Time-series anomaly detection in Industrial Internet of Things (IIoT) systems is crucial for ensuring operational safety and preventing catastrophic incidents in critical infrastructure. Diffusion models show promising potential in learning these complex anomaly patterns. However, there remain two limitations in implementing robust anomaly detection functionality in practical industrial scenarios: (1) they generate excessively diverse patterns without respecting the inherent constraints of industrial processes; and (2) they fail to effectively incorporate domain knowledge about system-specific causal relationships. To address these challenges, we propose CausEvDif, a novel IIoT anomaly detection method that integrates deep visual understanding with a diffusion model enhanced by mechanistic knowledge, bridging algorithmic and real-world process gaps. To achieve this form of understanding, pre-trained vision language models (VLMs) are employed as the core reasoning engine for the visual representation of time-series data. Our approach introduces three key innovations: (1) a conditional generation enhancement that transforms time-series data into visual formats and leverages pre-trained VLMs to guide the diffusion process towards realistic industrial patterns; (2) a dynamic evidential reasoning framework designed to reconcile potentially conflicting mechanistic knowledge and data-driven causal insights under uncertainty; and (3) a multi-scale spatiotemporal denoising network with topology-adaptive structures that captures features from local fluctuations to global trends. Extensive experiments on three public industrial benchmark datasets demonstrate that CausEvDif substantially outperforms state-of-the-art methods, particularly in complex anomaly scenarios. Additional experiments on a real-world industrial dataset further demonstrate that the proposed method can enable faster anomaly detection and support more effective anomaly handling.
Virtual reality (VR) technology has significant application potential for use in industrial settings. However, the existing VR target selection and interaction techniques often fail to meet the demands of complex industrial scenarios. In this study, a proxy-based interaction technique is employed, and a set of natural gestures is derived for performing region-of-interest proxy manipulation through user heuristic research. A system framework is subsequently designed to implement these gestural interactions: the “Take-A-Look” technique. A user study is then conducted to compare the proposed Take-A-Look technique with the existing methods across multiple dimensions, including performance, behaviour, and subjective evaluations. The results demonstrate its superiority in terms of selecting small-sized targets, with its application value further validated in a simulated virtual warehouse scenario. This research contributes a novel interaction technique that addresses the key challenges encountered in industrial VR applications, offering both improved performance and enhanced user experiences.
This study proposes a data-driven decision support framework for Maintenance, Repair, and Operation (MRO) inventory planning from a Lean, Agile, Resilient, and Green (LARG) perspective. The first stage of the proposed framework involves demand forecasting, which is implemented using a model that combines a genetic algorithm and an artificial neural network, initially focused on two of the most critical items used in railway track maintenance and using input indicators with a direct correlation with LARG. To allow the model to be replicated, the second stage of the framework begins, involving time series clustering with k-means and dynamic time-warping metrics, restarting the first phase, and replicating the model for items in the same cluster. Lastly, the proposed framework incorporates practical implications for each LARG paradigm precisely because of the connection between the first two stages of the framework and the LARG concept. The framework is applied to a case study using real data from a railway logistics operator. The research findings revealed superior demand forecasting performance compared to current company practices. The proposed framework is a crucial data-driven decision-support framework for enhancing MRO inventory planning within the LARG paradigm.
This paper presents a multi-stage diffusion approach for generating accessible and regulation-compliant layouts for multi-occupancy buildings. Existing generative design methods are often limited by rigid one-shot generation, weak constraint adherence, and lack of control over critical architectural features such as openings. To address these limitations, the proposed approach introduces a two-stage latent diffusion model guided by multi-modal input conditions, including boundary constraints, structural wall plans, room masks, spatial anchors, and room type lists, which enable fine-grained yet flexible design control. In addition to generating spatial layouts, a second diffusion model produces doors and windows, guided by rule-assisted annotations that ensure room connectivity and egress compliance. An iterative refinement workflow, supported by a custom web-based user interface featuring both rule-based accessibility checking and human-in-the-loop editing, enables regeneration of specific layout regions to satisfy accessibility requirements and user needs. Furthermore, the finalized 2D layouts are automatically converted into industry foundation class models, which enables direct integration with building information modeling workflows. Quantitative results demonstrate substantial improvements in layout quality, room type accuracy, and opening placement accuracy over existing baselines. Case studies highlight the approach’s capacity for iterative refinement, opening generation, and downstream interoperability. This work establishes a foundation for scalable, constraint-aware generative design in real-world multi-occupancy building scenarios.
Building Information Modeling (BIM) of various railway station facilities is crucial in the whole BIM implementation for a railway station. However, existing BIM methods and industrial standards mainly focus on expressing the geometric information of station facilities, while the essential interrelations among these facilities are neglected. Consequently, current BIM models lack critical semantic information, which hinders their applicability throughout the entire lifecycle of a station. To address this limitation, this study proposes a hypergraph-based relation model and a corresponding Industry Foundation Class (IFC) extension for comprehensively specifying the geometric and semantic information of railway station facilities. Based on the proposed model and IFC extension, a prototype system has been developed and validated through a real-world case study. The results demonstrate that, by efficiently retrieving the interrelations among facilities, the proposed approach enables automatic model updates and facilities related specialized computations with the BIM models.
Anomaly detection for aero-engine gas path performance is aimed at the prompt detection of early signs of failure, which is important for the health management of aero-engines. Currently, most methods for aero-engine anomaly detection focus primarily on the nonlinear mapping relationships within time series data and neglect the rich spatiotemporal features contained in aero-engine data. Therefore, this paper proposes a novel anomaly detection architecture, termed the STFFE-IIN, which is based on spatiotemporal feature fusion enhancement and an improved Informer. First, a feature extraction module based on a graph convolutional neural network and a gated temporal convolutional neural network is designed to model the aero-engine data from spatiotemporal dual dimensions, explore its features, and efficiently capture the anomalous variations in the aero-engine time series data under the framework of spatiotemporal dynamics. Second, a bidirectional long short-term memory module is introduced to enhance the local dependence of degraded feature maps within the Informer model, and a dynamic prediction model is constructed. Third, the residual signal is generated by using predicted and actual values of the exhaust gas temperature from the dynamic model, which is combined with a detection decision module to realize anomaly detection for the aero-engine gas path. Finally, the superiority and potential of the proposed method are verified by extensive operational data from airplanes.
The increasing digitalization of agricultural machinery within the Agriculture 4.0 paradigm has led to widespread deployment of smart tractors equipped with high-frequency, geo-referenced sensors. These systems generate large volumes of multivariate time-series data, creating new opportunities to enhance operational reliability. However, anomaly detection in this context remains challenging due to high data dimensionality, spatiotemporal complexity, strong context dependence of faults, and limited labeled failure data. To address these challenges, this work proposes a spatiotemporal anomaly detection framework that integrates terrain- and activity-specific autoencoders with a synthetic anomaly generation pipeline powered by Large Language Models (LLMs). Tractor sensor data are partitioned by terrain and clustered to uncover recurring activity patterns, enabling fine-grained, context-aware anomaly detection. To overcome the lack of real fault data, we introduce an LLM-driven, rule-based approach to generate realistic synthetic anomalies for systematic testing. The framework is validated on multivariate time-series data collected from the Controller Area Network bus of a New Holland T7.165 S tractor operating under diverse agricultural conditions. Experimental results show that context-specific autoencoders achieve high detection accuracy across a range of operational scenarios, while LLM-generated synthetic faults enable robust evaluation. Overall, by combining context-aware anomaly detection with LLM-driven fault generation, the proposed framework enhances predictive maintenance in precision agriculture and provides an effective approach to improving the reliability of modern agricultural machinery.
Retrieval-augmented generation systems are increasingly applied in the intellectual property industry, but they often fail to handle the diversity of real user queries. Spelling errors, colloquial expressions, and ambiguous terminology frequently cause retrievers to misinterpret semantically equivalent queries as unrelated, resulting in inaccurate responses. To address this problem, we propose the Multi-Angle Question Generation and Retrieval Fine-Tuning Method (MQG-RFM), a lightweight framework that leverages large language models to generate diverse paraphrased queries and uses hard negative examples to fine-tune retrievers. Rather than altering model architectures, MQG-RFM combines prompt engineering and fine-tuning to improve system robustness with low computational cost. Experiments on multiple intellectual property datasets demonstrate substantial gains in both retrieval accuracy and response quality compared with baselines. MQG-RFM offers a practical and scalable approach for small and medium-sized agencies. The proposed method has been successfully adopted by ScholarMate, the largest professional research networking platform in China, for real-world deployment.
Accurate pedestrian estimation is essential for developing sustainable, livable cities and supporting human-centered urban planning. However, existing studies often rely on coarse, point-based estimates that lack the spatial resolution needed to capture fine-scale pedestrian dynamics. This study proposes R-PIN (Regional Pedestrian Inpainting Network), a novel deep learning model that reformulates regional pedestrian estimation as an image inpainting task to generate high-resolution, spatially continuous pedestrian distributions. R-PIN integrates multi-source urban features through a dual-branch encoder and attention-based fusion block to capture both local details and global spatial dependencies. Applied to a real-world case of New York City, R-PIN exhibits robust performance across diverse urban landscapes, especially in Manhattan and Brooklyn with stable and homogeneous pedestrian patterns. Compared with representative deep learning baselines, R-PIN achieves 57.29% and 79.96% reductions in MAE (Mean Absolute Error) and MSE (Mean Squared Error), respectively. By explicitly leveraging the surrounding urban context, R-PIN captures spatial continuity and spillover effects, reducing MAE and MSE by 55.23% and 76.97%. Feature importance analysis highlights the key roles of streetscape design, road network, points of interest, and land-use diversity in shaping pedestrian distribution. Overall, this framework provides fine-grained insights into pedestrian dynamics and offers a robust analytical tool for street-level management and pedestrian-oriented urban analytics.
With additive manufacturing advancing toward higher precision and productivity, reliable fault diagnosis has become increasingly important for Powder Bed Fusion (PBF) equipment. PBF equipment generates high-dimensional, strongly correlated, and time-dependent multi-sensor data, which contain rich condition information but also introduce substantial redundancy and computational burden. Existing fusion methods mainly focus on accuracy improvement or redundancy reduction, yet often fail to effectively coordinate the trade-off between fusion cost and information gain, thereby limiting their applicability in dynamic industrial scenarios. Given that, a dynamic equilibrium game mechanism (DGAM) based multi-source time-series fusion method is proposed for fault diagnosis of PBF equipment. A data-driven fusion model is first developed to integrate heterogeneous signals from the sealed-chamber and recirculation filtration systems. Fusion cost and information gain are then jointly coordinated within a dynamic game-inspired allocation framework, while a Subgame Perfect Nash Equilibrium (SPNE)-inspired recursive strategy is introduced to guide stage-wise source allocation. Comparative experiments on multi-source monitoring data demonstrate that DGAM consistently improves the cost-performance balance of different fusion backbones without increasing model complexity. Among all compared methods, DGAM combined with Wavelet Local Response Convolutional Neural Network (WLRCNN) achieves the best overall performance, reducing fusion cost from 0.35 to 0.27, Mean Squared Error (MSE) from 0.052 to 0.041, and Root Mean Squared Error (RMSE) from 0.216 to 0.193, with only a slight decrease in information gain from 0.044 to 0.043. These results demonstrate that DGAM provides an effective and practical solution for fault diagnosis in PBF equipment.
In predicting the true remaining useful life (TRUL) of slurry pumps, directly adopting deep learning models faces several limitations: on one hand, the scarcity of degradation data requires the model to be trained and applied in a few-shot setting; on the other hand, due to the significant differences in degradation speed at different life cycles of slurry pumps, using TRUL as the direct label leads to a mismatch in the label ranges of the training and testing datasets, causing issues with label range extrapolation. To address these challenges, we present an innovative Meta-learning Wiener Framework (MW-F). This framework first introduces an adaptive meta-learning network (AMLN) for accurate prediction of the normalized remaining useful life (NRUL) in few-shot scenarios. Predictions from the AMLN are further integrated into a Wiener process-based state space model to track the trend of NRUL changes, deriving the probability density function for the time point at which NRUL first drops to zero, achieving effective estimation of TRUL and quantification of its uncertainty. Experiments on multiple industrial field datasets have verified the efficiency and accuracy of the proposed framework in predicting NRUL and TRUL in 0-shot, few-shot, and cross-machine few-shot scenarios. The results show that MW-F can effectively estimate NRUL and TRUL of slurry pumps in various scenarios, demonstrating its potential and practicality in complex industrial applications. Furthermore, test results on the Commercial Modular Aero-Propulsion System Simulation dataset demonstrate that MW-F is not limited to a specific equipment type and has the potential to serve as a more general RUL prediction tool.
Ensuring security and trustworthiness in large-scale Internet of Things (IoT) systems is critical, particularly with the increasing deployment of fog and edge computing (FEC) to support complex and distributed environments. While traditional security mechanisms effectively mitigate external threats, FEC-enabled IoT networks remain vulnerable to internal attacks that compromise system integrity and performance. This paper presents a unified, data-centric trust management framework featuring dynamic trust evaluation and propagation mechanisms. The framework establishes essential trust principles specifically designed for IoT environments, including dynamicity, context-dependence, fragility, objectivity, composability, and transitivity. It implements the principles through a layered trust model, assessing trust in entities based on their transmission behavior, data credibility, and historical records, thereby enabling a comprehensive and adaptive evaluation of trust. Additionally, we introduce the concept of community trust for edge nodes, which aggregates the trust of connected sensors using an optimized information entropy approach, ensuring robust inter-layer trust propagation. Simulations are conducted using a real-world dataset from an edge-based environmental monitoring system along the Wuyu Expressway in China. Results demonstrate that the proposed model significantly improves accuracy, reducing both false positive and false negative rates compared to existing methods. The proposed framework has been validated in real-world FEC-enabled IoT deployments, providing practical and interpretable trust metrics for industrial use. The quantitative analysis highlights the model's adaptability, sensitivity, and reliability against multiple attack scenarios, providing a practical solution for enhancing security and trust management in complex FEC-enabled IoT deployments.
High stacking precision is essential to prevent structural collisions, reduce safety accidents, and enhance project efficiency in modular integrated construction. However, current stacking precision is constrained by delayed hoisting control, limited perception, and fragmented information sharing. Therefore, this paper proposes a digital twin-driven pose estimation and trajectory prediction system via spatio-temporal feature fusion modeling, leveraging real-time information from high-fidelity digital twins. First, physical entities are transformed into smart stacking objects that combine static inherent properties with real-time dynamic activities captured by Ultra-Wideband sensors and Inertial Measurement Units. Then, a digital twin of stacking operations is used to interact with their physical counterparts in real time, enabling seamless mapping and visualization of the stacking process. Third, the PosTraFormer network is proposed for 6-dimensional module pose estimation and trajectory prediction based on spatio-temporal fusion modeling of trajectory point relations, to provide stacking guidance and improve installation precision. Both visualization and quantitative analysis from comparative experiments and case studies demonstrate the significant advantages of the proposed approach.
4D printing – a rapidly evolving field that integrates additive manufacturing with stimuli-responsive materials – enables the creation of structures that change shape or properties over time. As the field advances, it poses significant challenges in synthesizing cross-disciplinary knowledge spanning manufacturing, materials science, chemistry, physics, engineering, and design. To address this, we propose a novel framework that leverages multimodal large language models and retrieval-augmented generation for ontology learning tailored to the 4D printing domain. This framework systematically extracts, maps, and interconnects insights from diverse sources, including scientific literature and experimental datasets, therefore massively enriching and populating a comprehensive multi-domain ontology. This enables the construction of interconnected knowledge graphs that capture relationships among material compositions, properties, stimuli interactions, process parameters, and design strategies. By bridging semantic gaps across involved disciplines, the framework enhances interpretability, supports informed material selection and distribution, and accelerates innovation in the development of programmable and multifunctional structures by 4D printing. Overall, this integrative approach represents a foundational step toward structuring a rapidly growing body of scientific efforts to enable advanced active materials and intelligent structures.
As a core power transmission component in industrial equipment, the operational reliability of gearboxes directly influences the overall performance of machinery. Vibration-based fault diagnosis of gearboxes, however, is often challenged by noise interference, which obscures fault-related features and hinders accurate identification. To overcome this challenge, this study proposes an optimal enhanced envelope spectrum identification framework based on multiscale modulation frequency decoupling. The framework adopts a three-stage progressive strategy comprising demodulation, enhancement, and optimization: (1) a dual-frequency spectrum demodulation technique is introduced to decouple modulation features and reveal coupling characteristics in the two-dimensional frequency domain; (2) a marginal spectrum segmentation concept is developed to dynamically identify effective frequency bands based on signal energy distribution, thereby eliminating reliance on manually set thresholds; and (3) a multiscale fault-feature selection approach is implemented by integrating a multilevel segmentation framework with quantitative indicators. Experimental results demonstrate that the proposed framework substantially improves the extraction of weak fault features through the synergistic effects of demodulation and frequency-band optimization while reducing dependence on expert knowledge. Thus, this work offers a robust and generalizable solution for intelligent fault diagnosis in complex industrial environments.
Lithium-ion batteries are widely used as primary energy storage devices across diverse sectors, playing a crucial role in advancing emerging energy technologies. Nevertheless, battery production involves multiple instabilities, resulting in defective units with abnormal voltage differences. Furthermore, the conventional process of anomaly detection and consistency screening is time-consuming and complex, causing shipment delays and scheduling difficulties. To resolve these issues, we proposed a battery screening framework that integrates anomaly detection and consistency screening as a single step, thereby enhancing cell production efficiency. We developed the Hankel transformed multi-kernel Mobile-Net (HMM-Net), which combines Hankel transformation and multi-kernel convolution to improve feature extraction and performance for both anomaly detection and consistency screening. Initially, the charge and discharge time series data are transformed into a two-dimensional Hankel matrix. Subsequently, the input layer expands the channel dimension. Multiple blocks process the data with a multi-scale convolutional structure to extract features at various scales. Finally, the output layer assigns each battery as normal or anomalous. Experimental results demonstrate that the proposed HMM-Net method outperforms other methods in the combined task of anomaly detection and consistency screening, achieving an overall accuracy of 96.07% and exhibiting strong performance across additional metrics.
The construction industry faces great challenges in optimising fleet management under dynamic and uncertain operational conditions. Existing methods either rely on static schedules or heuristic rules, which lack adaptability to uncertain operational conditions and fail to capture the complexity of real-world construction site dynamics, leading to imbalanced resource allocation, disruptions in construction progress, and further deteriorating operational cost uncertainty. This study develops a novel digital twin framework that integrates modelling, simulation, and deep reinforcement learning, specifically the Deep Q-Network (DQN), for optimising fleet management under highly dynamical and uncertain conditions. A simulator of the earthmoving process is developed in the digital world, by modelling the cooperation between excavators and dumpers. By following the physical–virtual–physical workflow of digital twins, real-world data is used to quantify operational parameters and their associated uncertainties. The uncertainties are then integrated into DQN training to generate optimal real-time scheduling strategies, enabled by a new reward design that balances multiple objectives. Experimental results demonstrate that the proposed solution offers significant cost savings ranging from 1.41% to 9.71% while ensuring consistent decision-making, even under high uncertainty. This study also discusses how different levels of uncertainty impact operational productivity and costs. This work highlights the potential of the integration of deep reinforcement learning and digital twin approach, emphasising the necessity and impact of incorporating uncertainty in modelling and training to enhance predictive accuracy and decision-making. The proposed framework aims to advance sustainable and intelligent construction practices.
Automated fiber placement (AFP) quality control requires defect localization under limited annotation and diagnostic feedback grounded in process knowledge. This paper proposes a knowledge-enhanced multimodal agent that couples depth-map-based localization with ontology-constrained graph retrieval in a tri-space framework. To adapt Qwen3-VL-8B to AFP depth maps, we develop a two-stage post-training strategy. Supervised fine-tuning establishes modality alignment, defect-term grounding, and structured-output learning, while group relative policy optimization refines the language-side policy with detection-oriented rewards. On a production dataset of 1403 depth maps, the proposed method achieves 0.842 class-agnostic recall with only 40 labeled images and remains competitive in the 20-shot setting, surpassing single-stage baselines trained with 40 labels. Repeated experiments show stable performance under subset-sampling and optimization randomness. For diagnostic question answering on 50 multi-turn defect cases, graph-structured retrieval-augmented generation (RAG) improves claim-level evidence grounding over a text-chunk vector RAG baseline. It increases fully supported diagnostic claims from 70.6% to 80.4%, reduces unsupported claims from 13.4% to 7.9%, and improves citation precision from 84.6% to 91.4%, with 3.2 s average latency per question. These results show that metric-aligned post-training and ontology-constrained evidence retrieval support accurate AFP defect localization and auditable, mechanism-consistent diagnosis under limited supervision.
