This paper proposes a survey and critical analysis of recent advancements in antenna design for fifth-generation (5G) communication systems. It thoroughly examines the integration of machine learning (ML), deep learning (DL) and intelligent optimization techniques for enhancing antenna performance. Different architectures like Massive MIMO, phased arrays and beam forming frameworks are compared in structural design, operational mechanisms and implementation challenges. A quantitative comparison of various methods shows trends and differences in spectral efficiency, throughput, latency, connectivity and signal quality. The analysis reveals that AI-assisted optimization significantly improves adaptability, interference control and computational efficiency. The study concludes with insights into current limitations and provides directions for future research toward energy-efficient, self-adaptive and intelligence-driven antenna systems for next-generation 5G and beyond networks.
The rapid advancement of generative models has led to the widespread creation of deepfakes highly realistic but manipulated media posing serious risks to digital trust, security, and societal misinformation. Among various deepfake techniques, face-swapping methods are particularly concerning due to their visual realism and potential for misuse. This paper proposes a novel deepfake video detection framework called Feedback-Aware ViT-LSTM that integrates Vision Transformers (ViTs) with Long Short-Term Memory (LSTM) networks through a dynamic feedback mechanism. The architecture is designed to exploit the spatial attention capabilities of ViTs and the temporal modeling strength of LSTM to detect subtle manipulation artifacts in face-swapped videos. Each video is sampled using a hybrid uniform-random strategy, and facial regions are detected using a YOLOv8 model. These faces are processed by ViTs to extract spatial embeddings, which are then passed to an LSTM to model temporal dependencies across frames. A key innovation is the feedback mechanism: temporal insights from the LSTM are used to dynamically guide the ViT's attention toward regions exhibiting temporal inconsistency, thereby reinforcing sensitivity to manipulation cues. To enhance interpretability and transparency, the model incorporates Explainable AI techniques that highlight the facial patches influencing the classification decision. Experimental evaluation on benchmark deepfake datasets demonstrates that the Feedback-Aware ViT-LSTM achieves superior performance, confirming the effectiveness of the proposed Feedback-Aware ViT-LSTM in robust and interpretable deepfake detection.
Video-based anomaly detection seeks to discover anomalous events, such as crimes, fires, or medical emergencies, by utilizing both spatial and temporal features of video data. Traditional surveillance systems are frequently limited to minimal recording, requiring human analysts for post-event assessment, resulting in delayed responses during crucial occurrences. To address these issues, we present a multi-layered approach to detecting video anomalies that can deal with both temporal and spatial components of video data. The input video is initially obtained from the dataset and undergoes frame conversion. The extracted key frames are then preprocessed for further analysis. To obtain multi-scale spatial characteristics from each frame, the first layer uses a spatial Pyramid pooling network (SPP-Net) along with a convolutional neural network (CNN). These spatial features are then passed to an optimized bi-directional gated recurrent unit (Opt-Bi-GRU) enhanced with Multi-Head Self-Attention (MHSA), which analyzes the temporal dynamics and captures both forward and backward dependencies across frames. Finally, a capsule network (CapsNet) processes the output of the Bi-GRU, identifying complex patterns that may indicate abnormalities over time. The proposed method is implemented using Python. The proposed model performs better than existing methods in terms of F1-score, specificity, sensitivity, accuracy, recall, precision, FPR, and FNR. The proposed model achieves the highest accuracy of 98.2%, 98.87%, and 98.52%, respectively, utilizing the UBI-fights, UCF-crime, and UCSD pedestrian datasets. These results demonstrate that the proposed framework provides an automated, reliable, and effective solution for real-time anomaly detection in surveillance systems.
As the photovoltaic industry develops toward high efficiency and intelligent manufacturing, the requirements for defect detection accuracy and stability in the photovoltaic cell production process continue to increase. To address weak features and large-scale differences of multiple types of defects in photoluminescence images, this study proposes a defect detection method that combines physical consistency preprocessing and improved YOLO based on analyzing the physical characteristics and defect distribution rules of photoluminescence imaging. This method enhances fine-grained defect representation capabilities through full-dimensional dynamic convolution, introduces coordinate convolution to improve spatial positioning sensitivity, and combines a multi-threaded non-local attention mechanism to model the global correlation of cross-regional defects. Experimental results demonstrate that in comparative tests on PV-PL and SCPD datasets (Table 2), the method achieves high comprehensive performance in multi-type defect detection tasks, with an F1 score reaching 94.84%. Under conditions where 18% of the PV-PL data set is interference (Fig. 10), the false positive rate is approximately 7.0%, and the detection latency is as low as 9.7ms. In tests with varying illumination intensities (Table 6), the model achieves the lowest average localization error of 1.76px under normal lighting conditions. The proposed method has obvious advantages in detection precision and robustness, and can offer effective technical support for photovoltaic cell photoluminescence image defect detection.
One of the leading causes of impairment in the world is major depressive disorder (MDD), and a correct diagnosis and treatment plan depend on an accurate assessment of the disorder’s severity. Modern methods frequently use unimodal signals, which fall short in capturing the complex neurobehavioral foundations of depression. In order to overcome this limitation, we offer a multimodal deep fusion framework using graph neural networks (GNNs) that combines behavioral inputs, electroencephalography (EEG), and functional near-infrared spectroscopy (fNIRS) to predict the severity of depression. To illustrate physiologically significant interactions between and across modalities, we specifically create correlation-weighted graphs. This enables the network to learn multimodal embeddings that are both compact and discriminative. The framework is evaluated on a dataset comprising 100 participants (50 diagnosed with MDD and 50 healthy controls), integrating multimodal data from 32-channel EEG, 16-optode fNIRS, and behavioral assessments derived from cognitive tasks, thereby ensuring diversity in neurophysiological and behavioral representations. The suggested framework routinely outperforms unimodal models and traditional machine learning benchmarks in both regression and classification tasks, according to extensive experiments using a dataset of 100 individuals. Furthermore, the clinical relevance of this technique is highlighted by strong correlations between verified clinical metrics and expected scores. The results confirm the effectiveness of multimodal GNN-based fusion as a reliable and objective tool for improving computational psychiatry and strengthening the assessment of depression severity.
Liver segmentation is a challenging task for various clinical applications. The challenges include a weak boundary between the liver and its neighbors, speckle noise, partial volume effect, and large variations in both the shape and size of patient’s organs. The traditional deep learning models do not incorporate the biomechanical and physical characteristics of organs, and hence they may produce inaccurate shape deformation, which is useful much in clinical circumstances. Hence, this research work tackles the above problem using a physics-enhanced approach to detect the boundaries. The proposed model achieved a Dice score of 0.96, a 95th percentile Hausdorff distance of 2.9, an average surface distance of 0.65, an accuracy of 98.1%, a precision of 97.6%, a recall of 97.2% and an F1-score of 97.4%.
Current inspections of exterior wall insulation layers in high-rise buildings face challenges such as high risk, low efficiency, and difficulty in scaling up manual inspections, driving the integration of unmanned aerial vehicle (UAV) remote sensing and lightweight depth vision recognition technologies in building inspection. To achieve automatic identification and robust tracking of key defects such as insulation layer cracks, delamination, and moisture seepage, this paper proposes an automatic defect identification method for exterior wall insulation layers based on UAV image recognition technology. This method first acquires multi-source images using a UAV, then employs a lightweight exterior wall defect detection network model for accurate defect identification, and finally introduces a post-processing extension framework with spatiotemporal consistency and structural region constraints to achieve dual localization of two-dimensional defects based on structural attribution and three-dimensional coordinates. Results show that this method has low latency (< 40 ms) and low GPU usage (< 51%). In on-site inspection tests at a real building height of 60 m, the method achieves an engineering detection rate of 0.936. This demonstrates that the method can accurately identify exterior wall defects with low resource consumption, making it suitable for long-term, high-frequency, and low-consumption deployment of UAVs for edge-based inspections. The proposed method provides a low-altitude remote sensing-based intelligent inspection framework for UAV-based urban building facade monitoring, which can serve as a reference for engineering applications such as periodic inspection and large-scale screening.
Photovoltaic rapid shut-off devices operate for a long time in outdoor high-temperature, high-irradiation and humid-heat environments. Their early degradation characteristics are often weak and vary widely across sites. Traditional diagnostic methods based on a single electrical or thermal imaging signal are difficult to meet actual operation and maintenance needs. Therefore, this study proposes a photovoltaic rapid shutdown device health status assessment model based on multi-modal feature fusion and transfer learning. The model improves the complementary expression between different modalities through weak mode enhancement and cross-modal attention mechanisms, and it also uses a distribution alignment strategy to alleviate cross-site feature drift. Experiments showed that this method significantly improved recognition performance in multi-modal fusion experiments. The F1 value of the early fusion model increased from 82.21% to 89.54%, and the accuracy of the weak mode complete enhancement scheme increased from about 74% to nearly 86%, and remained the best in the four types of ablation comparisons. Furthermore, in the noise perturbation experiment, when the noise intensity reached 0.4, it still maintained an AUC of 0.89, which was 5-8% higher than the comparative method on average. At the same time, the performance retention rate still exceeded 89% when the mode was missing 30%, showing more stable robustness. The study verifies the stability of the proposed method under the conditions of weak degradation identification, noise immunity, and mode loss, and provides an effective technical approach for photovoltaic equipment health monitoring.
The concept of game-based learning has been discovered as a potent instrument for raising the level of engagement and motivation in learners. These systems are, however, not optimally effective in multicultural and multilingual environments because most of them do not incorporate the use of cultural symbols, which may connect with the identity of different learners. The proposed study seeks to fill this gap by developing an educational approach applied as a game-based learning program, which incorporates cultural symbols and the proposed game-based learning system enhances learner engagement by incorporating interactive game mechanics and culturally relevant elements. It increases motivation by providing personalized and meaningful learning experiences aligned with learners’ cultural backgrounds. In addition, the system improves language acquisition outcomes, particularly in vocabulary development, comprehension, and communication skills. The main aim of this paper is to examine the performance of a game-based learning system which encompasses the use of symbols relating to their culture as well as the use of adaptive learning controls, which will eventually result in an improvement in the performance of the learners in terms of acquisition of vocabulary, acquisition of communication skills and participation. A questionnaire-based data was gathered among 180 participants aged between 15 and 25 years in secondary and tertiary learning institutions. Adaptive difficulty, a feedback system, culturally relevant symbols like avatars and stories were integrated into the game-based system. The analysis was conducted using Structural Equation Modelling (SEM) to determine the connections among the elements of the game, cultural symbol integration, engagement, motivation and learning outcomes. The findings demonstrated that integration of cultural symbols had a strong positive influence on engaging and motivating learners, and it positively influenced the learning outcomes in terms of vocabulary acquisition and understanding. In addition, adaptive control, such as dynamic task difficulty and individualized feedback, was further applied to improve learner satisfaction and retention. Incorporation of cultural signs in the learning systems based on games can enhance the learning experience among different groups of people. This framework can provide practical lessons about developing culturally adaptive technological educational systems to create more contextual and personal learning environments.
With the rapid advancement of urbanization and the increasing density of population, urban public safety has become a critical focus of modern city governance. The integration of the Internet of Things (IoT) and Artificial Intelligence (AI) provides an innovative pathway for constructing intelligent public safety surveillance systems. This paper proposes a multi-layer urban surveillance framework that integrates an enhanced multi-label image recognition model, an adaptive pan-tilt camera positioning mechanism, and a collaborative cloud-edge control platform. To address challenges such as multi-label dependencies, real-time video segmentation, and heterogeneous system coordination, we design a hybrid architecture that combines a Graph Convolutional Network (ML-GCN), a Slice Recurrent Neural Network (SRNN), and convolutional visual features to model spatial-semantic dependencies in surveillance images. Furthermore, a real-time video localization and smart tracking system is implemented by incorporating task-driven frame classification, temporal filtering, and quadrant-based pan-tilt control guided by image block offsets. On the system level, we construct a cloud-edge integrated platform using Layer2/Layer3/SAN interconnection and deploy AI-based scheduling to manage distributed computational resources under dynamic constraints. Extensive experiments on urban surveillance datasets demonstrate that our method significantly outperforms traditional models in accuracy (+9.3%), micro-F1 (+10.4%), and frame-level anomaly detection delay (reduced by 74.7%). Moreover, the cloud-edge scheduling model reduces average processing latency by 69.7% and increases task throughput by 90.1% compared with baseline strategies.
Smart transportation is a burgeoning area of study that leverages technological advancements to enhance the safety and efficacy of transportation infrastructure. The application of sensors, data analytics, and communication technologies is utilized to augment the transportation experience holistically. Situational awareness is a crucial aspect of smart transportation, encompassing the ability to detect and anticipate potential hazards and obstacles on the road ahead. Maintaining situational awareness is of utmost importance in guaranteeing secure and adequate transportation. Nevertheless, its intricate and ever-changing characteristics make transportation systems a formidable task. Modern transportation systems generate copious amounts of data with various features, exceeding conventional algorithms' processing capabilities. In response to these challenges, the research suggests implementing an Advanced Randomized Algorithm-based Situational Awareness Model (ARA-SAM) to detect the congestion areas and route to free paths. The ARA-SAM algorithm employs a machine-learning approach. It utilizes a randomized algorithm, specifically the Hybrid Ant Colony Optimization (HACO), to forecast potential hazards and obstacles that arise on the roadway. The system is engineered to effectively manage large quantities of data and promptly adjust to evolving circumstances. ARA-SAM's performance was evaluated through simulations utilizing real-world data. The findings indicate that ARA-SAM can precisely predict potential hazards and obstacles while maintaining a low latency rate with a delay of 23.4s, a queue length of 44.4m, a travel time of 303.8s, a density of 44.8 vehicles per km, and a vehicle count of 1937.5 vehicles per hour. The model's efficacy in identifying lane changes, pedestrians, and vehicles, among other factors, is evidenced by the sample values. Enhancing the safety and efficiency of transportation systems can facilitate the emergence of novel advancements in the domain.
Enterprise process automation systems are more and more exposed to internal control deficiencies due to misconfigurations, resource bottlenecks, and software anomalies. The traditional fault detection and recovery systems are not scalable, interpretable, and not capable of providing real-time responses; hence, they cannot meet the requirements of the current cloud-based environments. The proposed methodology consists of three components: The predictive fault prediction by the bidirectional long short-term memory (Bi-LSTM) networks, the SHapley Additive exPlanations (SHAP)-based interpretability-based root cause analysis (RCA), and a hybrid self-healing engine which uses both rule-based logic and reinforcement learning (RL) for its operation. The whole setup is trained and tested on the Aliyun Cloud Fault Dataset, where detailed temporal and structural fault traces are provided from large-scale enterprise cloud clusters. The proposed solution involves a series of complex preprocessing techniques including KNN imputation, Min-Max scaling, and one-hot encoding which is then followed by statistical, temporal, event-pattern, and graph-based dimensions feature extraction. The Bi-LSTM model captures both forward and backward temporal dependencies that culminate in precise fault classification and severity scoring. To facilitate this, SHAP ranks the features by their contribution and the self-healing engine can run corrective actions automatically and also learns about new fault conditions through RL feedback loops. The results of the experiments show excellent fault detection accuracy (98%), precision (97%), recall (99%), and healing success rate (90%). The RL agent shows rapid convergence and generalization across different fault episodes. The system offers an auditable, adaptive, and scalable enterprise fault management solution that significantly reduces downtime and human effort. In the future, the solution will be extended to support multi-cloud environments and TinyML agents for edge deployment will be implemented.
Autonomous driving greatly relies on the correct object detection and scene comprehension. Nevertheless, it has been a continuous challenge to make sure that there is real-time accuracy in dynamic and unpredictable environments. The classical approaches have drawbacks in terms of detecting performance enhancement, computational effectiveness, and responsiveness, which degrade the security and reliability of self-driven cars based on the Internet of Things (IoT). To overcome those issues, in this paper, I have introduced YOLOSwinNet, a hybrid framework that will combine YOLOv8 with the Swin Transformer to improve perception in autonomous navigation. The proposed model takes advantage of the rapidity and lightweight characteristics of YOLOv8 and hierarchical feature acquisition of the Swin Transformer and enhances the detection performance in complex urban environments. YOLOSwinNet was developed to be effective in processing information produced by the IoT in order to allow a smoother navigation process and ensure successful obstacle avoidance. It is unique in that it integrates a convolutional architecture and transformer-based architecture to attain a context-aware perception with enhanced spatial awareness and flexibility to various road conditions. A superior attention process is also used to refine the outputs of the detection process by minimizing false positives and enhancing decision-making. Experimental results show superior performance, achieving mAP@50 of 0.98348 and mAP@50-95 of 0.9870. Comparative evaluations demonstrate YOLOSwinNet's advantages over state-of-the-art models in precision, recall, and inference speed, offering scalability, interpretability, and energy efficiency for next-generation autonomous vehicle systems.
In this study, a multi-modal driver behavior recognition framework aimed at improving the safety and reliability of autonomous vehicles utilizes sensor and vision data. The framework collects inertial signals (accelerometer and gyroscope) and visual images relating to driver activities to observe driver behaviors. The processing stage begins by preprocessing the inertial data using noise filtering, normalizing and interpolation, followed by feature extraction using Short-Time Fourier Transform (STFT) and Continuous Wavelet Transform (CWT), producing complementary time-frequency representations of the data. For the visual data, begin by contrast enhancement and normalization, before utilizing a Vision Transformer (ViT) to obtain spatial embeddings. The features from both modalities are encoded by a Swin Transformer to capture local and global dependencies, then fused using a combination of cross-attention and channel attention to provide improved interaction of the features. The fused representation is then processed to produce a final behavior representation utilizing ResNet-18, to classify into nine (9) categories. The results of the experimentation show outstanding performance at greater than 99.6% accuracy, F1-score, recall, and precision with a very low false positive and false negative rate, demonstrating its robustness and application for driver monitoring in manner in autonomous systems. The combination of STFT and CWT offers diverse fixed and multi-resolution representations in the frequency domain, whereas integrating ViT and Swin Transformer takes spatial-temporal encoding even further than previous multi-modal methods; the application of ResNet-18 classifier adds even more power to discriminative fusion, making it possible to perform behavior recognition that is more trustworthy than any existing single-or dual-branch systems combined.
Bridges are crucial for public safety and economic activities; thus, their structural integrity must be the priority concern; however, the current crack inspection methods are slow, depend on personal judgment, and do not 100% eliminate human error. The fading of old infrastructures, accompanied by the increase in traffic, has led to the need for an immediate, accurate, and efficient automatic crack detection system. The current aim of the project is the realization of a Bridge Structural Crack Detection System that is based on a Golden Tortoise Beetle Weighted Convolutional Neural Network (GTB-WCNN) to achieve credible and fast crack detection. In order to simulate a wide range of scenarios regarding light, material, and surface, high-quality images of the bridge surfaces were taken and supplemented by UAV-based field image capturing. In the pre-processing phase, Contrast-Limited Adaptive Histogram Equalization (CLAHE) was employed to enhance contrast and make the cracks more visible while simultaneously Gaussian filtering was used to minimize noise. The use of Histogram of Oriented Gradients (HOG) for feature extraction not only made it possible to collect the necessary edge and gradient features but also resulted in the very fine distinction between microcracks and background textures, thus leading to a better performance of crack identification in difficult-to-interpret photos of bridges. The GTB-WCNN model surpassed traditional ones by a significant margin in terms of the above-mentioned metrics, achieving 98.33% of detection accuracy, 97.61% of specificity, 98.21% of F1-score, 98.41% of precision, and 98.21% of recall. The model also expresses its strong generalization ability by being able to detect both small and large-scale cracks in different situations.
As the supply chain is growing exponentially owing to multimedia, the cost of logistics transportation has started to play a vital role in influencing the efficiency of the system. Conventional heuristic and evolutionary methods, including genetic algorithm (GA) and ant colony optimization (ACO), offer viable solutions to vehicle routing concerns; nevertheless, they experience difficulties in being flexible, efficient in computation, and sensitive to parameters in dynamic conditions. To overcome these limitations, this paper offers an adaptive Actor-Critic reinforcement learning system (ada-A2C) to optimize logistics distribution tracks. The model combines a convolutional encoder and a pointer network decoder to include attention mechanisms to capture spatial, temporal, and demand features of distribution nodes. Comparative experiments on synthetic datasets with 10, 20, and 50 nodes indicate that although Google OR-Tools can produce better quality paths in small-scale tasks, the proposed ada-A2C algorithm can converge much faster and possesses greater time efficiency, especially when the node size is larger. Moreover, the model does not need reconfiguration of the parameters many times, and it is adaptable to the input structures of variables. Cost analysis also proves that the ADA-A2C introduction lowers the overall logistics expenses by about forty percent, and the greatest effect was in the areas of manual labor, transportation, and inefficient inventory. A more precise definition of the audit control mechanism is as follows: (1) verification of each generated decision in real-time, which is constantly checked against capacity, distance and feasibility limitations; (2) dynamic penalty assignment, which is an evaluation that penalizes inefficient or infeasible routing behavior and is dynamically considered by the evaluating the agent to influence behavior toward optimal outcomes; and (3) feedback of corrective policies, which combines the assessments into the reinforcement learning process to steer the agent to good behaviors. This hierarchical process provides policy changes to be consistent and in line with the optimization goals during the training.
Background: Autonomy needs an intelligent interplay of controllers involving subsystems such as steering, braking and suspension to maintain safety, comfort, and efficiency. Rule-based controllers and single-agent reinforcement learning have yet to prove scalable, adaptable, and robust within a dynamical multi-agent environment. Methods: This paper proposes a vertical collaboration framework, tying MADDPG methods with dSPACE real-time simulation for chassis domain control. The subsystems are seen as decentralized actors with a centralized critic, operating as independent agents, to permit adaptability at the subsystem level and cooperative global optimization. The framework includes variable noise for exploration trade-off, prioritized replay to speed up learning and HIL testing in deployment. Results: As a simulated task, a success rate of 93% was ideal, and targets were not deviated from much (0.16-0.17m), and an average passing time of (approximately) 18.5s was optimal. The model however adjusts the comfort parameter that the brake-and-steer system offers by setting the reward of an episode constant to all the agents. The applied model is more noise-resistant and reactive, as well as less complex as compared to the classical methods. Conclusions: The MADDPG-dSPACE framework aims at balancing the reinforcement learning theory with real-life practice by offering a scalable, modular, and robust framework to future autonomous driving systems that would ensure safety, adaptability, and cooperative chassis control by providing a broad range of driving scenarios.
In order to improve the safety and energy utilization of vehicles, a combination of vehicle stability criterion models and traffic flow models is proposed to plan vehicle paths from the level of path planning to avoid extreme and inefficient working conditions, enabling fast and safe driving under slippery road conditions. The traffic flow model is used to predict the future changes in the traffic environment that the vehicle will face, and then the stability criterion model is used to assess the safety of future traffic in order to plan the fastest and safest path for the hybrid vehicle. Specifically, the generalized Aw-Rascle-Zhang (GARZ) macroscopic traffic flow model is solved using the flux vector splitting format in order to predict the future changes in speed and traffic density that the hybrid vehicle will face. In addition, the front-wheel steering angle responses given by the same driver at different speeds and different distances relative to the vehicle in front were collected using the driver-in-the-loop simulation platform Prescan. Simulink models based on front-wheel drive (FWD) front-wheel steering (FWS) vehicles and all-wheel steering (AWS) distributed drive vehicles (DDVs) give the force saturation factor (delta TFSC) response corresponding to different front-wheel steering angles. The stability criterion model of the vehicle was established by using artificial neural network (ANN) to train delta TFSC corresponding to different speeds and traffic densities. The parameters predicted by the traffic flow model (vehicle speed and traffic density) were evaluated for stability using the newly established stability criterion model. The vehicle traveling paths were optimized based on the above methods to ensure the safety of vehicle traveling on slippery road surfaces. Finally, real US-101 traffic flow data were used to verify the predictions of the traffic flow model.
In view of the complex snow and ice environment in the high cold region, the existing standard design parameters for vehicle-to-infrastructure cooperation application are not referenced. The working conditions of ice and snow roads are defined, namely three types of road surface, namely ice slab, compacting snow and melted snow. The application message set is delivered through the roadside terminal, and the early-warning model under the minimum safety distance is established to obtain the latest early warning time. Develop test cases and test schemes for vehicle-infrastructure cooperation early-warning application scenarios with characteristics of cold regions and complete real vehicle verification of 84 test cases in 16 test scenarios in Heihe Autonomous Driving Test field, providing practical reference for vehicle-infrastructure cooperation application development in snow and ice environment.
Probabilistic Self-Organizing Maps (PRSOM) are effective for visualizing complex patterns in large datasets due to their neural network-based structure and probabilistic reasoning capabilities. However, the use of ordinary derivatives in their learning rules limits their ability to capture underlying temporal or spatial dependencies. To address this, we propose a novel model; Fractional Probabilistic Self-Organizing Map (FRAC-PRSOM), which integrates fractional derivatives into the PRSOM framework. Specifically, the Caputo-Fabrizio derivative of order alpha is an element of (0,1) is adopted to introduce memory and non-local behavior into the learning process. The model reformulates the learning rule to simultaneously incorporate both fractional calculus and probabilistic density estimation, thereby enhancing the system's adaptability and depth of learning. We provide a theoretical analysis establishing the stability, sensitivity, and convergence of FRAC-PRSOM. The method exhibits near-linear scalability (O(N-0.96)). Comprehensive experiments on twelve benchmark datasets evaluate FRAC-PRSOM against seven state-of-the-art baselines using four clustering metrics. Results demonstrate statistically significant improvements (p<0.05), with average Silhouette gains of 5-15% and up to 187% improvement on complex datasets (e.g., Magic). Statistical validation via Friedman and Nemenyi tests confirms FRAC-PRSOM's superiority over all baselines in model fit while maintaining competitive cluster separation. Sensitivity analysis across alpha is an element of [0.1,0.95] reveals dataset-dependent optimal values while preserving numerical stability. Overall, FRAC-PRSOM provides a principled framework for memory-enhanced probabilistic clustering, offering particular promise for complex, high-dimensional datasets, where capturing non-local dependencies is crucial.
