When a carrier (such as drones, ships, and vehicles) moves in extreme environments, the visibility of satellites may be lost, leading to a temporary or prolonged loss of lock on Global Navigation Satellite System (GNSS) signals. In such scenarios, an integrated navigation system is forced to switch to a pure Inertial Navigation System (INS). However, prolonged reliance on inertial navigation alone results in the accumulation of errors and a rapid decline in navigation accuracy. To address the rapid decline in INS accuracy after GNSS signal loss, a fusion navigation technology of GNSS and INS assisted by Transformer networks is proposed. When the GNSS signal is locked, the Transformer network utilizes current INS information and GNSS incremental data (the change in GNSS position information between two adjacent time periods) to train a mapping relationship between the two. When the GNSS signal is lost, the Transformer network leverages the previously established mapping relationship to predict GNSS incremental information based on the current INS data, and then integrates the INS information with the predicted GNSS data for navigation. Simulation results demonstrate that the Transformer network-assisted GNSS/INS fusion navigation technology can provide stable and reliable navigation signals even under conditions of temporary or prolonged GNSS signal loss. Furthermore, the Transformer network-assisted fusion navigation method offers a reference for other network-assisted fusion implementations.

In smart grids, the presence of numerous high-power electrical devices and communication sensing equipment severely hinders signal transmission in positioning systems. To address the challenge of accurately locating weak signals in a Reconfigurable Intelligent Surface (RIS)-assisted Non-Orthogonal Multiple Access (NOMA) system under interference from multiple base stations and communication users, this paper considers the impact of multiple small base stations and multiple users in a smart grid environment. A horizontal positioning error of the target user is used as the evaluation metric. While ensuring the Quality of Service (QoS) for communication users, the proposed method jointly optimizes base station power, multi-user interference, and power allocation factors. The Lagrangian dual method and sub-gradient approach are employed to solve the non-convex optimization problem caused by multiple users and small base stations. Simulation results demonstrate that, under the same resource allocation, the proposed RISNOMA integrated sensing and communication system significantly reduces the average positioning error compared to traditional NOMA-based system.

To address the issue of current command and control network key node recognition methods relying on expert knowledge, a method based on convolutional neural networks from the perspective of communication reconnaissance is proposed. Powerful feature extraction capabilities of convolutional neural networks are leveraged to develop an intelligent paradigm for key node recognition. First, the communication relationship information between nodes is transformed into a multi-dimensional information matrix using feature engineering. Then, inspired by the Finite Impulse Response (FIR) filter structure, a Finite Impulse Response Squeeze and Excitation (FIRSE) neural network is proposed. Finally, a dynamic peak detection method is introduced to improve the training strategies and obtain optimal neural network parameters. Experimental results show that compared with typical machine learning and deep learning-based recognition methods, the proposed method offers higher identification accuracy.

In industrial networks, more and more intelligent applications put forward new requirements for deterministic guarantee ability of end-to-end transmission. Most of current research works on deterministic technologies focus on guarantee of their own network certainty, but ignore end-to-end determinism requiring cooperation of multiple deterministic technologies. Based on application scenario of nuclear power industry, a wide-area end-to-end deterministic network architecture with Flexible Ethernet (FlexE) and Time-Sensitive Networking (TSN) fusion is proposed, and an architecture is elaborated from two aspects of control plane function and data plane cooperative scheduling. In real wide area network environment, based on the dual-motor collaboration and industrial machine vision application scenario, the performance of the architecture is tested and verified. Results show that the architecture can meet the needs of applications for wide-area end-to-end deterministic networks, and offer good application value.

In recent years, Transformer-based visual models (e. g. , Swin Transformer) show good prospects in visual tasks, however, these methods usually focus on reducing signal distortion between original and reconstructed data, while ignoring perceptual quality. Considering that the conventional Mean Square Error (MSE) loss fails to reflect perceptual and semantic quality effectively, we propose a weighted loss function combining MSE and Learned Perceptual Image Patch Similarity (LPIPS), and accordingly construct a Swin Transformer-based semantic communication framework, called Swin Transformer with LPIPS-based Joint Source-Channel Coding (STL-JSCC) method, which significantly enhances image reconstruction quality and semantic consistency. For performance evaluation, two semantic-aware metrics are introduced: the Images Semantic Deviation (ISD) value and Iamges Semantic Similarity(ISS). These indicators form a joint perceptual-semantic evaluation system, which breaks through the limitations of traditional evaluation methods. Experimental results show that the proposed STL-JSCC outperforms other models in all the indexes, verifying the significant potential and advantages of the proposed method in improving the image reconstruction quality and semantic extraction capability.

Seepage monitoring is crucial for the safe operation and maintenance of dams. Traditional dam observation methods suffer from significant random errors and insufficient inspection frequency during flood seasons. To address these limitations, this study proposes an infrared thermography-based unmanned aerial vehicle inspection system for detecting surface seepage on dam bodies. First, an image dataset of seepage-affected areas on the dam surface was collected and established using an infrared camera. Then, an improved Mask Region-based Convolutional Neural Network(Mask R-CNN) framework was employed to extract seepage region data, enabling rapid detection of surface seepage. Subsequently, binary processing was applied to quantify the seepage area. Finally, the proposed method was validated on the downstream face of a hydropower station. Experimental results demonstrate that the proposed approach reduces the inspection cycle by 80% compared to traditional methods while maintaining sufficient accuracy for routine dam monitoring. This study provides a novel technique for seepage detection and quantitative analysis, offering a new solution for dam leakage inspection and seepage-related damage assessment.

Compared with a different-color backgrounds, recognizing and detecting cucumber fruits under uniform-color backgrounds remains a key challenge due to limited distinguishing features and increased susceptibility to occlusion and background interference. To address this, we propose YOLO-ACG, a detection network based on YOLOv11n. An Adaptive Dynamic Downsample (A-Down) module is introduced, combining deformable convolution and channel attention to achieve adaptive cross-scale feature sampling. A Ghost_HGNetV2 architecture is designed, where the High-resolution Group Stem (HGStem) reduces input channels to extract efficient intrinsic features, and the Ghost_HGBlock applies knowledge distillation to enhance feature representation. A Context and Spatial Feature Calibration Network (CSFCN) network structure is introduced, which includes Context Feature Calibration (CFC) and Spatial Feature Calibration (SFC). The CFC module aggregates context information relevant to each pixel, while the SFC module leverages calibrated spatial features to ensure accurate understanding of spatial layout the image. Together, they enable the network to more precisely distinguish cucumber fruits from backgrounds with similar colors. Experimental results show that the improved model achieves 4.64 percentage points increase in precision, recall by 5.07 percentage points, F1 by 4.89 percentage points, and mAP by 4.48 percentage points. Ablation and comparative experiments confirm that YOLO-ACG significantly reduces false positives and missed detections, offering effective technical support for cucumber fruits recognition in complex, uniform-color environments.