To handle the main problem of double-spending attacks in blockchain networks, this paper introduces a new, Light-weight Graph Neural Network (LGNN) approach named Dynamic Sparse Graph Attention Network (DSGAT). To effectively detect double spending behavior, DSGAT method integrates adaptive graph sparsification with attention based on the fundamental graph-structured nature of blockchain transactions. Unlike computationally intensive GNNs, DSGAT may be implemented on edge devices or distributed monitoring systems with low-tech, low-cost hardware since it is optimized for resource-limited environments and doesn't need much processing capacity. To detect double-spending attack, this paper explains building blockchain transaction graphs from a large set of node and edge features. A set of simulated transactions involving double-spending attack is generated using large-scale simulations with the BCASim blockchain simulator, and the performance of DSGAT is compared with normal baselines. The experiment's outcomes prove that DSGAT is able to reduce model sizes and inference latency while keeping high detection rates, proving its feasibility and effectiveness for real-time double spending detection in low-resource environments. To improve blockchain security against double-spending attacks, this paper introduces a novel and realistic alternative.
The application of Artificial Intelligence (AI) technologies is speeding up in educational institutions, particularly in the realm of student engagement monitoring and adjusting classroom environments. Current systems are incapable of combining various modules for attendance tracking, perception of the environment, and behavior tracking, and keeping such modules independent while tracking in a classroom session. This study proposes a transformer-based modular deep learning system (Fusion-AttendNet) combining various environmental sensors, including temperature, humidity, dust, and CO₂, and visual behavioral cues from face detection and temporal signal, to provide real-time student engagement assessment, classroom comfort prediction, and actuator operations. The system utilizes cross-modal attention mechanisms and multitask learning to adaptively switch between model attention on different modalities and tasks. Experimental results show the proposed model outperforms the baseline LSTM, CNN, and BiLSTM models in terms of engagement classification performance, with 94.7% accuracy, and comfort prediction performance, with 0.121 RMSE, while processing in less than 10 milliseconds and achieving 96.2% actuator control operation precision. The framework successfully fills the gap in the related literature that lacks a deployment-ready solution which combines heterogeneous data streams for autonomous classroom management. This study proposes a single platform that combines sensory and behavioral data, using transformer architectures to create a basis for advanced environments that learn with AI.
Private higher technical education requires admission processes that connect student attraction, guidance, follow-up, and enrolment closing within an organized customer management route. This study proposes and evaluates CIDEM-EDU, a student-centric integrated digital marketing model aimed at strengthening customer management in the admission pathway of private higher education. A quantitative applied approach is used through a quasi-experimental design with non-equivalent groups and independent cohorts. The study includes 1331 enrolled students from SENATI Zonal Piura, with Talara as the experimental campus and Sechura as the control campus. Customer management is measured through a structured five-point Likert questionnaire organized into customer attraction, customer service and guidance, and customer follow-up. The instrument presents excellent reliability and a coherent three-factor structure. The results show that the experimental post-test group reaches the highest customer management level, with a mean of 65.6 and a median of 67.0. Welch’s ANOVA confirms statistically significant differences among cohorts, F(3, 310) = 313.0, p < .001, and Games-Howell comparisons indicate that the experimental post-test group differs from the remaining groups. These findings show that CIDEM-EDU strengthens customer management by articulating digital contact, communication channels, commercial follow-up, remarketing, enrolment closing, and feedback as connected moments of the same admission process. The model contributes an integrated digital marketing framework for customer management in private higher technical education.
Crop yield forecasting is an essential approach for providing accurate and interpretable information to facilitate food security, resource allocation, and Policy-Making in agriculture. Limitations in capacity for regional calibration, transparency, and generalisation are common issues for many traditional statistical methods and Black-Box deep learning models, hindering proper adaptability across heterogeneous Agro-Climatic zones. In addition, the predictions of most current models are not interpretable, therefore less trustworthy and unusable in practice. To fill these gaps, we proposed a Two-Stage deep learning framework, called XCalibYieldAI, which aims to maximise both prediction accuracy and interpretability for crop yield. The proposed method combines attention-based temporal modelling with Region-Specific calibration and explainable AI methods, including SHAP value analysis, temporal attention visualisation, and geospatial overlays. They collectively maximise spectral, atmospheric, and vegetation information while also accommodating regional differences in crop phenology and climate. Multi-season remote sensing datasets, with a case study over the continental USA, show that the proposed framework outperforms state-of-the-art CNN, LSTM, and Transformer baselines, resulting in statistically significant improvements in macro accuracy (up to 92.4%) and R² scores. This enhances stakeholder confidence by providing actionable visual explainability through measures of the temporal importance of crop growth stages and the spatial heterogeneity of yield across yield zones. This new XCalibYieldAI framework addresses the critical Trade-Off between crop yield models that are scalable across large areas and highly generalisable, and those that enable complex deep learning predictions to drive agronomic decisions. This acts as an accessible, Open-Source, and Data-Driven platform for contemporary precision agriculture use cases.
Based on artificial intelligence, the paper proposes a framework for ethically and effectively controlling the process of Surrogacy in India, which includes statistical analysis, machine learning, fairness assessment, and explainable AI techniques. Data were processed using descriptive and inferential statistics in a synthetic data set of socio-economic, clinical and procedural variables, and the approval time, maternal health, age and education level were found to be significant predictors of ethical risk. Machine learning (Random Forest and XGBoost) models achieved good predictive performance (ROC-AUC > 0.88), and feature contributions were also given by explainability (SHAP), which helped to make the contribution of features to the model's performance more transparent for policy makers and clinicians. The fairness analysis has showed little bias in relation to income groups, and this has led to an equitable risk analysis and compliance to regulations. The framework provides practical suggestions on refining the approval process, focus on high-risk cases, and improving ethical surrogacy management. Limited availability of data and the fact that regional diversity is not considered represent weaknesses, with future research that should continue to use real-world datasets, federated learning and IoT-based monitoring as risk assessment methods on an ongoing basis. The research adds to AI-assisted governance, policy-making, and ethical decision-making of reproductive healthcare.
Internet of Things (IoT) devices with no battery or energy-harvesting capabilities, like passive RFID tags, NFC modules, and intermittently-powered sensors, have extremely constrained requirements on both power and memory as well as on hardware area. Providing data confidentiality in these setups has been difficult because the traditional cryptographic primitives have energy and computational overheads that are prohibitive. In this paper, FeatherCrypt, a new ultra-lightweight block cipher, is presented that is specifically targeted at providing security in communication between battery-less IoT systems. FeatherCrypt uses a small Substitution-Permutation Network (SPN) architecture that is designed to address low switching, low pillar, and deterministic implementation in intermittency power. The cipher will include a block size of 64bits and either key sizes of 64 bits or 80 bits, and use specifically designed 4-bit substitution boxes and lightweight nonlinearity-based permutation layers to reach significant diffusion and nonlinearity with a minimal hardware footprint. Hardware implementation with a 90-nm CMOS process shows that FeatherCrypt consists of only 1180 gate equivalents and only 0.87 microjoules/encryption, which runs a full encryption in 824 clock cycles at 200 kHz. Implementations on low-power microcontrollers evidenced a small memory footprint of 980 bytes of program code and 456 bytes of RAM, which can be deployed on ultra-constrained environments. Security analysis indicates a high level of resistance to different and linear cryptanalysis, a good avalanche property, and none of them exhibit fixed points. FeatherCrypt provides a trade-off between the energy consumption, hardware area, latency, and cryptographic strength that is optimized in comparison to the current lightweight block ciphers. These findings indicate that FeatherCrypt can be effectively used in the next-generation batteryless IoT applications with security, sustainability, and privacy.
Accurate grading of tumors in the bones is very important in delivering treatment methods and also giving a prognosis to the patient. Conventional classification techniques usually cannot cope with the heterogeneity of the tumor picture and the absence of annotated data. This research work presents a new lightweight deep learning architecture called Radiomic-Modulated Deep Network (RMD-Net) that is aimed at improving the performance of tumor grading based on the combination of radiomic and deep visual image representations. The model utilizes radiomic descriptors in the form of shape, intensity, and texture measurements of segmented tumor regions, and is trained to modulate deep feature activations produced by a shallow convolutional or transformer-based backbone dynamically as symbols of target changes. The resulting radiomic-guided modulation will provide interpretability of the features and better Modularity of the classes because the learning process encapsulates clinically useful properties. A vast amount of experiments on a curated set of CT and MRI scans show that RMD-Net is better at multi-class bone tumor grading tasks, making it more accurate and generalizing with much less parameters. The suggested framework is an effective, interpretable, and clinically flexible solution to assist radiologists in the non-invasive measures of the severity of bone tumors.
Air pollution is a global environmental issues with severe threats and damage to human health and ecosystems. Air pollutants like Particulate Matter (𝑃𝑀2.5), Nitrogen Dioxide (NO2), Sulfur Dioxide (SO2), Carbon Monoxide (CO), and Ozone (O3) has been linked to adverse effects on the environment, including climate change and greenhouse gas emissions. Many researchers carried out their research on machine learning and artificial intelligence techniques for predicting environmental air pollution and for increasing the air quality. But the researcher’s faces the problem of meeting the growing energy demand with minimum greenhouse gas emission and environmental vulnerability. Therefore, a large amount of air quality data is used for improving green energy efficiency. In this paper, the Maximum Normalized Orthogonal Lemma Projective Extreme Learning Machine Classification (MNOLPELMC) Method is introduced for sustainable environmental air pollution prediction with high accuracy and less time complexity. An experimental evaluation of the MNOLPELMC method is carried out with performance metrics such as prediction accuracy, precision, recall, root mean square error, and prediction time.
Multicast transmission is a fundamental mechanism used to deliver content to multiple receivers/users and is still an essential method in television broadcasting services and data distribution networks. Efficient management of multicast traffic has a direct impact on bandwidth consumption and helps to prevent network congestion, especially during high-demand periods that are common in the broadcasting industry. This article investigates the effectiveness of the Internet Group Management Protocol (IGMP) on optimizing the multicast traffic and the utilization of resources by creating a representative simulated network in Cisco Modeling Labs (CML), based on previous experiments and extended versions of them that are carried out on physical devices. The simulated model allows the evaluation and monitoring of two key performance metrics, including bit rate and packet rate, for two operating conditions: case one, with IGMP enabled, and case two, with IGMP disabled. Results show an average of 79.4% reduction in forwarded multicast traffic on active ports when IGMP is enabled, and a 4.85× increase in bandwidth usage when IGMP is disabled. These results confirm how much control IGMP provides in reducing unnecessary data replication and its direct impact on improving network efficiency. The alignment between the simulated and real-world results strengthens the reliability of the results. The combined evidence provides a solid foundation for understanding the role of IGMP in controlling and improving multicast traffic by offering practical guidance for designing and implementing multicast communication systems in networking and broadcasting domains.
The management of small-scale fisheries poses several challenges given the need for biodiversity and climate resiliency in the face of current crises. PELAGIC is a hybrid web-mobile application that aims to streamline pelagic fisheries terminal management. Using StormGlass API, users can post pictures, report a species profile, and receive weather warnings on wind, waves, and El Niño. PELAGIC is data-driven, needs offline capabilities, and uses GPS to relay data. Through a developmental-descriptive process and a variant of SDLC Waterfall model, PELAGIC is designed using Flutter/Dart mobile and Svelte/JavaScript web, and MongoDB, Tailwind CSS, and Leaflet.js. ISO/IEC 25010 outlined questionnaires assessed PELAGIC and rated it “Very Effective” (overall mean = 4.62/5.00). Using quantitative measures, Pelagic transformed a 70% current baseline MFRS transcription error rate to 3.2%, shortened the current 30-90 days report lag to a near real-time (<2 sec query) framework, and showed 95.2% success rate for offline (no-data internet) users. While existing platforms - PeskAAS (requires constant connection to the internet, no weather feature), and ABALOBI (indifferent to weather conditions and grocer owner interfaces, does not feature national languages, including Hiligaynon) show working capacities, Pelagic is unique in its offline-first synchronization and localized language capacity.
In an electrical power system, the operating engineers are more concerned with keeping the frequency at the desirable value. If it deviates from the standard value, it will cause a reduction in operating life and efficiency of the component. This paper presents cascaded 1+PI with PID (1+PI)-PID optimized using the Kookaburra Optimization Algorithm (KOA) for Load Frequency Control (LFC) action, tested on a Two-Area Renewable Penetrated Multi-Source (TARPMS) power system. Nonlinearities are also imposed to model a real system with constraints. The cascaded controller performance is compared with PID using the honey badger algorithm, KOA, and WCA optimized fuzzy PID controller. Furthermore, the controller has also been examined on a Dual Area Thermal-Hydro system (DATH) well-analyzed in the contemporary work. The simulation results demonstrated the superior performance of the presented controller in settling down system deviations compared to the already stated control techniques. However, it is shown that nonlinearity constraints do affect the TARPMS system considerably, and they do need to be accounted. Moreover, it has also been analyzed when TCSC-BES devices have been coupled with the TARPMS system for supplementary control action, which gives more improvement to the system. Finally, robustness analysis is carried out on the TARMPS secondary and supplementary control level.
The DDoS attacks continue to be one of the most widespread menace to the modern network infrastructure, since they flood systems with harmful traffic and disabling of the valid services. Considering the continuously changing nature of the attack pattern, detection through a machine-learning method has been a mandatory requirement to observe the slight differences that are present between the different kinds of DDoS attacks. The paper trains and tests a multi-class XGBoost detection model on a mixed dataset with eleven different categories of DDoS attacks. Precision, Recall, F1 -score, ROC -like interpretations, Precision -Recall curves, confusion matrices, and computational efficiency metrics were used to evaluate the model. The results indicate that there is a high level of detectability variation among classes. High-performance attacks, including DrDoS NTP, TFTP, and DrDoS MSSQL, had almost perfect F1 -scores, which means that they were very separable and displayed consistent patterns of features. NetBIOS, SNMP, and Syn were moderate in their performance, with some overlap in the distributions of features. Classes with the worst performance of LDAP, SSDP, DrDoS UDP, DrDoS DNS, and UDPLag had more misclassification rates and difficulty scores, indicating complex or noisy traffic characteristics that do not help with accurate classification. The model was also found to be very computationally efficient, and inference times were fast enough to allow the model to be used in a real or near-real-time setting. These findings highlight the importance of using hybrid datasets and diagnostics of class-level performance to reveal variability of attack detection. The study concludes that XGBoost has strong flexibility in the accuracy, stability, and working efficiency in multi-class DDoS detection. The next improvements can include the addition of deep-learning frameworks, adversarial training, and real-time threat feeds, which can be used to enhance detection and other network defense against the most difficult types of attacks.
Brassica juncea (Mustard) is one of the most economic seed vegetable crops of the world, playing a major role in the production of world edible oil and the agricultural economy. The third-largest producer is India, which had an area under cultivation of about 8.6 million hectares of Mustard in 2021 22, and annual revenue of over USD 5 billion in 2021 22. Nevertheless, the presence of diseases like Alternaria Leaf Spot, White Rust, Powdery Mildew, and Septoria Leaf Spot threatens yield and quality by up to an estimated 2070% loss every year, based on the severity of the disease, and thus economic losses are estimated at over USD 1.5 billion per year in India alone. Traditional diagnostic systems are based on a manual examination of an agronomist trained to look at the sample and make a judgment, which is time-consuming, subjective, and subject to human error. Current deep learning methods of automated disease detection, promising as they are, are prone to inaccuracies on complex disease patterns, poor uncertainty estimation that is essential in real-world implementation, and poor generalizability to different field conditions. In response to these drawbacks, Swin-BNN-RF, a hybrid framework that combines Swin Transformer as a hierarchical attention-based feature extractor, Bayesian Neural Network (BNN) with symmetrized posterior as a probabilistic learner, and a Random Forest (RF) as an ensemble classifier, is proposed in this study. The Swin Transformer also harnesses local and global spatial biases with its shifted window self-attention network, and it is able to extract better features on leaf images of high-resolution. The uncertainty estimates of the BNN component are trusted, and unambiguous predictions are highlighted to get the opinion of the human expert. Random Forest classifier uses the bagging and boosting ensemble methods to improve stability and the robustness of the classification. A large dataset was experimented with; it consisted of more than 10,000 samples per category of disease in four diseases. In the case of binary classification, the proposed model was 98.32% accurate, 98.52% precise, 98.70% recall, and 98.36% F1. On multi-class classification, it obtained 97.50, 97.82, 98.51, and 97.46 accuracy, precision, recall, and F1 score, respectively, which showed consistent performance in comparison with state-of-the-art models such as EfficientNet, MobileNet, and Residual Networks. The contribution of each component is verified by the Ablation studies and statistical analysis of significance (p < 0.001). The suggested framework is a major achievement of scaling up real-time disease diagnosis in mustard crops, with possibilities of mobile and edge implementation in precision agriculture systems.
The expansive soils have extreme swelling-shrinkage characteristics that undermine the foundations and pavements' serviceability. Simultaneously, ceramic and stone-processing sectors produce high amounts of marble dust, the disposal of which is not planned and is an environmental and land-management problem. This research paper will assess the sustainable re-utilization of ceramic marble dust waste as a microfiller in the stabilization of expansive soil, but to reinforce it further using synthetic polyester fibers. A systematic program that included compaction, shear strength, and California Bearing Ratio (CBR) and model footing tests was undertaken in order to measure density, stiffness, and bearing performance improvements. Scanning Electron Microscopy (SEM), Transmission Electron Microscopy (TEM), and X-Ray Diffraction (XRD) of microstructure and minerals provided the means to explain the stabilization mechanism, and finite-element analyses of PLAXIS 2D were employed to confirm the behaviour of load-settlement. The findings indicate that marble dust acts mainly as a calcium-carbonate microfiller, pore structure refiner, and enhances particle packing, and polyester fibers act as tensile bridging of particles and deformation control. The use of new crystalline phases was not observed, which proved the improvement mechanism to be mainly physical but not chemical. The hybrid mixture, which had 30 wt.% marble dust and 1.0 wt.% fiber, gave the best performance, showing high density, strength, and settlement resistance. The numerical predictions were in good agreement with experimental trends. On the whole, the research shows that there is an ecologically friendly avenue of valorizing the waste of ceramic marble dust and enhancing the working of expansive soils based on the principle of microstructural densification and mechanical reinforcement.
Rapid growth in Internet of Things (IoT) devices has expanded the attack surface of modern-day networks. Therefore, it is imperative to deploy an IoT Intrusion Detection System (IDS) for secure communication and reliable operation. However, traditional IDS systems are often inefficient when dealing with high-dimensional data, new attack strategies, and severe imbalanced data problems, leading to low detection performance and high false alarms. In order to solve these problems, this paper introduces a novel intelligent IDS system using the Feature Selection technique inspired by Bowerbird Courtship (BBFS) and Long Short-Term Memory Autoencoder (LSTM-AE) using Seagull Optimizer (SGO). The most important objective here is to build an effective and scalable IDS that can perform efficient feature selection, learn deep temporal dependencies, and tune its hyperparameters to classify malicious and benign traffic more effectively. In this way, we consider all aspects of data preprocessing, feature optimization, balancing, and classification, overcoming the limitations of existing techniques. The evaluation of the CIC IoT 2023 intrusion dataset proves the efficiency of the model, which is shown by high scores: 99.63% of accuracy, 99.55% of detection rate, 99.71% of precision, and 99.59% of F1 score. As seen from the comparison with other models, the BBFS-LSTM-AE-SGO model is better than the compared model in terms of all metrics. It means that the proposed IDS can detect various types of attacks with minimum errors. All through this research has made way for the design of a novel, optimized IoT-IDS model that thereby strengthens cybersecurity resilience, supports real-time monitoring, and hence advances the intrusion detection for IoT-enabled environments.
The study employed an Artificial Neural Network in combination with the optimized Adaptive Moment Estimation (Adam) algorithm, currently the only AQI forecasting model available in the Philippines. The modified QHAdamW - Quasi-Hyperbolic Momentum (QHAdam) and Adam with decoupled weight decay (AdamW) were both extensions of the Adam optimizer, and both offer unique advantages for training ANN. The proposed QHAdamW optimizer addresses the issues on convergence, generalization, and forecasting performance of Adam. Hyperparameter tuning results revealed that 0.01 and 0.001 were the most effective optimal values for the generalization performance of QHAdamW. The comparative analysis results using seven evaluation metrics revealed that the error value range is lower, and the regression coefficient, having a value approximately equal to 1, improved the model accuracy performance. Likewise, the model converges to a satisfactory level of performance with the convergence performance results of lower loss values as obtained from training and validation losses. Based on data from a real-time air quality tracking station in Manila, a feed-forward neural network is used to predict the AQI of PM2.5 and PM10 separately. This model can be used to forecast Particulate Matter (PM), to help the Department of Environment and Natural Resources - Environmental Monitoring Bureau (DENR-EMB) implement a comprehensive air quality management.
Critical vulnerabilities have been made public by the fast expansion of Internet of Things (IoT) devices, making these networks easy target for cyber-attacks. While security solutions based on Machine Learning (ML) have shown potential, they often encounter issues including slow detection times, scaling issues, and a lack of generalisability when it comes to diverse IoT devices. To work with these issues, this paper introduces an innovative ML-based security paradigm. The proposed framework improves the attack detection accuracy by combining adaptive feature extraction techniques with a context-attentive hybrid mechanism. The new paradigm maximizes detection accuracy and computational efficiency. This is achieved through real-time dynamic adjustment of feature selection against network conditions, rather than traditional hybrid approaches. Furthermore, a lightweight and scalable detection method fit for execution on low-resource IoT devices is offered. It is apt for several IoT environments. The proposed framework beats several current models by 15% in accuracy, 25% in the reduction of false positive rates, and 30% in detection times, according to experimental tests carried out on numerous IoT datasets.
The increasing influence of generative Artificial Intelligence (AI), especially in all aspects related to language in higher education, has developed at an increasingly rapid pace. There are no national policies that support the use of AI in higher education in the Philippines; therefore, there is little to no consistency in how different Higher Education Institutions (HEIs) will employ AI within their respective systems. This study evaluates the extent to which AI technology was used by four HEIs in Bulacan, Philippines, for both oral and written tasks among students, teachers, and administrators. Utilizing both the qualitative grounded theory methodology as well as the Delphi method, this study found out that many AI technologies have been utilized by the student, teacher, and administrators for brainstorming ideas, editing documents, developing lessons, preparing materials, and documentation. However, the study also reveals numerous issues associated with the use of AI, including plagiarism, excessive reliance on AI, loss of original thought and creativity. Ultimately, the findings from this study demonstrate a significant need for comprehensive national policies governing the use of AI, consistent and fair enforcement of these policies, as well as significantly increased awareness among educators, students, and administrators regarding the appropriate use of AI. Based on the findings of this study, it is recommended to develop a context-specific national policy framework that provides reasonable and balanced governance and regulation over the development, implementation and use of AI technologies within the Philippines' system of higher education.
Reading proficiency is considered a critical educational challenge in a highly multilingual nation such as the Philippines. Digital literacy tools available on the market and those that are found in the literature are mostly English-centric and often lack interactive mechanisms. This study shows the design, technical validation, and implementation of the iRead mobile application software. It is a multilingual mobile reading platform with offline speech recognition function available for three languages, specifically English, Filipino, and Hiligaynon. The mobile application was developed specifically for the Android Operating System using the Flutter framework, while the Vosk API was used for the speech recognition engine. Publicly available pretrained speech recognition models were utilized for English and Filipino languages, while a novel baseline small-vocabulary speech recognition model for Hiligaynon was developed and trained from scratch. A Gaussian Mixture Model-Hidden Markov Model (GMM-HMM) pipeline within the Kaldi framework was then used to form the Hiligaynon speech recognition model. Recognition vocabulary was limited to a 380-word phonics-based lexicon that is aligned with early literacy instruction. Cross-speaker generalization for Hiligaynon was evaluated using a leave-one-speaker-out cross-validation technique across four speakers. Recognition stability was further assessed using standard deviation and confidence interval analysis. The overall system evaluation was conducted using 540 utterances across the three languages under controlled conditions. Recognition performance achieved average accuracies of 92.8% for English, 88.3% for Filipino, and 85.6% for Hiligaynon. Category-level analysis demonstrated the highest performance for vowels, followed by consonants, then consonant–vowel blends. Results suggest that a classical small-vocabulary acoustic model combined with grammar-constrained decoding is technically viable and deployment-ready in a multilingual offline speech-supported literacy app for low-resource educational settings.
This research focuses on developing a system using deep learning techniques for the automatic detection of products at the company Negociaciones 7 E.I.R.L., with the aim of reducing errors and providing a clear, efficient alternative to traditional record-keeping. To this end, three architectural models were evaluated: YOLO11n, YOLOv8s, and YOLOv8n, taking into account metrics such as accuracy, recall, mAP, and detection time. The results show that YOLOv8s performs best, achieving an accuracy of 92%, a recall of 90%, and an mAP50 of 91%, standing out for its balance between accuracy and speed. In contrast, YOLOv8n demonstrated a rapid response, albeit with intermediate performance (85% accuracy and an mAP50 of 87%), whilst YOLO11n showed low accuracy and stability at 80%. The implementation of the system enabled a 80% reduction in registration errors and optimised operational efficiency. The process was guided by the CRISP-DM methodology, covering everything from image processing and data collection to integration with SQL Server. Furthermore, statistical validation, using non-parametric tests such as the Mann-Whitney U, Kolmogorov-Smirnov, and Welch’s t-tests, confirmed the significance of the improvements. The results demonstrate that the YOLO architecture optimizes productivity, aids inventory management, and contributes to digital transformation in the retail sector. Furthermore, in the context of Peru, particularly in Trujillo, there are few similar studies, which highlights its innovative nature.
