Fault diagnosis of electromechanical equipment faces three major challenges. At the signal level, the multidirectional and multimodal characteristics of equipment signals prevent a single sensor from fully characterizing equipment health states. At the engineering level, sharp variations in rotational speed induce domain shift phenomena, which further degrade diagnostic accuracy. At the application level, fused multi-source features provide strong representations but often exhibit limited generalizability and entail high deployment costs. In contrast, current signals are readily available and inexpensive to acquire, but their capability to represent mechanical faults is relatively weak. To address these challenges, this paper proposes the Synchronization Induction Network Based on Progressive Microelement Strategy and Hierarchical Dimensions (SI-CDN). SI-CDN incorporates a mutual-attention-driven information coupler that dynamically models correlations among signals collected at different measurement locations, thereby enabling collaborative representation learning from multiple signal sources. To mitigate the effects of rotational-speed variations, a progressive infinitesimal embedding and fractal reconstruction strategy is developed to iteratively refine feature representations through convolution and reconstruction operations. This strategy captures temporal patterns associated with rotational-speed variations while suppressing their adverse effects on feature extraction. Furthermore, an advanced similarity loss function is integrated into a cross-modal synchronization-induced distillation framework to dynamically transfer knowledge from multimodal fused features to current-signal features, thereby enhancing the representation capability of the current model. Experiments on two representative electromechanical systems demonstrate that SI-CDN achieves state-of-the-art fault diagnosis performance.
Single-pass linear array synthetic aperture radar (SAR) systems face a fundamental challenge in three-dimensional imaging: severely insufficient resolution in the elevation direction due to physical aperture limitations. To overcome this bottleneck, this study proposes a hybrid imaging framework combining the Back-Projection (BP) algorithm with Graph Signal Processing (GSP). This framework first employs the BP algorithm to achieve two-dimensional focusing in the range-azimuth plane [1]. Subsequently, a novel GSP method is introduced to process signals within each two-dimensional resolution cell, enabling super-resolution direction-of-arrival (DOA) estimation in the elevation direction [2, 3]. Analysis of real experimental data from a Texas Instruments IWR1843 sensor demonstrates that the proposed GSP method exhibits multiple advantages. It shows superior robustness in challenging scenarios with low SNR or closely spaced target angles. The study concludes that the GSP framework introduces a new physics-driven paradigm for signal modeling in radar computational imaging, offering an effective and promising complement to traditional sparse recovery methods [4].
To overcome the limitations of insufficient information utilization, suboptimal imaging quality, and limited scene adaptability inherent in existing end-to-end sparse signal imaging within synthetic aperture radar (SAR) imaging networks, this paper proposes a SAR imaging method that integrates deep unfolding-based sparse echo recovery with data-driven modeling. In contrast to conventional unrolled networks for SAR imaging that directly reconstruct images from sparse echo signals, the proposed approach utilizes a deep unfolding framework to first recover the sparse echoes into fully sampled signals, followed by high-resolution imaging based on the completed data. This methodology effectively harnesses the intrinsic information embedded in the raw echo signals, thereby significantly improving both signal utilization efficiency and final image quality. Experimental evaluations using a 25
In this study, a composite catalytic layer (CL) featuring multiple bubble nucleation sites was constructed by two-step electrodeposition and phosphorus doping to regulate bubble behavior on the electrode surface during oxygen evolution reaction (OER). Firstly, the porous nickel framework layer constructed in the first electrodeposition provided ample anchoring sites for nickel-iron layered double hydroxide precursors in the second electrodeposition, thereby increasing the loading and accessibility of active species. Secondly, the resulting CL exhibited an abundant microporous structure. On one hand, this structure was expected to enhance capillary-driven electrolyte infiltration into the CL pores, which could improve electrolyte contact with active sites, increase site exposure, and enable efficient electrolyte renewal. On the other hand, the high tortuosity of the CL may create a spatial confinement effect on dissolved oxygen (DO), slowing DO diffusion into the bulk solution during OER. This could favor local DO accumulation near the electrode surface, providing a possible condition for bubble formation. Additionally, the porous structure with ample exposed edges and defects forms discontinuous "gas-liquid-solid" three-phase interfaces, which are likely to provide favorable interfacial regions for bubble formation, weaken bubble retention, and facilitate gas release. Finally, Ni2P and Fe2P heterojunctions formed by phosphorus doping enhanced CL conductivity and facilitated electron transport during OER. Ultimately, the optimized NiFeP@NF-skeleton exhibited overpotentials of 306 and 418 mV at 100 and 300 mA cm(-2), respectively, and showed acceptable operational durability under the tested conditions, outperforming the commercial IrO2-loaded electrode.
Recently, 3D Gaussian Splatting explicitly represents scenes and synthesizes high-quality novel views with impressive performance. However, reconstructing accurate Gaussian geometry becomes extremely challenging when using pure RGB images with few-shot inputs. We propose 2.5D-GS, which projects Gaussians into structured 2D spaces and utilizes the 2.5D representations from monocular models to separately optimize the projected depth and normal maps, ultimately achieving consistent and accurate Gaussian geometry. First, we ensure the spatial accuracy of Gaussians with Depth Plane Constraints. Since monocular depth maps construct only rough shapes, Normal Plane Constraints are then applied to refine the orientations of the Gaussians and enhance surface connectivity. Additionally, we introduce Density Ratio-Based Pruning to eliminate redundant Gaussians generated during optimization, leading to compact and efficient scene representations. Extensive experiments on the LLFF, DTU, Blender, and Mip-NeRF360 datasets demonstrate that 2.5D-GS accurately reconstructs scene geometry and renders high-quality novel views with sparse inputs.
