Complex system reliability modeling and analysis based on energy landscape

Abstract

Complex system reliability modeling and analysis are valuable to forecast large-scale failures and locate key elements for in-time interventions. Reductionist methods are challenging to emulate underlying mechanisms, while data-driven methods ignore causality. To bridge the gap between mechanistic interpretability and data-driven adaptability, a method based on statistical physics is proposed. A maximum entropy model is built to quantify system states’ probabilities, and material implication logic is introduced to represent bidirectional asymmetric causalities. Mapping probabilities to energies, all states form an energy landscape, and the state transitions, steady states, and attractive basins are identified. Further, critical elements are located, whose failures switch attractive basins and potential steady states from reliability to a large-scale failure. In practice, the proposed method can predict system states and guide the interventions on critical elements. In the case of the cascading failures in a network, with the observed nodes’ states, we can reconstruct failure propagations and locate hubs, showing the feasibility to balance physical explainability and data-based adaptability. In the case of the data center suffering burst traffic, the cascading failures caused by migrations and the critical states before collapses are identified from the statistical physical machines’ states, giving an insight to understand complex engineering systems.