ABSTRACT. Dynamic Visual Sensors (DVS) output pixel information asynchronously in the form of address-event, only transmitting information for pixels where the intensity change exceeds a threshold, thereby reducing redundant data generation at the source. They possess advantages such as low-latency response and high efficiency with low power consumption. However, factors such as thermal noise and leakage current can lead to noise events even when the light intensity remains constant, which affects image quality and subsequent processing. In this paper, we propose a new method for calculating event density, called Local Density Segmentation(LSD), which divides events and their surrounding neighborhoods into different regions for independent event density calculation. Based on LSD-calculated event density, we further propose a denoising method named Local Density Segmentation K-means(LSD-K), which utilizes event density as a feature to select and remove noise events. Finally, through experiments, we validate the performance of our algorithm.

ABSTRACT. This study aims to investigate cyberbullying incidents on Weibo by constructing a comprehensive dataset of labeled conversations. We collect 89K social media sessions from 10K user profiles and manually annotated the data to identify instances of cyberbullying. We analyze cyberbullying based on three fundamental characteristics, examining its distribution across different topics and the relationship between posting times and cyberbullying incidents. Additionally, we explore the influence of user popularity on the similarity of receiving bullying comments. Our study provides valuable insights into the characteristics of cyberbullying on Chinese social media and highlights the need for culturally sensitive detection models.

ABSTRACT. In current online education, Knowledge Tracing (KT) technology plays a pivotal role, in monitoring and updating students' evolving knowledge states throughout their educational journey. Previous knowledge tracing models like DKT and DKVMN, have made some progress in personalised learning. However, they fall short in fully capturing the dynamic changes in the learning process, including forgetting behaviour and individual learning gains. To address these issues, this paper introduces a novel knowledge tracing model named dynamic forgetting and enhanced gain mechanisms (DF-EGM), which builds on Deep-IRT. The model initially utilises a CART decision tree classifier to extract pre-classification labels by analyzing response times and learning abilities. Subsequently, the model incorporates forgetting factors and personalised learning gains. It introduces attempt counts and time intervals to simulate knowledge forgetting. and the realization of personalised learning gains is achieved through a deeply refined knowledge gain module. This module models the perceived difficulty of exercises by students and dynamically adjusts learning gains based on their answer performance and current knowledge state. Experiments on public datasets indicate that the DF-EGM model can more accurately track students' knowledge states.

ABSTRACT. Water quality prediction is an important tool for environmen-tal protection and public health. It can prevent water pollution, maintain ecological balance, and provide crucial scientific evidence for water resource management. To address the low prediction accuracy of traditional methods, this study uses data from the water quality monitoring station on the Huang-pu River in the Taihu Basin of Shanghai to predict dissolved oxygen levels. We propose a water quality prediction method based on feature engineering and the BiGRU-Attention-TCN model. The proposed method employs feature engineering to extract monitoring station status information, improving mod-el robustness and stability. It uses Bidirectional Gated Recur-rent Unit (BiGRU) to process time series data, capturing bidi-rectional dependencies, and applies the Attention mechanism to enhance the model's focus on key time steps and features. Additionally, Temporal Convolutional Network (TCN) is in-troduced to further extract long-term dependencies and global features. Our method was compared with several commonly used baseline methods, and experimental results show that the proposed method outperforms other advanced baseline meth-ods.

ABSTRACT. Despite modern artificial intelligence has achieved significant success in static data learning, it still faces challenges in continual learning. The stability-plasticity dilemma is a key challenge in continual learning. To address this problem, we propose a novel continual learning network, DTCL, that integrates three dynamic training strategies: adaptive learning rate, experience replay, and knowledge distillation. These strategies collectively enhance the ability of the network to learn new tasks while preserving information from previous tasks. Experimental results on the CIFAR-100 dataset demonstrate that DTCL significantly outperforms existing baselines in average accuracy. The dynamic training strategy of DTCL effectively balances stability and plasticity, thereby improving the performance of continual learning.

ABSTRACT. The paper relies on the team's "WeChat Mini Program Development—From Basics to Practice" course at Chinese universities' MOOCs. By analyzing approximately 60,000 learning data points in course development, the study aims to predict students' grades. This paper addresses traditional models' complexity and subpar performance in prediction processes by utilizing and optimizing the time-series model Nbeats. It compares the predictive performance of this model with mainstream forecasting models to explore its superiority. In addition to the data collected by the team, three publicly available datasets are incorporated to eliminate randomness in model predictions. Through ablation experiments, it demonstrates the Nbeats model's greater robustness and effectiveness.

ABSTRACT. Kubernetes has become the basic platform for building cloud native applications. However, existing horizontal scaling methods based on Kubernetes have problems with resource redundancy. Furthermore, the combined horizontal and vertical scaling based on Kubernetes takes a long time. Regarding the issue above, first, we describe the execution processes of three scaling techniques and compare them in different dimensions;second, we construct a formal elastic scaling model for expansion scenarios, with the goal of minimizing the total cost; and third, we introduce a heuristic expansion algorithm after balancing resource redundancy and execution time as the heuristic information. The experiment results show that the elastic scaling algorithm proposed in this paper achieves better performance than other algorithms.

ABSTRACT. With the rapid development of the 6G network, the proliferation of Internet of Things (IoT) devices and demand for high-bandwidth and low-latency communications are driving the need for more efficient paradigms. However, due to inherent limitations, IoT devices cannot handle complex tasks independently, necessitating offloading to more powerful resources. As an extension of 6G, the Space-Air-Ground Integrated Networks (SAGIN) achieve broader coverage and efficient resource utilization. Nevertheless, task offloading to SAGIN faces challenges in inefficient resource utilization and security concerns. To address these issues, this paper proposes a novel framework integrating blockchain technology. First, we present a joint task offloading and resource allocation framework in a blockchain-empowered SAIGIN (BE-SAGIN). Based on network topology and communication characteristics, a customized blockchain consensus mechanism is designed. Furthermore, the joint task offloading and resource allocation problem is formulated as an optimization problem, considering the cost of tasks and the blockchain system, including delay and energy consumption. Extensive experiments demonstrate that the proposed scheme can significantly reduce system costs, improve resource utilization, and enhance system efficiency compared to baseline approaches.

ABSTRACT. Banks form interconnected networks through interbank lending. These network connections enhance interbank relationships but also provide channels for risk contagion. When risks spread within the interbank lending network, the government must promptly offer necessary controls and bailouts based on the contagion situation. However, most current studies on bailout strategies only consider the static environment and lack the ability to address dynamic changes during the contagion process. Therefore, this paper constructs a government bailout model for dynamic risk contagion processes. Using the greedy algorithm, the model transforms the dynamic bailout problem into a static multi-objective optimization problem. When contagion risk is detected and external assistance is deemed necessary, the model can derive the optimal bailout plan targeting the dual objectives of minimizing both the bailout funds and the systemic risk. Finally, experiments on a real-world contagion bailout scenario confirm the proposed model's effectiveness and superiority.

ABSTRACT. Smartphones are one of the most widely used carriers in network technology applications, and the information security is of paramount importance. To achieve secure, unobtrusive, and innovative continuous authentication in mobile scenarios, further research into implicit biometric authentication techniques is necessary. This paper proposed a continuous authentication method based on Transformer, which enhances multimodal biometric recognition capabilities by learning user behavior patterns from keystrokes and IMU data on mobile devices. Specifically, this study optimizes the multi-head attention mechanism by constructing dynamic behavior weight module and employs the quadruple loss function to improve the accuracy of biometric recognition. Through experimental results analysis, the effectiveness of the proposed model is demonstrated. The model shows high usability and a low misclassification rate on multimodal biometric data, making it suitable for implicit and continuous identity authentication tasks on smartphones.

ABSTRACT. Due to the need to handle a huge size of image data and limited battery power at each node, visual sensor networks (VSNs) suffer from short network lifetime and long network delay. Consequently, how to reduce network traffic is very important in VSNs. The deployment of processing nodes for feature extraction is an effective way to reduce it. Deploying multiple types of camera nodes should help do so as well. Therefore, in this work we investigate how to deploy a heterogeneous VSN for target coverage and feature aggregation with lifetime guarantee at the minimum cost. Since this problem is NP-hard, we propose two heuristics, i.e., separate and joint node deployment algorithms to solve it. A large number of experimental results expose that the latter has higher success rate and can obtain lower network deployment cost than the former at the expense of extra computation.

ABSTRACT. As the Internet and big data technologies advance, a tremendous amount of data is generated daily. Efficient network operations are essential for handling this data. Accurately predicting future network traffic in real-time enables prompt response from cloud infrastructure and efficient traffic scheduling and allocation, ultimately reducing costs, preventing economic losses, and optimizing the performance of downstream facilities. However, predicting network traffic in large-scale data centers is challenging due to the multidimensional, nonlinear, and high-volatility natures of the time series. Traditional prediction methods, e.g., regression algorithms, struggle to capture nonlinear features effectively. Many deep learning models face issues such as gradient explosion or vanishing during training. Current commonly used prediction methods do not fully uncover vital information about the frequency domain features in the time series. Consequently, this work proposes a novel network traffic prediction model named SDMTAB that combines a Savitzky Golay filter, Sequence Decomposition, Multi-Scale Attention, Temporal convolutional network, an Autocorrelation mechanism, and Bidirectional Long Short-Term Memory (Bi-LSTM). SDMTAB accurately predicts future network traffic by adaptively extracting important features without requiring excessive feature engineering in the original data. It performs end-to-end sequence prediction, avoiding redundant data processing, feature selection, and automatic learning of data features and temporal dependencies to achieve high-quality time series prediction. Experiments involving ablation and comparisons with advanced prediction models are performed with three datasets from the Google cluster trace. Experimental results with realistic data show that SDMTAB improves the prediction accuracy by at least 38.52% compared to state-of-the-art models and its components, with each module contributing to performance improvement.

ABSTRACT. This work studies the distributed permutation flowshop scheduling problems (DPFSP) for minimizing the maximum completion time (makespan). First, a mathematical model is established for the DPFSP. Second, four meta-heuristics, named genetic algorithm (GA), particle swarm optimization (PSO), artificial bee colony (ABC), and differential evolution (DE), are used and improved to solve the problems. Third, the Nawaz-Enscore-Ham (NEH) heuristic is employed to initialize the factory allocation scheme. Then, six local search operators are designed based on problem-specific knowledge. Two Sarsa-based strategies are developed to enhance the exploitation of meta-heuristics by optimizing factory allocation and choosing appropriate local search operators, respectively. Finally, to validate the effectiveness of the proposed algorithms, we solve 72 benchmark instances. The performance of the two Sarsa-based strategies is evaluated. Comparisons and discussions show that the improved ABC with the Sarsa strategy for local search selection is the best one among the 16 algorithms.

ABSTRACT. Besiege and Conquer Algorithm (BCA) is a novel evolutionary algorithm recently proposed, inspired by human warfare. It outperforms other algorithms in both single-objective and multi-objective optimization scenarios. However, its design has certain defects and cannot balance exploration and exploitation. In response to this problem, we are inspired by the guerrilla warfare principle and propose some mechanisms to balance exploration and exploitation and propose an improved BCA algorithm Guerrilla Warfare Strategy Ameliorate Besiege and Conquer Algorithm (GW-BCA). First, soldiers’ ambush actions are increased to wait for potential enemies, thereby reducing the likelihood of the algorithm being trapped by local optima. Second, during the soldiers' updating process, valuable information is collected to build a backstage battlefield, helping the algorithm escape local optima and achieve global optima. Third, following the idea of guerrilla warfare, GW-BCA dynamically selects different best armies to besiege and conquer, which effectively increases the mobility of the army. The GW-BCA algorithm was compared with several state-of-the-art evolutionary algorithms on the CEC 2017 benchmark functions. Experimental results show that the proposed algorithm effectively addresses the shortcomings of BCA and demonstrates excellent capabilities compared to other algorithms. The source code will be available at www.jianhuajiang.com.

ABSTRACT. Production scheduling performance can not only be improved by good algorithms, but also by resource augmentation. This paper studies identical parallel batch machines scheduling to minimize total completion time with machine capacity augmentation. The parallel batch machine can process multiple jobs as a batch simultaneously. We quantitatively analyze the impact of machine capacity augmentation on total completion time for unit processing time. We derive an upper bound and lower bound of the impact ratio. This research provides new insights into production scheduling performance improvement with machine capacity augmentation.

ABSTRACT. Distributed flow shop scheduling problems considering dynamic job arrival are important and challenging issues in manufacturing systems. To address the problem, a mixed integer programming model is formulated to minimize the maximum completion time. Then, an enhanced genetic algorithm is proposed to solve the model. Genetic operators are performed to update population and quickly search the global solution space. Meanwhile, a local search method is applied to refine the best solutions in population. Finally, the developed method is compared against two well-known meta-heuristics on a set of test instances. Experimental results demonstrate that the proposed method exhibits superior performance in addressing distributed flow shop scheduling problems with dynamic job arrival.

ABSTRACT. The latent factor analysis (LFA) is an important and effective matrix decomposition model which has important applications in recommended system. Traditional LFA models based on the stochastic gradient descent (SDG) algorithm are difficult to adjust the learning rate which is a critical parameter in the model training process. To address this problem, the particle swarm optimization (PSO) algorithm is often used to adjust the learning rate of LFA model due to its powerful global search ability. However, the standard PSO algorithm easily falls into a local optimal solution, causing accuracy loss. In this paper, the PSO based on Gaussian mutation (DG-PSO) is innovatively proposed by incorporating the elite particle selection strategy to improve its convergence speed and accuracy. Then the DG-PSO based LFA (DGPL) model is constructed by employing the DG-PSO to adaptively tune the learning rate of SDG in the training process. The proposed model is tested on dataset and compared with the state-of-the-art models. Experimental results show that the proposed DGPL is superior than existing LFA models.

ABSTRACT. In decomposition-based evolutionary multi-objective optimization algorithms (MOEA/Ds) with adaptive strategies for weight vectors, the vectors are updated periodically. Their updates’ timing and frequency significantly influence optimization outcomes. Better results are promising to be obtained when weight vectors are updated at the time that the population reaches stability. Therefore, this work proposes a solution stability judgment mechanism based on a dynamic grid system to analyze whether the solutions in objective space have reached a state of stable convergence. To verify the effectiveness of our proposed mechanism, we add it to three MOEA/Ds with adaptive weight vector strategies and conduct experiments on 11 benchmarks compared with these algorithms’ original forms. The results demonstrate that the proposed mechanism can significantly enhance the optimization performance of these algorithms. It can find better weight update times, which helps to result in better solutions in limited iterations with a faster speed.

ABSTRACT. Human-robot collaboration in manufacturing is an increasingly important topic in the field of robotics. However, the unresolved mechanical and stiff behaviors of the current collaborative robot make its interaction with humans extremely dull, especially for an extended period of time. Such interaction patterns could further deter the human’s willingness to work with such robots. To solve this question and be inspired by human-human communication, this paper aims to enable collaborative robots’ social-emotional competence by adding visual information based bidirectional emotion interaction (VI-BEI) in the collaboration process and uses manufacturing co-assembly tasks as working scenarios to test the developed interactive interface. The developed bidirectional emotion interface enables the robot to not only recognize human emotions visually but also provide artificial emotion feedback via 3D simulation technology which has flexible and quick prototypes, customization, and upgrading advantages compared to hardware design. Specifically, this paper introduces (1) the development of a 3D digital human interface that not only monitors human facial expressions but also produces artificial emotion feedback. (2) integrating the 3D digital human interface enables a collaborative manufacturing robot to express real-time emotions in addition to performing actions during co-assembly tasks, facilitating a friendly collaboration process. (3) validation experiments and analysis to evaluate the effectiveness and performance of the updated collaborative robot with facial expressions through real-world assembly tasks. The experimental results and analysis demonstrate the effectiveness and advantages of the current system, as well as guide the future improvement of the developed collaborative robots to be more empathic and friendly.

ABSTRACT. Blockchain technology, the foundation of cryptocurrencies like Bitcoin, has utility beyond finance due to its decentralized and secure transactional nature. However, today’s blockchain networks face the challenge of low transaction throughput due to block propagation delays. In this study, we propose a novel approach that divides large blocks into segments and leverages multi-origin pipelining for segment dissemination, thereby expediting block propagation. The scheme also integrates the use of a Bloom filter for lightweight segment integrity verification to prevent the reception of tampered segments. Simulation results reveal that our scheme outperforms Bitcoin’s legacy approach, demonstrating significant improvement in propagation speed, particularly for large blocks, while maintaining a low fork rate.

ABSTRACT. WiFi backscatter communication has gained many applications, but its performance characteristics remain to be analyzed. While existing research has investigated the success probability of backscatter tags in contention-based WiFi backscatter networks (CWBNs), it has focused solely on the first-order statistic of the signal-to-interference-plus-noise ratio (SINR). In this paper, we present a meta distribution-based fine-grained analysis that provides high-order statistics of SINR and characterizes the disparity among backscatter transmission links in CWBNs. Leveraging stochastic geometry, we, for the first time, derive mathematical expressions for the $b$-th moments of conditional success probability and its meta distribution. The extensive Monte-Carlo simulation results validate the accuracy of our proposed theoretical model and demonstrate its outstanding value to help us understand the overall performance of CWBNs.

ABSTRACT. Credit card fraud detection (CCFD) is essential for the integrity of modern financial systems and is significantly challenged by class imbalances, where the volume of legitimate transactions substantially exceeds that of fraudulent ones. This imbalance severely limits the ability of CCFD models in accurately learning and detecting fraudulent patterns. Although existing techniques adeptly mitigate class imbalances, their application in CCFD is compromised by the adept concealment tactics of fraudsters, which enhance the behavioral overlap between legitimate and fraudulent transactions. This study introduces a novel time-aware density-based generative network model method (TDGN) incorporating a density-based wasserstein generative adversarial network (DWGAN) with a time-aware gated network (TaGN). The density-based wasserstein generative adversarial network penalizes synthetic samples that closely mimic legitimate transactions and refines the generator's focus on distinct fraudulent characteristics, thus minimizing the detrimental effects of behavioral overlap and producing high-quality synthetic samples. Subsequently, TaGN is employed to transform transactional features into new representations that more effectively delineate between fraudulent and legitimate transactions, thus improving the CCFD model’s ability to manage class imbalance. Experimental results from a large-scale real-world credit card transaction dataset demonstrate the superior performance of our proposed method compared to its competitors, offering a robust solution to the persistent challenge of class imbalance in CCFD.

ABSTRACT. Due to the suddenness of disasters and the presence of multiple uncontrollable factors, such as the number of casualties, the availability of emergency relief supplies, and the responsiveness of hospitals, it is crucial to ensure the timely delivery of emergency relief supplies and the promptness of medical services under conditions of significant uncertainty. Therefore, this paper, based on the myriad uncertainties caused by disasters, adopts a priori optimization methods, incorporates multiple random variables and discrete stochastic scenarios to enhance model reliability, and proposes a two-stage stochastic model for the location-allocation of emergency facilities. This model is further transformed into an equivalent deterministic model based on discrete scenarios, and integrates queueing theory to calculate the casualties' waiting penalty cost. For solving the model, this paper first employs an outer approximation method to simplify the nonlinear parts of the model, facilitating the resolution of small-scale problems; for large-scale problems, a Kernel search algorithm is proposed, enhanced by a sorting algorithm. Finally, the effectiveness of the model and algorithms is validated through test instances.

ABSTRACT. In this paper, an optimization method based on a group cooperation strategy is proposed to make reasonable improvements to the problems of low transformation efficiency and high operational loss in the formation process of a multi-vehicle cooperative system. Three objective functions (cost loss degree, obstacle avoidance constraint, queue compactness) are constructed by analysing the multi-vehicle system. Secondly, the grey wolf optimization introduces a multi-period cosine convergence factor and a hunting search strategy based on dimensional learning, which effectively improves the algorithm's tendency to fall into the phenomenon of locally optimal solutions. Finally, the multi-objective grey wolf optimization is used to iterate to obtain the corresponding Pareto frontiers and simulation results, and the value coupling function is established to find out the optimal path that reduces the distance by 60% compared with the original, which reasonably achieves the path improvement of the multi-vehicle system at the overall level.

ABSTRACT. In the cellular-based vehicular communication network, a V2V link shares the same spectrum resources with a V2I link, resulting in mutual interference between them, and adversely affecting the vehicular communication performance. The existing resource allocation strategies in cellular-based vehicular communication networks rely on individual links to adjust network parameters, leading to unfair resource allocation. In order to resolve this problem, we propose a joint spectrum sharing and power control strategy based on deep reinforcement learning is proposed. Firstly, a coordination graph is designed to describe the dependencies among V2V links. Then, a graph attention mechanism is proposed to extract the relevant information from neighboring V2V links. Finally, based on the multi-agent deep deterministic policy gradient method, a coordination learning algorithm for wireless resource allocation in the cellular-based vehicular network is established. Experimental results show that the proposed method can improve the transmission rate of V2I links and the successful transmission rate of V2V links.

ABSTRACT. Traffic congestion has become an increasingly severe problem and needs to be solved urgently. Reinforcement learning (RL) is an advanced data-driven approach for large-scale adaptive traffic signal control (TSC) in the complex urban traffic network. The decentralized multi-agent RL (MARL) framework is feasible for multi-intersection scenarios. However, there is the lack of effective behavior policy of the agent to explore actions in TSC scenario. In this paper, a novel behavior policy is proposed to introduce the Max-Pressure control method in transportation field into the ϵ-greedy behavior policy. The agent selects the action with the maximum pressure phase to maximize the traffic throughput of the intersection. It enables each agent to choose the appropriate action according to the real-time traffic condition at the intersection in TSC scenario. The novel behavior policy is applied to a multi-agent Q-learning algorithm for TSC at multiple intersections. Experiments are carried out on a large synthetic grid under simulated peak-hour traffic dynamics. Results demonstrate that our approach outperforms the other existing decentralized MARL algorithms.

ABSTRACT. Emergency vehicles play an important role in responding to time-critical accidents such as traffic accidents, medical emergencies, and urban fires. The increase in car ownership and the lag of road infrastructure construction have led to increasing urban traffic congestion, which seriously affects the rescue efficiency of EVs. Existing emergency traffic control strategies are often designed on the premise that a single EV goes to a single accident point, ignoring the complex environment of multiple EVs and multiple accident points. Therefore, this paper proposes a multi-vehicle coordinated emergency traffic guidance heuristic strategy for the simultaneous presence of multiple EVs and multiple accident points in the traffic scene. This strategy ensures that the EVs can reach the emergency scene in time, and reduces the negative impact on other traffic participants (other EVs and regular vehicles). Firstly, the priority of an accident emergency degree is determined based on fuzzy logic. Then, two traffic guidance schemes are designed respectively for the situation that multiple EVs go to the same or different accident point from different starting points, so as to reduce the conflict in the process of traffic guidance. Finally, in order to test the effectiveness of the proposed strategy, it takes the traffic network of Manhattan Island in New York as a scene and uses the traffic simulation software SUMO to carry out simulation experiments. Experimental results show that the proposed strategy is superior to the existing methods in terms of the emergency service response time of EVs and the normal traffic time delay caused by EVs.

ABSTRACT. As disassembly lines evolve, keeping workers safe from all potential dangers has become paramount. This work proposes an improved Proximal Policy Optimization for Disassembly Line(PPO-DL) approach to address Multi-objective Disassembly Line Balancing Problem(MDLBP) that integrates task switching time and hazardous tasks. The MDLBP model aims to optimize disassembly net profit, balance workstation utilization, and mitigate penalties from hazardous tasks. The proposed PPO-DL modifies the action space of PPO to accommodate MDLBP characteristics, ensuring efficient learning by restricting actions to valid ranges through action masking. Comparative experiments with PPO, Q-learning, and A2C demonstrate PPODL’s superiority in finding near-optimal solutions across various scenarios.

ABSTRACT. In this paper, we study the problem of discovering new types of defects on steel surfaces. Steel is an indispensable and important material in modern industry, making steel surface inspection critically important. However, existing methods are insufficient for discovering new types of steel surface defects. A method based on UNified Objective function (UNO) was developed to address these limitations. UNO unifies all objectives under a single cross-entropy loss using multi-view self-labeling, which simplifies the integration of supervised and unsupervised learning. Additionally, multi-head clustering and overclustering strategies are integrated to improve clustering performance and representation quality. Experimental validation demonstrates that our method significantly outperforms existing methods on the NEU dataset, achieving an ARI of 0.9750 and an NMI of 0.9614.

ABSTRACT. The global water environment confronts numerous challenges, e.g., water pollution, overexploitation, and ecological degradation. Comprehensive protection and management are imperative for sustainable water resource utilization. Water quality predictions provide timely warning of future water quality problems and enable early action to avoid deterioration. As science and technology are increasingly applied in comprehensive water environment management, a diverse array of multimodal data is gathered from various sources, including remote sensing images and hydrological time series. However, current water quality prediction methods, e.g., statistical, machine learning, and deep learning methods fail to utilize multimodal data to enhance their accuracy of water quality prediction. To solve the above problem, this work proposes a multi-factor and long-term water quality prediction model based on multimodal data fusion named Low-rank Multimodal Fusion TimesNet (LMF-TimesNet). It first extracts features from hydrological time series and remote sensing images, respectively. Then, they are fused with the low-rank multimodal fusion network to extract diverse information. Finally, TimesNet is adopted to integrate fused multimodal water environment information for water quality prediction. Experimental results on a real-world dataset show that LMF-TimesNet achieves higher prediction accuracy and generalization ability than its state-of-the-art peers.

ABSTRACT. With the increasing role of distributed message systems (DMS) represented by Kafka in handling large-scale data streams, it is critical to improve system performance by optimizing configuration parameters in different production environments. Due to the target system having numerous configuration parameters and the complex interactions between them, parameter optimization becomes a complex and time-consuming process that relies heavily on human effort and an iterative trial-and-error approach. At present, this kind of configuration parameters problem has been classified as the Configuration of Distributed Message Systems (CDMS) problem, whose goal is to obtain the optimal set of parameters for best performance. In order to deal with this complex problem, an automatic optimization method is proposed in this paper, which combines an automatic script-based deployment and running framework with an improved Genetic Algorithm (GA) that is cooperated with the Random Forest (RF) and statistical methods, named Directional-Mutation Genetic Algorithm (DMGA). The framework and algorithm are designed to find the best configuration efficiently, obtaining optimal parameters accurately without human supervision. The real results of comparison experiments based on the Kafka system show that the DMGA algorithm significantly improves the data throughput performance of Kafka, and its optimization speed is more efficient than the basic GA, highlighting its potential to quickly optimize Kafka's operational efficiency in different scenarios.

ABSTRACT. The Internet of Behaviors (IoB) aims to leverage advanced sensing and computing technologies to understand human behaviors, and Human Action Recognition is an essential task in IoB. However, recognizing human actions across varied and unseen domains remains a challenging problem, largely due to domain shifts that degrade the performance of conventional models. To address the aforementioned issue, a Deep Causal Domain Generalization Network (DCDGNet) is designed to enhance the generalizability of the human action recognition model. Specifically, a structured causal model for the data generation of human motion is constructed, where human motion signals are decoupled into action class-related causal factors and domain shift-related non-causal factors. DCDGNet integrates a human domain encoder, causal encoder, domain discriminator, motion reconstruction decoder, and action classifier, contributing to the model’s ability to disentangle causal factors from non-causal ones from domain-specific features and achieve domain-invariant representations. Extensive experiments demonstrate that DCDGNet significantly outperforms existing state-of-the-art methods in terms of both accuracy and generalizability.

ABSTRACT. Anomaly detection of business process can monitor the business process in real time to help enterprises to intervene before the problem occurs, to ensure the security and continuity of business process. In this paper, we propose an anomaly detection method for business process based on graph convolutional neural network. The traces in the offline event logs are first extracted from the prefix traces to construct a trace dataset containing anomaly labels; then the Word2Vec model and graph convolutional network model are used to complete the embedding of local and global features of the traces and get the final embedding of the traces through the fusion of the GCN-Word2Vec structure; finally, anomaly detection of the business process is accomplished with the help of the neural network classifiers. Experiments are conducted on four real event log datasets, and the experimental results show that the GCN-Word2Vec structural anomaly detection model proposed in this paper has a certain degree of improvement in the accuracy of anomaly detection compared to the BERT algorithm and the Transformer algorithm, with the highest accuracy rate reaching 93.75% and the average accuracy rate being 92.85%.

ABSTRACT. Glaucoma is one of the leading causes of blindness worldwide. Patients with open-angle glaucoma have difficulty realizing the graduate vision loss. However, the morphological change of the optic nerve head can be manifested in the retinal image. Therefore, there are many studies offering computer-aided diagnosis for glaucoma based on retinal image analysis. With the trend of deep learning, there have been several attempts to utilize deep-learning-based glaucoma classification in practice. However, the limited size of the dataset is a problem when adapting deep-learning models to glaucoma classification. Therefore, we test three vision transformer models to handle this problem and demonstrate their performance in glaucoma classification. Also, based on their performance, we give suggestions for their usage in clinical practice.

In this paper, we fine-tuned three transformer models for glaucoma classification using two datasets, RIM-ONE DL and Drishti-GS. Based on our experiments, the pyramid vision transformer (PVT) model achieves the best performance, with good scores in terms of accuracy and sensitivity. This suggests that the PVT can catch the missing detail and can potentially be used to support the glaucoma diagnosis.

ABSTRACT. Exploiting contextual information is important for road segmentation tasks. In contrast to existing work primarily focusing on expanding the receptive field and stacking network layers, we propose a novel Graph-representation-based Multi-scale Feature Fusion U-net model (GMFFUnet) to address this problem. Our method enhances global context extraction by establishing connections between features of varying types and scales. Specifically, our model projects the entire feature pyramid into multiple interaction spaces and models the complex relationships among them to construct a graph and facilitate information exchange. The enhanced features are then fused back into the original feature pyramid. Our method provides substantial benefits on road segmentation tasks with complex backgrounds. Comparative experiments were carried out to evaluate the performance of the proposed method, and the results demonstrated that our proposed method outperforms all SOTA methods on the Massachusetts road dataset.

ABSTRACT. Due to their ability to operate effectively in a variety of environments, including unstructured terrains, humanoid robots have undergone significant advancement. The type of actuation employed for such robots is crucial to ensure the required force while respecting the volume constraints. In this context, a servo electro-hydraulic actuator, characterized by its power-to-weight and power-to-volume ratios, has been developed for humanoid robots. This actuator is distinguished by an integrated force compensation module, which enhances safety. Nevertheless, a stable control architecture that operates in real-time without system failure is still crucial for ensuring the robot's safety and stability. For a stable and safe functioning of the humanoid, the key requirement of the control architecture is to ensure stable communication between the different components with high update rates while respecting time constraints. The proposed solution has the advantage of being a distributed, real-time, open-source, modular, and adaptable control architecture. The contribution of this paper is an enhancement of the updated rate compared to other humanoids by 20% and by 40% in the latency of the master. The results demonstrate the potential of using EtherCAT fieldbus and open-source software to develop a stable robot control architecture.

ABSTRACT. Virtually coupled train set (VCTS) has become a popular research topic in the railway industry in recent years. It is considered one of the solutions to the uneven spatial and temporal distribution of passenger flow in urban rail transit. A VCTS consists of multiple train units (TUs). The overall operation of the VCTS relies on the coordinated efforts of each TU. A feasible approach involves pre-planning reference curves for the TUs, with the Automatic Train Operation (ATO) system tracking these curves. However, as the TUs operate at close intervals, the control strategy needs to ensure safety and precise control. Consequently, the reference curve of the TU is characterized by frequent changes in acceleration. As a result, the ATO system is prone to deviating from the curve. This paper builds a time-varying vehicle dynamics model and a controller performance model based on data learning. Using these models, we predict deviations and construct an online optimization model. A mesh adaptive direct search (MASD) algorithm is designed to solve for online adjustment curves, promptly correcting any deviations. Simulation experiments demonstrate the effectiveness of the proposed online optimization method when combined with PID and MPC algorithms.

ABSTRACT. Transaction Fraud Detection (TFD) is a vital aspect of contemporary financial security, substantially challenged by class imbalance, with legitimate transactions far outnumbering fraudulent ones. This imbalance challenges TFD models in effectively learning fraudulent patterns. Traditional over-sampling techniques, while effective in addressing class imbalances, encounter difficulties in TFD due to the concealment of fraudsters, resulting in increased behavioral overlap between fraudulent and legitimate transaction patterns. In this work, we introduce a novel generative adversarial network-based hybrid sampling method (GANHS) that first employs over-sampling through a dual-discriminator generative adversarial network (D2GAN) to generate a balanced transaction dataset. Subsequently, an adaptive neighborhood-based under-sampling method is employed to refine the generated samples, eliminating those that exacerbate behavioral overlap. Our method not only generates samples that closely resemble the actual distribution of fraudulent patterns but also markedly reduces the impact of behavioral overlap, thereby enhancing the detection of highly concealed fraudulent transactions. Experimental results on a large-scale real transaction dataset demonstrate that our proposed method outperforms its competitive peers, highlighting its potential to advance the effectiveness of TFD systems.

ABSTRACT. Collaboration enhances work performance and efficiency, yet certain collaborative actions may induce stress in individuals. This study aims to examine variations in stress levels among participants during collaborative tasks and explore changes in their emotional status based on different collaboration styles with their co-workers (such as slow response, poor success rate, and lack of clarity). To achieve this, a Long Short-Term Memory (LSTM) network is trained in this paper to predict human stress status. For the experimental validation, participants are invited to safely collaborate on finishing a shared task in virtual reality (VR). Physiological data, including blood volume pulse (BVP), electrodermal activity (EDA), skin temperature, and motion intensity signals, are streamed through a medical wearable watch for stress analysis by the developed LSTM model. The experimental results demonstrate the effectiveness of the system in identifying stress during collaborative VR tasks. In addition to the implemented VR game, the research findings from this paper can be extended to a broader area, such as human-machine interaction, intelligent medical assistance, and psychological studies.

ABSTRACT. This paper investigates a resource constrained excavator procurement problem (RCEPP), in which the budget is restricted and the performance of the excavators in collaboration with the pre-purchased mine trucks is considered. The RCEPP consists in deciding the purchase quantity of excavators to maximize the profit from the mining system composed of the imported excavators and the existing mining trucks. A mixed integer programming model (MIP) is established to address the RCEPP. In line with the specifications of quantum computers, a quadratic unconstrained binary optimization model is developed via reformulating the MIP. The scheduling scheme is coded into the ground state of Hamiltonian operator, and the problem is solved by using optical coherent Ising machine (CIM). The experimental results indicate that, compared to traditional computation methods, quantum computers are capable of achieving better global optimal solutions. It has great application potential.

ABSTRACT. With the rapid growth of transportation passenger flow in recent years, the research direction of Virtual Coupling (VC) in railways has become hot. VC increases capacity by tracking shorter distances. The traditional train emergency braking (EB) model can no longer meet the high operating efficiency requirements of VC. VC urgently needs a more efficient safety EB model. Therefore, to improve efficiency, a measurement of the EB model's rationality is defined considering safety and efficiency. Based on the defined model measurement, an EB model optimization method and a new EB model are proposed. The new model is closer to the actual braking process of the train, and the model's safety is guaranteed from the distance level. Finally, the safety and efficiency of the new model were verified through simulation experiments. The new model can avoid train collisions during the entire EB process. Experimental results show that compared with the traditional old EB model of VC based on space-time interval, the new model's minimum safe distance (MSD) between two VC train units can be reduced by 39%.

ABSTRACT. Virtual coupling (VC) is an emerging technology that enhances railway operations by dynamically and accurately matching line capacity with passenger demand. However,challenges arise in coordinating the virtually coupled train set (VCTS) during inter-station operations and maintaining accurate control of the VCTS. To address this issue, we propose a multi-phase trajectory planning control approach to generate the reference trajectories for VCTS to enhance the synchronicity of inter-station operations. However, during speed-varying inter-station operations, it remains challenging to control VCTS accurately under disturbances while adhering to safety constraints. the control layer employs a model predictive control (MPC) approach to ensure precise reference tracking, based on the optimized reference trajectories for all train units. The effectiveness of this approach is validated through experimental results.

ABSTRACT. This work is an extension of the previous work [8]; instead of offline tuning of PID controller using Reinforcement Learning (RL), the tuning will be online and ensure stability guarantees for RL behavior in real-world scenarios, especially during training. LQG cost function is used as a reward function to optimize the PID parameters. The cost combines tracking error reduction and actuator effort limitation. Two learning agent performances are explored: the simple Deep Deterministic Policy Gradient (DDPG), and more advanced agent Twin Delayed Deep Deterministic (TD3). The learning procedure checks the closed-loop stability before the parameters update. DDPG and TD3 are compared in terms of performance to MATLAB Simulink control parameter tuner.

ABSTRACT. This article focuses on enhancing the resilience of autonomous vehicles (AVs) against cyberattacks. It integrates Model Predictive Control (MPC) and Lyapunov-based Economic Model Predictive Control (LEMPC) to manage vehicle dynamics and detect potential cyber threats. The research comprehensively covers vehicle modeling, the implementation of MPC for vehicle control, and a detailed analysis of various cyberattacks targeting autonomous vehicles. Detection mechanisms using MPC are developed to identify and mitigate these attacks, ensuring the stability and reliability of AV operations. These mechanisms are benchmarked against traditional detection methods like ChiSquare, Cumulative Sum (CUSUM), and Multivariate Exponentially Weighted Moving Average (MEWMA). Additionally, the Kalman Filter (KF) is employed for state estimation from noisy measurements, enhancing the accuracy of the detection system. Extensive simulation studies validate the effectiveness of the proposed methods in a lane change scenario, demonstrating significant improvements in the resilience and operational safety of autonomous vehicles.

ABSTRACT. Script event prediction selects the most likely event from a set of candidates based on the sequence of events. Previous studies primarily relied on event pairs, chains, and evolution diagrams, using probabilistic theories to determine event transfer direction. However, as data volume increases, the transfer direction becomes unpredictable, leading to reduced model accuracy. Additionally, these studies often overlook the inclusion of background knowledge and conditional constraints, resulting in a broader solution space for event prediction. To address these issues, we propose a novel framework called Script Event Prediction Based on Knowledge Distillation of Large Language Models (SEKD). We establish a formal definition for script event data and utilize LLMs's reasoning capabilities to guide internal thinking and logical explanations through a thinking chain. These explanations serve as additional supervision for a smaller model trained through knowledge distillation, enabling it to inherit and learn reasoning abilities from LLMs. Experimental results on two datasets demonstrate the strong performance of our proposed model, showcasing robust understanding and learning of reasoning cognition and logic from LLMs, thus possessing fundamental reasoning capabilities.

ABSTRACT. Collaborative robots play a significant role in the Industry 5.0 revolution. In this work, we present the development of a new dual-arm all-terrain Multimodal Collaborative RObot System (MCROS), which has dynamic decision-making capabilities in addition to a physical capacity to be controlled and directed for human-centered tasks. MCROS consists of two UR10e collaborative robots, an all-terrain mobile base, a computing and control unit, and a set of multimodal sensory systems for human-robot interaction. These sensory systems include 3D LIDARs, Force-Torque sensors, 3D/security cameras, an inertial measurement unit, and a global position system. Such a design merits a broader range of applications for MCROS. The Robot Operating System is employed for MCROS programming and control, which enables MCROS to be an open-source agent to seamlessly integrate with other cyber-physical systems in different tasks. A specification discussion of MCROS and other existing systems is illustrated. We carry out several applications and analyze the results of MCROS in different real-world work contexts. The scalability of MCROS is also discussed.