Vocational education assessments are vital for maintaining vocational education quality, but manual analysis is slow and error-prone. Leveraging deep learning (DL) enables proactive monitoring, faster feedback loops, and data-driven quality control in vocational education institutions, improving both instructional effectiveness and student learning experiences. This research introduces a DL-based system to automate vocational education evaluations and issue early warnings by analyzing multimodal instructional and student feedback data. A hybrid Frog Leaping Optimization-driven BERT-Attention-tuned Bi-LSTM (FLO-BERT-Attention Bi-LSTM) method is developed. Bidirectional Encoder Representations from Transformers (BERT) extracts semantic features from feedback, Attention enhances relevant context, Bidirectional Long Short-Term Memory Network (Bi-LSTM) captures sequential dependencies, and Frog Leaping optimizes (FLO) model parameters, resulting in accurate, real-time classification of instructional quality risks. Teaching quality and student engagement dataset includes feedback scores, attendance, assignment submission rates, quiz, and exam averages, and LMS activity. The collected data undergoes rigorous cleaning, min-max normalization, and tokenization, and feature extraction is conducted using Term Frequency-Inverse Document Frequency (TF–IDF). Python-based experiments show significant accuracy improvements using the proposed hybrid model. The proposed FLO-BERT-Attention Bi-LSTM achieved high performance accuracy (0.9524), precision (0.938), recall (0.94), and F1-score (0.9041), outperforming baseline methods. The proposed FLO-BERT-Attention Bi-LSTM method effectively automates quality assessment, detects instructional risks early, and supports evidence-based interventions. Its integration of DL and optimization techniques demonstrates strong potential for scalable vocational education analytics and enhances the responsiveness of academic quality assurance processes.