To address the issues of poor adaptability to static parameters and high missed detection rates for low-density attacks in traditional information entropy methods for in-vehicle CAN network intrusion detection, this paper proposes a real-time detection framework based on simulated annealing optimization. The framework dynamically analyzes the distribution characteristics of CAN message IDs through windowing techniques, constructing multi-dimensional anomaly perception indicators by combining information entropy variance. A multi-objective optimization function is designed to adaptively search for optimal window sizes and sensitivity thresholds using simulated annealing algorithms. A two-stage detection architecture is introduced, achieving rapid response through coarse-grained entropy screening and fine-grained variance classification. Experimental validation on the Car-Hacking dataset demonstrates that the proposed method achieves detection accuracies of 95.13% for DoS attacks, with an average single-message detection latency of only 1.62 ms. Compared to traditional fixed-window information entropy methods, this approach improves detection precision by 8.79% and reduces false positive rates by 68.7%, while meeting real-time requirements for automotive edge devices. The research outcomes provide new optimization perspectives for efficient intrusion detection in dynamic in-vehicle network environments.