Traditional numerical prediction of multiple environmental parameters in marine systems relies on complex physical numerical simulations. In recent years, the emergence of deep learning in marine environmental forecasting has challenged conventional physics-based approaches. Previous deep learning prediction models typically employ a single architecture to forecast multiple marine environmental indicators simultaneously, assuming homogeneous data distributions while neglecting two critical issues: 1) temporal inconsistencies in data collection timestamps across variables, and 2) distributional shifts caused by extreme weather events. Additionally, single-model frameworks suffer from prediction lag and associated temporal errors. To address these limitations, this study proposes DA-HAM, a novel hybrid model for marine environmental forecasting. The model implements data inversion tailored to individual marine variables to resolve timestamp misalignment in multi-source data collection, enabling variable-centric learning of intrinsic inter-variable dependencies. A distribution-aware attention mechanism dynamically adjusts attention scoring based on global data distribution patterns learned by the model, thereby compensating for over-unified distribution assumptions induced by normalization. Concurrently, a linear component captures cross-variable dependencies and disentangles future seasonal trends from general trend patterns. Final predictions are generated through adaptive integration of these components via learnable weighting factors. Experimental results demonstrate that DA-HAM achieves significant improvements in accuracy, stability , and trend fidelity compared to state-of-the-art Transformer-based models when processing multi-indicator datasets from designated marine regions. Notably, it exhibits enhanced robustness against data fluctuations induced by complex climatic variations and superior performance in long-term forecasting. These findings not only validate DA-HAM's potential in multi-indicator marine environmental prediction but also establish a methodological foundation for future research in marine multivariate systems.

To address the challenges of traditional marine meteorological prediction methods, which struggle to effectively capture intervariable correlations in multivariate time series data and suffer from insufficient prediction accuracy, this article proposes a multivariate short-term marine meteorological prediction model. First, an intelligent marine prediction fusion architecture is constructed, which is well-suited to artificial intelligence (AI) technologies. This architecture optimizes the process of marine meteorological data collection, processing, and analysis, providing a flexible and efficient infrastructure for short-term marine prediction. Second, an influence-based importance attention mechanism for meteorological variables is designed. By exploiting the differences in interactions among meteorological variables, it selects significant attention heads for computation, effectively reducing the model’s computational complexity and enhancing its response speed. Finally, a multivariate dimension prediction method for marine meteorology is proposed. By independently processing the time series of each variable, it enhances the capability to capture interactions among meteorological variables, thus improving the model’s understanding of and predictions for dynamic changes in marine meteorology. The experimental results show that the model can fully capture and analyze the complex relationship between variables in a multivariable marine meteorological environment, effectively improve the accuracy and efficiency of the prediction, and verify its application potential in marine meteorological prediction.

In recent years, the development of satellite IoT technology has been driving the continuous improvement of global IoT communication capabilities. However, the inherent characteristics of satellite networks, such as high latency, limited bandwidth, and unstable link quality, make traditional QoS management methods difficult to effectively adapt to their highly dynamic and frequently changing topological features. To address these challenges, we propose LDQoS, an innovative lightweight dynamic QoS optimization framework. This framework integrates a multi-constraint link evaluation model and a distributed dynamic routing optimization algorithm, which enhances end-to-end QoS performance while significantly reducing control overhead. Moreover, this paper introduces an innovative dynamic compression technique for segment routing headers (SRH) based on region partitioning, achieving a 50% path compression ratio and minimizing data plane overhead while ensuring route reachability. Experiments show that, compared with Dijkstra and SWAY, LDQoS reduces the average delay by 47.3% and 29.7% and increases throughput by 16.9% and 8.2%, respectively. These improvements provide important references for the practical deployment and operation of satellite IoT.

Abstract The high energy density of solar energy gives wireless sensor networks advantages in outdoor monitoring applications. However, long-term stable monitoring is challenging due to frequent weather changes, shading by buildings and trees, etc. The existing research usually uses two technologies to solve the above problems: (1) the energy prediction algorithm, and (2) the energy-aware routing strategy. However, in an actual deployment, frequent weather changes can significantly reduce the accuracy of the existing prediction algorithms. When using the algorithms as the support for energy-aware routing, the network lifetime is less than ideal. The existing routing strategies are in need of further improvement. Because of its lack of environmental adaptability, nodes consume energy quickly and have a high mortality rate. Therefore, aiming at the long-term stability of solar wireless sensor networks, this paper proposes a prediction algorithm based on classification and recurrent neural networks, and integrates the shadow judgement method from our previous research to correct the predicted values. Furthermore, we propose a routing optimization model that can flexibly adjust the target according to the solar intensity. The experimental results show that the prediction and routing scheduling algorithm can significantly improve the energy prediction accuracy (30–50%) and prolong the network lifetime (10–42%) in outdoor small sensor scenarios.

The scale of Internet routing is large. Satellite resources are limited and cannot afford millions of Internet routes. The method of planning new semantics or embedding locations for addresses is only applicable to IPv6. Aiming at the on-satellite deployment of existing IPv4 prefixes, for the first time, this paper converts the expression of address sets from traditional binary mask prefixes to decimal integer ranges, which can support precise control. Furthermore, a geographically based routing compression scheme (GRC) is proposed: Iteratively aggregates prefixes with the ground station (GS) as the center, and then corrects errors by sorting and constructing non-overlapping intervals. In addition, we present a fast interval lookup (FIL) algorithm that improves Binary Search. The results of processing all real prefixes on the Internet show that the average compression ratio of GRC can reach 17%, and the number of prefixes in 2022 drops from 942,000 to 153,000. The average lookup rate of FIL can reach 18.8Mlps. This can meet the requirements of wire-speed forwarding under common space link bandwidths.

Intrusion detection system (IDS) is an active defense mechanism implemented by the Internet of Things (IoT), which can identify the intrusion behavior and initiate alarms. However, there are concerns regarding the sustainability and feasibility to existing schemes when facing the increasing of threats in IoT. In particular, these concerns in terms of the increasing levels of adaptive performance and the insufficient levels of detection accuracy. In this paper, we present a novel deep learning method to address the aforementioned concerns. We detail the proposed convolution neural network model based on the developed feature fusion mechanism. Furthermore, we also propose a Symmetric logarithmic loss function based on categorical cross entropy. In addition, the proposed detection framework has been applied to GPU-enabled TensorFlow, and evaluated using the benchmark of NSL-KDD datasets. Extensive experimental results indicate that the developed model outperforms traditional approaches and has great potential to be applied for attacks detection in IoTs.

Recent years, a large number of open routing software architectures have emerged, such as XORP, ForCES, Ryu in SDN, reconfigurable platform, and so on. All of these architectures are open and comprising of components. They use components as the smallest resources, through the support of different assembly and release, to construct the complex routing and switching system. This paper take reconfigurable routing platforms for example, proposes a reliability valuating method based on Petri net and Fokker-Plank equations. Through the theory, it can verify the software component robustness ahead of time. Experiments show that the reliability evaluating results are coincident. The method of this paper solved the function robustness testing and reliability estimation problems of asynchronous interacted software in the reconfigurable routing platform with components dynamic loading and unloading. The practical significance is to reduce the workload of the testing staff and provides guidance to the routing software design to a certain extent.