Deep learning-based object detection models have shown strong potential for automated plant disease detection from leaf images. Among these models, YOLO architectures are widely used due to their ability to achieve high detection accuracy while maintaining real-time performance. However, despite these advantages, the adoption of such systems in agricultural practice remains limited. One of the main reasons is that their predictions are often difficult to interpret, which can reduce the confidence of farmers and agricultural experts, who need to understand the basis of model predictions before relying on them for decision-making. To address this issue, this paper proposes an explainable deep learning framework that combines a YOLOv12-based detection model with explainable artificial intelligence techniques. The proposed approach is evaluated on a self-generated dataset of plant leaf images. The experimental results show that the proposed YOLO model achieves satisfactory detection performance, confirming its suitability for plant disease detection tasks. In addition to performance evaluation and to improve transparency, Gradient-weighted Class Activation Mapping is employed to generate visual explanations of the model's predictions. The resulting heatmaps reveal that the network consistently concentrates on relevant diseased regions of the leaves, indicating that the detection decisions are guided by meaningful visual features. By combining accurate detection performance with visual explanations, the proposed framework aims to provide a more transparent and trustworthy solution for deep learning-based plant disease detection.
Network Intrusion Detection Systems (NIDS) deal with class imbalance in network traffic data, where minority attack classes are underestimated. FCM-Cosine, a modified Fuzzy C-Means clustering algorithm, replaces Euclidean distance in the objective function with Cosine distance to better capture directional similarity in high-dimensional feature spaces. The cluster-then-classify framework decomposes the global intrusion detection problem into localized classification sub-problems to detect minority attack classes. Five classifiers have been examined on the CICIoT2023 dataset at two scales (16,100 and 465,000 samples). FCM-Cosine had an average F1-Macro of 69.36%, while Decision Tree had 86.79%, resulting in a 37.97% improvement over direct training. The framework is ten times faster than SMOTE (19.18s vs. 189.73s average training time) and scales nearly linearly with dataset size. Results demonstrate that FCM-Cosine offers competitive classification performance with computational efficiency for large-scale NIDS deployments.
In many engineering, computer science and economic applications, the model, or the system design has to satisfy the condition that it precisely (or approximately) goes through a set of specified points. This means that the system has to perform as an (approximate) interpolator (interpolating function, also named interpolants). The first question asked is if there is a simple way to design good interpolators with FLSs. We answer this first question in a positive way by providing a systematic and simple method for interpolation with FLSs and proving that FLSs are universal exact interpolators. Fuzzy logic systems are known to be universal approximators; that guarantees that they can also perform as universal approximate interpolators. Finding a good approximator with fuzzy logic systems (FLSs) may be difficult and computationally demanding. A procedure is provided for automatically building guaranteed approximators for real valued functions defined on the line. Also, a method for building interpolators for functions with jump discontinuities is shown. The next question asked is: Is it possible to define interpolators with logic systems under various other logics determined by t-norms and related s-co-norms? A method for interpolation and approximation in Rn is proposed, based on t-Logic Systems (t-LS, TLS). t-Logic Systems are systems from an n-dimensional space to the real line, based on a logic defined by t-norms and co-norms. They include a first stage of assigning to the real-valued inputs a t-distributions, a second stage where an inference is performed in the framework of the chosen logic, and an estimation stage, where the result of the inference is converted into a real number. Typical examples include fuzzy systems, Z-number based system, and probabilistic systems. Similar to neural networks (NN), fuzzy logic systems (FLSs) with center of gravity (c.o.g.) estimator (named defuzzifier for FLSs) are universal approximators. Even more, FLSs with a c.o.g. defuzzifier are interpolators for wellbehaved functions. The general t-logic systems have similar properties. These general properties have wide applicability in economics, engineering, and decision making. In many cases, the type of t-logic, in particular the type of fuzzy logic used in the FLS does not play an important role in the universal approximation or interpolation properties, thus leaving much space for the tLS and FLSs optimization. The paper provides results about interpolation, and in subsidiary, for approximation with tLSs, in particular with FLSs, and discusses various applications. The results may find applications in various fields.
Cryptocurrency systems integrate computing, communication, and control mechanisms; therefore, they create a need for structured education to support understanding and innovation. Public readiness for crypto currency education has been investigated across behavioral, technological, and regulatory dimensions. Previous studies have approached these dimensions in isolation, without a unified framework to explain their combined influence on readiness and behavioral intention. To address this gap, this paper examines the conditions that shape readiness for institutionalized cryptocurrency education and explores their effects on behavioral intention. The Theory of Planned Behavior was integrated with a Computing-Communications-Control structure to establish the proposed research model. Readiness is viewed as the product of computing-related capabilities, communication processes, and regulatory conditions. A survey was conducted with 476 adults in Romania. Based on Partial Least Squares Structural Equation Modeling, the findings reveal that perceived behavioral control is positively influenced by education, financial literacy, and system familiarity. Behavioral intention is positively predicted by perceived behavioral control and regulatory clarity. Within the communication dimension, social media influence positively predicts subjective norms. However, subjective norms show a negative relationship with intention, pointing to a counterintuitive dynamic in how communication conditions operate in this context. This work introduces a system-level framework that connects capability formation, communication dynamics, and control conditions within a unified structure. It provides a foundation for designing cryptocurrency education programs that integrate technical competence, information evaluation, and regulatory awareness.
To address the challenge of potential container source identification in sea-rail intermo dal transport scenarios, which arises from data silos and privacy barriers among multiple stakeholders, this paper proposes FSL-Qwen, a Federated Split Learning framework integrated with Large Language Models. Innovative to this framework is the vertical partitioning of the Qwen model at the embedding layer: clients (e.g., ports, railways, customs) deploy only lightweight embedding layers for local feature extraction, while the server retains the Transformer backbone for centralized reasoning. This architecture decouples local computation from inference, theoretically reducing client-side complexity to O (1) and drastically minimizing communication overhead compared to standard Federated Learning. To resolve cross-domain semantic heterogeneity, a ChatML-based semantic alignment mechanism is introduced, enabling collaborative inference without sharing raw records. Privacy analysis demonstrates that the framework achieves inherent structural isolation, converting data reconstruction attacks into blind inverse problems. Experiments on a dataset of 48,800 SRIT container data confirm that FSL-Qwen achieves a predictive accuracy of 94.0% and an F1-score of 94.1%, effectively matching the centralized upper bound while limiting client-sidespacememory usage to merely 0.26 GB. These results validate FSL-Qwen as a robust, efficient, and privacy-preserving paradigm for intelligent logistics decision-making.
Chinese spelling correction (CSC) in legal texts presents unique challenges due to specialized terminology, complex error patterns, and the high accuracy requirements of legal documentation. To address these issues, we propose Multi-granularity Curriculum Learning (mgCL), a novel two-level adaptive training framework integrating batch-level and instance-level curricula. At the batch-level, mgCL dynamically prioritizes training samples based on cross-entropy-derived difficulty, ensuring the model is exposed to increasingly complex examples as its competence evolves. At the instance-level, it leverages Monte Carlo Dropout to quantify prediction uncertainty and adopts a weighted cross-entropy loss function with both sentence level weights and token level weights (weighted pinyin and glyph similarity) to guide the model in adapting its learning to individual samples, directing greater learning focus to ambiguous characters-especially domain-specific legal terms. To support research in legal-domain CSC, we also introduce CNLAW, a novel benchmark dataset featuring diverse error patterns and extensive legal terminology. Experimental results confirm mgCL's effectiveness: on CNLAW, it achieves a 98.02% F1 score (outperforming the strong Rephrasing Language Model (ReLM) baseline of 96.75%) and dramatically reduces the False Positive Rate (FPR) from 1.60% to 0.16%. Additionally, robust performance on the general-domain SIGHAN15 benchmark validates its cross-domain generalization. These findings demonstrate mgCL's value as an effective, scalable framework for specialized-domain CSC, with potential extensions to medical and financial text processing.
Traditional machine learning models typically address closed-set recognition problems. However, in real-world scenarios, open-set recognition challenges frequently arise, where the accuracy of existing methods tends to be low. To address this issue, this paper proposes a versatile classification framework that integrates federated learning and meta-learning to tackle open-set recognition in federated environments. In the proposed methodology, federated meta-learning is first employed to train a global model with strong generalization capabilities across different clients, while preserving data privacy. Then, each local client extracts features and applies a relational network-based classifier to perform classification, obtaining the final prediction results. The local models are updated accordingly, and a coefficient-based aggregation algorithm is designed to update the global model, considering both the aggregation coefficients and classification accuracy. Finally, the server distributes the updated global model back to the clients, who update their local models and proceed to the next training round. To validate the effectiveness of the proposed approach, open-set recognition experiments are conducted on the MNIST, CIFAR-100, and Omniglot datasets. Experimental results demonstrate that the proposed method not only ensures local data privacy but also achieves higher accuracy compared to baseline algorithms such as FedAvg.
The widespread adoption of light electric vehicles (LEVs) is constrained by excessively long battery charging times, which reduce user convenience and practicality. This paper presents the design, implementation, and experimental validation of a PID-controlled constant current-constant voltage (CC-CV) fast-charging system tailored for a 12V, 6Ah lithium iron phosphate (LiFePO4) battery, a typical energy storage unit in LEVs. The system integrates a buck converter for power regulation, an Arduino Uno microcontroller executing a parallel PID algorithm, and real-time feedback from voltage and current sensors. Charging tests were conducted from 90% depth of discharge (12V) using three protocols: conventional constant voltage (CV), CC-CV at 0.5C (3A), and CC-CV at 1C (6A). The PID-controlled CC-CV method drastically reduced charging time, achieving full charge in 2 hours and 1 minute at 0.5C (78.8% reduction) and 1h7min at 1C (88.2% reduction) compared to 9h 33min for CV charging. The dynamic behavior during the constant-voltage phase was further analyzed via closed-loop transfer function modeling. Using a first-order plus dead time (FOPTD) plant identified from step response data (tau = 39 s, L = 9 s) and PID parameters (Kp = 0.3, Ki = 0.02, Kd = 0.05), the step response exhibits a rise time of 43.7 s, settling time of 151.3 s, and overshoot of 14.7%. Closed-loop poles at 0.2078 and 0.0142 +/- j0.0342 confirm stability and predict a lightly damped oscillatory mode consistent with experimental observations. The consistency between experimental and simulated responses validates the modeling approach and controller tuning. This research provides a validated framework for low-cost, embedded fast-charging solutions, contributing to the acceleration of LEV adoption and sustainable urban mobility.
Automatic modulation recognition has a wide range of applications in the field of signal processing. Real-world signal environments are complex and variable, and multiple datasets with domain differences are formed due to different sampling frequencies. However, existing methods usually rely on a single data domain for training, which makes it difficult to adapt to domains with inconsistent distributions. To address this, this paper proposes an attention-enhanced domain adversarial training (AM-DAT) method. Initially, radar signals are transformed into two-dimensional timefrequency images via the Smoothed Pseudo Wigner-Ville Distribution (SPWVD). Subsequently, discriminative and robust features are effectively extracted by the attention-enhanced neural network; while domain adversarial learning strategy is combined to achieve the consistency of feature distributions in the source and target domains, thus improving the generalization ability of the model to data domains with different distributions. Experimental results show that AM-DAT achieves superior classification accuracy across signals-to-noise ratio (SNRs), and its performance is much higher than the methods lacking adversarial training mechanism under low SNR (-2 dB). Our approach demonstrates strong potential for practical radar signal classification applications.
Intrusion Detection Systems (IDS) are critical to ensuring cybersecurity in complex, dynamic, and data-intensive network environments. Traditional IDS, whether signature-based or classical machine learning (ML)-based, struggle to adapt to evolving attack patterns and to provide explainable decisions in real time. This paper presents a comprehensive evolutionary framework leading to a new unified model: the Neuro-Fuzzy Reinforcement Transformer Intrusion Detection System (NFRT-IDS). Three intermediate hybrid algorithms, a Transformer-CNN (Convolutional Neural Network) IDS, a Fuzzy-Ensemble IDS, and a Deep Q-Learning based Artificial Neural Network (DQL-ANN) IDS, are first proposed, rigorously optimized through cross-validation, and extensively evaluated on benchmark datasets (CICIDS2017, UNSW-NB15, and BoT-IoT). These models respectively address deep feature extraction, interpretability, and adaptive decision optimization challenges in IDS, while providing complementary architectural and learning advantages. Their integration inspired the unified NFRT-IDS framework, which combines global attention-based feature learning, fuzzy inference for uncertainty modeling and rule-based explainability, and reinforcement learning (DQL agent) for dynamic parameter adaptation and performance-driven optimization. Experimental results demonstrate that NFRT-IDS achieves superior performance, reaching 99.98% accuracy and F1-score on CICIDS2017, with a 0.31% False Alarm Rate (FAR) and 0.999 AUC, outperforming state-of-the-art hybrid models. Beyond single-dataset evaluation, NFRT-IDS exhibits strong cross-dataset generalization, maintaining consistent accuracy and F1-scores when trained on CICIDS2017 and evaluated on heterogeneous datasets such as UNSW-NB15 and BoT-IoT. Furthermore, the framework ensures scalability, robustness, and interpretability, enabling efficient real-time intrusion detection in modern IoT and cloud environments.
Considering the challenge in hyperspectral imaging of developing new computational methods that strike a balance between accurate material classification and computational complexity, this work proposes the design and tunability of a model based on a sequential artificial neural network (ANN) to classify vegetation in hyperspectral images with 380 bands. To carry out this research, an adaptation of the CRISP-DM methodology was used, structured into four phases: P1. Business and data understanding, P2. Data preparation, P3. Modeling and evaluation, and P4. Modl application. As a result, a sequential ANN model was developed, featuring 380 input layers and a single output layer, along with a set of dense layers containing 12, 8 and 4 artificial neurons. After 20 epochs, the model showed high performance and consistent behavior in the training and test sets under the experimental setup considered. The model was applied to a hyperspectral image of the Manga neighborhood in Cartagena, classifying 41.921% of the image pixels as vegetation. This percentage of points exceeds by 12.941% the percentage obtained by the spectral differential similarity method, in which less continuous point detections were observed. This method is a viable alternative for use in environmental monitoring systems, especially when applied in parallel
Acute hypotension is a common and severe clinical condition, closely associated with adverse outcomes such as acute kidney injury and myocardial infarction. Existing research primarily relies on vital sign data to develop predictive models; however, relying solely on these data is insufficient to fully capture the progression of the condition. Integrating laboratory biomarkers can provide a more comprehensive assessment of patient status, thereby significantly improving the model's accuracy and clinical utility. However, the vast number of available laboratory biomarkers introduces data redundancy and increases complexity, necessitating an efficient feature selection method to identify the most relevant indicators. To overcome limitations of existing methods, this study introduces a novel integration of multi-objective optimization with quantum particle swarm optimization (QPSO), significantly improving indicator selection and prediction accuracy compared to traditional methods. This method optimizes the number of indicators, patient status identification accuracy, sensitivity, and specificity, while setting different priorities for various clinical scenarios. The method was applied to the emergency database of infection patients from the Chinese People's Liberation Army General Hospital to identify key indicators that align with current clinical needs. The selected indicators demonstrated strong alignment with clinical experience and established guidelines, highlighting their clinical relevance. In the full feature scenario, the model exhibited significant improvements in accuracy, F1 score, and specificity, with accuracy ranging from 89.75% to 94.07% and a peak specificity of 98.51%, outperforming other feature selection methods. Even in high-restriction scenarios, where fewer indicators were available, the model maintained strong performance, with accuracy ranging from 88.66% to 93.38%,demonstrating the method's resilience and reliability. The proposed multi-objective QPSO method performed excellently across different scenarios, providing an effective tool for indicator selection in acute hypotension prediction and complex medical data analysis. This method enhances the model's accuracy and clinical applicability, supporting more precise predictions and practical implementation in diverse clinical environments.
To address the limitations of traditional dominance relation-based rough sets in handling information systems with partial order relations, this paper proposes a multi-attribute decision-making method based on entanglement dominance relation. First, the definition of the existing entanglement domain is revised, and the concept of entanglement dominance relation is introduced. The properties of the entanglement domain and its computation method-the improved Bron-Kerbosch algorithm are discussed. On this basis, a secondary ranking method is proposed, in which the subsets of the entanglement domain are first externally ordered, and then the objects within each subset are internally ordered according to specified rules. Finally, an empirical analysis of logistics supply chain performance evaluation is conducted to verify the effectiveness and dominance of the proposed method. The results demonstrate that this method can more accurately reflect the superiority and inferiority relation among objects, providing a new perspective and solution for multi-attribute decision-making problems.
High-accuracy indoor positioning is essential for reliable wireless services and drone monitoring, yet the performance remains highly sensitive to access point (AP) failures. This paper proposes a fingerprinting-based localization framework utilizing Channel Impulse Response (CIR) features and a deep convolutional neural network (DCNN) to ensure robust positioning under both nominal and degraded operational conditions. Two scenarios are examined. In the first scenario, where all four APs operate normally, the influence of network depth and hyperparameter settings on localization accuracy is systematically evaluated. The network block count is progressively reduced and optimal hyperparameters are selected via the Hyperband algorithm, followed by parameter refinement through simulated annealing in a two-stage optimization strategy. Second, in single-AP failure conditions, a mitigation approach is employed wherein a single-input DCNN is retrained for each failure case, and its hyperparameters are independently optimized using Hyperband to compensate for reduced signal diversity. Experimental validation demonstrates that the proposed system achieves an average distance error (ADE) of 0.552 m under nominal conditions, outperforming existing methods. Under single-AP failure scenarios, the framework maintains strong positioning performance, with ADE ranging from 0.776 m to 1.019 m. These results confirm the effectiveness and resilience of the proposed DCNN architecture, highlighting its suitability for reliable indoor localization in realistic and fault-prone environments.
With the increasing complexity and frequency of computer network faults, efficient fault diagnosis is crucial to the reliability and security of the network. To this end, this paper proposes an advanced fault diagnosis method based on the improved Teaching Learning-based Optimization (TLBO) algorithm and the Mini Batch Gradient Descent (MBGD) algorithm under the framework of Convolutional Neural Network (CNN). Different from the traditional CNN-based methods, this method innovatively integrates the TLBO algorithm with the Differential Evolution (DE) strategy, optimizes hyper-parameters and training convergence, and significantly improves the detection accuracy and speed. Meanwhile, MBGD can effectively refine the model parameters and prevent convergence to local minima. The experimental results using the public dataset prove the effectiveness of this method, achieving a high classification accuracy rate of up to 88.3% and significantly reducing the false detection rate to below 0.20%. Compared with the traditional CNN model and the latest methods, this method has a faster convergence speed, model stability and fault diagnosis performance. This research provides a robust solution for real-time fault detection and diagnosis, significantly enhancing the reliability and security of complex network systems.
This paper presents a machine learning (ML)-based framework to predict performance degradation in Wi-Fi indoor localization systems (ILSs) under varying moving human crowd densities. While indoor localization can be performed in both mobile and fixed wireless settings, the majority of prior research emphasizes mobile devices in motion. In contrast, this study adopts a fixed-wireless configuration, where a smartphone node was held stationary while moving human density varied around it. This design particularly isolates the effect of human crowd-induced interference on received signal strength indicator (RSSI) fluctuations, enabling a controlled evaluation of ML-based error compensation, which is a perspective rarely explored in the literature. Accelerometer-derived motion features were integrated with RSSI measurements, and baseline localization errors were calculated using the conventional Weighted Least Squares (WLS) indoor localization algorithm. Three main ML regression models namely Random Forest, CatBoost, and XGBoost were trained and evaluated. Among them, CatBoost demonstrated the best performance, achieving a root mean squared error (RMSE) of 0.331 m compared to the WLS baseline error of 1.405 m, corresponding to a 76.47% improvement in localization accuracy. The evaluation was intentionally limited to a single indoor layout with a stationary device to isolate crowd-induced RSSI distortions, and multi-position validation and mobile-user scenarios are reserved for future work. The findings confirm that smartphone sensor-fused ML models can anticipate human crowd-induced localization errors and enhance the robustness of multilateration-based ILSs.
In this research, facial features are identified from face templates using deep neural network models. Facial templates are widely used in modern biometric systems, enabling efficient and reliable individual recognition. Such templates are compact, easy to process, and are often used in real-time systems. During the research, a dataset with 115,023 items (face templates), used by the Neurotechnology company to identify persons, was prepared as a new dataset. The dataset has been prepared to study three facial features from face templates: gender, race, and age. The facial templates are lighter-weight than real photos and could be reused to estimate the facial features. However, it is not yet known how effective facial feature identification will be, as some data are lost when converting a photo into a face template. Several methods have been proposed for classifying original image attributes solely from face templates. Both deep neural networks and classical machine learning algorithms were used in the experiments. The experiments revealed that gender is the most accurately predictable attribute, with the best model achieving 93 % accuracy. Classification of people's race and age from face templates is more challenging, likely because face template generation models are designed to eliminate information unrelated to identity. A comparison of machine learning methods showed that deep neural networks are better suited to this task than classical classification algorithms.
The identification of key nodes within multi-relational networks presents significant challenges due to the heterogeneity of node attributes and the varying significance of different attributes across distinct relationships. Conventional methods often fail to effectively capture these complexities, leading to suboptimal mining outcomes. To address this issue, a passive clustering-based approach is introduced to enhance the identification of important nodes in multi-relational networks. By constructing an adjacency matrix framework, the network structure is systematically represented, encapsulating the connectivity relationships among nodes. The comprehensive centrality of entity nodes is then evaluated to preliminarily select candidates with substantial network influence. Subsequently, a passive clustering algorithm is applied to categorize nodes into clusters based on attribute similarities, enabling a refined analysis within each cluster. The principle of node centrality metrics is further adapted to assess node importance within and across clusters, thereby mitigating the impact of attribute heterogeneity. Nodes exhibiting weak intra-cluster associations are eliminated, ensuring the robustness of the clustering process. The proposed method demonstrates superior efficiency and scalability, requiring a memory footprint below 160 KB. Furthermore, the computational efficiency of node degree centrality, median centrality, and proximity centrality is improved, with relative computational time ratios of 14.2%, 8.9%,and 8.6%,respec-tively. These results indicate that the proposed approach effectively captures complex dynamic interactions within undirected and unprivileged multi-relational networks, offering a scalable and computationally efficient solution for important node mining.
Transport system efficiency is a fundamental and strategic issue for all transport companies. The ability to adapt transport networks reliably is crucial as demand fluctuates, specifications shift and traffic specificities cannot be neglected. Uncertainty, ubiquitous in rail transport networks, complicate this task even further. These uncertainties can manifest themselves in a variety of ways: unexpected fluctuations in journey times, rolling stock failures or the emergence of additional traffic tasks that could not have been anticipated in the initial scheduling process. Each type of uncertainty creates a potential risk related to system imbalance, which requires rapid and complex adjustments to guarantee rail traffic availability and stability. These maintenance scheduling issues in rail transport systems demand planning approaches that extend beyond traditional techniques. It is crucial to evolve maintenance scheduling tools able to manage scheduling under stable conditions, as well as to effectively respond to unexpected disruptions and quickly shifting traffic conditions. This paper addresses these challenges problem and proposes a reliable and robust maintenance policy taking account of tasks imprecision and human expertise. The maintenance model is designed to assist decision making systems to increase traffic safety significantly, while saving time and money. To resolve this problem, a fuzzy inference system is used to appropriately deal with uncertainties using Colored Petri nets and fuzzy logic. The findings indicate that the adaptive fuzzy model developed has an excellent ability to precisely learn and predict traffic constraints and lead to significant changes in decision making and the incorporation of feedback into the management and support system.
Modern multi-criteria decision-making (MCDM) approaches must handle ambiguous and imprecise initial information about decision-making phenomena in an efficient way. Application of the Dempster-Shafer theory (DST) provides a strong mathematical basis for processing uncertain information. Due to the close relationship between fuzzy sets and DST, various fuzzy set environments are applied to model the DST system to solve multi-criteria decision-making problems. To work on this challenge, a picture fuzzy set (PFS) can be applied to model the vague information. This paper presents a novel MCDM approach, namely the hybrid SWARA DS and CoCoSo DS method, developed within Dempster-Shafer's framework. The benefits of the new proposed approach in handling MCDM problems involving uncertain criteria and expert weightings are demonstrated through its application in selecting the optimal roof shape for a renovated single-family house. Sensitivity and comparative analyses validate the method's reliability and effectiveness.
