Jin-Gyun Kim

Thermal anomaly detection in safety-critical engineering systems is often constrained by sparse sensing, which limits direct observation of internal temperature fields. This study proposes a low-latency anomaly detection framework for finite element method (FEM)-based thermal data reconstructed from a small number of optimally placed sensors. Building on a prior denoising autoencoder-based sensor placement and reconstruction framework, the present work focuses on downstream online anomaly scoring. A shared-parameter long short-term memory autoencoder computes node-wise temporal anomaly scores from sliding windows of reconstructed temperature sequences, and a graph-based Laplacian-of-Gaussian refinement with score fusion enhances spatial coherence on irregular FEM meshes. The framework is evaluated on three FEM thermal cases---a heat plate with a cavity, a battery module, and a brushless direct current motor---and compared with autoencoder and Transformer-autoencoder baselines under matched training and evaluation protocols. The proposed pipeline produces anomaly maps with higher spatial selectivity than temporal scoring alone and delivers competitive or superior detection performance depending on the dataset, with the clearest gains observed in systems with gradually evolving spatiotemporal anomalies. Across the evaluated cases, the method maintains per-step anomaly-scoring latency below the corresponding simulation sampling interval, supporting its suitability for real-time monitoring. These results show that reconstructed FEM thermal fields obtained under sparse sensing can support reliable online anomaly detection when sequence-aware temporal modeling is combined with mesh-aware spatial refinement.

Clearance is an important phenomenon in mechanisms that stems from manufacturing imperfections and wear and tear. Undetected clearance can compromise machine operations, negatively impacting its performance, and cause premature damage that requires maintenance actions. Monitoring clearance growth is useful to improve maintenance plans and helps reduce the number of unwanted interruptions during machine operation. In this work, a prediction method that targets the determination of the clearance size based on minimal, raw sensor data and machine learning is proposed. The data in this study are generated using multibody dynamics simulations based on planar mechanisms and used to train and compare several types of neural networks in terms of their ability to assess clearance size. Results show that clearance parameters can be reliably estimated with appropriate combinations of sensor locations and type of neural network. The developed method offers reliable clearance detection based on measurements that can be obtained from physical systems and used to monitor the state of their clearance defects.

Long-term spatio-temporal forecasting and reconstruction of complex physical systems are fundamental to digital-twin development, fault diagnosis, and control strategy design, yet they remain major challenges due to limited sensor availability and high computational costs. Existing pixel-based sensor placement optimization methods attempt to address data sparsity but determine sensor locations via Proper Orthogonal Decomposition-Discrete Empirical Interpolation Method (POD-DEIM) as a fixed pre-processing step, after which forecasting and reconstruction are trained. As a result, sensor selection is not jointly optimized with the predictive model, allowing suboptimal choices to propagate through the pipeline and degrade overall performance. Moreover, pixel-level inputs do not directly correspond to physical sensor locations, limiting physical interpretability. To overcome these limitations, this study proposes Sensor Placement Optimization (SPO)-Informer, a Transformer-based framework that jointly optimizes sensor selection, forecasting, and reconstruction within a unified end-to-end architecture through a learnable parameter layer. By employing node-level inputs that directly correspond to physical sensor coordinates, SPO-Informer enhances both physical interpretability and practical applicability. Experiments on Circular Membrane Vibration System, Hot Plate Thermal System, Battery Module Thermal System, and Motor Thermal System demonstrate that SPO-Informer achieves superior long-term forecasting performance compared with existing methods in terms of Mean Absolute Error (MAE), while maintaining stable accuracy even with significantly reduced numbers of sensors. These results confirm the robustness and generalization capability of the proposed approach and support its suitability for digital-twin applications requiring accurate and interpretable long-term predictions.

As multibody dynamic systems grow in size and complexity, applying machine learning becomes challenging due to limited training data and high computational cost. To overcome these limitations, this study employs the Subsystem Synthesis Method, which decomposes the overall system into subsystem-level components that can be analyzed independently and efficiently synthesized into the complete system model. Building upon this framework, a data-driven Subsystem Synthesis approach is proposed, in which each subsystem is replaced with a subsystem-wise deep neural network metamodel trained on numerically generated data from the Subsystem Synthesis Method formulation. This subsystem-level modeling strategy not only reduces the data and training requirements of machine learning-based approaches but also preserves the modularity and physical interpretability of the overall multibody dynamic system. The proposed framework is validated through numerical simulations of representative multibody systems, including a cart–single pendulum, a cart–multiple pendulum, and planar quarter-car and half-car models. The results confirm that the proposed method achieves computational efficiency while maintaining high predictive accuracy, demonstrating its potential for scalable, real-time simulation of complex multibody systems.

Inverse force identification is beneficial for the operational state estimate of structural dynamic systems. In this problem, Tikhonov regularization with a 𝐿2-norm penalty parameter, known as Ridge regression, has been widely used to control the trade-off between suppressing noise amplification and maintaining accurate force estimation. Recently, real-time inverse force identification, which is a sequential time method, has been accomplished by implicit time integrators with single-parameter Tikhonov regularization techniques. However, a lower accuracy issue has newly emerged due to solution error propagation originated from measurement noise through iterations in the time integration scheme. To address this issue, we propose a recursive multi-parameter regularization framework for real-time implicit inverse force identification, expressed in a Maximum a posteriori (MAP) estimation with a joint prior. In the sequential time method formulation, the additional regularization mechanism couples successive force identifications and mitigates the propagation of solution errors across time iterations. The resulting framework therefore attenuates solution error propagation and ensures identification accuracy in real-time force identification. The performance of the proposed method is investigated through numerical and experimental examples.