Internet Technology Letters最新文献

In Industrial Internet of Things (IIoT) environments, smart manufacturing systems generate massive textual data (equipment logs, maintenance reports, etc.) requiring accurate similarity analysis for fault diagnosis and predictive maintenance. Traditional methods underperform in Industry 5.0 scenarios due to technical vocabulary and domain-specific language. This paper presents RoBERTaSN, an enhanced model combining RoBERTa with a Siamese network, featuring self-attention and dual pooling optimized for industrial texts. It enables precise similarity calculations between fault descriptions and historical records. Experiments on industrial datasets (e.g., equipment fault logs, maintenance reports) yield 94.2% accuracy in fault diagnosis text matching—7.8% higher than traditional TF-IDF (86.4%) and 6.0% higher than mainstream pretrained models (BERT: 88.2% accuracy; BiMPM: 84.67% F1-score), addressing semantic challenges in smart factories and advancing Industry 5.0's human–machine collaboration and intelligent decision-making goals.

In the era of pervasive wireless communication, the need for efficient and accurate text recognition systems is growing, especially for applications involving edge devices in resource-constrained environments. This paper proposes a lightweight AI-based approach for English text recognition, leveraging a hybrid model combining convolutional neural networks (CNN) and gated recurrent units (GRU). The model effectively handles noisy wireless signals by capturing both spatial and temporal features from modulated signals. We incorporate techniques such as pruning and depthwise separable convolution (DSC) to reduce the model's size, making it suitable for deployment in wireless communication systems. Experimental results demonstrate that the proposed model outperforms several state-of-the-art methods, including traditional modulation recognition and deep learning-based alternatives, in terms of both recognition accuracy and model efficiency, even in low signal-to-noise ratio (SNR) conditions. The proposed model offers a promising solution for real-time text recognition in wireless communication environments.

Within the industrial internet of things (IIoT) ecosystems, apparel manufacturers face the dual challenge of integrating high-velocity consumer feedback streams from e-commerce platforms and translating them into real-time, high-fidelity demand forecasts. This study presents an IIoT-native framework that employs random forest regression (RFR) to fuse multi-modal review features—sentiment polarity, key phrases, and aggregate ratings—collected via edge gateways from 1100 men's garments on JD.com. Innovatively, the proposed framework not only outperforms traditional linear models such as ordinary least squares (OLS) and multiple linear regression (MLR) in terms of predictive accuracy but also demonstrates robustness to noise and outliers across heterogeneous product categories. The cloud-hosted RFR model achieves an R² of 0.9442 and root mean square error (RMSE) of 105.76, representing a 5.6% and 35.9% improvement over MLR and OLS in RMSE, respectively. This study provides the first multi-category empirical evidence that fusing review-level sentiment, key phrases, and ratings via RFR yields significant enhancements in IIoT-scale apparel demand forecasting.

The next generation of Wireless Communication Networks (WCNs), such as 5G and 6G, require highly adaptive, intelligent, and efficient transmission mechanisms to meet the demands of low latency, high throughput, and robust Quality of Experience (QoE). This paper introduces a novel framework that integrates Large Models (LMs), particularly transformer-based deep learning architectures, with Digital Twin Networks (DTNs) for predictive and real-time optimization in WCNs. The proposed LM-enhanced DTN architecture enables advanced capabilities such as traffic classification, predictive scheduling, quality-aware transmission, and failure forecasting. Experimental evaluations using real-world telemetry datasets demonstrate the superiority of the LM-powered system in achieving over 98% classification accuracy and enhancing 12.4% improvement in QoE in congested scenarios. Additionally, a case study in industrial networks illustrates the effectiveness of this approach in predictive maintenance and adaptive traffic management. This work paves the way for self-optimizing, intelligent wireless networks by harnessing the cognitive power of large AI models in virtual network replicas.

In modern intelligent construction automation systems, due to the interference of network delay, task synchronization between devices is hindered, resulting in uncoordinated operations between robots and collisions, and task conflicts. This paper builds an intelligent construction automation system based on a 6G network, using the low latency and high bandwidth characteristics of a 6G network to effectively solve the delay problem in task synchronization and collaborative work. Its innovative combination of network slicing technology and edge computing methods customizes specific network resources for different application scenarios to minimize latency. The fusion of convolutional neural network (CNN) and long short-term memory (LSTM) models can make better predictions, and combined with the deep reinforcement learning model (DRL), a path planning plan can be formulated based on the prediction results to avoid collision problems in the robot's work. The experimental results show that after the 6G network optimization system, the task scheduling rate of the robot can reach 0.95, compared with 5G network optimization, which only reaches 0.90, and the collision problem of the robot can be well avoided. The collision rate after optimization can approach 0, which can ensure the smooth progress of the construction process and the safety and reliability of task execution.

Speech emotion recognition based on edge computing technology and deep learning can effectively assist in improving the quality of English short passage reading instruction. Restricted by limited computing resources of different edge devices, existing deep models pose a huge challenge for mobile deployment. To alleviate this issue, this paper proposes a novel hybrid speech emotion recognition model in multi-access edge intelligence scenarios. Firstly, we extract the Log Mel features from the speech signal collected by different clients' microphone sensors. Then, on the cloud platform, we deploy an efficient feature extraction backbone by exploiting 1D convolution operations, a minimal gated unit (MGU) module, and a Mamba module, which is introduced for exploiting long-range dependencies with linear computational complexity. We conducted extensive comparative experiments on the public dataset and our own English reading sentiment dataset, and our proposed model achieved the highest recognition performance.

The increasing ubiquity of 5G/6G wireless communication networks has created unprecedented opportunities for multi-community collaboration and virtual engagement. However, existing platforms often fail to support scalable, context-aware, and efficient interaction across decentralized social groups. In this paper, we propose LM-MCVIS (Large Model-Based Multi-Community Virtual Interaction Scheme), a novel framework designed to facilitate personalized content exchange and context-aware message routing in wireless community networks. LM-MCVIS integrates three key components: (1) an Edge-Aware Prompt Compression (EPC) module that semantically distills conversation inputs to reduce wireless transmission overhead; (2) a Community State Encoder (CSE) that models dynamic group structures and latent social contexts; and (3) a Federated Reinforcement Optimizer (FRO) that enables privacy-preserving, feedback-driven content routing. We evaluate LM-MCVIS on three public multi-community datasets using a high-fidelity 5G/6G emulator and benchmark its performance against five strong baselines. Results demonstrate significant gains in engagement depth, interaction diversity, response latency, and bandwidth savings. Ablation studies further validate the individual impact of each module. LM-MCVIS offers a scalable and modular paradigm for intelligent community interaction over wireless networks, with broad implications for collaborative learning, digital governance, and decentralized social ecosystems.

Wireless communication has been in high demand over the last decades. Soon, the globe will be equipped with fifth-generation (5G) communications, which provide an incredible number of additional capabilities compared to fourth-generation communications. An innovative paradigm has evolved with the combination of artificial intelligence (AI) with sixth-generation (6G) communication networks in response to the increasing need for intelligent communication and seamless connection. This integration enables optimum resource allocation and greater efficiency. It also enhances adaptive system performance by incorporating AI across multiple network layers. The next generation of wireless networks must address many fundamental issues, including increasing system capacity, data throughput, latency, security, and quality of service in comparison to 5G. This article provides a through review of the vision of future 6G network AI and wireless communication architecture, touching on their conceptual foundations, inherent difficulties, and potential fields for further study. Some new technologies discussed in this article include AI, terahertz communications, free-space optical networks, blockchain, quantum communications, drones, mobile free communications, integrated sensing and communication, dynamic network slicing, big data analytics, and wireless optical technology. This could all be useful in ensuring the quality of service in the 6G architecture development. Furthermore, we provide a concise overview of the AI standardization process for wireless networks, focusing on essential achievements and current initiatives. We also examine the significant obstacles that 6G's AI and communication integration encountered. Lastly, we provide an overview of prospective future studies that ideally promote advancing and improving AI and 6G communications by describing possible obstacles and possibilities.

In this era, Wireless Communication Networks (WCNs) need dynamic and adaptive resource allocation approaches to handle user association and power allocation specifically under multi-connectivity and diverse traffic conditions. However, the conventional approaches struggle due to high computational cost, poor adaptability, and limited generalization. Therefore, this research proposes a large Spatio-Temporal Graph Transformer-based Reinforcement Learning (STGT-RL) model to jointly optimize user association and power allocation in large-scale WCNs. Initially, the network topology is designed using graph representations and incorporates a hybrid encoder that integrates Graph Transformers for spatial user-Base Station (BS) relationships and Spatio-Temporal Transformers for capturing time-varying traffic and channel states. Further, to ensure adaptive decision-making, a Transformer-RL policy agent is trained through a multi-objective reward function that assists in balancing throughput maximization and power efficiency. Furthermore, to enable stable policy learning, the model is initially trained using high-quality supervision from CRFSMA-generated labels, followed by reinforcement-based policy refinement. Hence, the experimental results are simulated on WCN environments to demonstrate that the proposed STGT-RL significantly outperforms baseline deep learning and heuristic-based methods in terms of throughput, energy efficiency, and fairness.