Search

Accurate defect detection is crucial for ensuring the trustworthiness of intelligent railway systems. Current approaches either rely on static ensembling of Convolutional Neural Networks (CNNs) or on single deep-learning CNN models. Traditional methods use a large volume of data to identify underlying patterns. However, training a new defect classifier with limited samples often results in overfitting and poor performance on unseen images. To overcome these challenges, we introduce a dynamic weight adaptation mechanism within an ensemble learning framework, enhanced by state-of-the-art transfer learning models (VGG-19, MobileNetV3, and ResNet50). We obtained a good test and validation accuracy of 99%, accompanied by a precision of 0.99, and recall and F1-score of 0.98. Unlike previous methods, our approach demonstrates consistency. Particularly, techniques that involve selecting predictions based on minimum loss previously required a significant number of epochs to stabilize. Data fusion techniques are even more demanding, requiring a substantially higher number of epochs for stabilization. Conversely, our proposed method achieves stable performance and rapid convergence within just 10 epochs, and it shows minimal fluctuation in the training curve (0–5 epochs). This visibly contrasts with earlier methods, which, despite reaching similar levels of accuracy, required significantly more epochs and exhibited greater fluctuations in their training curves. Through these findings, we anticipate that the proposed dynamic weight adaptation-based ensemble approach will further research and development of reusable AI-enabled solutions for rail defect classification. [ABSTRACT FROM AUTHOR]

The education sector is a rapidly growing global infrastructure for human resources. It is an essential and sensitive data-containing sector where sensitive private information about the user is included in the data. The incorporation of digital twins (DT) into virtual representations is used in higher education to enhance data analytic performance. However, the DT-enabled industry generates enormous amounts of data, making it challenging to manage security and privacy. The management of personal data by human resources involved in critical and resource-constrained tasks has increased since the globalization of Industry 5.0 and the higher education sector. Managing the security and privacy of users' private data through global data collection methods poses significant challenges. We present a framework for delivering decentralized data processing and learning in blockchain-based industry 5.0, with a use-case for higher education that integrates blockchain and deep learning (DL). The framework begins by introducing a new DT model that makes creating a virtual environment to model and replicate security-critical processes in higher education easier. Second, we propose a higher education data analysis model, so that blockchain-based data are ready for the fully functional higher education DT model. At last, the security and cost analyses of the proposed system are presented. Long short-term memory (LSTM) is used for the deep learning model. A custom Python-based blockchain with a proof-of-authentication consensus algorithm is used to implement the blockchain. Our proposed system is designed in such a way that it includes the higher education sector with the concepts of blockchain technology, digital twin, and Industry 5.0. The efficiency and throughput results of the blockchain show that the proposed system is better than the digital twin-based developed systems. The novelty of our proposed system is that educational blockchain with Industry 5.0 is conceptualized and implemented with deep learning models. [ABSTRACT FROM AUTHOR]

In this work, the authors present a digital twin (DT) to analyse the manufacturing process of structural composites by using resin transfer moulding (RTM). During RTM, a dry textile preform is impregnated with a polymer resin injected in a closed mould. RTM is one of the most used production methods for high-performance structural composites. The DT is focused on detecting in-homogeneous resin flow produced by race-tracking channels that divert resin flow to the outlet gates of the mould, producing dry spots and lack of impregnation. The DT core contains two surrogate models based on encoder/decoder deep learning architectures, providing the fast/accurate response necessary for interrogation during manufacturing. The first surrogate acts as the disturbance detector, providing the on-the-fly representation of the fabric permeability with the only information gathered by a set of five pressure sensors distributed over the mould surface. The second offers real-time representation of a set of quantities of interests (QoI): namely, the flow progress and the pressure field inside the mould. Training of both surrogates was performed with synthetic data generated by high-fidelity multi-physics simulations of the flow progress in a porous preform by following Darcy's law. Errors in the pressure field predictions of the surrogates are lower than 1 % with consultation time <50 ms, enabling encapsulation in the digital twin. The DT performance was evaluated by comparing the response against a set of RTM experiments for different race-tracking scenarios. The two most relevant novelties of the work are the use of the concept of the instantiated DT, which provides information on the current state of the process from data provided by a network of distributed sensors, and that this DT has been trained exclusively with synthetic data from multiphysics simulations while evaluated against experimental data from injection tests. [ABSTRACT FROM AUTHOR]

With the development of collaborative edge computing (CEC), the manufacturing market is gradually moving toward large-scale, multi-scenario, and dynamic directions. The existing scheduling strategies based on machine learning or deep learning are only applicable to specific scenarios, which is difficult to meet the requirements of dynamic real-time scheduling in multiple scenarios. The proposed digital twin technology provides a new solution for real-time scheduling of multiple scenarios. In this paper, a digital twin-oriented multi-scene real-time scheduler (GNN-RL) is proposed. This scheduler converts task sequences into node trees and sets up two learning layers. The first layer is an online learning representation layer, which uses GNN to learn node features of embedded structures in real time to boost large instances without additional training. The second layer is the online learning policy layer, which introduces imitation learning mappings into optimal scheduling behavior policies adapted to multiple scenarios. Finally, our approach is validated in several scenarios in 3D digital twin factories, such as computationally intensive, delay-sensitive, and task-urgent scenarios. Since the scheduler proposed in this paper learns general features of the embedding graph rather than instance-specific features, it has good generality and scalability, with good generalization and scalability, outperforming other scheduling rules and schedulers on various benchmarks. [ABSTRACT FROM AUTHOR]

The rapid advancement of deep learning and computer vision has driven the intelligent evolution of fire detection, quantification and fighting, although most AI models remain opaque black boxes. This work applies explainable deep learning methods to quantify fire power by flame images and aims to elucidate the underlying mechanism of computer-vision fire calorimetry. The process begins with the utilization of a pre-trained fire segmentation model to create a flame image database in various formats: (1) original RGB, (2) background-free RGB, (3) background-free grey, and (4) background-free binary. This diverse database accounts for factors such as background, color, and brightness. Then, the synthetic database is employed to train and test the fire-calorimetry AI model. Results highlight the dominant role of determining flame area in fire calorimetry, while other factors displaying minimal influence. Enhancing the accuracy of flame segmentation significantly reduces error in computer-vision fire calorimetry to less than 20%. Finally, the study incorporates the Gradient-weighted Class Activation Mapping (Grad-CAM) method to visualize the pixel-level contribution to fire image identification. This research deepens the understanding of vision-based fire calorimetry and providing scientific support for future AI applications in fire monitoring, digital twin, and smart firefighting. [ABSTRACT FROM AUTHOR]

This document discusses the emergence of deep learning as a transformative tool in the field of operations research (OR). Deep learning, a specialized category of machine learning algorithms, uses multiple layers to uncover hidden patterns and insights from large datasets. It has gained popularity due to its superior predictive performance compared to traditional machine learning models. The integration of deep learning applications into OR has the potential to enhance decision-making, productivity, sustainability, and optimization processes. However, there are challenges to address, such as data availability, interpretability, and computational cost. The document highlights the need for practical research and provides future research directions in areas such as supply chain management, demand forecasting, inventory optimization, and decision-making processes. Overall, deep learning has the potential to transform the field of OR, and this document aims to provide new perspectives for further research and practical applications. [Extracted from the article]

Social bot detection is crucial for ensuring the active participation of digital twins and edge intelligence in future social media platforms. Nevertheless, the performance of existing detection methods is impeded by the limited availability of labeled accounts. Despite the notable progress made in some fields by deep semi-supervised learning with label propagation, which utilizes unlabeled data to enhance method performance, its effectiveness is significantly hindered in social bot detection due to the misdistribution of individuation users (MIU). To address these challenges, we propose a novel deep semi-supervised bot detection method, which adopts a coarse-to-fine label propagation (LP-CF) with the hybridized representation models over multi-relational graphs (HR-MRG) to enhance the accuracy of label propagation, thereby improving the effectiveness of unlabeled data in supporting the detection task. Specifically, considering the potential confusion among accounts in the MIU phenomenon, we utilize HR-MRG to obtain high-quality user representations. Subsequently, we introduce a sample selection strategy to partition unlabeled samples into two subsets and apply LP-CF to generate pseudo labels for each subset. Finally, the predicted pseudo labels of unlabeled samples, combined with labeled samples, are used to fine-tune the detection models. Comprehensive experiments on two widely used real datasets demonstrate that our method outperforms other semi-supervised approaches and achieves comparable performance to the fully supervised social bot detection method. [ABSTRACT FROM AUTHOR]