Knowledge tracing (KT) is essential in intelligent education for predicting students’ future performance based on their historical interactions. However, current deep learning-based KT models face two major challenges: inadequate consideration of global and local information in interaction records, limiting the ability to capture both long-term trends and short-term changes; and insufficient utilization of latent features, resulting in incomplete feature extraction and limited personalization. To address these issues, this paper proposes the Multi-feature Global-Local Information Fusion Knowledge Tracing model (MFGL-KT). This model integrates explicit and latent features while capturing global and local information in learning sequences. The model uses multi-head attention with relative position encoding for global dependencies and depthwise separable convolution for local information. To fuse global and local information, this paper introduces an online knowledge distillation mechanism. The two mechanisms in the dual-scale interaction sequence modeling module are treated as two student models, which are adaptively integrated through online knowledge distillation to form a more robust teacher model. Comparative experiments with other fusion strategies validate the effectiveness of this method. Additionally, the teacher model provides extra guidance to student models during training. Experimental results on four public datasets show that MFGL-KT consistently outperforms other state-of-the-art models.