Aiming at the problems of authorization conflict, condition overlap and priority anomaly when heterogeneous access control policies run across systems, this paper constructs a formal model of multi-modal rules, maps policy text, permission table, access log and environment context into rule feature vectors, and designs algorithms for semantic matching, constraint judgment and risk classification. On this basis, a rule conflict self-repair system is developed, which implements repair strategy generation, candidate version verification, exception rollback and log tracking. The experiment is carried out based on 12000 access control rules, 18 000 access logs and 2860 conflict rule pairs. The results show that the Precision, Recall and F1 of the complete method reach 95.4%, 93.1%and 94.2%respectively, which is 19.9 percentage points higher than that of the field matching method. After automatic repair, the residual conflict rate is reduced to 4.9%, and the average response time is 138.7 ms under 12000 rules and 1000 concurrent requests, which verifies the detection accuracy of the method and the stability of the system.
