To address the coupled degradation of tulip contacts in 10 kV withdrawable switchgear under the combined effects of insertion–withdrawal wear, spring relaxation, increasing contact resistance, and localized overheating, this paper proposes a physics-informed thermal–electrical–mechanical feature fusion network (PI-TEMFNet). The method extracts multi-rate features, including temperature rise, contact resistance, contact force, and vibration, and employs reliability-aware masked cross-modal attention to fuse asynchronous and missing information. Contact-resistance relationships, the heat-balance equation, and a monotonic degradation constraint are embedded into the joint prediction of degradation stage, health index, and probabilistic remaining useful life (RUL). Accelerated degradation tests on 12 GC5 specimens and leave-one-specimen-out cross-validation show that the model achieves an Accuracy of 96.4%, a Macro-F1 of 96.1%, and an AUC of 0.988, with a health-index RMSE of 0.041. The MAE, RMSE, and MAPE for RUL prediction are 21.6×103 cycles, 29.4×103 cycles, and 5.8%, respectively, while the coverage probability of the 90% prediction interval is 91.3%. Under random missing rates of 20% and 40%, the MAPE remains 6.3% and 7.2%, respectively, demonstrating high accuracy and strong robustness.
Cross-border e-commerce requires dynamic, heterogeneous credit evaluation, but static models fail to capture temporal and network interactions. This paper proposes an Improved Temporal Graph Convolutional Network (ITGCN) for dynamic credit assessment. ITGCN constructs a dynamic heterogeneous graph integrating sellers, buyers, logistics, and payment nodes, with a relationship-aware graph convolution operator and a temporal decay attention mechanism to model heterogeneous dependencies. A multi-scale time series module captures both short-term fluctuations and long-term credit trends, while a cross-domain feature alignment strategy handles inconsistencies in multi-source data granularities. The experimental results on two public fraud detection datasets show that the ITGCN achieves an AUC of 0.938 (D1) and 0.934 (D2), outperforming CAREGNN by 1.8 per cent and 2.1 per cent, and an F1 score of 0.876 and 0.871, representing improvements of 2.2 per cent and 2.6 per cent. Ablation and robustness tests confirm its effectiveness and resilience. ITGCN offers a reliable solution for cross-border credit assessment.
Finger vein recognition is a promising biometric technology for identity authentication, but practical systems are often constrained by limited computing and storage resources. To improve the recognition performance of lightweight models under these constraints, this paper proposes PSE-MobileNetV2, a MobileNetV2-based finger vein recognition network with partially embedded Squeeze-and-Excitation (SE) channel-attention blocks. Instead of adding SE blocks to every inverted residual stage, the proposed model inserts them only into stages 4-7, where the feature maps encode higher-level vein representations. This design strengthens channel-wise feature recalibration while avoiding unnecessary early-stage computational overhead. In the reported closed-set, image-disjoint classification experiments on the FV-USM and SDUMLA public datasets, PSE-MobileNetV2 achieves accuracies of 99.49% and 97.06%, respectively, exceeding the MobileNetV2 baseline by 0.31 and 0.84 percentage points. These results suggest that selective SE placement can improve the accuracy-efficiency trade-off of MobileNetV2. However, the current study does not establish verification performance or statistical significance across repeated runs.
With the rapid iteration of ubiquitous sensing, edge computing and artificial intelligence technologies, intelligent perception has become the core driving force for accelerating the digital transformation and intelligent upgrading of various social scenarios. Traditional smart scenario construction schemes suffer from low data fusion efficiency, high resource occupation, poor scene adaptability and unclear standardized application modes, which severely restrict the large-scale popularization and practical landing of smart service systems. This paper takes intelligent perception technology as the core enabling carrier, systematically explores the core bottlenecks restricting the construction efficiency of smart scenarios, proposes a multi-layer collaborative efficiency optimization framework oriented to heterogeneous perception data, constructs a standardized and scalable smart scenario application paradigm integrating perception, analysis, decision-making and execution. On the basis of sorting out the existing research achievements of intelligent perception and smart scene construction, this paper innovatively designs a weighted adaptive perception data fusion algorithm and an edge-cloud hierarchical resource scheduling optimization model, and derives complete mathematical derivation formulas. Combined with business logic, the overall technical flow chart and module deployment architecture diagram of the optimized system are constructed. A multi-group comparative experiment is carried out based on actual urban community, industrial park and campus smart scenario datasets. Experimental results verify that the proposed optimization strategy can effectively reduce system delay, reduce computing resource consumption and improve perception recognition accuracy. The constructed universal application paradigm has strong universality and expansibility, which can provide theoretical reference and technical support for the efficient construction, rapid deployment and standardized promotion of diversified smart scenarios, and lay a solid foundation for the deep integration of intelligent perception technology and smart social governance, smart industry and smart life fields.
Nonlinear resonance of CNT-reinforced cantilever beams carrying a concentrated tip mass is important forlightweight pipeline-type structures, but the combined influence of finite rotations, CNT reinforcement, and terminal inertia remains insufficiently clarified. This study develops a geometrically exact reduced-order model for a CNT-reinforced cantilever beam subjected to harmonic transverse excitation. The effective elastic properties are evaluated using the Mori–Tanaka homogenization scheme, while the governing nonlinear equations are nondimensionalized, reduced by Galerkin projection, and analyzed near primary resonance using the method of multiple scales. For the benchmark case, the first three dimensionless natural frequencies obtained by the present model are 19.2565, 64.1797, and 117.5724. The nonlinear analysis shows a mode-dependent response: the first isolated resonance branch exhibits hardening behavior, whereas the second and third branches exhibit softening behavior due to the sign change of the effective nonlinear frequency-shift coefficient. Increasing the excitation amplitude from F¯=0.1 to 0.6 enlarges the resonance amplitude and promotes jump type multi-valued responses. Increasing CNT content shifts the resonance to higher frequencies and reduces peak amplitudes, whereas increasing the tip-mass ratio from ∆ = 0 to 0.3 lowers the resonance frequencies and amplifies the response, particularly in the lower modes. These results clarify how CNT reinforcement and terminal inertia tune the nonlinear frequency response characteristics of cantilever nanocomposite beams.
