Predicting Dye Formulations and Ultrasonic Pretreatment Power in Wood Staining with a SAT Fusion-Gated BiLSTM Model Optimized by Improved Dung Beetle Algorithm

Wood dyeing plays a crucial role in enhancing the value of plantation wood and addressing the imbalance between supply and demand in the wood industry. However, challenges such as low dye uptake and inaccurate color matching persist. This study introduces a novel Separable Attention Feature Fusion-Gated Bidirectional Long Short-Term Memory (SAT Fusion-Gated BiLSTM) model to more accurately predict dye concentration and ultrasonic pretreatment power to address these issues. The model integrates multispectral data and uses ultrasonic treatment to open fiber gaps and decompose wood components, thereby improving dye uptake. Unlike traditional models, SAT Fusion-Gated BiLSTM combines a Maximum Information Coefficient (MIC)-based feature selection module, gated linear units for enhanced feature extraction, and an improved cross-attention mechanism for efficient data fusion. Compared to traditional deep learning models, the model’s performance is significantly improved, including an R2 of 0.9274, an RMSE of 0.0470, an RPIQ of 5.5825, an RPD of 5.0400, and a MAE of 0.0229. RPIQ and RPD improve by over 100.8% compared to traditional deep learning models, and the runtime is reduced by 83.94% when compared to Convolutional Neural Networks (CNNS). The Shapley Additive Explanation (SHAP) reveals the key spectral features that drive these predictions. These advancements in accuracy, efficiency, and interpretability make the model particularly promising for industrial applications, offering both improved operational efficiency and more precise resource management in the sustainable use of timber.