Knowledge tracing (KT) is a method used to evaluate learners' learning states in personalized learning environments. However, traditional KT faces several challenges, including the cold-start problem, inadequate modeling of individual differences, and noise in real-world learning data. To address these issues, we propose a generalized and personalized enhanced deep knowledge tracing (GPE-DKT) model. First, we extracted key behavioral features from student interaction data using decision tree analysis and Pearson correlation. We then combined these features, which include both static and dynamic features, and fed them into the model. Furthermore, we compressed the input and label spaces via dimensionality reduction to improve behavior pattern identification. To reduce data noise, we established an anomaly rate threshold and introduced a “Mini-Test Model” to identify untrustworthy records. The training process was innovatively decomposed into two sequential stages: generalized modeling to mitigate missing knowledge points in learning records, and personalized modeling to capture learner-specific behavioral patterns. Experiments on real-world learning data demonstrate the effectiveness of GPE-DKT in evaluating learning states and forecasting student performance, especially in data-limited scenarios, showcasing higher stability and accuracy. We further validate the proposed model through a system deployed for learning evaluation.