FE-Based Data Generation for Structural Integrity Verification with Neural Networks

Abstract

The application of deep learning within Prognostics and Health Management (PHM) frameworks offers immense potential for real-time structural health monitoring. However, in high-consequence applications such as high-velocity collisions, algorithmic deployment is severely bottlenecked by a fundamental scarcity of training data. Because destructive physical testing is prohibitively expensive and real-world incident telemetry is rare, acquiring the massive, diverse datasets required to prevent neural network overfitting remains a critical operational challenge. To overcome this data scarcity, this paper presents a scalable methodology that transforms high-fidelity Finite Element (FE) analysis into an automated Physics-of-Failure "data engine". Using the safety-critical underrun guard of a high-speed intercity train as a primary case study, constructing a complete Digital Thread that bridges quantitative continuum mechanics with qualitative human operational logic.

To prevent the curse of dimensionality during automated data generation, this framework implements rigorous physical and algorithmic boundaries. Complex multi-body impact kinematics are mathematically distilled, and six non-linear material parameters are algorithmically condensed into a singular "toughness" variable to honor the physical inverse correlation between material strength and ductility. Utilizing a low-discrepancy Sobol sequence, the simulation parameters, are sampled uniformly with 256 transient-dynamic simulations, ensuring the highly sensitive boundary transitions of structural failure are densely captured without statistical clustering.

To simulate real-world hardware constraints, a virtual sensor network of accelerometers and strain gauges is deployed. Multi-axial strain tensors are mathematically reduced to Von Mises scalar invariants, optimizing the AI input layer and preventing feature bloat. Concurrently, standardized multi-angle visual renderings of the deformed FE mesh are generated, enabling domain experts to label the physical damage into discrete, actionable operation and maintenance directives (continue operation, reduced operation, immediate operational stop and reuse, repair, exchange).

Following the experimental validation of the digital twin against physical destructive laboratory tests, the resulting synthetic dataset was utilized to train a baseline classification neural network. Despite the severely constrained sample size ($n = 256$), the diagnostic AI achieved an outstanding predictive accuracy of 82 \%. This confirms that mathematically optimized, physics-grounded data generation can successfully bridge the data gap in modern PHM frameworks, providing a robust pathway for real-time diagnostic deployment.