataset validate the effectiveness of the proposed method.