In the hospitality and tourism (HT) industry, Employee Service Innovative Behavior (ESIB) is crucial for organizational competitiveness, service quality, and long-term sustainability. However, the factors influencing ESIB are diverse, interrelated, and often characterized by uncertainty in expert judgements. The present study identified and prioritized the main drivers that influence employees’ innovative behavior in HT businesses by applying the Fuzzy Analytic Hierarchy Process (Fuzzy AHP). The study investigated five main drivers (Work Attitudes, Individual Characteristics, Organizational Characteristics, Leadership Style, Team Characteristics) and twenty related sub-drivers, after a thorough examination of the literature and expert evaluation. The findings of the study show that ‘Work Attitude’ is the most significant predictor of innovative behavior (IB) in HT services, and among the sub-factors, ‘Job Security’ plays an important role in encouraging IB among the employees. The study offers valuable insights for the HT industries by promoting an innovative culture and improving their ability to compete in the rapidly changing service market.
Water absorbs and scatters light in a wavelength-dependent way, and the result for underwater imagery is a predictable set of problems: color cast, washed-out contrast, and degraded structure. Physics++ addresses these through a deterministic pipeline built from closed-form stages: depth-aware attenuation modeling, edge-preserving guided-filter transmission refinement, and a residual-constrained fusion step that blends a direct radiance estimate against a dark-channel-guided prior. None of these stages involve learned parameters, so identical input always produces identical output and no training data is required. We evaluated the method on four public benchmarks (LSUI, UIEB, EUVP, and SQUID) against seven other physics-based configurations. Physics++ posted the best SSIM on LSUI (0.7944), the best PSNR (17.55 dB) and UCIQE (52.89) on UIEB, and the best average rank across all four datasets (3.00 of 8); the gains over the UDCP and WCID baselines were statistically significant ( p < 0.001), with large effect sizes on UCIQE and UIQM. We do not obscure the method's ceiling: its absolute PSNR sits 3–8 dB below current learning-based state of the art, and we position it deliberately as a training-free, interpretable alternative rather than a fidelity-maximizing competitor. To test whether that ceiling can be narrowed without abandoning the physics, we also trained a compact hybrid variant, Physics++-Hybrid, conditioning a learned network on the same transmission signal; on a held-out EUVP split it reaches 23.35 dB mean PSNR, an encouraging result, though measured under a different protocol and not directly comparable to the core result above.
With suitable kernels and parameters, support vector machines (SVMs) usually outperform other classification methods in terms of accuracy and generalization ability. However, finding appropriate kernels and hyperparameters is time-consuming, especially for large-scale classification tasks. In this paper, we demonstrate that clustering—when designed with classification in mind—can effectively replace kernelization. We propose a novel class-aware clustering method, SEKmeans, which jointly optimizes intra-cluster compactness, cluster purity, and intra-cluster linear separability within a unified objective. This design explicitly facilitates downstream linear classification and enables linear classifiers to capture nonlinear patterns without using any kernel function.We conduct extensive comparisons with linear SVM, linear Twin-SVM (TWSVM), kernel SVM, kernel Twin-SVM, the clustering-based CSVM, and two K-means-based counterparts of SVM and TWSVM, namely K_SVM and K_TWSVM. Experimental results show that our SEKmeans-based variants of SVM and TWSVM(SEK_SVM and SEK_TWSVM) achieve competitive accuracy and F1-scores comparable to those of kernel-based methods, while requiring substantially less computational time.
Abstract This study analyzes and ranks life satisfaction across multiple domains for individuals living in Türkiye using fuzzy multi-criteria decision-making approaches. The analysis is based on microdata obtained from the Turkish Statistical Institute for the period 2020–2024 and covers fifteen life domains, including health, marriage, education, housing, neighborhood, employment, income, salary, social life, time allocated to personal care, time spent in traffic, and interpersonal relationships. Although the original survey data are collected using a Likert-type scale, this study accounts for the inherent uncertainty and subjectivity of individual responses by transforming ordinal evaluations into fuzzy numbers. Specifically, Picture Fuzzy Sets are employed to represent uncertainty in survey responses, while Picture Fuzzy Entropy, TOPSIS, and WASPAS methods are integrated to evaluate and rank life satisfaction domains. Fuzzy multi-criteria decision-making methods are employed to reduce uncertainty in the decision-making process and to ensure an effective and reliable decision framework. The findings consistently indicate that economic domains, particularly income and salary, are associated with the lowest levels of life satisfaction. Social life, which is closely linked to economic conditions, ranks second in terms of dissatisfaction. In contrast, interpersonal domains, especially friendships and neighborly relations, exhibit the highest levels of satisfaction. Overall, the results highlight the central role of economic well-being in shaping subjective life satisfaction and illustrate the value of fuzzy approaches in the analysis of ordinal well-being data. As an application contribution, the study demonstrates the usefulness of a picture fuzzy multi-criteria decision-making framework for evaluating life satisfaction dimensions using large-scale national survey microdata.
In this paper, a fundamental structural asymmetry in dengue fever epidemic control approaches is identified and formalized. We show that treatment and quarantine interventions must influence the sick compartment, whereas the majority of epidemic game models treat interventions as affecting the susceptible compartment (e.g., vaccination). A fundamental imbalance is produced by this structural difference: public health interventions (downstream) alter the linear elimination term, whereas behavioral control (upstream) alters the bi linear incidence term. We demonstrate through careful mathematical analysis that behavioral control has a greater global impact through multiplicative effects over time, while public health measures dominate in local immediate sensitivity. For a differential game model, we calculate the open-loop Nash equilibrium and demonstrate an 86.4% decrease in peak infections when compared to no control. Our findings show that where control enters the epidemic system is just as important as how much control is applied, resolving the apparent dichotomy between local sensitivity and global impact. The best dengue management methods are significantly impacted by this structural asymmetry.
The development of effective intrusion detection systems for renewable-integrated smart grids necessitates models that are capable of distinguishing between legitimate operational variations and genuine cyberattacks. In this study, we present a systematic evaluation of the Sherlock dataset, a recent benchmark for process-aware security research. The investigation encompasses three distinct learning paradigms: unsupervised anomaly detection, binary supervised classification, and a proposed multi-class supervised formulation. The results demonstrate that anomaly detection methods fail to separate attacks from benign operational anomalies due to structural characteristics of the dataset, including benign anomalous behavior and substantial drift between training and testing distributions. While binary supervised classification appears effective under a single split, its performance collapses once cross-validation is applied, with accuracy fluctuating significantly between 35% and 96% (σ ≈ 0.19), which indicates that its apparent success relies on favorable sampling rather than meaningful generalization. Conversely, the multi-class formulation, combined with duplicate removal, variance filtering, and top-80 feature selection, provides a stable and context-aware solution. A tuned eXtreme Gradient Boosting classifier achieves a mean accuracy of 99.81% ± 0.04% across 5 independent seeds successfully identifying all fifteen classes, including rare maintenance events and four distinct attack types. Feature importance analysis confirms that the model primarily relies on physically meaningful variables such as voltage, reactive power, and line currents, thereby demonstrating process-aware decision-making. The results establish that, when high-fidelity event labels are available, supervised multi-class learning provides a robust, accurate, and engineering-valid foundation for intrusion detection in the modern, renewable-integrated grids.
Image enhancement in low-light conditions is a challenging problem within the field of computer vision, since underexposed images generally lead to poor visibility, low contrast, noise amplification, and color distortion. Recent deep learning approaches have indeed shown promising performance, yet most of them adopt computation-heavy architectures and do not treat the luminance enhancement and color restoration separately; as a result cause unnatural restorations. To this end, in this paper, we present an efficient low-light image enhancement framework based on the Deep White-Balance (DWB) with Dark Channel Prior guidance in the YCbCr color space. The proposed framework separates luminance and chrominance components, making it easier to manage brightness enhancement and color restoration separately. This method aims to generate visually consistent enhanced images while preserving color fidelity and avoiding common enhancement artifacts. We evaluate the performance of the proposed method on reference benchmark datasets (LOL, LOLv2-Synthetic, and LIME) as well as a no-reference benchmark dataset (DICM). The experimental results demonstrate that, although state-of-the-art deep learning methods achieve higher numerical scores, qualitatively, the framework produces enhanced images with consistent contrast and illumination, while retaining color fidelity, all requiring less computational resources. Moreover, it does not require any further training or fine-tuning since the proposed approach is based on a pretrained Deep White Balance model and only uses inference. Experimental results show that the proposed method provides a convincing quality-efficiency compromise for low-light image enhancement.
Flying ad hoc networks—composed of self-organizing unmanned aerial vehicles (UAVs)—offer numerous applications in various fields, including military operations, industry, and agriculture. Due to the UAV network's unique characteristics, such as high node velocity, sparse UAV distributions, and frequent topology changes, their data routing encounters significant challenges, compromising the quality of service aspects. We introduce a hierarchical type-II fuzzy logic system integrated with a particle swarm optimization algorithm aimed at enhancing the quality of service parameters. The UAVs’ link quality, residual energy, distance, neighboring nodes’ degree, movement direction, and relative velocity are the fuzzy system inputs to compute UAV nodes’ utility, forming the most suitable multiple relay nodes in the optimized link state routing protocol. Meanwhile, our approach employs a hierarchical fuzzy structure to address the curse of dimensionality caused by the exponential growth of fuzzy rules. Furthermore, the particle swarm optimization algorithm adjusts the fuzzy membership functions to tackle ambiguity and uncertainty in the UAV environment. We simulate our approach using NS-3 and compare it with traditional methods under varying node densities and mobility models. The simulation outcomes demonstrate that our approach enhances end-to-end delay, packet delivery ratio, network throughput, and energy consumption compared to rival schemes.
Social media have emerged as some of the most mainstream media communication platforms yet they are also the platforms where different forms of online toxicity exist; hate speech, harassment, cyberbullying, threats, and obscene content. Such content is difficult to detect because of overlapping labels, ambiguity of contexts and class imbalance in a multi-label environment. This paper introduces a hybrid architecture, which is called Identification of Multi-Label Toxicity Classification (IMLTC), to fine-grained multi-label toxicity detection in texts. The method uses contextual embeddings obtained through transformer models ALBERT, RoBERTa and DistilBERT in order to identify subtle semantic relations in comments. Rather than using standard end to end transformer classification heads, which share the same output of sigmoids, the suggested technique divides the representation learning and decision modelling. The obtained embeddings are then used to classify them with an XGBoost algorithm in a one-vs-rest setup, which allows the label-specific nonlinear decision boundary and offers better performance on overlapping and imbalanced groups of toxicity. The dataset on Google Jigsaw Toxic Comment Classification Challenge was experimented. The experimental design involves comparative analysis against various baseline models, including BERT with Multinomial Naïve Bayes (MNB), FastText with Random Forest (RF), and GloVe with Support Vector Machine (SVM), and makes sure that there is a balanced benchmarking of the classical, hybrid, and transformer-based paradigm. As standard multi-label measures, accuracy and F1-score were used to determine performance. The suggested IMLTC framework had a competitive and strong score of 95.34 percent and 95.39 percent accuracy and F1-score respectively. These findings are shown by the fact that the designed combination of contextual embeddings and gradient boosted label specific classifiers provide an effective and scalable solution to intelligent content moderation in the online setting.
Abstract This study examines structural alignment between human and transformer-based multi-label emotion annotation in a low-resource language setting. A corpus of 408 Serbian narratives was manually annotated using soft-label scores mapped to the 28-category GoEmotions taxonomy. Inter-annotator agreement was moderate (micro F1 = 0.58), reflecting the inherent subjectivity of fine-grained emotion labeling. Four modeling paradigms were compared: a zero-shot English model applied to machine-translated texts, a domain-adapted fine-tuned model trained on the translated corpus, multilingual XLM-RoBERTa models trained under hard- and soft-label supervision, and regional South-Slavic models trained on original Serbian texts. The highest overall performance was achieved by the domain-adapted GoEmotions model fine-tuned on the translated corpus (micro F1 = 0.50), substantially outperforming zero-shot cross-lingual transfer. Among multilingual and regional models not originally trained on emotion-specific data, XLM-R-BERTić achieved the strongest micro-level performance (micro F1 = 0.36), indicating the contribution of regional language pre-training to human–machine alignment. Polarity analysis conducted on the pre-trained baseline model indicated low disagreement (≈6%), whereas divergences remained pronounced at the fine-grained category level. The intensity-weighted soft-label annotations reflect the fuzzy boundaries of affective categories, highlighting the importance of domain adaptation and emotion-specific training for improving human–machine alignment.
Dry and wet machining each offer distinct trade-offs in surface quality, tool life, cost and environmental impact, and selecting between them is a multi-objective decision that is difficult to model with classical empirical equations. This study experimentally compares dry and wet turning of aluminium alloy Al6061-T6 and develops two complementary predictive models an Artificial Neural Network (ANN) and a Mamdani-type Fuzzy Inference System (FIS) to predict surface roughness (Ra) and material removal rate (MRR) under both conditions. A Taguchi L9 orthogonal array (three factors spindle speed, feed rate, depth of cut each at three levels) was used on a CNC lathe, giving 9 runs per condition (18 observations total: 9 dry, 9 wet); Ra was measured with a Mitutoyo SJ-210 surface tester and MRR was calculated gravimetrically. Both ANN models (dry and wet) used an identical architecture 3 input neurons, 1 hidden layer of 8 neurons, 2 output neurons trained with the Levenberg–Marquardt backpropagation algorithm on min-max normalised data, with a 70:15:15 train/validation/test split and 5-fold cross-validation. The FIS used triangular/trapezoidal membership functions on the same three inputs and a 27-rule Mamdani rule base with centroid defuzzification. The ANN achieved strong predictive accuracy for dry machining (Ra: MAPE 3.6%, R 2 = 0.972; MRR: MAPE 5.9%, R 2 = 0.961) and somewhat lower but still strong accuracy for wet machining (Ra: MAPE 6.2%, R 2 = 0.951; MRR: MAPE 6.7%, R 2 = 0.944), with 5-fold cross-validation confirming generalisation (mean test R 2 = 0.966). The FIS was more balanced across conditions, in particular for MRR (MAPE ≈ 4.5–4.8% versus 5.9–6.7% for the ANN), though it was less accurate than the ANN for Ra prediction. A paired t-test showed a statistically significant reduction in Ra under wet machining relative to dry ( t = 2.71, df = 8, p = 0.031; mean improvement 12.4%), whereas the difference in MRR between conditions was not significant ( p = 0.74), consistent with MRR being calculated gravimetrically from material removed rather than from surface-level thermal effects. Overall, the ANN is recommended for dry-Ra and MRR-critical applications requiring the highest predictive accuracy, while the FIS is recommended where interpretable, rule-based decision support is preferred, particularly for MRR prediction. These findings give manufacturing engineers a validated, quantitative basis for choosing between dry and wet machining strategies and between ANN- and fuzzy-logic-based predictive tools for Al6061 turning.
Grapevine foliar diseases require timely scouting, yet many image-based diagnostic models remain too demanding for routine edge deployment. This study presents MOHEOA–SDM-LiteNet , a lightweight, explainable classification framework for Black Rot, Esca, Leaf Blight, and Healthy grape leaves, evaluated on a public dataset of 9,027 images. The framework combines a MobileNetV2-derived search space, an auxiliary training exit with logit-level self-distillation, and multi-objective human evolutionary optimization over validation accuracy, inference latency, and parameter count. Methodological safeguards include perceptual-hash grouping, a post-split Hamming-distance audit, and an independent calibration split for temperature scaling. Representative pretrained backbones and focused ablations are evaluated using the same leakage-resistant partitions. The selected self-distilled model contains 0.54 million parameters and achieves 98.78% test accuracy, 0.9882 macro-F1, and an expected calibration error (ECE) of 0.0055, reduced to 0.0048 after temperature scaling. Across three matched seeds, the compact configuration without the auxiliary exit or self-distillation and the full model attain the same mean test accuracy (0.9893). The compact variant uses 0.338 million parameters and is therefore the recommended resource-efficient operating point, whereas the full model provides lower calibrated test ECE (0.00490 versus 0.00792) in every paired seed and is retained as a secondary calibration-oriented variant. Thus, self-distillation is not claimed to improve accuracy. Grad-CAM++ provides qualitative evidence that predictions often focus on symptomatic regions, without establishing formal lesion localization. Overall, the workflow offers a reproducible within-dataset framework for uncertainty-aware grape leaf decision support. Reported latency is a controlled CUDA-based efficiency proxy, not a device-specific deployment guarantee.
The sustainable alternative fuels show strong potential to replace fossil fuels in compression ignition (CI) engines. In this study, the feasibility of plastic pyrolysis oil as a CI engine fuel was evaluated through detailed engine tests, focusing on key performance and emission characteristics. Identifying optimal fuel parameters is complex due to the large number of variables and outputs involved. To support effective fuel selection, a fuzzy multi-criteria decision-making (MCDM) model was proposed. To more accurately represent decision preferences, picture fuzzy sets were incorporated. A knowledge-based quantitative technique was then applied to determine criterion weights using a knowledge measure for fuzzy numbers. Furthermore, an enhanced multi-objective optimization by ratio analysis plus the full multiplicative form (MULTIMOORA) method-supported by a unique score function and the Borda rule-was used to rank fuel candidates objectively. Experiments were conducted on a 3.5 kW CI engine operating at 1500 rpm to compare various blends: PO25/D75, PO50/D50, PO75/D25, and PO100. Results showed that the PO25/D75 blend achieved the highest overall performance index, demonstrating superior efficiency and balanced behaviour among the tested fuels. The sensitivity analysis shows that the top-ranked alternative, PO25D75 remained unchanged, confirming the robustness of the proposed model in identifying the optimal fuel. The comparative analysis with established MCDM methods i.e., SAW, TOPSIS, and COPRAS methods was found to be identical (Spearman's rank correlation coefficient ρ = 1.00), and highly consistent with VIKOR (ρ = 0.90), with only a minor exchange between the ranked alternatives. Importantly, all four methods identified PO25D75 as the best-performing fuel and Diesel as the second-best, confirming that the optimal fuel selection obtained by the proposed framework. This study highlights the promising potential of plastic oil blends, particularly PO25/D75, as sustainable and effective alternatives for CI engine applications.
Web security has become a critical domain as modern applications increasingly rely on dynamic user-generated content, making them highly vulnerable to Cross-Site Scripting (XSS) attacks. Traditional detection systems struggle to cope with evolving payload patterns, limited generalisation across institutions, and strict privacy restrictions that prevent sharing of sensitive request logs. To address these challenges, this work proposes a Zero-Knowledge Federated Sequence Learning (ZK-FSL) framework that enables collaborative XSS detection without exposing raw data or intermediate gradients. The model integrates attention-based deep sequence learning with zero-knowledge proof validation, ensuring both strong predictive capability and verifiable trust among participating clients. Experimental evaluation demonstrates that ZK-FSL achieves superior performance compared to centralised and federated baselines, reaching 96.3% accuracy , 96.7% precision , 95.9% recall , 96.3% F1-score , and an AUC of 0.98 . These results confirm that the proposed framework effectively enhances privacy-preserving threat detection while maintaining high robustness against diverse and sophisticated XSS attack patterns.
Path planning and navigation control of humanoid robots in uneven and unstructured environments are challenging due to terrain uncertainty, stability requirements, and real-time decision-making constraints. This study presents an intelligent path-planning and navigation technique using a modified African vulture-assisted fuzzy logic controller (MAV-FLC). The proposed method combines the global search capability of a modified African vulture (MAV) optimization with the adaptive reasoning of fuzzy logic control (FLC) to generate smooth, collision-free, and energy-efficient paths. The vulture-based strategy improves exploration-exploitation balance, while the fuzzy controller effectively manages local terrain variations and motion uncertainties. The MATLAB and WEBOTs software are used to simulate path generation and route navigation for the humanoid robot within the designed environment. The performance of MAV-FLC is evaluated through extensive simulation and real-time experimentation on a humanoid robot navigating uneven terrain. Later, the outputs are compared with particle swarm optimization (PSO), ant colony optimization (ACO) and African vulture optimization algorithm (AVOA) using path distance and navigation time as performance metrics. Finally, experimental outputs show that MAV-FLC achieves over 10% improvement in both path optimality and time efficiency compared to PSO, ACO and AVOA. Moreover, the deviation between simulated and experimental outputs remains within 5%, confirming the robustness and practical applicability of the proposed approach. The outputs demonstrate the effectiveness of MAV-FLC for reliable humanoid robot navigation in real-world environments.
Emotion detection from textual data is a key challenge in natural language processing (NLP), playing an important role in applications such as sentiment analysis, human-computer interaction, and psychological evaluation. Growing utilization of social networks and online portals leads to the creation of a huge amount of reviews and ratings. Analyzing users’ and customers’ reviews and opinions are so important for governments and businesses. While recent advances have primarily leveraged transformer-based architectures for this task, we propose a novel approach by employing Deep Operator Networks (DeepONets), originally designed for learning operators in scientific computing, to model the mapping between textual representations and emotional states. In this study, we extract high-dimensional semantic embeddings from text using a pre-trained sentence transformer model and feed these embeddings into a DeepONet architecture for emotion classification. The primary contribution of this work lies in architectural innovation. To the best of our knowledge, this is the first application of DeepONet to textual emotion recognition, introducing a fundamentally different perspective on function approximation in language understanding tasks. Results obtained from experiments on benchmark emotion-labeled datasets indicate that our proposed model attains performance and results comparable to related baselines, with notable generalization capabilities across emotion categories. To ensure a fair and comprehensive evaluation, we assessed the performance of the proposed model using widely adopted classification metrics, including accuracy, precision, recall, and F1-score. We used two datasets and the results were around 80 percent on one dataset for all named metrics, and around 88 percent on the other. The findings suggest that DeepONet can serve as a robust alternative framework for capturing and modeling complex relationships in natural language processing tasks and opens new avenues for operator-based learning in text analysis.
Cyberbullying poses a severe threat to individual mental health and social stability. Existing detection methods suffer from poor domain adaptability, a contradiction between accuracy and lightweight performance, and insufficient interpretability. To address these issues, this study proposed the Chain-of-Thought Optimized LoRA-Quantized DeepSeek-8B (CLQP) Model, a scenario-adaptive joint optimization framework integrating Chain-of-Thought (CoT), Low-Rank Adaptation (LoRA), and quantization for efficient cross-scene detection. Specifically, CoT enhances interpretability and implicit bullying identification; LoRA fine-tunes core attention layers to reduce trainable parameters; NF4 quantization balances lightweight deployment and semantic retention. By integrating and unifying annotations for an external database and adopting targeted data augmentation, CLQP was deployed on an RTX 4060 (8GB VRAM). Experimental results showed that CLQP's F1-score exceeded 85% in four scenarios; the false negative rate (FN) decreased by 26%; the false positive rate (FPR) was controlled within 5.2%; the inference speed reached 2 s per sample; and the GPU memory usage was only 5.3GB.
Heart rate (HR) detection, as an important physiological monitoring technique, has been widely adopted in clinical care and health management due to its non-invasive nature and continuity. Currently, most radar-based heart rate detection methods require the subject to remain stationary and process radar echoes via phase analysis, yet they fail to handle radar signals under exercise conditions. This study proposes a non-contact millimeter-wave radar heart rate detection method based on a power-aware adaptive self-attention transformer (ASAT). First, a motion power detection module is constructed to establish a motion–heart rate data correlation model. Then, an innovative integration of a channel-partitioned attention refinement feedforward network and Efficient Channel Attention (ECA) is implemented to effectively suppress motion artifacts and enhance critical features. Finally, heart rate is estimated based on power data. To validate the model's effectiveness, heart rate data under three different types of exercise were collected. The results show that the mean absolute error of heart rate estimation via power data is approximately 9.24 beat per minute (bpm), which is 4.05 bpm lower than that of Long Short-Term Memory (LSTM) models. This study establishes a novel paradigm for continuous health monitoring in complex exercise environments, offers an effective solution for radar technology applications in health monitoring, and lays a foundation for future research.
Fuzzy C-Means (FCM) and its variants are prominent unsupervised clustering techniques, renowned for their robustness in handling data ambiguity. Such methods can also be applied to the design of supervised classifiers by involving the participation of real labels in the optimization of the algorithm. This study develops a novel prototype optimization strategy for enhancing the performance of the fuzzy classification. The developed strategy systematically refines the initial cluster prototypes and concurrently optimizes the associated fuzzy coefficients, enabling the learning of data-specific prototype configurations that directly maximize class separability. By iteratively adjusting both the prototypes and the fuzzy membership coefficients, the algorithm achieves a more discriminative feature space where data points from different classes are better separated. This refinement process is guided by a supervised perturbation mechanism that balances the compactness of clusters with the separation between them, ensuring robustness against noise and outliers. Furthermore, the proposed strategy leverages the inherent uncertainty in fuzzy clustering to explore a wider range of potential prototype configurations, thereby enhancing the performance of the method to discover optimal class boundaries. The model is rigorously evaluated on several benchmark datasets, and the results show an improvement in classification performance.
Radial Basis Function Neural Networks (RBFNNs) are frequently used in UAV disturbance estimation. However, the number of center points and the width values of the activation function significantly affect estimation accuracy, and selecting appropriate values empirically is challenging. Therefore, evolutionary optimization algorithms such as Particle Swarm Optimization (PSO) are often employed to determine the optimal parameters. Nevertheless, these evolutionary algorithms may suffer from premature convergence or entrapment in local optima, causing the fitness function value to stagnate. To address this issue, this study proposes a K-Means initialization-based two-stage particle swarm optimization algorithm (KTS-PSO-RBFNN) for RBFNN parameter optimization. The proposed method initializes the basis function centers and widths via K-Means clustering and optimizes them in two stages: first, jointly optimizing the centers and widths to locate the global optimal region; then, fine-tuning the widths while fixing the centers to eliminate parameter coupling interference. To comprehensively evaluate the proposed method, synthetic data comparative experiments, Dryden wind disturbance experiments, cross-validation experiments, and real-world dataset experiments were conducted. The results demonstrate that the KTS-PSO-RBFNN consistently outperforms traditional baseline algorithms across all test scenarios. The proposed method achieves lower fitness function values, demonstrating superior prediction accuracy in UAV disturbance estimation.
