Abstract: The knowledge tracing task in smart education aims to predict students' future answering performance based on students' historical answering behavior. Existing knowledge tracing models ignore the influence of short-term learning state fluctuations on answering behavior, resulting in bias in estimating students' knowledge status. In view of the above situation, a personalized knowledge tracing model (PKT-ISALS) that integrates students' ability and lear-ning status is proposed: Firstly, the short-term learning state is extracted based on the interactive sequence of students' historical answers, and the phased fluctuations of students in the learning process are described, so as to realize the modeling of students' long-term learning state; Secondly, the purpose is to estimate the possibility of short-term state fluctuations in wrong answers by comprehensively considering students' ability, learning status and question characteristics. Finally, the update amplitude of knowledge state is adjusted in incremental form to make the know- ledge state update more in line with the cumulative characteristics of the real learning process. On the four public datasets of ASSIST09, ASSIST12, ASSISTChall and algebra05, the PKT-ISALS model was compared with six mainstream knowledge tracing models such as DKT, DKVMN, SAKT, AKT, DTransformer and FKT, and the ablation experiment was further carried out to verify the effectiveness of each key module. The comparative experimental results show that the PKT-ISALS model achieves better prediction performance than the baseline model on different datasets. On the ASSIST12, ASSISTChall, and algebra05 datasets, compared with the optimal baseline model FKT, AUC of PKT-ISALS model increases by approximately 1.0%, 3.8%, and 0.6%, respectively. ACC of PKT-ISALS model increases by approximately 0.7%, 2.6%, and 0.2%, while RMSE of PKT-ISALS model decreases by approximately 1.1%, 2.6%, and 1.0%, respectively. On the ASSIST09 dataset, compared with the optimal baseline model DTransformer, AUC and ACC of PKT-ISALS model increase by approximately 2.1% and 1.2%, respectively, while RMSE of PKT-ISALS model decreases by approximately 3.0%. In addition, the ablation experiment results further show that long-term ability modeling, short-term learning state modeling and error cause discrimination mechanism can effectively improve model performance, among which the short-term learning state module plays a more significant role in long-sequence learning behavior modeling. In general, PKT-ISALS can describe the changes in students' knowledge state in the learning process at a more granular level, and improve the prediction accuracy of the model.