Abstract: In classroom teaching scenarios, the sequence of students' learning behaviours during the duration of a knowledge point's presentation reflects their cognitive level regarding that content. To address issues of periodic data redundancy and long-sequence noise interference when classifying knowledge point cognitive levels based on classroom learning behaviour sequences, thereby enhancing the model's ability to represent student learning behaviour sequences and improve classification accuracy, we propose the Model-based Cognitive Level Classification Model for Knowledge Points (MCMPNet). This model employs MPTSNet as its backbone architecture, incorporating a ConvNeXt-CBAM module within its structure. By integrating convolutional structures with channel-time attention mechanisms, it enhances the model's ability to express key sequence features within periodic intervals. Concurrently, the introduction of the Mambaformer module improves the model's global modelling capabilities for long-sequence data and its ability to capture dynamic characteristics, thereby mitigating the impact of long-sequence noise. Experimental results demonstrate that the proposed model achieves superior recognition accuracy compared to baseline models on both public time-series classification datasets and the learning behaviour sequence-based knowledge point cognitive level classification dataset in classroom settings. This validates the model's effectiveness, offering an efficient approach for classroom learning behaviour analysis and intelligent educational applications.