Efficiently solving large-scale linear systems is a critical challenge in electromagnetic simulations, particularly when using the Crank-Nicolson Finite-Difference Time-Domain method. Existing iterative solvers are commonly employed to handle the resulting sparse systems but suffer from slow convergence due to the ill-conditioned nature of the double-curl operator. Approximate preconditioners, like SOR and Incomplete LU decomposition (ILU) provide insufficient convergence, while direct solvers are impractical due to excessive memory requirements. To address this, we propose FlashMP, a novel preconditioning system that designs a subdomain exact solver based on discrete transforms. FlashMP provides an efficient GPU implementation that achieves multi-GPU scalability through domain decomposition. Evaluations on AMD MI60 GPU clusters (up to$\mathbf{1 0 0 0 ~ G P U s}$) show that FlashMP reduces iteration counts by up to$16 \times$and achieves speedups of$2.5 \times$to$4.9 \times$compared to baseline implementations in state-of-the-art libraries Hypre. Weak scalability tests show parallel efficiencies up to 84.1 %.

Planetary gearboxes are critical components in wind turbine systems, where accurate fault diagnosis is essential for maintaining operational stability. However, current signal demodulation methods rely heavily on expert knowledge for demodulated component selection and feature extraction, often struggling to precisely identify fault types and severity levels. To overcome these limitations, this article proposes a novel demodulated component enhanced learning framework, which incorporates three key innovations: first, a proper rotation component reconstruction strategy guided by instantaneous frequency fluctuations to preserve key fault information; second, a method for constructing a feature vector by using characteristic frequencies derived from Amplitude modulation/frequency modulation models to extract key fault amplitudes; third, a hybrid architecture combining feature and model design, in which the long short term memory-convolutional neural network captures spatial and temporal patterns for fault diagnosis across different platforms. Experiments were conducted on two independent test platforms to demonstrate the framework’s superior performance, achieving 99.64% diagnostic accuracy across four gear fault types in case study 1, and 94.48% diagnostic accuracy across four gear fault types and three severity levels in case study 2. The results highlight the method’s strong generalization capability and its potential for practical application in wind turbine condition monitoring systems.

Exact sparse-triangular solvers are highly sequential and difficult to implement efficiently on GPUs with ILU preconditioning. Lower precision is crucial for reducing data movement and storage demands in memory-bound problems. However, current mixed-precision systems struggle to achieve performance gains for ILU preconditioning on multi-GPU platforms due to challenges in (1) utilizing two levels of parallelism and (2) minimizing off-chip memory bandwidth while maintaining accuracy. Additionally, these systems focus on scalar operations and lack support for point-block matrices, which arise naturally in multiphysics problems and require tailored algorithm designs. To address these challenges, we propose MIST, a novel Mixed-precision Iterative Sparse-Triangular solver optimized for GPUs to accelerate preconditioning in Krylov methods. We (1) implement an efficient mixed-precision Jacobi iterative local solver to harness single-GPU parallelism and scale it to multi- GPU via domain decomposition, and (2) design a BSpMVA kernel to reduce bandwidth while achieving high double-precision accuracy. Integrated into a widely-used numerical library, MIST offers end-to-end support for solving sparse linear systems, balancing efficiency and convergence. Experimental results show that MIST provides a 3.38× average speedup over cuSPARSE�s exact sparse-triangular solver, with an additional 1.37× speedup when using low-precision, while maintaining robustness

Multimodal image fusion is one of the important research directions in the field of multimodal fusion. This technique can realize image and data enhancement by using complementary multimodal images and be widely used in medicine, industry, security and fire protection, automatic driving and consumer electronics. In this work, we propose a transformer-based universal fusion (TUFusion) algorithm, and it has a multidomain fusion capability. The advantage of TUFusion algorithm is the design of hybrid transformer and convolutional neural network (CNN) encoder structure and a new composite attention fusion strategy, which has the ability of global and local information integration. Compared with the classical state-of-the-art multimodal image fusion methods, the experimental result on multidomain data sets showed that the TUFusion algorithm has certain universality in image fusion. Meanwhile, the TUFusion algorithm we proposed achieves good values on peak signal to noise ratio (PSNR), root mean square error (RMSE) and structural similarity index measure (SSIM). The code of the TUFusion algorithm in this article is available athttps://github.com/windrunners/TUFusion.

Abstract Multi‐view image fusion can be used to extract features from redundant and complementary multisource images. And the technique of obtaining high quality fusion images has become one of the research hotspots for image processing. In order to realize defect detection and intelligent grinding smoothly, multi‐view fusion technology was applied in the field of overexposure and underexposure industrial welds, achieving high quality image enhancement. When preparing the data set of multi‐view images, a hybrid registration algorithm with high matching ability is proposed. The data set of overexposure and underexposure weld images was obtained successfully by using the registration algorithm. In order to improve the fusion ability of overexposure and underexposure industrial welds, we propose a novel multi‐view image fusion algorithm based on deep learning. The multi‐view fusion algorithm uses an autoencoder network structure, and its innovation lies in a parallel branch network with lightweight structure and strong generalization ability. The experimental results demonstrate that compared with other classical multi‐view algorithms, our proposed algorithm gets the best parameters on the industrial weld data set in peak signal to noise ratio (PSNR) and root mean square error (RMSE), reaching 59.12 and 0.084, respectively. And the ablation and performance comparison experiments verify that the proposed parallel branch network has better generalization ability and fusion accuracy than other classical multi branch networks.

A novel network paradigm of edge computing, namely, two-tiered wireless sensor networks (TWSNs), has been proposed by researchers in recent years for its high scalability and robustness. However, in the TWSNs-based edge computing systems, the storage nodes, which are located at the upper layer of the systems, are prone to be attacked by adversaries because they play a key role in bridging sensor nodes and Sink, which may lead to the disclosure of all the data stored on them as well as some other potentially devastating results. In this article, we study the integrity-and-privacy preservation problem for spatial- temporal Top-k queries in the TWSNs-based edge computing systems and propose a sequence-encryption-based lightweight scheme named sequence-encryption-based lightweight scheme for securing spatial-temporal Top-k queries (SLS-STQ) to solve the problem. In SLS-STQ, three algorithms, namely, the report preparation algorithm, the query processing algorithm, and the integrity verification algorithm, are designed for the sensor nodes, the storage nodes, and Sink, respectively. The theoretical analysis shows that SLS-STQ is able to achieve both integrity validation and privacy preservation with low computational complexity, and the simulation results show that SLS-STQ is much more efficient than the related state-of-the-art schemes.

The existing search engine system almost based on keywords from users inputting. In a network environment, people can make use of the pc or mobile device for information retrieval. The way of inputting keywords has lasted for more than 20 years until Siri appeared in 2011. Siri can do information retrieval and process by voice actions. Compared to the traditional search mode, this is a huge step forward. Siri solved some limitations of the current search engines. But people need a new system, at any time and place, and fast search and process information. In this paper, we described new search engine system based on computer vision. Containing the information search and problem solve two basic parts. And for the needs of the people in daily lives, designed a number of convenience functions. System can use web camera to read and identify information. With the aid of network, maintain a real-time search and feedback. It is noteworthy that mobile devices include the existing smart phones and tablet PCs for now. These mobile devices are an appropriate choice. Like Google Glass. However, the best choice is new concept of mobile devices in the near future. After all, the existing equipment are still many restrictions.