Abstract
This study addresses persistent challenges in traditional talent cultivation models, including misalignment with industry demands, outdated instructional content, and limited depth in school-enterprise collaboration. To overcome these issues, the study proposes an innovative approach by integrating Back Propagation Neural Network (BPNN) technology from the field of artificial intelligence to construct a multimodal, integrated teaching quality evaluation system. By aggregating diverse data sources-including classroom performance, online learning behaviors, and hands-on practice videos-the study builds a comprehensive digital representation of the teaching process. An enhanced BPNN model is introduced, optimized by determining the ideal number of hidden nodes to improve evaluation performance. The method employs a hybrid fusion strategy, in which features from different data modalities are extracted independently, processed through dedicated neural networks, and fused at the decision level. This design enhances the model's adaptability to complex instructional environments and improves predictive accuracy. Empirical validation uses 1,500 samples collected from 18 public courses offered by Nanjing Institute of Industry Technology. The experimental results show that the improved BPNN model significantly outperforms baseline models in both prediction accuracy and convergence speed. Specifically, it achieves a prediction accuracy of 92% for teaching quality rated as "excellent" and 90% for those rated as "very good." This study provides a novel and effective solution for evaluating teaching quality in vocational education, fostering deeper integration between industry and education, advancing talent development, and supporting broader goals of industrial transformation and upgrading.