Autonomous vehicles (AVs) rely heavily on the controller area network (CAN) for internal communication among sensors and actuators. However, this protocol lacks built-in security mechanisms, making AVs vulnerable to cyber-attacks such as denial of service (DoS), fuzzy, and impersonation attacks. This work proposes a hybrid intrusion detection system (IDS) that integrates Rainbow deep Q-network (DQN), a reinforcement learning technique, with a random forest classifier to improve the detection accuracy and robustness against such attacks in real-time AV environments. The methodology involves preprocessing CAN traffic datasets into state-action representations suitable for reinforcement learning. Rainbow DQN learns optimal policies by leveraging prioritized experience replay, duelling architecture, and multistep bootstrapping to predict attack probabilities. The output Q-values are then passed to a random forest classifier, which refines the decision boundaries using ensemble voting, thus enhancing classification precision. The implementation was carried out using Python (PyTorch and scikit-learn) on a labelled CAN intrusion dataset comprising attack-free and malicious traffic. Experimental results demonstrate that the proposed hybrid model outperforms standalone rainbow DQN in terms of classification performance, achieving an accuracy of 87.93
Electricity distribution networks are increasingly affected by extreme meteorological conditions, which pose significant risks to above-ground energy infrastructure and the reliability of power supply. This study develops a machine learning based approach to predict weather-induced power line damages using more than 16 years of daily observations (2009–2025) from Latvia. The analysis evaluates feature engineering and three spatial data partitioning strategies, applying both binary and multiclass classification to assess how spatial granularity influences predictive performance. Results show a clear inverse relationship between spatial resolution and model sensitivity. The best performance was achieved at the macro spatial scale (two cells), where a neural network model reached an F1-score of 0.803, demonstrating the potential of predictive analytics to support proactive grid maintenance and emergency response planning.
Fuzzy programming has shown great vitality in the fields of management, optimal control, economics, power dispatch, and other applications. However, the algorithm that solves such problems has not made significant progress due to the injection of fuzzy parameters. This study aimed to propose a novel intelligent algorithm combining fuzzy simulation, multilayer backpropagation (BP) artificial neural networks, and the stochastic particle swarm optimization (SPSO) algorithm to solve the fuzzy dependent-chance programming (FDCP) model problem in the fuzzy programming field more efficiently. In this computational method, fuzzy simulation was used to acquire training samples for multilayer BP artificial neural networks, and the artificial neural networks were used to approximate the fuzzy chance function. Also, SPSO was used to complete iterative optimization. Meanwhile, in SPSO, the neural networks were used to compute the fuzzy chance function. Our algorithm addressed several constraints such as a long run time, intricate computation, and a tendency to get trapped in a locally optimal value in the mixed algorithm with intelligence based on the classic genetic algorithm. The results for the issues related to the fuzzy dependent-chance programming model were validated by some cases. Also, the validity and reliability of the algorithm were verified, highlighting its application value.
Human–computer interaction (HCI) has advanced into another age arising from developments in artificial intelligence that allow machines to comprehend and interpret human emotions dynamically. Modern systems confront challenges such as noisy input data and emotion classification accuracy, impeding adaptability in interactive environments. This work proposes a design of an intelligent HCI system for interpreting human emotions to improve better decision-making in intervention applications for human–computer collaboration in virtual reality (VR) applications. Initially, data acquisition of facial emotion expressions was done, which is followed by preprocessing techniques, including Savitzky–Golay filtering for noise removal and quadratic contrast stretching for contrast enhancement to improve clarity of the image. Subsequently, feature extraction via GhostNet along with the emotion classification through ShuffleNetV2, the classified emotions are integrated into the HCI system for dynamic adjustments based on user emotion, thereby fostering immersion and engagement in VR. Experimental evaluation shows the efficiently of proposed model performance with an accuracy of 99.80
Fraud incidents involving credit cards pose a persistent and escalating threat to financial institutions and consumers globally, necessitating the development of sophisticated and efficient detection mechanisms. This study examines the effectiveness of contemporary machine learning methods in identifying fraudulent credit card transactions, addressing the prevalent issue of class imbalance in fraud datasets (284,807 transactions; only 0.17
