Artificial neural network-based finite element model updating for electrical cabinet in power plant

• A DNN-based framework predicts density distributions of the 3D electrical cabinet directly from modal data. • Ridge regression identifies influential structural subsets to enhance DNN training stability and interpretability. • The data-driven approach replaces costly iterative optimization, offering efficient updating for large-scale models. • Validation confirmed reduced frequency errors and improved mode shape correlation. • The method supports real-time diagnostics and digital twin integration for critical non-structural components. Accurate representation of structural dynamics through finite element models is essential for ensuring safety and reliability in engineering systems, particularly for critical non-structural components such as electrical cabinets which are susceptible to seismic damage. However, discrepancies frequently arise between numerical predictions and experimental responses because of uncertainties in material properties, boundary conditions, and degradation effects over time. These inconsistencies necessitate finite element model updating (FEMU) techniques to refine model parameters and align analytical results with physical behavior. Despite significant advances, conventional optimization-based FEMU approaches often become computationally inefficient when applied to complex three-dimensional (3D) solid structures with high-dimensional parameter spaces. To address these limitations, this study presents a deep neural network (DNN)-based FEMU framework, validated using an electrical cabinet-type non-structure component, that predicts spatial density distributions of a solid structure directly from modal data. In addition, a ridge regression–based feature selection method is integrated to identify the most influential structural subsets, enhancing both interpretability and training stability of the DNN. The proposed framework provides a scalable and data-driven alternative to traditional iterative updating methods, capable of capturing complex nonlinear relationships between modal responses and material parameters. Beyond its computational efficiency, this approach establishes a foundation for real-time structural diagnostics and potential integration with digital twin platforms, enabling intelligent monitoring and adaptive maintenance of future structural systems.