To address practical challenges in FCM clustering—including data complexity, heterogeneity, variations in model architectures, and device communication interruptions—this paper proposes a global model optimization strategy based on mKD to obtain high-quality global clustering prototypes in heterogeneous environments. FCM performs clustering by constructing IGs, whose quality directly determines clustering performance; thus, the principle of justifiable granularity is introduced to optimize IGs by balancing coverage and specificity. To cope with data complexity and non-IID issues, AEs are employed to extract deep features, and a regularization term is incorporated into the granularity-optimized FCM objective function to constrain deviations in local clustering centers. To handle variations in model architectures, the server adopts an mKD mechanism that enables bidirectional knowledge interaction between clients, allowing collaborative optimization of local and global models through clustering prediction results—thereby improving global clustering performance while supporting personalized client training. For device communication interruptions, the server retains historical parameter copies and uses the most recent parameters for updates before connection restoration, enhancing system robustness. Additionally, to enable a more flexible and stable solving process, parametric strategies and gradient-based methods are adopted, leveraging automatic differentiation for end-to-end joint optimization of membership degrees and cluster centers. Experimental results on multiple public datasets demonstrate the competitive clustering performance of the proposed method, and sensitivity analysis further validates its robustness in heterogeneous scenarios, personalization capability, and comprehensive advantages over comparative methods.