As a key component of the industrial intranet, the production network is the source of data generation and the object of intelligent decision-making. Therefore, it is very important for the management and control of the production network. Currently, software-defined network, as one of the key technologies to break the "two-level and three-level" networking model of factory intranet, provides a centralized control and programmable network management capability for the production network. However, as the number of sensor devices in the production network continues to increase, the current single controller deployed at the industrial intranet router may encounter control latency, single points of failure, and uneven load. For this reason, this paper proposes a three-level hierarchical architecture for Software-Defined Network(SDN) controllers in industrial production networks called TLHAC. TLHAC consists of three levels of hierarchy, with the first level being the primary controller deployed on the router of the production network backbone, the second level being the secondary controllers deployed on the edge gateways of the workshop network, and the third level being the sub-controllers deployed on the wireless sensor nodes in the field. When a secondary controller fails, a control latency optimal migration algorithm based on load capacity limitation called LCL_CDOM is proposed to migrate industrial equipment. In addition, to optimize the deployment of sub-controllers, this paper also proposes a sub-controller deployment strategy based on node importance. The strategy first uses the Technique for Order Preference by Similarity to Ideal Solution(TOPSIS) analysis based on multi-attribute decision-making to comprehensively evaluate the importance of wireless sensor nodes, then uses the improved fuzzy multi-objective particle swarm algorithm (called IFMOBPSO) to optimize the solution and select the optimal deployment position of the sub-controller. This paper conducts simulation experiments on the three-level hierarchical deployment architecture and the optimal deployment strategy of the sub-controller. Simulation results demonstrate that TLHAC reduces the average control latency by 42 %-48 % and the average synchronization latency by 19 %-22 % compared to traditional two-level and Edge-SDN architectures. While IFMOBPSO achieves 8 %-14 % lower average control latency of important nodes and than 9 %-12 % lower average synchronization latency between secondary controllers compare to other meta-heuristic algorithms.

The integration of IT and OT networks in smart factories has expanded the attack surface, necessitating proactive and precise defense strategies. This article presents a dual-layer prediction framework targeting two core challenges: (1) macro-level attack path identification across converged IT/OT networks, and (2) micro-level attack technique evolution prediction. First, a multi-source network security knowledge base is constructed to model the threat ontology, and device threat levels are quantified based on the Common Vulnerability Scoring System (CVSS). Macro attack paths are predicted using threat propagation algorithms. Second, micro-level attack techniques are predicted along the macro paths by integrating Firewall-Intrusion Detection System(FW-IDS) configurations and Adversarial Tactics, Techniques, and Common Knowledge(ATT&CK) techniques into an evolution path prediction algorithm. The dual-layer prediction results are visualized to enhance interpretability. They demonstrate improved accuracy in attack path and technique prediction as well as real-time threat perception. Simulation experiments show that this approach significantly enhances the security posture of factory networks. Compared to existing methods, the dual-layer architecture proposed in this paper reduces the Weighted Mean Residual Paths (WMRP) by over 30%. This metric is used to quantify the number of potential attack paths remaining during the repair process. This significant reduction demonstrates that our method can more effectively contain threats at an early stage and minimize the attack surface more efficiently.This enables earlier detection and response to complex threats, thereby strengthening the protection of critical systems and devices in industrial environments.Beyond the implemented prediction tasks, the ontology-based semantic integration and cross-layer prediction structure also provide a foundational pathway toward future (self-*)adaptive security in Industry 5.0 smart factories, where online knowledge evolution, closed-loop feedback, and adaptive response orchestration are required to cope with dynamic and uncertain industrial threat environments.

Efficient load balancing is a key challenge in cloud computing to ensure high resource utilization, scalability, and service reliability. However, existing strategies face three major limitations: (a) static algorithms cannot adapt to dynamic and bursty workloads, resulting in severe resource imbalance; (b) single-metric evaluation (e.g., CPU utilization) fails to capture the multi-dimensional nature of heterogeneous cloud nodes; and (c) centralized scheduling suffers from global information acquisition latency, limiting responsiveness under dynamic conditions. To address these challenges, this paper proposes a Dynamic Feedback-driven Adaptive Queue-based Load Balancing (DFAQ-LB) algorithm. The core idea is to achieve self-adaptive task scheduling through dynamic feedback from queue states and system metrics. Specifically, two mechanisms are introduced: (a) an elastic queue capacity regulation model that uses short-term load prediction to adjust queue length and scheduling frequency dynamically; and (b) a multi-dimensional node health evaluation model that integrates CPU, MIPS, memory, and bandwidth with a dynamic weight allocation strategy to enhance real-time decision accuracy. Extensive simulations on CloudSim demonstrate that DFAQ-LB significantly outperforms traditional static and heuristic methods. Compared with ACO and PSO, DFAQ-LB reduces the average waiting time by 38.5% and 53.2% and the makespan by 30.7% and 35.2%, respectively. When compared with CDLB and DDLB, the average waiting time decreases by 53.2% and 75%, the average completion time by 28.3% and 63.4%, and the makespan by 35.2% and 40.7%. These results confirm that DFAQ-LB effectively alleviates task congestion, minimizes migration overhead, and achieves higher adaptability and scalability, providing a robust and intelligent paradigm for dynamic load balancing in cloud environments.

Trajectory similarity computation is crucial for many trajectory data mining applications. However, existing methods primarily focus on the sequential arrangement of trajectory points to represent movement and behavior patterns, overlooking the rich intrinsic spatial structure and topological relationships inherent in trajectories. This limitation prevents fully capturing deeper information, such as spatial associations between trajectory segments and complex behavioral motifs. To address this, we propose GraphTraj, a representation learning framework for structure-aware trajectory similarity computation. Our core motivation and novelty are two-fold: first, we convert trajectories into structured graphs to introduce their inherent spatial structure and topological information, without relying on any external information such as predefined road networks or map matching. Second, we preserve the original sequential order as a complementary feature, ensuring no information loss. To fuse these modalities, we employ a dual-stream architecture that unifies the structural graph representation with sequential trajectory information, which helps address the limitations of traditional sequence-only modeling. Specifically, GraphTraj partitions each trajectory into sub-trajectories and constructs undirected weighted graphs (UWGs) where nodes represent sub-trajectory units and directed edges encode their spatial relationships with weight attributes. This graph representation fundamentally transcends sequence-based paradigms by holistically encoding global morphology, local motion patterns, and spatio-behavioral correlations. The UWGs then serve as core structural inputs for representation learning in our ranking-loss-optimized graph convolutional neural network, while raw trajectory sequences are processed in parallel as sequence-enhanced inputs. GraphTraj operates without prior knowledge, integrating graph structures with raw trajectory sequences through a dedicated dual-stream graph-sequence neural network to learn trajectory similarity. Optimized via a ranking loss function, this architecture explicitly prioritizes relative similarity relationships to align learned representations with any user-defined similarity metric. Extensive experiments on three public trajectory datasets demonstrate that the proposed approach outperforms state-of-the-art methods for approximating the metric space.

Ensuring reliable perception under adverse weather is critical for Cyber-Physical Systems (CPS) in intelligent railway applications. LiDAR, as a key sensor, provides high-precision 3D data but suffers from severe degradation in fog due to signal scattering and attenuation, introducing noise into point clouds and disrupting CPS feedback loops. This paper proposes a CPS-oriented point cloud defogging framework that estimates fog distribution based on attenuation modeling and local density analysis. A multi-scale adaptive thresholding method is introduced, dynamically adjusting denoising parameters by combining a base multiscale factor with density-driven adaptive scaling. Dynamic neighborhood statistics and height-based corrections further refine the noise removal process, enhancing the distinction between fog noise and true object structures. Experimental evaluations on synthetic and real-world foggy datasets show that the proposed method outperforms traditional techniques such as SOR, DSOR, DROR in terms of precision, recall and structural preservation. By improving the reliability of the perception layer, our approach reinforces the overall robustness and safety of CPS-driven railway systems operating in foggy environments, providing a lightweight and adaptable solution suitable for real-time deployment.

Data publishing of the intelligent logistics big data platform is an important part of data management and sharing on this platform which regularly publishes logistics related data, mainly logistics order data, to the public. Although sensitive identifiers such as logistics order numbers, names, phone numbers, etc. in logistics order data have been desensitized using private desensitization technologies such as generalization, now a more effective privacy protection still has not been provided to the location data in the logistic order datasets. Attackers can combine the location data, use data mining, machine learning, and other methods to infer the sensitive information of the shipper and recipient, resulting in the leakage of sensitive information. Therefore, we propose a differential privacy protection scheme for shipper and receiver location data in logistics orders. Our approach first use the non-uniform Quadtree method to partition the location data space, and use the hierarchical progressive privacy budget to allocate for each node. Then, an L-staircase data perturbation algorithm based on a staircase noise mechanism and a epsilon-LLCDP privacy protection algorithm suitable for publishing logistics order position data were designed. Finally, the privacy and time complexity of the epsilon-LLCDP algorithm were theoretically analyzed. The experimental results indicate that the location data of the logistics order data in the intelligent logistics big data platform processed by epsilon-LLCDP not only retains its usability, but also avoids the leakage of the location privacy information related to the shipper and payee in the publishing of the logistics order data.

Due to its ability to generate high-quality 3D point cloud data, lidar sensors are widely used in many fields, especially in unmanned driving. However, in bad weather conditions, such as fog, snow, and so on, the sensor is easy to produce large noise, which affects the quality of 3D point cloud data. At present, there are statistical methods, such as SOR, and ROR, and methods based on deep learning, such as WeatherNet, which have more or less their defects. In this study, a 3D point cloud data snow removal technology based on Graph Convolutional Neural Network (GCN) is proposed. Compared with ordinary deep learning, this study does feature extraction in the data preprocessing step. It introduces an attention mechanism into the GCN network for optimization, to remove noise points more effectively. We do a large number of experiments on the real WADS dataset to verify that our snow removal method significantly outperforms the baseline methods, which proves that our method achieves good results in point cloud denoising. The proposed method is designed with embedded and cyber-physical systems (CPS) in mind, offering the potential for real-time deployment in autonomous driving platforms. By integrating multi-scale graph aggregation strategies with hardware-aware optimizations, the method establishes a new paradigm for resource-constrained cyber-physical systems.

The rapid growth of the Internet of Things (IoT) has significantly shaped the Smart City paradigm by enabling efficient data collection and resource management. Secure firmware updates and distribution mechanisms are crucial stages in the lifecycle of IoT device management. Traditional mechanisms, however, are vulnerable to unauthorized access, tampering, and single points of failure, which expose IoT devices to security threats. They also often fail to account for the dynamic nature of IoT environments and the varying criticality of devices for updates distribution, leading to inefficiencies and potential vulnerabilities. This work, proposes a blockchainenabled firmware update framework that addresses these limitations by employing a Merkle tree-based chunking approach for firmware data integrity assurance, and the blockchain decentralization for a secure, tamper-proof update mechanism. It incorporates smart contracts to enable automatic validation and authorization of firmware updates, mitigating the risks of malicious attacks and unauthorized access. Additionally, it utilizes peer-to-peer storage for firmware update distribution, eliminating reliance on centralized servers and resolving the issue of author disappearance. It introduces a machine learning (ML)-based method, the Adaptive Criticality-Based Distribution (ACBD), which dynamically adjusts firmware update distribution priorities based device criticality, defined by application domain, operational impact, and prevailing external conditions, a key gap in prior works. This ensures an optimized distribution strategy. Finally, it introduces a third-party creator delegation support which facilitates firmware updates delegation to multiple manufacturers, ensuring scalability and interoperability. Extensive experiments demonstrate robust security, high efficiency and reduces computational overhead, essential for sustainable smart cities.

The efficient deployment of Big Data processing tasks in cloud environments is the basic core function of Big Data processing, which refers to the effective deployment of tasks to the computing resources of cloud platforms, achieving high-performance and high-throughput data processing. In this process, task deployment needs to consider load balancing on the cloud platform to ensure that tasks can be evenly deployed to each computing node. However, currently in the process of providing services on cloud platforms, the available resources of all hosts will be automatically and dynamically readjusted, and it cannot be guaranteed that each task will be deployed to the host with the most remaining resources. This load imbalance in the platform will result in computational results that cannot be returned to users in a timely and effective manner. So, a heuristic multi-task efficient deployment approach for Big Data processing based on QoS awareness and Bayesian classification in cloud environments called QBC is proposed. The QBC first performs long-run QoS awareness on hosts in the cloud; Then, based on user task requirements, selects host nodes that meet QoS constraints to form a candidate set, and performs Bayesian classification to find the host node which has highest a posteriori probability to serve as the clustering center; third, designs an objective function based on euclidean spatial distance to acquire the optimum host clustering set in the candidate set; Finally, deploys the user's tasks to this optimal host cluster set. The experimental results show that this approach implements optimization of long-run load balancing in Big Data cloud platforms with minimal resource consumption, enhances the ability of the cloud platform to provide external support, and thus promotes efficient deployment of multitasking in Big Data processing under cloud computing. The proposed QBC based framework reduces the overall Energy Consumption by an average of 47.98%, MakeSpan by an average of 24.42%, Total Cost by an average of 30.17%, Average Waiting Time by an average of 36.92%, and the Throughput is increased by an average of 41.93% as compared to the existing algorithms.

As the foundation of distributed systems, consensus mechanisms play a crucial role in ensuring the proper operation of the system. In a distributed network composed of trusted peer nodes, nodes may need to perform operations on random Boolean values (0 or 1) during their execution. To address this, this paper, inspired by Paxos, Raft, and their derivative algorithms, proposes a random Boolean consensus algorithm based on the principle of the Probabilistic Propagation Model. In each consensus round, a Boolean value is selected with a smaller probability in the first round of communication, and the final decision is made with a larger probability in the second round, ensuring that the probabilities of the system obtaining 0 or 1 are approximately equal. The running time is further reduced by minimizing the participation of nodes. Experimental results show that this method enables nodes in the system to quickly reach a consensus on a random Boolean value, and as the number of nodes increases, the savings in system overhead become more apparent.

In geo-distributed deployments, sensor data are collected under the coordination of smart terminals and relayed on-chain via decentralized oracles. A motivating scenario involves healthcare networks where regional hospitals submit aggregated medical data to blockchain systems while maintaining strict information security—often designating one gateway per region for external communication. Long geographical distances between smart terminals stress traditional consensus with excessive network overhead and limited efficiency. To address this, we propose a layered BFT consensus method, 2-layer EaZy BFT (2EZBFT). The system forms multiple independent groups of smart terminals and builds a two-layer consensus architecture—“intra-group synchronization, inter-group consensus”—to complete cross-group data aggregation and final on-chain consensus. This layered design reduces intra-group communication complexity by lowering the number of nodes per group and reduces cross-group interactions via leader-side aggregation, thereby lowering overall network overhead. Compared with other BFT algorithms, the proposed scheme improves the efficiency of data collection and on-chain reporting while ensuring consensus security and consistency. Experiments show improvements in metrics such as network overhead and consensus latency. In a discrete-event simulation with an asymmetric WAN latency matrix and geo-partitioned groups, 2EZBFT achieves up to 45% higher throughput than flat BFT algorithms such as PBFT and HotStuff under high load. It provides a practical path for efficient data interaction in decentralized oracles and offers guidance for improving the performance of blockchain–real-world data exchange.

Concurrent wireless charging offers significant performance improvements for Wireless Rechargeable Sensor Networks (WRSNs). However, wave interference, arising from interactions between electromagnetic waves from multiple chargers, disrupts this process. This results in uneven power distribution, potentially leading to significantly attenuated or even negligible energy reception at certain locations. This paper addresses this challenge by introducing the Concurrent cHarging with Efficient phaSe Scheduling (CHESS) problem. CHESS maximizes the energy received by critical sensors through a novel on-demand phase scheduling approach. To achieve this, we propose a practical charging model with charger phases and wave interference effects. Subsequently, a charger grouping algorithm reduces computational complexity, followed by a phase vector searching algorithm to identify optimal phases for maximizing sensor energy reception. Finally, a phase scheduling algorithm enables dynamic adaptation to the evolving energy demands. Simulations show significant efficiency improvements, outperforming baseline algorithms by an average of 89.8%.

Nowadays, advancements in facial recognition technology necessitate robust solutions to address challenges in real-world scenarios, including lighting variations and facial position discrepancies. We introduce a novel deep neural network framework that significantly enhances facial recognition accuracy through multi-scale feature fusion and spatial attention mechanisms. Leveraging techniques from FaceNet and incorporating atrous spatial pyramid pooling and squeeze-excitation modules, our approach achieves superior accuracy, surpassing 99% even under challenging conditions. Through meticulous experimentation and ablation studies, we demonstrate the efficacy of each component, highlighting notable improvements in noise resilience and recall rates. Moreover, the introduction of the Feature Generative Spatial Attention Adversarial Network (FFSSA-GAN) model further advances the field, exhibiting exceptional performance across various domains and datasets. Looking forward, our research emphasizes the importance of ethical considerations and transparent methodologies in facial recognition technology, paving the way for responsible deployment and widespread adoption in the security, healthcare, and retail industries.

This article examines the mechanisms by which aperiodic denial-of-service (DoS) attacks can exploit vulnerabilities in the TCP/IP transport protocol and its three-way handshake during communication data transmission to hack and cause data loss in networked control systems (NCSs). Such data loss caused by DoS attacks can eventually lead to system performance degradation and impose network resource constraints on the system. Therefore, estimating system performance degradation is of practical importance. By formulating the problem as an ellipsoid-constrained performance error estimation (PEE) problem, we can estimate the system performance degradation caused by DoS attacks. We propose a new Lyapunov-Krasovskii function (LKF) using the fractional weight segmentation method (FWSM) to examine the sampling interval and introduce a relaxed, positive definite constraint to optimize the control algorithm. We also propose a relaxed, positive definite constraint that reduces the initial constraints to optimize the control algorithm. Next, we introduce an alternate direction algorithm (ADA) to solve the optimal trigger threshold and design an integral-based event-triggered controller (IETC) to estimate the error performance of NCSs with limited network resources. Finally, we verify the effectiveness and feasibility of the proposed method using the Simulink joint platform autonomous ground vehicle (AGV) model.

The use of privacy-enhanced facial recognition has increased in response to growing concerns about data security and privacy in the digital age.This trend is spurred by rising demand for face recognition technology in a variety of industries, including access control, law enforcement, surveillance, and internet communication.However, the growing usage of face recognition technology has created serious concerns about data monitoring and user privacy preferences, especially in context-aware systems.In response to these problems, this study provides a novel framework that integrates sophisticated approaches such as Generative Adversarial Networks (GANs), Blockchain, and distributed computing to solve privacy concerns while maintaining exact face recognition.The framework's painstaking design and execution strive to strike a compromise between precise face recognition and protecting personal data integrity in an increasingly interconnected environment.Using cutting-edge tools like Dlib for face analysis, Ray Cluster for distributed computing, and Blockchain for decentralized identity verification, the proposed system provides scalable and secure facial analysis while protecting user privacy.The study's contributions include the creation of a sustainable and scalable solution for privacy-aware face recognition, the implementation of flexible privacy computing approaches based on Blockchain networks, and the demonstration of higher performance over previous methods.Specifically, the proposed StyleGAN model has an outstanding accuracy rate of 93.84% while processing high-resolution images from the CelebA-HQ dataset, beating other evaluated models such as Progressive GAN 90.27%, CycleGAN 89.80%, and MGAN 80.80%.With improvements in accuracy, speed, and privacy protection, the framework has great promise for practical use in a variety of fields that need face recognition technology.This study paves the way for future research in privacy-enhanced face recognition systems, emphasizing the significance of using cutting-edge technology to meet rising privacy issues in digital identity.

In recent years, the extensive use of facial recognition technology has raised concerns about data privacy and security for various applications, such as improving security and streamlining attendance systems and smartphone access. In this study, a blockchain-based decentralized facial recognition system (DFRS) that has been designed to overcome the complexities of technology. The DFRS takes a trailblazing approach, focusing on finding a critical balance between the benefits of facial recognition and the protection of individuals' private rights in an era of increasing monitoring. First, the facial traits are segmented into separate clusters which are maintained by the specialized node that maintains the data privacy and security. After that, the data obfuscation is done by using generative adversarial networks. To ensure the security and authenticity of the data, the facial data is encoded and stored in the blockchain. The proposed system achieves significant results on the CelebA dataset, which shows the effectiveness of the proposed approach. The proposed model has demonstrated enhanced efficacy over existing methods, attaining 99.80% accuracy on the dataset. The study's results emphasize the system's efficacy, especially in biometrics and privacy-focused applications, demonstrating outstanding precision and efficiency during its implementation. This research provides a complete and novel solution for secure facial recognition and data security for privacy protection.

In recent years, facial recognition technology has become increasingly integrated into society, making privacy protection crucial. Previous techniques offered minimal secrecy safeguards through simple obscuration methods. This paper addresses the strict privacy requirements of face image data by developing a novel framework that synergistically integrates Generative Adversarial Networks (GANs), clustering algorithms, and Blockchain technology. The methodology proposes a cutting-edge Privacy-Preserving Self-Attention GAN (PPSA-GAN) to generate realistic synthetic facial imagery. An integrated mini-batch K-means clustering algorithm anonymizes these images into distinct groupings, maximizing privacy preservation. Blockchain integration complements the system by fortifying trust through decentralized ledgers for transparent yet secure data storage and auditing. Rigorous benchmarking on the CelebA dataset confirms the PPSA-GAN architecture’s state-of-the-art performance, attaining an impressive Inception Score of 13.99 and a Fréchet Inception Distance of 35.50. The mini-batch clustering forms 125 distinct clusters, effectively anonymizing facial attributes within the synthetic images. Blockchain integration further bolsters privacy assurances via tamper-proof historical records, showcasing precision, recall, F1-score, and accuracy values of 0.948, 0.938, 0.943, and 0.947, respectively. This multifunctional framework represents a novel contribution, fostering an ethical technological ecosystem that balances progress and privacy. Prospective deployment horizons encompass identity verification, surveillance infrastructure, and augmentation of medical image repositories, seeding an enlightening future for facial recognition domains.

In this article, by incorporating with the skills of fuzzy modeling method and fuzzy switching control technique, the nonfragile sampled-data control (SDC) synthesis problem for the ship electric propulsion systems (SEPSs) with lag transmission signal and stochastic actuator failures (SAFs) has been addressed. Firstly, to delineate the SAFs, the control input model is incorporated with semi-Markovian jump faulty coefficients for depicting various stochastically occurring actuator failures, which has stronger modeling ability and exhibits the practical scenario more accurately. Secondly, in contrast with the conventional SDC mechanism, an invariant lag signal is introduced in the control rule and the sampling pattern is assumed to be heterogeneous within an interval, which is exploited innovatively to tackle the control synthesis issue. Thirdly, by fully capturing the information of the sampling intervals [tk,tk+1) and [tk−τ,tk+1−τ) as well as the available information of fuzzy membership functionals (FMFs) simultaneously, a refined looped Lyapunov functional called dual-sided sampling-interval-dependent and looped FMFs-dependent Lyapunov functional is raised, which is deliberately tailored for the utilization of N-order generalized free-matrix-based integral inequality and canonical Bessel-Legendre integral inequality. Afterwards, by resorting to the FMFs and utilizing the time derivative information of FMFs, several sufficient conditions are established in the shape of multi-parameterized linear matrix inequalities (LMIs) such that the underlying fuzzy systems can achieve the globally asymptotically stable based upon the switched control technique. The desired nonfragile control gain matrices can be subsequently determined by checking the developed LMIs. Besides that, the theoretical developments show the hierarchical character, and the conservatism decreases with the increase of the level of hierarchy. Conclusively, the proposed control technique for the stabilization issue is validated by a numerical simulation performed on the SEPSs.

Computing trajectory similarity plays a critical role in various spatio-temporal applications that involve trajectory analysis. In recent years, trajectory representation learning has been extensively studied and applied for trajectory similarity calculation. However the majority of existing algorithms for trajectory representation generally have two problems. The first problem is the emphasis of spatial similarity over temporal similarity, and even to discard the temporal dimension of spatio-temporal trajectories. As a result, the outputs of these approaches cannot fully represent the similarity of spatio-temporal trajectories. The second problem is the introduction of additional information, such as the topology of the road network, which increases the uncertainty of capturing the spatio-temporal correlation of trajectories and prevents their application in scenarios where it is difficult to obtain such information. This poses a significant challenge when dealing with complex and time-varying traffic networks. This paper proposes a novel method, named STTraj2Vec (Spatio-temporal Trajectory 2 Vector), which relies only on spatio-temporal trajectories to capture their similarity without spatio-temporal separation. This takes into account the whole spatio-temporal trajectory information. In this method, an extended clustering algorithm is introduced, which maps the trajectory into a point-region quadtree, and constructs a time-varying virtual network structure based on the point-region quadtree. In this method, an extended clustering algorithm is introduced, which maps each trajectory into a point-region quadtree, and then completes density clustering through the adjacency relation of leaf nodes to construct a time-varying virtual network structure. This virtual network structure not only considers the spatial proximity, but also the time, so as to reflect the spatio-temporal characteristics of the trajectory more accurately. Then, a novel spatially and temporally integrated random walk algorithm is designed, which carries out spatiotemporal random walk on the virtual network structure, to capture the spatiotemporal characteristics of the trajectory, and thus obtains the representation of all nodes in the virtual network. Furthermore, each trajectory is converted into a sequence of vectors on the virtual road network according to the latitude, longitude, and time of the trajectory points. Finally, based on these node representations and trajectories, a transformer model with ranking loss is employed to capture the distinct contributions of the various locations and times to the similarity computation and encode each sequence of vectors into target vectors. Experiments on two public datasets show that STTraj2Vec is superior to the state-of-the-art methods in terms of effectiveness for top-k trajectory similarity search and trajectory clustering, while exhibiting low parameter sensitivity and high model robustness.