Huafon Group (China)

Inspired by recent success of foundation models in vision and language domains, time series foundation models (TSFMs) have garnered increasing attention in general time series analysis tasks like finance, weather, healthcare, and power. However, given high heterogeneity and severe annotation scarcity in time series sensor data, how to unlock the potential of large-scale general-purpose TSFMs for downstream activity recognition tasks remains yet unexplored? This paper makes the first attempt to address this timely challenge by adapting the self-supervised pre-trained TSFM (i.e., MOMENT) to few-shot activity recognition. We introduce a simple and efficient Sensor-Prompt Tuning (SPT) strategy, which employs multiple convolution-based sensor-friendly filters with a gating mechanism to act as learnable soft prompts, which can dynamically adapt sensor input space to the frozen TSFM backbone, effectively bridging domain gap between pre-training general time series data with wearable sensor stream. Extensive experiments across three public activity recognition benchmarks demonstrate that our SPT achieves up to 15.5% performance gains over existing state-of-the-art baselines under few-shot scenarios, while considerably outperforming other mainstream fine-tuning strategies with smaller than 1% of backbone parameters. Practical cloud-edge inference latencies are measured. This work offers a new prompt-tuning perspective on how to adapt pre-trained TSFMs for wearable activity recognition tasks. Code will be released.

Từ tháng 2/2024 đến tháng 4/2025, tiến hành điều tra và thu thập 742 mẫu vòi voi đục thân chuối (Odoiporus longicollis) bị nấm ký sinh (29 mẫu) tại 5 xã/phường (Thanh Oai, Bình Minh, Quản Bị, Chương Mỹ và Yên Nghĩa) vùng trồng chuối ven sông Đáy thuộc Thành phố Hà Nội, trên các giống chuối ngự, chuối lá và chuối tiêu. Tỷ lệ sâu bị nấm ký sinh ghi nhận ở giai đoạn sâu non là 5,9-9,4%, nhộng 2,4-4,2% và trưởng thành 1,7-2,4%. Kết quả phân lập, nuôi cấy, mô tả hình thái và phân tích sinh học phân tử xác định loài nấm ký sinh là Metarhizium anisopliae. Loài này xuất hiện ở hầu hết tại các điểm khảo sát từ tháng 9 đến tháng 2 năm sau. Trong điều kiện phòng thí nghiệm, M. anisopliae gây nhiễm hiệu quả hơn so với ở ngoài tự nhiên, với tỷ lệ nhiễm sau 14 ngày đạt 18,5% (sâu non tuổi 3-4), 31,9% (tuổi 5-6), 12,6% (nhộng) và 10,4% (trưởng thành). Kết quả cho thấy, nấm M. anisopliae là tác nhân sinh học tiềm năng trong phòng chống đục thân chuối tại Hà Nội.

ABSTRACT It is well established that the mixture of sand and rubber powder (sand-rubber mixture [SRM]) is commonly used in embankments and foundations. The lightweight and high damping ratio of rubber can reduce the vibration in engineering practices. In this article, numerous simple shear tests involving unidirectional and bidirectional shear stress are conducted using the bidirectional simple shear apparatus to investigate the monotonic shear behavior of SRM. The specimens are consolidated under different conditions and are sheared monotonically. The results indicate that under drained conditions, the higher rubber content leads to a smaller normalized shear stress and a higher degree of noncoaxial angle during the early shearing stage, but the confining pressure slightly affects the normalized shear stress and degree of noncoaxial angle. In addition, the direction of consolidation shear stress (CSS) also affects the normalized shear stress and noncoaxial angle. The smaller angle between the CSS direction and the following shearing direction leads to a higher shear stress and a smaller noncoaxial angle. Under undrained conditions, the higher rubber contents and confining pressure reduce the stiffness of SRM and the development of pore water pressure.

Image classification, a core computer vision task, faces challenges in balancing long-range dependency capture, computational efficiency, and local detail preservation. Though Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) have seen great success, they suffer from limited receptive fields or quadratic computational complexity. Recent State Space Models (SSMs) enable efficient long-sequence modeling with linear complexity, but they overlook key local structural information. To solve this, we propose Local Information Awareness Vision Mamba (LIA-VMamba), a new image classification network integrating a Bidirectional State Propagation Module (BSPM) and an Adaptive Graph Convolution Enhancement (AGCE) module. BSPM aims to boost global context capture via bidirectional spatial information flow in the SSM framework. AGCE builds dynamic graph structures over image patches, using adaptive graph convolution and multi-head attention to model local dependency and fine-grained features. Extensive experiments on CIFAR-10 and ImageNet-1K benchmarks show a favorable accuracy-efficiency trade-off of our model, and clear ablation studies validate the effectiveness of our key modules.

Abstract In this study, FeSiAl powder was mixed with nano-TiO 2 powder to prepare composite powder cores. The influence of different TiO 2 addition content (0–4 wt%) on the performance of FeSiAl soft magnetic powder cores was systematically investigated via scanning electron microscopy (SEM), B-H analyzer, and other characterization techniques. The experimental results show that the total loss of the coated composite powder core is greatly reduced. With increasing TiO 2 content, the total core loss initially decreased and then increased, reaching an optimal value at 3 wt% TiO 2 addition, where the lowest total core loss ( P cv = 557 mW cm −3 , at 20 mT,1000 kHz) and best coating effect were achieved. The DC bias performance was significantly enhanced with increasing TiO 2 coating content, demonstrating a 28.65% improvement in percentage permeability ( μ %) under a 100 Oe external field, and achieving a maximum μ % of 83.57% (4 wt%) under the same condition. This work provides valuable insights for reducing high-frequency losses and enhancing DC bias characteristics in FeSiAl-based soft magnetic materials, offering a promising reference for their application in high-frequency scenarios.