The paper proposes a generalized method for constructing graph-logical models of the failure behavior of fault-tolerant multiprocessor systems. Such models are used for evaluating system reliability by means of statistical experiments based on simulation of failure behavior. The method is applicable to non-basic systems whose failure behavior cannot be characterized solely by the number of failed components and therefore requires more flexible modeling approaches. The proposed method is based on combining auxiliary models corresponding to different operating conditions of the system into a single model that correctly represents the overall failure behavior. In contrast to existing approaches, it imposes no restrictions on the graph structures of the auxiliary models and does not depend on the specific procedures used for their construction. The key idea of the approach is to integrate such models under a set of mutually exclusive logical conditions, each of which determines the applicability of a particular auxiliary model for a given system state. A set of model transformations is introduced, and it is shown that these transformations preserve model equivalence, that is, correspondence to the same failure behavior of the system. It is demonstrated that these transformations are sufficient to transform auxiliary models to graphs with identical structures, which is a necessary condition for their combination within the proposed framework. Several illustrative examples of the application of the method are provided. The correctness of the constructed models is validated through analysis of representative system states and through exhaustive evaluation over all possible states. The results for the considered example indicate that the proposed method can construct models that accurately represent system failure behavior while yielding more compact graph structures compared to the existing approach. At the same time, comparable logical complexity is maintained. The evaluation is limited to a representative example and exhaustive analysis of system states, and further validation on a broader class of systems remains a direction for future work.