Chunxiao Song

A series of experiments were carried out within a scaled-down, elongated, and narrow space featuring multiple lateral doors. These experiments aimed to thoroughly examine the fire characteristics both inside the enclosed compartment and within the connected tunnel, which were induced by the fire occurring in the compartment. It shows that the classic maximum temperature model obtained from a single confined space underestimates or overestimates the values in the carriage and tunnel, and the three-region plume law cannot effectively cover the temperature values inside the carriage, the prediction relations for the double confined space scenario are proposed. The rate of temperature decrease beneath the ceiling in the double long-narrow confined space is greater than in a single confined space, with a significantly reduced impact of HRR on the decay rate. The study also provides a segmented characterization of the influence of HRR on temperature attenuation. Additionally, distinct flame extension and spillover patterns are observed at different fire source locations. These include ceiling jet flame overflow primarily from the upper part of the lateral door and direct overflow from the entire door height. The study's findings offer valuable insights for fire assessment and control in complex confined rail systems.

Network traffic classification is the basic block for network management services such as network security and QoS control. When facing with the emergence of stronger encryption protocols like TLS 3.0, we need more advanced technologies, such as classical machine learning and deep learning methods to solve the problems. Every moment there are thousands of applications are sending the packets in the network, it's difficult for us to label all the sessions, we only have few labelled samples to train our classifiers. In this paper, we propose a semi supervised network traffic classification method. The basic idea is to use the samples segmented by encoder-decoder networks and unlabeled data to improve the performance of the classifier trained based on a small number of labeled samples. To explore the correlation of the continuously arriving packets, we segmented the sequence of packets' features parsed from raw traffic files (PCAPs), and using the segmentations to identify the traffic types. Then we adopt an Adversarial Auto-encoder to perform semi-supervised phase. The experimental study on VPN-nonVPN datasets shows that the proposed method is effective compared with the baseline method when there are relatively few available labeled training data.

Abstract This study investigated the efficacy of the full transverse exhaust method for smoke extraction in tunnel fires. It examines factors such as the number and layout of air supply and exhaust outlets, analyzing their impact on smoke spread, tunnel temperature, visibility, and airflow. The results demonstrate that the full transverse exhaust method effectively controls smoke emissions in raised highway tunnels. It limits smoke spread, reduces tunnel temperature, and effectively controls the fire‐affected area. The number and layout of outlets significantly influence smoke dispersion, with fewer exhaust outlets providing better smoke control and optimizing the tunnel environment. However, insufficient outlets disrupt gas flow stability. The position of exhaust outlets affects smoke distribution, and caution is advised to prevent directing fresh air flow toward the fire. Opening an equal number of exhaust outlets on one side of the fire source yields superior smoke extraction results, reducing tunnel ceiling temperatures and minimizing risks to personnel and structures. Though stabilization may take longer, this configuration proves advantageous. The study offers valuable insights and practical guidelines for implementing the full transverse smoke control method in real‐world scenarios.

Massive amounts of data have given a huge boost to artificial intelligence for communication networks, for instance, intelligent inspection, power IoT management, but they have also brought problems. The original data generated at the edge of the mobile communication network and imported into the core network not only takes up a lot of bandwidth resources, but also poses a great challenge to the fast and reliable transmission and computing. Traditional cloud-based machine learning methods require data to be centralized in cloud servers or data centers. However, in edge networks, due to limited network resources, direct transmission of centrally learned data will lead to unacceptable communication delays, resulting in low system efficiency, and may lead to serious privacy problems. In order to solve these problems, federal learning technology is attracting people's attention. This paper first analyzes the factors that affect the efficiency of federated learning system, establishes a federated learning system model, then uses DDPG to design and implement a node selection algorithm, the goal is to reduce the federated learning time to the maximum and improve the learning accuracy. Finally, under the condition of different node quality, the simulation experiment verifies that the algorithm can shorten 40% of the model training stability time, thus proving the effectiveness and feasibility of the proposed algorithm, indicating that the federated learning system can effectively select nodes in this way.