AI-Driven Failure Analysis in Engineering Systems: LSTM-Based Failure Prediction with Efficient Dependency-Graph Transitive-Closure Updates

Abstract

Failure analysis is essential for ensuring the reliability and safety of engineering systems. However, conventional approaches are often too resource-intensive to keep pace with the growing complexity and interconnectivity of modern engineering systems. This research proposes a novel, AI-driven combinatorial approach to failure analysis. By integrating natural language processing (NLP), machine learning (ML), and graph-theoretic methods, the approach aims to improve component-failure prediction, generate actionable insights, and help safety and reliability engineers identify failure patterns that inform more resilient system designs. The proposed approach uses a long short-term memory (LSTM) network to predict component failure probabilities and represents complex interdependencies among system components with a dependency graph. The graph’s transitive closure is cached and incrementally maintained using efficient algorithms, enabling on-demand recalculation of estimated component failure probabilities as the system design changes. An aerospace engineering case study is used to demonstrate the approach’s application in a real-world setting. The proposed AI-driven approach offers a promising path to improving system safety and reliability across a range of engineering domains, particularly in industries that operate highly complex, safety-critical systems. Future work should train the LSTM on larger, more representative datasets to improve the accuracy of failure-probability predictions.