Lightweight Correlation Inference and Prediction of Asphalt Mixture Image Processing Technology Based on Dual-Tower Qwen

To address the bottlenecks of low accuracy in cross-domain technological association recognition and difficulties in quantitative prediction of technological trends encountered when applying large language models to literature mining of asphalt mixture image processing technology, this paper proposes a multi-module collaborative dual-tower Qwen model (Qwen-STGF). A dataset is constructed from the academic literature in this interdisciplinary field. First, an improved dual-tower network is built by incorporating nonlinear activation functions and Dropout regularization, using the Qwen-1.8B-Chat large model as the shared semantic backbone to achieve unified mapping of the semantic features of keywords and titles. Second, learnable bias vectors are introduced to optimize the GraphSAGE network, thereby enhancing its capability for extracting spatial structural features. Meanwhile, a CNN-LSTM hybrid architecture is employed to construct a three-dimensional fused feature representation system encompassing semantic, spatial, and temporal dimensions, and a weighted random forest classifier with a class weight balancing strategy is adopted. Finally, a lightweight framework is developed for keyword co-occurrence prediction and quantitative technological-trend forecasting. The results demonstrate that compared with five types of traditional machine learning models, two types of classical graph neural network models, five native Chinese representation models, and four open-source large models of similar scale, the Qwen-STGF model achieves Precision, Recall, F1-score, AUC, and AUPR of 0.8673, 0.8162, 0.841, 0.8786, and 0.9322, respectively, among which AUC, AUPR, and F1-score rank as the best among all comparative models. This study reveals that keyword co-occurrence frequency can reflect the maturity of technology convergence to a certain extent, and the predicted trends correspond to the potential breakthrough prospects of relevant technologies. Three major future research directions are forecasted: establishing a unified mesoscopic feature parameter system, developing digital twin technology, and constructing reliable macro-mesoscopic performance correlation models. These directions provide technical support for asphalt mixture image processing.