Abstract Automated clash detection in Building Information Modeling (BIM) often produces an excessive number of results, many of which are irrelevant or non-critical to constructability. This burden requires project teams to manually filter clashes, introducing inefficiency and subjectivity. To address this limitation, this study proposed a deep learning–based framework for classifying irrelevant clashes into penetration categories that reflect their constructability implications. A taxonomy of 15 categories was developed through expert consultation, differentiating penetrations by orientation, size, and shape. Two multi-view architectures were evaluated: the Multi-View Convolutional Neural Network (MVCNN), a widely adopted baseline, and the Multi-View Vision Transformer (MVT), a state-of-the-art architecture designed to capture inter-view dependencies through attention mechanisms. A dataset generated from a federated BIM model was used to train and test both models. Results showed that MVT achieved superior performance across accuracy and F1-score, with particular improvements in minority categories involving small or diagonal penetrations. IoU-based analysis further demonstrated that MVT attended more precisely to clash regions, enhancing interpretability. The findings confirmed that Transformer-based multi-view learning offers significant advantages for clash classification. By linking automated classification with reinforcement and constructability requirements, the proposed framework supports more reliable constructability analysis, cost estimation, and project planning.

Abstract Object detection is a critical task in drone vision perception systems. However, in complex low-altitude environments, the recognition of tiny, dense, and occluded objects presents significant challenges. To overcome these limitations, this paper proposes an enhanced lightweight object detection model, YOLO-BWS, based on the YOLO11 architecture, aiming to improve tiny object detection performance while maintaining computational efficiency for drone vision. First, we introduce a higher-resolution detection head network to the YOLO11 architecture to incorporate finer spatial feature representations. Second, to optimize information exchange across multi-scale features, we integrate a bidirectional feature pyramid network into the neck structure. This network utilizes its learnable feature fusion mechanism and bidirectional path propagation to effectively enhance semantic expression across different layers. In the bounding box regression branch, we employ a more rigorous Wise Intersection over Union loss function to dynamically adjust the backpropagation gradients based on sample quality, improving the stability of localization and accelerating the convergence of the training process. Experimental results on three datasets, COCO, RSOD, and VisDrone, demonstrate that YOLO-BWS significantly outperforms the baseline YOLO11 across all scenarios, while maintaining a computational cost of 11.6 GFLOPs and only 2.7 million parameters. Compared with YOLO11, YOLO-BWS achieves a 2.37% improvement in mAP@0.5 and a 1.81% improvement in mAP@0.5:0.95 on the COCO dataset. Similarly, on the RSOD dataset, YOLO-BWS improves mAP@0.5 by 3.5% and mAP@0.5:0.95 by 1.1%. On the VisDrone dataset, the mAP@0.5 increases by 4.8%, while the mAP@0.5:0.95 improves by 3.0%. Additionally, visual analysis reveals that YOLO-BWS exhibits exceptional object focusing ability, particularly for tiny and occluded objects, and provides more accurate bounding box regression.

Abstract Spatial layout planning is a core issue in architectural design with a significant impact on building performance. Although various artificial intelligence methods have advanced automated spatial layout planning in the construction area, applications of reinforcement learning for existing building renovation, particularly those with performance-based aims, remain limited. This study proposes an automated spatial layout planning framework using deep reinforcement learning to evaluate its viability and efficiency in functional layout renovation targeting sustainable objectives. The agent-based simulation with the queuing theory was employed to capture the occupant patterns, providing a foundation for the performance indicators of the reinforcement learning model. The model was trained and applied to a real case to generate optimized layouts, achieving improvements in energy and occupant circulation compared to the original layout plan. It should be clarified that the above energy-optimization estimation of the model relied on a simplified approximation rather than a dynamic energy simulation. The results demonstrate the advantages of this framework in advancing AI-aided design and indicate future research directions for enhancing the design efficiency and quality in automated spatial layout planning.

Abstract Supervised super-resolution methods estimate high-resolution simulation results from low-resolution inputs. These methods typically train the neural networks using pairs of high-resolution data and their down-sampled low-resolution counterparts, implicitly assuming the global similarity between the low-resolution and high-resolution models. This assumption, however, often fails as the low-resolution models usually exhibit stiffer behavior than high-resolution models due to numerical stiffening. This paper proposes a novel supervised super-resolution method to mitigate this problem for linear deformable object simulation. The method constructs training data pairs by matching low-resolution and high-resolution simulation snapshots based on the similarity of the normalized strain energy, the normalized temporal change rate of strain energy, and the modal contributions to the overall strain energy. The loss function also incorporates the equation residuals derived from the finite element method. The time integration scheme is selected by examining the eigenvalue distributions of the neural tangent kernel associated with the finite element residuals. Compared with previous down-sampling methods, the proposed method reduces the maximum relative displacement error by 25%, 54% and 38% for the beam, elephant and armadillo models, respectively.

Abstract We introduce TDiff-HSI, a diffusion-based model that can generate hyperspectral images (HSIs) directly from RGB images and material-wise segmentation masks. HSI provides both spatial (u, v) and spectral (λ) information. The accompanying dataset that we are releasing spans wavelengths in the range from 420 to 1728 nm, digitized into 512 channels. Directly handling this immense three-dimensional dataset is computationally prohibitive and often leads to numerical errors. To address this challenge, TDiff-HSI leverages Tucker decomposition to reduce dimensionality, enabling more stable and efficient processing. Moreover, spectral precision is enhanced by combining RGB channels with a material segmentation mask. To support this research, we constructed a new dataset using a hyperspectral camera. The dataset comprises 40 014 RGB-HSI pairs across 78 scenes, featuring 12 objects with corresponding polygonal segmentation labels. Experimental evaluation demonstrates that TDiff-HSI achieves state-of-the-art performance verified on the existing dataset. For the new dataset that we are releasing, we establish new benchmarks of MRAE 0.2169, RMSE 0.0192, PSNR 36.46 dB, SAM 0.0424, and SSIM 0.9327 Project and dataset are available at https://github.com/JaeikBae/TDiff-HSI

Abstract Large-scale multi-objective optimization problems are prevalent in practical applications. Consequently, researchers have proposed numerous large-scale multi-objective evolutionary algorithms (LMOEAs) to address these problems. Among them, neural network (NN)-based LMOEAs have garnered significant attention owing to their superior search efficiency. Regrettably, existing NN-based LMOEAs are limited by a single source of training data and pre-fixed network topology, resulting in insufficient generalization and weak adaptability. We propose a heterogeneous operator mechanism and stochastic configuration network-enhanced LMOEA (HMSCN-LMOEA) to overcome these issues. Specifically, a heterogeneous operator mechanism is introduced to address the limited data sources in NN model training, thereby effectively enhancing the quality of the training set. A stochastic configuration network (SCN) model with dynamic structural adaptability is designed to replace the NN model with a pre-fixed network topology, thereby enhancing the algorithm’s adaptability. Finally, to evaluate the algorithm’s performance, four large-scale multi-objective optimization problems—UF, WFG, LSMOP, and ZCAT—are adopted. Experimental results based on IGD, IGD+, Spacing, and HV metrics demonstrate the significant advantages and competitiveness of HMSCN-LMOEA over state-of-the-art LMOEAs. Furthermore, HMSCN-LMOEA is employed to solve the cloud task scheduling problem across different task scales, and the experimental results show that it achieves the top rank in 75% of the evaluated scenarios, further demonstrating its promising potential.

Abstract Transformers have achieved remarkable success in natural language processing and computer vision, motivating their extensive research into intelligent fault diagnosis. The multi-head self-attention mechanism in Transformers effectively extracts global features and emphasizes important feature maps. However, directly applying Transformers to vibration signals can incur substantial computation and training effort, which often makes convolutional neural network (CNN)-based approaches more practical. This paper, thus, proposes a new architecture, called Multi-scale Dilated Contracted sub-Transformer (MDCSformer), specifically designed to efficiently process noisy vibration signals for fault diagnosis. First, multi-scale dilated convolution layers capture features at various scales, extracting both local and global characteristics from the signals. Next, a contracted multi-head sub-self-attention module refines these features by selectively emphasizing relevant information while maintaining computational cost. The proposed method is validated with two gearbox vibration datasets and a bearing vibration dataset, demonstrating that MDCSformer outperforms comparable CNN-based, Transformer-based, and hybrid models by achieving higher accuracy and robustness in fault diagnosis. This superior performance stems from its integration of rich multi-scale feature extraction with contracted subspace attention.

Abstract The intelligent scheduling of autonomous trucks at intersections in open-pit coal mines poses a fundamental challenge due to nonlinear dynamics and complex multi-truck interactions under gradient–load coupling. Unlike existing intersection control studies that are mainly designed for flat urban roads and neglect mining-specific physical constraints, this study proposes a Gradient-Load Aware Multi-Agent Attention-Enhanced DDPG (GLA-MA-MADDPG) approach that explicitly embeds mining-domain physical laws into multi-agent reinforcement learning for intersection scheduling. A gradient–load dynamic coupling model is developed to inject physics fidelity into MARL-based intersection scheduling, capturing nonlinear and load-sensitive kinematics overlooked by flat-road assumptions in prior studies. To handle coordination under these constraints, we design a task-oriented multi-dimensional attention mechanism that jointly interprets physical heterogeneity, asymmetric dynamics, priorities, and collision risks. Additionally, a gradient-aware adaptive priority strategy redefines right-of-way as a physics-grounded, state-dependent process, ensuring safe and preferential passage for loaded and downhill trucks. In the simulated four-branch gradient intersection environment, the proposed GLA-MA-MADDPG method achieves substantial performance gains over traditional MADDPG across 20 independent runs. Specifically, it improves throughput by 23.5%, reduces average transit time by 8.3%, decreases waiting time by 31.2%, and lowers the collision rate from 3.47‰ to 1.32%, achieving efficiency gains of 39.2% in mixed uphill–downhill scenarios. Overall, this study contributes a first-of-its-kind integration of physics-informed gradient–load modeling with attention-driven MARL, providing a generalizable computational framework for intelligent intersection scheduling in open-pit mining and other large-scale, safety-critical engineering systems.

Abstract This paper presents a novel algorithm that enhances Structure-from-Motion (SfM) pipelines by integrating efficient novel view synthesis (NVS) using 3D Gaussian Splatting (3DGS). Traditional SfM methods often fail in challenging scenarios with repetitive structures, outliers or misaligned images, leading to ghost and doppelganger artifacts. To address this, we propose an NVS-guided evaluation and correction framework that leverages 3DGS-rendered views to detect and rectify misaligned images, improving reconstruction accuracy. Our method not only enhances SfM outputs but also benefits downstream tasks such as dense reconstruction [multi-view stereo (MVS)], NVS, and visual localization. Experiments on public datasets and our real-world indoor dataset show that state-of-the-art baselines fail in several ambiguous scenes where our approach succeeds. For example, on the Cambridge Landmarks dataset (Shop Facade), our method reduces localization error from 5.7 to 5.4 cm and improves SSIM from 0.735 to 0.787. These results confirm the robustness and broad applicability of our framework, demonstrating that while 3DGS is an effective tool, the core contribution lies in the proposed SfM re-alignment strategy.

Abstract As the core component responsible for high-frequency power switching in photovoltaic inverters, accurately predicting the remaining useful life (RUL) of insulated gate bipolar transistors (IGBTs) has become a key factor in ensuring the stable operation of photovoltaic systems. However, existing methods struggle to precisely characterize the degradation characteristics and processes of IGBTs at different time points. To address these issues, this paper proposes a MIG-PI-Pathformer RUL prediction method that integrates physical degradation models with deep learning. This method establishes a multi-stage Inverse Gaussian degradation model based on the physical failure mechanisms of IGBTs and couples it with the dual attention mechanism of the Pathformer model to capture complex degradation features, adaptively divide time scales, and thereby correct prediction errors in the physical model; Additionally, physical rule constraints are incorporated into the Pathformer loss function to ensure that RUL predictions align with degradation mechanisms. Simulation results show that, on NASA’s IGBT aging dataset, compared to the single Pathformer, the proposed method reduces MSE and MAE by 70.21% and 17.84%, respectively, and improves R2 by 7.66%. This method provides more accurate and physically interpretable technical support for fault warning and optimized maintenance of photovoltaic inverters.