An Evidence-Centered Decision-Making Framework for Explainable Autonomous Recovery in Self-Healing Software Systems

Self-healing software systems have become an important approach for improving the reliability and availability of modern distributed applications. However, existing autonomous recovery methods are predominantly based on predefined rules or data-driven models, which often provide limited transparency, weak evidence traceability, and insufficient justification of recovery decisions. This paper proposes an Evidence-Centered Decision-Making (ECDM) framework for explainable autonomous recovery in self-healing software systems. The proposed framework transforms runtime observations into structured evidence, constructs an Evidence Graph, generates recovery hypotheses, and evaluates alternative recovery actions using the proposed Hierarchical Decision Confidence Model (HDCM). The framework was evaluated through an experimental campaign consisting of 110 recovery-decision episodes across five representative failure scenarios and multiple fault severity levels. In the confirmatory evaluation of 110 recovery-decision episodes, ECDM achieved an overall decision accuracy of 85.45%, with the highest scenario-level accuracy observed for CPU Resource Exhaustion and Service Dependency Failure. The Safety Gate admitted 49 decisions for autonomous handling, including 34 active recovery actions. All 34 executed active recovery actions satisfied the predefined recovery-verification criteria, with a median Recovery Verification Latency of 1048.9 ms, while autonomous coverage varied substantially across scenarios. The proposed ECDM framework provides a traceable recovery-decision framework by combining evidence modeling, hierarchical confidence estimation, and structured decision justification.