Using a dataset of 114 meteorological stations in the Yangtze River Basin from 1980–2019, the standardized precipitation evapotranspiration index (SPEI) was calculated based on the Penman-Monteith evapotranspiration model for multiple time scales, and the spatial and temporal evolution characteristics and driving factors of drought in the Yangtze River Basin were analyzed by combining spatial and temporal analysis methods as well as geodetector. The main results obtained are as follows: (1) The climate of the Yangtze River Basin is an overall wet trend, and the trend of summer drought is more similar to the annual scale trend. (2) Most areas in the Yangtze River Basin showed mild drought or no drought, and there is little difference in drought condition among the Yangtze River Basin regions. The areas with drought conditions are mainly distributed in the southwest and east of the Yangtze River Basin. (3) There are significant seasonal differences in drought conditions in all regions, and the drought condition is more different in autumn compared to spring, summer and winter. (4) The average annual precipitation and elevation factors are the dominant driving factors of drought in the Yangtze River Basin, and the double-factor interaction has a greater influence on the drought variation in the Yangtze River Basin than the single-factor effect, indicating that the difference of drought condition in the Yangtze River Basin is the result of the combination of multiple factors.

Software trustworthiness allocation and reallocation are the symmetry of software trustworthiness measure. They can provide the optimization scheme for trustworthiness development and improvement, according to the requirements. The existing allocation and reallocation models do not consider the absolute majority of software trustworthiness classification; therefore, they cannot be very accurate in the allocation and reallocation of software trustworthiness. In this paper, improved allocation and reallocation models are constructed, which can resolve the above problem, and their polynomial solving algorithms are designed. At the same time, a demonstration application of the improved models and algorithms is given, and the trustworthiness enhancement specification of spacecraft software, based on factory reports, is established, including trustworthiness development specification and trustworthiness improvement specification. This enhancement specification provides a scientific and reasonable theory and method for the delivery acceptance of spacecraft software, from qualitative to quantitative grading acceptance, and furnishes a standard guarantee for the trustworthy development and improvement of such software.

MRI (magnetic resonance imaging) technology aims to map the internal structure image of organisms. It is an important application scenario of Non-Uniform Fast Fourier Transform (NUFFT), which can help doctors quickly locate the lesion site of patients. However, in practical application, it has disadvantages such as large computation and difficulty in parallel. Under the architecture of multi-core shared memory, using block pretreatment, color block scheduling NUFFT convolution interpolation offers a parallel solution, and then using a static linked list solves the problem of large memory requirements after the parallel solution on the basis of multithreading to cycle through more source code versions. Then, manual vectorization, such as processing, using short vector components, further accelerates the process. Through a series of optimizations, the final Random, Radial, and Spiral dataset obtained an acceleration effect of 273.8×, 291.8× and 251.7×, respectively.

Accurate sleep staging plays a pivotal role in the diagnosis and treatment of sleep-related disorders, yet manual annotation remains a costly task. This study introduces a self-supervised learning approach for sleep staging, requiring minimal labeled data. It is further supported by a novel convolutional neural network framework, utilizing single-channel EEG data for accurate sleep stage identification. Extensive experiments conducted on three datasets—DOD-O, sleep_edf, and challenge2018—demonstrate the efficacy of the proposed method. The results indicate that the self-supervised learning method based on single-channel EEG signals notably enhances the network's feature learning capabilities, endowing the model with greater generalizability and robustness. The performance in sleep staging tasks surpasses fully supervised methods, achieving accuracies of 71.873%, 84.02%, and 79.08% on the three datasets, respectively.

Currently, research on Chinese electronic medical record named entity recognition mainly focuses on the impact of Chinese characters on recognition, and enhances the recognition ability of models with external information such as dictionaries, ignoring the role of mutual positional information between characters. A new method that that utilizes the RoFormer model combined with multi-feature fusion is proposed to address issues such as fuzzy boundaries and strong terminology in medical entities. Use the RoFormer model in the embedding layer to capture the relative positional information between words, while extracting radical and lexical information to enhance semantic representation. Send the fused vector representation to the BiLSTM layer and CRF layer for training. Experiments were conducted on the publicly available medical named entity recognition datasets CCKS2019 and CCKS2020, with F1 values reaching 89.38% and 88.52%, respectively. The results indicate that compared with existing medical entity recognition models, the method of combining relative position information with external information significantly improves the recognition performance.

While the Triton compiler has revolutionized GPU kernel development, its deployment on general-purpose CPUs struggles to fully utilize the underlying hardware capabilities. This is primarily due to the semantic gap between Triton’s SPMD execution model and CPU vector architectures, which leads to suboptimal utilization of vector units during complex memory accesses. In this paper, we present a comprehensive compilation and runtime co-optimization framework for Triton-CPU, specifically targeting Vector Length Agnostic architectures (VLA) like ARM SVE. At the compiler level, we propose a novel semantic reconstruction and explicit base-offset decoupling strategy, enabling native VLA gather/scatter generation and eliminating scalar loop overheads. At the runtime level, we introduce a Machine Learning-driven thread scheduling model to optimally orchestrate the synergy between Thread-Level Parallelism and Vector-Level Parallelism. Extensive evaluations on an ARM-based multi-core processor demonstrate that our framework achieves up to a 2.0× throughput improvement for compute-bound GEMM operators (peaking at 346 GFLOPS), notably outperforming the hand-optimized OpenBLAS library by up to 1.54× at small-to-medium scales. Additionally, it delivers a 1.7× speedup for element-wise workloads. Furthermore, our optimizations saturate memory bandwidth (up to 55 GB/s) for memory-bound operators with zero compilation bloat, establishing a robust, high-performance foundation for deploying deep learning models on general-purpose CPUs.