The application of virtual reality (VR) in industrial training and safety emergency needs to reflect realistic changes in physical object properties. However, existing VR systems still lack fast and accurate simulation of complex, high-fidelity dynamic display of physical object evolution. To enhance the application of VR, a real-time VR visualization method is introduced, which adopts a pre-trained deep learning model to construct high-fidelity physical dynamic changes. This method firstly integrates data dimensionality reduction and temporal convolutional network (TCN) to pre-capture time-series data from numerical simulation results, and then employs Kolmogorov–Arnold Networks (KAN) to approximate nonlinear characteristics to improved Long Short-Term Memory (LSTM) network, thereby predict time-series simulation data accurately to achieves realistic and responsive dynamic displays. The experimental results of predicting time-series numerical simulation data demonstrate that the method balances computational efficiency and achieves good prediction accuracy, with Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) values increased to 0.0087 and 0.0063, respectively. These studies indicate that the proposed method significantly enhances VR’s capability for realistic physical modeling, paving the way for its broader application in high-stakes industrial training and emergency training environments.

Accurate prediction of concrete compressive strength is essential for ensuring structural safety in civil engineering, particularly in road and bridge construction, where inadequate strength can lead to deformation, cracking, or collapse. Traditional non-destructive testing (NDT) methods, such as the Rebound Hammer Test, estimate strength using regression-based formulas fitted with measurement data; however, these formulas, typically optimized via the least squares method, are highly sensitive to initial parameter settings and exhibit low robustness, especially for nonlinear relationships. Meanwhile, AI-based models, such as neural networks, require extensive datasets for training, which poses a significant challenge in real-world engineering scenarios with limited or unevenly distributed data. To address these issues, this study proposes a gradient-boosting artificial bee colony (GB-ABC) algorithm for robust regression curve fitting. The method integrates two novel mechanisms: gradient descent to accelerate convergence and prevent entrapment in local optima, and a point density-weighted strategy using Gaussian Kernel Density Estimation (GKDE) to assign higher weights to sparse data regions, enhancing adaptability to field data irregularities without necessitating large datasets. Following data preprocessing with Local Outlier Factor (LOF) to remove outliers, validation on 600 real-world samples demonstrates that GB-ABC outperforms conventional methods by minimizing mean relative error rate (RER) and achieving precise rebound-strength correlations. These advancements establish GB-ABC as a practical, data-efficient solution for on-site concrete strength estimation.

Shot peening is a critical surface treatment for improving the fatigue resistance of aircraft propeller blades operating under complex cyclic loads. While accurate coverage evaluation is essential for quality assurance, its development is severely hindered by a fundamental bottleneck: the extreme scarcity of annotated datasets in this niche aerospace domain, where data collection is costly and low-frequency, as each acquisition requires the actual peening of high-value components. Consequently, existing practices are restricted to subjective manual inspection or conventional segmentation methods that lack robustness under complex textures. To bridge this gap, this study develops an integrated automated surface evaluation framework, termed DT-ZSAM (Dual-Threshold Zero-shot Assessment Model), which circumvents the data-dependency bottleneck by leveraging the zero-shot capabilities of the Segment Anything Model (SAM) within a custom-designed prompt-generation pipeline. To ensure end-to-end automation without manual intervention, the framework identifies candidate regions via a dual-threshold scheme in grayscale and brightness domains and extracts representative prompt points through density-based analysis refined by DBSCAN clustering. Experimental results demonstrate that the proposed framework achieves precise segmentation without requiring any pixel-level annotated training data. Notably, the proposed framework yielded a coverage rate of 30.57%, aligning closely with the expert visual consensus (25–35%), whereas the standard commercial instrument (TCV-2A) significantly overestimated the coverage at 62.33% due to its sensitivity to surface textures and fixed calibration logic. This framework provides a robust and pragmatic solution for high-stakes industrial quality control, offering a reliable path for automating inspection in domains where large-scale data acquisition is practically unfeasible.