With the rapid development of the Internet of Things (IoT) and communication technologies, for the latency-sensitive and high-computing-load applications of terminals, the cloud servers in traditional cloud computing are located far away from the terminal devices, resulting in high communication delays and difficulty in meeting the requirements of low-latency tasks. While the edge node resources in traditional edge computing are limited, making it difficult to meet the growing high-computing-load requirements of terminals. This paper proposes a cloud-edge collaborative system architecture to solve the problem of task offloading and computing resource allocation including communication model and latency model, and proposes the problem formulation by minimizing the average task execution delay and maximizing the task success rate. Furthermore, this paper proposes a task offloading and resource allocation scheme based on Genetic Algorithm (GA) and Deep Deterministic Policy Gradient (DDPG). The global optimal learning rates of the Actor and Critic network of DDPG are obtained through GA pre-training. The global optimal learning rate of the Actor and Critic networks of DDPG is obtained through GA pre-training. Then, the optimal task offloading and resource allocation strategy is obtained through DDPG. Lastly, the effectiveness and superiority of the proposed scheme are verified through simulation experiments.

Large Language Models (LLMs) and Retrieval-Augmented Generation (RAG) technologies have been widely applied in vertical domains such as healthcare. But traditional RAG systems struggle with active dialogues for tasks, in which they often fail to proactively guide interactions, leading to information overload or inaccuracies, especially for non-professional users. To address this, we leverage knowledge graphs (KGs) to propose the Knowledge Graph–Guided Dialogue Augmented Generation (K2GDAG) system, which mimics human reasoning (e.g., doctor-patient consultations) and divides the question-answering process into two stages: possibility exploration and KG-based screening. In the exploration stage, K2GDAG decomposes user questions to extract key entities and reasoning paths, retrieving relevant information from KGs to construct the potential KGs that cover high-probability scenarios. In the screening stage, leveraging the structural information of KGs, the system employs a dichotomous approach to analyze key nodes, guiding LLMs to actively query users for more information. By querying about critical node information, K2GDAG refines user inputs iteratively to obtain precise and comprehensive details. Experiments on two healthcare datasets show that K2GDAG improves performance by 49.5

Remaining useful life (RUL) estimation for aero-engines becomes unreliable under mixed operating conditions, where onboard sensor measurements are non-stationary and entangle two concurrent drivers: irreversible degradation evolution and reversible operating-response dynamics. When these effects are mixed in the measurement stream, data-driven estimators may overfit condition-induced fluctuations, causing biased RUL estimates after regime shifts. To mitigate such coupling-induced bias, we propose a lightweight coupling-sensitive feature decoupling network (CFDN) whose architecture encourages partial separation between degradation-related trends and condition-related fluctuations at the representation level, rather than enforcing a strict factorial decomposition. CFDN adopts a dual-stream backbone: a Trend Stream tracks slowly varying degradation representations, while a Fluctuation Stream summarizes short-term operating responses through attention pooling. We further introduce a zero-initialized residual fusion scheme that starts from a trend-only estimator and progressively incorporates fluctuation information to improve optimization stability under strong coupling. Experiments on the NASA Commercial Modular Aero-Propulsion System Simulation benchmark, supported by controlled ablations, branch-level quantitative analyses, and same-protocol multi-seed comparisons with paired Wilcoxon signed-rank tests and bootstrap 95% confidence intervals, show that CFDN is particularly effective on mixed-condition subsets and exhibits the smallest seed-level variability among the compared baselines. A supplementary pilot study on N-CMAPSS provides feasibility evidence on a more flight-realistic dataset, rather than a comprehensive cross-dataset generalization benchmark.

Unmanned Aerial Vehicle (UAV)-assisted edge computing supports latency-sensitive applications by offloading computational tasks to ground-based servers. However, determining optimal resource allocation under strict latency constraints and stochastic channel conditions remains challenging. This paper addresses the joint computation partitioning and power allocation problem for UAV-assisted edge computing systems. We formulate the problem as a Constrained Markov Decision Process (CMDP) that explicitly models latency constraints, rather than relying on implicit reward shaping. To solve this CMDP, we propose Constrained Soft Actor-Critic (C-SAC), a deep reinforcement learning algorithm that combines maximum-entropy policy optimization with Lagrangian dual methods. C-SAC employs a dedicated constraint critic network to estimate long-term constraint violations and an adaptive Lagrange multiplier that automatically balances energy efficiency against latency satisfaction without manual tuning. Extensive experiments demonstrate that C-SAC achieves an 18.9% constraint violation rate. This represents a 60.6-percentage-point improvement compared to unconstrained Soft Actor-Critic, with 79.5%, and a 22.4-percentage-point improvement over deterministic TD3-Lagrangian, achieving 41.3%. The learned policies exhibit strong channel-adaptive behavior with a correlation coefficient of -0.894 between the local computation ratio and channel quality, despite the absence of explicit channel modeling in the reward function. Ablation studies confirm that both adaptive mechanisms are essential, while sensitivity analyses show that C-SAC maintains robust performance with violation rates varying by less than 2 percentage points even as channel variability triples. These results establish constrained reinforcement learning as an effective approach for reliable UAV edge computing under stringent quality-of-service requirements.

Network intrusion detection in actual production environments is very challenging because network traffic data contain high-dimensional features, the distribution of attack classes is highly imbalanced, and spatial and temporal features provide complementary information. To address these problems, this paper proposes a hybrid ensemble learning intrusion detection method based on a spatiotemporal adaptive gated fusion framework (STAG-IDS). The method first constructs a balanced training set through a phased sampling strategy, which combines majority-class downsampling and data cleaning with oversampling of synthetic minority-class samples. It then uses a two-stream feature extractor: a convolutional neural network with a convolutional block attention module to capture local feature interactions, and a bidirectional long short-term memory network to model temporal dependencies. An adaptive gated fusion unit is then used to dynamically weight the spatial and temporal representations for feature fusion. Finally, the fused deep-learning branch and the machine-learning branch are combined through soft voting. Experiments on CIC-IDS2017 show that STAG-IDS achieves an accuracy of 99.33% and a Macro-F1 score of 0.9074. The results show that the proposed framework can improve the detection performance of minority attack classes while maintaining strong overall classification performance.

the rapid growth of the Industrial Internet of Things (IIoT), more and more devices are connecting to the network, generating vast amounts of data, including sensors, actuators, and controllers. Traditional anomaly detection methods often rely on centralized data collection, leading to concerns about privacy leakage and data centralization. To address these challenges, approaches that combine federated learning (FL) with blockchain technology offer an efficient, scalable solution. These methods enable automatic scaling based on system size, adapting to increasing devices and data traffic. However, the limitations of fixed shards and security risks associated with the aggregation of data from different shards introduce new issues. To overcome these challenges, this paper proposes a self-adaptive blockchain sharding strategy, based on IIoT device grouping, which jointly optimizes the number of shards and the security of model updates. This optimization is modeled as a Markov Decision Process (MDP), with deep reinforcement learning (DRL) used to determine the optimal device sharding parameters. Furthermore, a joint committee mechanism is introduced to ensure secure cross-shard transactions, while a federated learning filtering mechanism (Fed-Filt) is applied to enhance the accuracy and security of global model aggregation by screening out malicious nodes. In the clustering experiments, the proposed method shows significant improvements in clustering quality metrics such as Normalized Mutual Information (NMI), Adjusted Rand Index (ARI), Adjusted Mutual Information (AMI), and silhouette score, with faster convergence. Experimental results demonstrate that under 25% and 50% malicious node scenarios, the proposed algorithm can effectively resist poisoning attacks and achieve stable convergence, outperforming the traditional FedAvg algorithm. Specifically, with 50% malicious nodes, the accuracy is improved by approximately 27.8%, and the method exhibits strong resistance and recovery capabilities.

Integrating federated learning (FL) with blockchain technology provides an effective solution for enhancing data security and privacy in decentralized Internet of Things (IoT) ecosystems. However, challenges arise in efficiently achieving consensus and balancing the load across multiple shards, primarily due to the resource constraints and heterogeneity of IoT devices. This article introduces SynergyMining, a novel collaborative sharding consensus mechanism designed specifically for blockchain-based FL in IoT. SynergyMining leverages reinforcement learning to dynamically optimize consensus group (CG) selection, ensuring balanced workloads across shards and efficient resource utilization. Additionally, we propose a freshness and quality-aware FL framework with asynchronous model aggregation called FedFQ that dynamically adjusts aggregation weights based on the recency and quality of local models. This approach mitigates client instability and improves the efficiency of the global model aggregation process. Experimental results demonstrate that SynergyMining outperforms leading algorithms, including Monoxide, hidden Markov model (HMM), Elastic and optimal resource scheduling policies (ORSP) across key performance metrics. Specifically, compared to these algorithms, SynergyMining improves system throughput by 9.97% to 72.16%, final model accuracy by 1.18% to 5.53%, and reduces load imbalance by 6.65% to 16.29%. These advancements, combined with the freshness-aware aggregation, make SynergyMining a robust and scalable solution for IoT-based FL applications, offering significant improvements in efficiency, scalability, and security.

Driven by technologies such as deep learning, online detection equipment can perform comprehensive and continuous monitoring of high-speed railways (HSR). However, these detection tasks in the railway Internet of Things (IoT) are typically computation-intensive and delay-sensitive, that makes task processing challenging. Meanwhile, the dynamic and resource-constrained nature of HSR scenarios poses significant challenges for effective resource allocation. In this paper, we propose a cloud-edge collaboration architecture for deep learning-based detection tasks in railway IoT. Within this system model, we introduce a distributed inference mode that partitions tasks into two parts, offloading task processing to the edge side. Then we jointly optimize the computing offloading strategy and model partitioning strategy to minimize the average delay while ensuring accuracy requirements. However, this optimization problem is a complex mixed-integer nonlinear programming (MINLP) issue. We divide it into two sub-problems: computing offloading decisions and model partitioning decisions. For model partitioning, we propose a Partition Point Selection (PPS) algorithm; for computing offloading decisions, we formulate it as a Markov Decision Process (MDP) and solve it using DDPG. Simulation results demonstrate that PPS can rapidly select the globally optimal partition points, and combined with DDPG, it can better adapt to the offloading challenges of detection tasks in HSR scenarios.

Online detection devices, powered by artificial intelligence technologies, enable the comprehensive and continuous detection of high-speed railways (HSRs). However, the computation-intensive and latency-sensitive nature of these detection tasks often exceeds local processing capabilities. Mobile Edge Computing (MEC) emerges as a key solution in the railway Internet of Things (IoT) scenario to address these challenges. Nevertheless, the rapidly varying channel conditions in HSR scenarios pose significant challenges for efficient resource allocation. In this paper, a computation offloading system model for detection tasks in the railway IoT scenario is proposed. This system includes direct and relay transmission models, incorporating Non-Orthogonal Multiple Access (NOMA) technology. This paper focuses on the offloading strategy for subcarrier assignment, mode selection, relay power allocation, and computing resource management within this system to minimize the average delay ratio (the ratio of delay to the maximum tolerable delay). However, this optimization problem is a complex Mixed-Integer Non-Linear Programming (MINLP) problem. To address this, we present a low-complexity subcarrier allocation algorithm to reduce the dimensionality of decision-making actions. Furthermore, we propose an improved Deep Deterministic Policy Gradient (DDPG) algorithm that represents discrete variables using selection probabilities to handle the hybrid action space problem. Our results indicate that the proposed system model adapts well to the offloading issues of detection tasks in HSR scenarios, and the improved DDPG algorithm efficiently identifies optimal computation offloading strategies.

In developing the sixth-generation (6G) system, integrated sensing and communication technology is becoming increasingly essential, especially for applications like autonomous driving. This paper develops an architecture for integrated sensing, communication, and computation (ISCC) in the vehicular network, where vehicles perform environment sensing, sensing data computation, and transmission. To support low-latency cooperation between vehicles and extend vehicles’ sensing range, over-air-computation federated learning is employed. The optimization problem of joint beamforming design and power resource allocation in the ISCC scenario is formulated to maximize the achievable data rate while ensuring sensing and computing performance. However, solving this joint optimization problem is a great challenge due to the high coupling resource and time-varying channel environment. Therefore, a hybrid reinforcement learning scheme is proposed in this work. First, the semidefinite relaxation and Gaussian randomization techniques are leveraged to obtain the approximate solution of the aggregation beamformer. Then, the deep deterministic policy gradient algorithm is proposed to tackle the transmit beamforming design and resource allocation problem in continuous action space. Extensive simulation results validated the admirable performance of the proposed scheme in convergence and achievable sum rate compared with the benchmark schemes. In addition, the impact of variables on the optimization performance is demonstrated via numerical results.

This paper introduces BiLSTM-MLAM, a novel multi-scale time series prediction model. Initially, the approach utilizes bidirectional long short-term memory to capture information from both forward and backward directions in time series data. Subsequently, a multi-scale patch segmentation module generates various long sequences composed of equal-length segments, enabling the model to capture data patterns across multiple time scales by adjusting segment lengths. Finally, the local attention mechanism enhances feature extraction by accurately identifying and weighting important time segments, thereby strengthening the model’s understanding of the local features of the time series, followed by feature fusion. The model demonstrates outstanding performance in time series prediction tasks by effectively capturing sequence information across various time scales. Experimental validation illustrates the superior performance of BiLSTM-MLAM compared to six baseline methods across multiple datasets. When predicting the remaining life of aircraft engines, BiLSTM-MLAM outperforms the best baseline model by 6.66% in RMSE and 11.50% in MAE. In the LTE dataset, it achieves RMSE improvements of 12.77% and MAE enhancements of 3.06%, while in the load dataset, it demonstrates RMSE enhancements of 17.96% and MAE improvements of 30.39%. Additionally, ablation experiments confirm the positive impact of each module on prediction accuracy. Through segment length parameter tuning experiments, combining different segment lengths has resulted in lower prediction errors, affirming the effectiveness of the multi-scale fusion strategy in enhancing prediction accuracy by integrating information from multiple time scales.

Federated learning enables data owners in the Internet of Things (IoT) to collaborate in training models without sharing private data, creating new business opportunities for building a data market. However, in practical operation, there are still some problems with federated learning applications. Blockchain has the characteristics of decentralization, distribution, and security. The blockchain-enabled federated learning further improve the security and performance of model training, while also expanding the application scope of federated learning. Blockchain has natural financial attributes that help establish a federated learning data market. However, the data of federated learning tasks may be distributed across a large number of resource-constrained IoT devices, which have different computing, communication, and storage resources, and the data quality of each device may also vary. Therefore, how to effectively select the clients with the data required for federated learning task is a research hotspot. In this paper, a two-stage client selection scheme for blockchain-enabled federated learning is proposed, which first selects clients that satisfy federated learning task through attribute-based encryption, protecting the attribute privacy of clients. Then blockchain nodes select some clients for local model aggregation by proximal policy optimization algorithm. Experiments show that the model performance of our two-stage client selection scheme is higher than that of other client selection algorithms when some clients are offline and the data quality is poor.

The field of computer vision is currently transitioning from closed-set to open-set tasks. Vision foundation models have already demonstrated success in open-set scenarios. Building on these models, the utilization of a feature supervision framework can further enhance results. Our paper introduces a new method called SAMS (Segment Anything Model using Similar Images), which is a type of feature supervision framework. It is designed to segment specific masks from visual supervision features. Our framework comprises a pre-trained segmentation model and an efficient, novel prompt generation model capable of generating new prompts based on pre-extracted image features. This innovation eliminates the need for manually crafted prompts in the mask generation phase by integrating the principles of one-shot or few-shot learning with visual instructions from similar images. The effectiveness of the SAMS method is evident in its performance across various tasks, particularly in open-set tasks where traditional models tend to struggle. The pretrained model not only achieves impressive mean Intersection over Union (mIOU) scores without incurring additional time loss, but also demonstrates potential for further improvement through targeted module training.

Anomaly detection in the industrial internet of things (IIoT) devices is significant due to its fundamental role in protecting modern critical infrastructure. In the IIoT, anomaly detection can be carried out by training machine learning models. Data sharing between factories can expand the data from which the model is trained, thus improving the performance of the model. However, due to the sensitivity and privacy of IIoT data, it is also difficult to build a high-performance anomaly detection model between factories. To address this problem, we design an anomaly detection method for IIoT devices combined blockchain of main-side structure and federated learning. We store the global model on the main-chain while the side-chain records the hash value of the global models and local models, which updated by participating nodes, controlling nodes access to the global model through the main-side blockchain and the smart contracts. Only the nodes participating in the current federated learning training can get the latest global model, so as to encourage the nodes to take part in the training of the global model. We designed a proof of accuracy consensus algorithm, and select the nodes to participate in training according to the accuracy of the local model on the test dataset to resist the poisoning attack of the models. We also use the local differential privacy (LDP) algorithm to protect user data privacy from model inference attacks by adding noise to the local model. Finally, we propose a new algorithm named Fed_Acc to keep the accuracy of the global model stable when the users add a lot of noise to their local models.

In recent years, with the deepening of power system architecture adjustment and market-oriented reform, smart microgrid has become the main development direction of power grid architecture. In order to meet the higher requirements of decentralization, autonomy, intelligence and marketization in the process of power grid reform, digital technologies such as artificial intelligence and block chain should be introduced as the key support. In view of the consistency between the technical characteristics of block chain and the development demand of power grid, a power trading system model based on alliance chain is proposed. In order to realize the power balance and maximize the benefits of the microgrid within the power grid, a trading strategy optimization scheme based on MADDPG algorithm is proposed. The experimental results show that the algorithm can help microgrids to formulate the trading strategy which is most in line with the overall benefits of the grid and maximize the total revenue of the system. The performance of this algorithm is better than DDPG algorithm and random trading method.

Based on the service-oriented nature that 5G possesses, network slicing can build logical virtual networks in which various services are supplied by a common physical infrastructure. The function of network slicing is decided by the management of multidimensional resources that multi-access edge computing (MEC) offers. This paper presents the optimization of communication and computing resources in blockchain-enabled MEC network slicing. Given its immutability and traceability, blockchain is used to keep records of information concerning resource optimization, including resource information, system utility and slice strategy. The optimization goal of the resource allocation problem is to ensure the quality of service while maximizing the utility gained by the mobile network operator. The optimal resource allocation strategy is obtained using the deep deterministic policy gradient algorithm. The availability of the algorithm is proved through numerical simulations.

In the era of big data, data has gradually become an important asset. In order to fully protect the value of data and ensure the effective conduct of data transactions, this paper combines bulletproof and blockchain to ensure that data transactions can be carried out in a safe and reliable environment, while also hiding the value of data itself. The commitment of bulletproof ensures that all users on the blockchain can verify the validity of the transaction but not the details of the transaction. In this paper, the bulletproof is simulated, and the simulation results verify the feasibility and practicability of the scheme.

The arrival of the Internet of Things and 5G technology enables users to rely on edge computing platforms to process massive data.Data sharing based on edge computing refines the efficiency of data collection and analysis, saves the communication cost of data transmission back and forth, but also causes the privacy leakage of a lot of user data.Based on attribute-based encryption and blockchain technology, we design a fine-grained access control scheme for data in edge computing, which has the characteristics of verifiability, support for outsourcing decryption and user attribute revocation.User attributes are authorized by multi-attribute authorization, and the calculation of outsourcing decryption in attribute encryption is completed by edge server, which reduces the computing cost of end users.Meanwhile, We implemented the user's attribute revocation process through the dual encryption process of attribute authority and blockchain.Compared with other schemes, our scheme can manage users' attributes more flexibly.Blockchain technology also ensures the verifiability in the process of outsourcing decryption, which reduces the space occupied by ciphertext compared with other schemes.Meanwhile, the user attribute revocation scheme realizes the dynamic management of user attribute and protects the privacy of user attribute.

: With the increasing popularity of distance education and online learning in colleges and universities, the network video conference and live streaming system with multiple scenes has become an important part of the information construction of college education. In this paper, for the problems of the current network video conference and live streaming system in colleges and universities, we design and implement a network video conference and live streaming system supporting multiple scenarios. The system adopts the audio and video transmission scheme based on WebRTC technology, supports multiple platforms such as Web, Android and iOS, and integrates a variety of functional modules, including screen sharing, interactive Q&A, voting, recording and playback, etc. The experimental results show that the system is stable and efficient, and it can meet the demand for web video conferencing and live streaming in multiple scenarios in universities.

Multi-access edge computing (MEC) enables the computation-intensive and latency-critical application to be processed at the network edge, which reduces the transmission latency and energy consumption. The quality of the wireless channel seriously affects the performance of the edge network. Consequently, the performance of the edge network can be significantly improved from the perspective of communication. The recently advocated intelligent reflecting surface (IRS) intelligently controls the radio propagation environment to improve the quality of wireless communication links. This paper proposes an edge heterogeneous network with the assistance of intelligent reflecting surface. Specifically, the macro base station and small base stations are equipped with MEC servers, and IRS is adopted to provide an additional computation offloading link. The user association, computation offloading and resource allocation, as well as IRS phase shift design are optimized with the aim of minimizing the long-term energy consumption subject to the constraints imposed on quality of service (QoS) and available resources. The challenge of the optimization problem is rooted from the fact that update timescale of user association is different from others. Hence, a two-timescale mechanism is invoked by marrying tools from matching theory and deep reinforcement learning. More specifically, the user association decision takes place in the long timescale. In the short timescale, the computation offloading, resource allocation and IRS phase shift design strategy is performed. The effectiveness of the proposed two-timescale mechanism is verified by the simulation results.