Facing the stringent communication requirements of emerging fields, Cyclic Queuing and Forwarding (CQF) mechanism stands out among Time-Sensitive Networking (TSN) shaping mechanisms for its simplicity and effectiveness. CQF scheduling is crucial for ensuring its performance. However, existing studies primarily focus on packet scheduling under fixed timeslot lengths, while neglecting the effects of message-packet relationship and timeslot length. This omission can lead to transmission congestion and even packet loss, thereby compromising the applicability and reliability of their scheduling schemes in real-world deployments. To address these limitations, we propose a novel Timeslot Aware Message Scheduling (TAMS) algorithm. It incorporates message delay constraint and packet timing constraint to prevent message deadline violations and packet injection blocking. It also conducts a theoretical analysis based on the complete constraint set to support effective timeslot length configuration. Furthermore, it establishes a Deep Reinforcement Learning (DRL) framework to enhance the scheduling efficiency in complex CQF-based TSN networks. Extensive experimental results demonstrate the significant advantages of the proposed TAMS algorithm. Compared with the state-of-the-art DRL-based scheduling algorithm, it improves the schedulable scale by 37% and reduces the solving time by 17%.

Time-Sensitive Networking (TSN) with Cyclic Queuing and Forwarding (CQF) enables deterministic communication in distributed cyber-physical systems. However, existing CQF schedulers model the timeslot as the minimum unit and ignore intra-timeslot dynamics; then, packets arriving within the same timeslot may contend for serialization, leading to temporal conflicts that cause unexpected packet loss and an overestimation of schedulability. To address this issue, we propose a fine-grained CQF scheduling framework that integrates both a refined model and an efficient algorithm. The model replaces timeslot-level injection with precise instants and introduces an intra-timeslot conflict-free constraint to explicitly capture and resolve temporal queue resource conflicts. Building on this model, we further develop a lightweight Fine-Grained Temporal Resource (FGTR) scheduling algorithm, which intelligently allocates injection instants and exploits residual temporal vacancies to improve CQF scheduling performance. Comparative experiments show that FGTR increases the average schedulable scale by 46.1 % and 32.9 % over existing constraint-solving and Deep Reinforcement Learning-based algorithms, respectively, while maintaining high computational efficiency, thereby demonstrating its effectiveness and practicality.

ABSTRACT Wireless avionics intra‐communications (WAIC) technologies play an important role in the real‐time transmission between airborne equipment. As the number of deployed WAIC nodes increases, severe inter‐user and inter‐cabin interference occurs, degrading system performance and fairness. In this paper, a deep reinforcement learning‐based interference alignment and power allocation (DRL‐IAPA) scheme is proposed for multi‐user WAIC system. By integrating interference alignment with a deep deterministic policy gradient (DDPG) algorithm, the DRL‐IAPA scheme can mitigate interference, optimize power allocation, and ensure fairness among users under stringent latency and power constraints. Simulation results show that DRL‐IAPA improves spectral efficiency and fairness over traditional and heuristic‐based methods, and demonstrates scalability and consistent performance across various network configurations. Compared with other representative DRL algorithms, such as proximal policy optimization and soft actor–critic, our proposed DDPG‐based approach exhibits faster convergence, higher achievable reward, and better applicability in dynamic WAIC environments.

Real-time networks employ traffic shaping and scheduling mechanisms to provide bounded and low-latency services for safety-critical traffic. However, strict priority (SP) scheduling, while commonly used to meet stringent timing requirements, can lead to a coupling between priority assignment and real-time assurance. This coupling means that the deadline requirements of different priority classes do not always align with the actual criticality of the traffic, potentially compromising the performance of lower-priority safety-critical applications. To address this issue, we explore the priority-coupling problem and propose a Credit Rate Round-robin (CRR) scheduling method that simultaneously guarantees real-time performance and reserves bandwidth for safety-critical traffic. CRR utilizes the logical bandwidthidleSlopeas a weight factor for round-robin queues and is analyzed under two configurations: with and without an upper-bound credit limit. Evaluation experiments demonstrate that CRR effectively decouples priority from real-time performance. Specifically, compared to SP+CBS, CRR without an upper-bound credit reduces the average delay for medium- and low-priority traffic by 18% and 47%, respectively, with a corresponding 13% increase in the average delay for high-priority traffic. These results demonstrate that the proposed CRR scheduling effectively circumvents priority assignments and mitigates the priority-coupling issue, thereby ensuring reliable real-time transmission for safety-critical applications.

Time-sensitive networking is a promising real-time network protocol, especially because it can utilize a time awareness shaping mechanism to realize very low delay and jitter transmission of control data traffic (CDT). Nowadays, deep learning methods have been widely applied in CDT scheduling to improve the scheduling performance. However, these trained models can only schedule for fixed networking with limited change. They have low generalization and cannot adapt to diverse application situations. To address this challenge, we propose GTSNet, a novel scheduler with strong generalization capabilities. We design the deep learning model based on a graph neural network to overcome the lack of generalization performance. Moreover, we transform the problem into a continuous node classification problem to enhance the scheduling ability and generalization. Compared with existing methods(heuristic and reinforcement learning methods), GTSNet demonstrates strong generalizability for the first time and avoids more than 50% nonschedulable flows in many generalization ability evaluation cases.

Time-sensitive networking (TSN) is an evolving set of Ethernet standards designed to ensure deterministic communication. Several schedulers can be applied in TSN to support the transmission of mixed time-critical traffic, among which deficit round robin (DRR) is a potential candidate since it can achieve fair queuing with very low complexity. The real-time quality of service (QoS) performance of DRR is determined by the quantum assigned to each queue, yet obtaining configurations with QoS guarantess at the network level is challenging due to the complex interactions of traffic flows. Conventional configuration methods typically adopt the framework based on ex-post schedulability verification, using heuristic algorithms to search for feasible configurations. However, these methods incur significant time overhead during the solution searching and performance verification stage and often lead to bandwidth wastage. This paper proposes an innovative DRR quantum assignment method that maximizes bandwidth utilization while providing prior realtime QoS guarantees. We use network calculus (NC) to build the optimization model and develop a numerical approximation approach based on local analytic solutions to solve for the optimal solution, making our method efficient and scalable. The correctness and optimality of our method are formally proven. Through experiments with network cases of various scales, we comprehensively evaluate the performance of our method. The results show that our method exhibits significant advantages over the conventional methods in terms of both bandwidth utilization and time complexity. In industrial-sized cases, our method saves around 31% of transmission bandwidth compared to the state-of-the-art with a runtime within milliseconds, which is several orders of magnitude faster.

Time-sensitive networking (TSN) has great potential as an airborne network to interconnect modules in integrated modular avionics (IMA) system. For TSN-based IMA system, the hybrid scheduling of processing in modules and communication in TSN can guarantee its real-time performance. However, traditional task-message scheduling methods still lack applicability and scalability due to their incompatibility with the partition-task hierarchical architecture in modules and high complexity brought by the tight coupling of tasks and messages. Partition-message scheduling methods can ensure this applicability and scalability, but cannot coordinate tasks with messages well, thus sacrificing real-time guarantee capabilities. Namely, existing methods cannot comprehensively ensure the scheduling performance, including applicability, real-time, and scalability. Therefore, we propose a novel loosely coupled partition-(task)-message scheduling framework. It takes partitions and messages as scheduling objects and uses tasks as their coordination medium, to overcome the dependencies of existing methods on time-triggered tasks and guarantee applicability. Besides, it can also enhance real-time performance by analyzing task execution boundaries and application-layer end-to-end delays, and improve scalability through parallel optimizing and the incremental solving with block identification and adaptive adjustment. Experiments validate that it can schedule complex systems with up to 150 partitions, 1000 tasks, and 600 messages. Compared with the existing methods, it can accelerate the solving speed by 41% and reduce end-to-end delays by 27%.

Time-sensitive networking (TSN) has been considered as a promising networking candidate for avionics systems due to its capability of deterministic communication. In such TSN-based avionics systems, the network scheduling enables the timely transmission of messages. However, this is insufficient to satisfy the real-time requirements of functions since functions involve the chain execution of several tasks where messages only serve for the inter-task communication. In order to enhance the functionality of TSN-based avionics systems, scheduling should be extended from the network level to the system level to coordinate message transmission with task execution. Then, how to efficiently implement the hybrid scheduling of tasks and messages becomes an important issue. In this article, we construct a novel hybrid scheduling framework for TSN-based avionics systems, which consists of system consistency constraints, in-loop function delay calculation, and two hybrid scheduling algorithms. Consistency constraints restrict the unexpected interaction of messages and tasks for hybrid scheduling to guarantee the system-level determinism. Function delay is the end-to-end delay of the task chain, and its calculation provides the optimization objective for hybrid scheduling indicated by two metrics. Scheduling algorithms improve solving efficiency and functional performance through the incremental strategies of message dynamic ordering and task-message rescheduling. Experimental results verify that, compared with existing scheduling work that considers the dependency of messages on tasks, our work can complete complex scheduling for large systems even with hundreds of functions and can reduce function delays by 69% in reaction delay and 37% in age delay.

IEEE 802.1 Time-Sensitive Networking (TSN) comprises a set of real-time sub-standards that are rapidly becoming the de facto standard in numerous safety-critical domains. Nevertheless, existing TSN standards face challenges when encountering runtime reconfigurations, such as the efficient and safe real-time reconstruction of networks, which has become a current focal point of research. Although certain studies have commenced on dynamic scheduling and routing for TSN networks, a research gap persists in the online schedulability verification of mixed-critical messages. The present real-time performance analysis method is oriented toward assessing the schedulability of entire network flows. Consequently, even minor alterations made during reconfiguration necessitate the re-evaluation of all flows to verify their adherence to deadlines. In the paper, we propose an incremental performance analysis method (iPAM) to mitigate the cost of schedulability re-verification when reconfiguring the network of the TSN/TAS+CBS hybrid architecture. Extensive experiments conducted on a large-scale adapted realistic test case, demonstrate that, in comparison to the state-of-the-art full-performance analysis method, our iPAM substantially reduces evaluation time by approximately 75%-95% for large-scale networks. This advantage becomes increasingly pronounced as the scale of the network expands.

Abstract The real‐time performance is a major concern of complex real‐time systems during information acquisition, transmission, processing, and application. The interactions among information within these systems constitute complicated information exchange networks. Network topology plays an important role in the behaviors of information interactions and interferes, which severely affects the real‐time performance of the whole systems. Methods such as synchronization‐based method or traffic balancing strategy are the main ideas that guide the construction of network topology. However, these methods could not consider the real‐time performance in the process of topology construction. This article proposes an automatic topology generation algorithm with real‐time guarantee. Specifically speaking, the nodes in the network are clustered to generate the network topology by introducing the concepts of node degree and eigenvector centrality to ensure the real‐time performance of the network. The position of the nodes in the network is determined by comparing the node degree and the communication factors. Analytic method and simulation method are used to verify the real‐time performance of the proposed algorithm. Results show that the real‐time performance of more than half of the total information exchanges is improved compared with the traffic balancing method and topology grouping method for an industrial‐scale networking case.

Time-Sensitive Networking (TSN) evolves from Ethernet AVB in which the Credit-Based Shaping (CBS) is employed to guarantee the deterministic transmission of traffic. In CBS, the real-time performance of traffic associated with the credits can practicably settle the uncertainty of queueing and forwarding for stream-reservation traffic. Hence, the optimization of bandwidth allocation under the CBS becomes a key point to ensure guaranteed performance in the network. In this paper, a bandwidth allocation method is proposed for stream-reservation traffic while the real-time requirements are satisfied. We first illustrate the significance of appropriate bandwidth allocation for traffic under the CBS and construct a general schema for its bandwidth allocation. Also, the influences of the protected window for Control Data Traffic (CDT) on stream-reservation traffic are considered in our method. Further, mathematical models are constructed to analyze the delay bounds and backlogs of traffic to form the feedback for the optimal bandwidth allocation process. Finally, two node-level cases with different bandwidth utilization and a synthetic industrial networking scenario are carried out to demonstrate our method. The results confirm that the excessive reserved bandwidth does not necessarily decrease the delay bound, especially under the high traffic loads scenario. A more desirable bandwidth allocation strategy under the CBS mechanism is that the reserved bandwidth should be just enough according to the traffic loads to ensure the deterministic transmission of traffic.

The past decade has witnessed the use of highlevel features in saliency prediction for both videos and images. Unfortunately, the existing saliency prediction methods only handle high-level static features, such as face. In fact, high-level dynamic features (also called actions), such as speaking or head turning, are also extremely attractive to visual attention in videos. Thus, in this paper, we propose a data-driven method for learning to predict the saliency of multiple-face videos, by leveraging both static and dynamic features at high-level. Specifically, we introduce an eye-tracking database, collecting the fixations of 39 subjects viewing 65 multiple-face videos. Through analysis on our database, we find a set of high-level features that cause a face to receive extensive visual attention. These high-level features include the static features of face size, center-bias and head pose, as well as the dynamic features of speaking and head turning. Then, we present the techniques for extracting these high-level features. Afterwards, a novel model, namely multiple hidden Markov model (M-HMM), is developed in our method to enable the transition of saliency among faces. In our MHMM, the saliency transition takes into account both the state of saliency at previous frames and the observed high-level features at the current frame. The experimental results show that the proposed method is superior to other state-of-the-art methods in predicting visual attention on multiple-face videos. Finally, we shed light on a promising implementation of our saliency prediction method in locating the region-of-interest (ROI), for video conference compression with high efficiency video coding (HEVC).