Fusion: Practice and Applications

Infrastructure-based autonomous perception operates under fundamentally different constraints than vehicle mounted systems: elevated-mounting geometries producing depression-angle-dependent sparse point clouds, a 12.95 W IEEE 802.3af Power-over- Ethernet (PoE) power ceiling, and distributed asynchronous sensing governed by IEEE 1588v2 precision time protocol (PTP) synchronization uncertainty. Existing automotive radar–LiDAR fusion frameworks assume abundant power, dense sensing, and synchronous measurements — assumptions that all fail in fixed infrastructure deployments. This paper presents XADAR, an uncertainty-aware multi-modal fusion framework designed for these infrastructure-specific constraints. XADAR makes three princi-pal contributions: (1) a covariance inflation mechanism that propagates PTP synchronization uncertainty continuously through the fusion pipeline, replacing hard synchronization thresholds with a smooth degradation curve proportional to temporal offset; (2) adap-tive sensor-specific fusion weights derived from modality covariance matrices that account for IWR6843 77 GHz FMCW radar Doppler ambiguity and ground-reflection multipath, and TFS20-L ToF LiDAR atmospheric scattering and range-zone limitations; and (3) a complete reproducible architecture including an IEEE 802.3af-compliant power budget (5.78 W maximum concurrent load; 41.6% PoE safety margin), quantitative 77 GHz propagation analysis based on ITU-R P.676-12 and P.838-3 (10.7 dB fade margin at 100 m under 50 mm/hr rain), and an MIL-STD-1629 FMEA covering twelve failure modes with severity classifications. A structured five-stage validation pathway from synthetic temporal-offset experiments to sixmonth field trials is defined for future empirical work.

Black fungus disease (mucormycosis) has emerged as a critical health threat, particularly during the COVID-19 pandemic, where immunosuppressed individuals have shown increased susceptibility to opportunistic fungal infections. The aggressive progression of mucormycosis and its high mortality rate, exacerbated by diagnostic delays, underscore the urgent need for accurate and automated detection systems. In this study, a deep learning-based diagnostic framework is proposed for the early identification of black fungus infection using convolutional neural networks (CNNs). Experimental pipelines were developed and evaluated. Several deep learning models based traditional CNN architectures including VGG16, VGG19, InceptionV3, and MobileNetV2 have been study on a structured dataset comprising high-resolution mucormycosis images. Comparative evaluations across both pipelines revealed that the MobileNetV2 architecture consistently outperformed other models, with accuracy reaching 99.86%, F1-score of 0.98, and minimal overfitting across validation datasets. The proposed system holds strong potential for real-world clinical deployment, particularly in resource-limited healthcare settings, offering rapid, scalable, and explainable AI-driven diagnostics to combat the rising threat of black fungus infections.

SARS-CoV2 virus has affected the peoples in worldwide with several issues, like health and economy. Moreover, mathematical definition of fractal dimension affords a method for calculating the non-linear dynamic behaviour difficulty revealed through time series of countries. The fuzzy logic model illustrates and manages the characteristic uncertainty of classification issue. In this paper, an effectual SARS-CoV2model is developed using optimized Deep learning model through time series data. The derived features are derived from the input sequential data for disease forecasting. Moreover, over sampling scheme is exploited for data augmentation, which enhances the prediction process. Fuzzy systems and various distance measures are calculated for choosing most significant features. The Deep Recurrent Neural network (DRNN) is applied for performing SARS-CoV2prediction, in which DRNN is trained through designed Fractional Water Poor and Rich Optimization (FrWPRO) method. Meanwhile, the training process of DRNN using hybrid optimization model from scratch proves that, the designed SARS-CoV2prediction method accomplishes better performance compared to other existing approaches with Mean Square Error (MSE), Root MSE (RMSE), and Mean Absolute Percentage Error (MAPE) of 0.1425, and 0.3775, and 0.3467 respectively.

MLOps, short for Machine Learning Operations, is a practice that aims to streamline and automate the process of deploying, monitoring, and managing machine learning models in production. In the context of educational technology, MLOps can help optimize the performance of learning algorithms, ensure scalability and reliability. By implementing MLOps, educators can utilize real-time data to identify patterns of behavior that may indicate a student is struggling. This proactive approach allows timely interventions to be put in place, addressing issues before they escalate and potentially lead to academic failure. Additionally, MLOps can also help educators personalize learning experiences for students, catering to their individual needs and preferences. The participants were 60 learners enrolled in the Ready-Made Garment Manufacturing Technologies course, part of the Fashion Manufacturing Technology specialization in the Faculty of Human Sciences and Design at King Abdulaziz University. The findings of research found that integration of MLOps in educational technology has the potential to support and guide students in their learning through detecting undesirable student behaviors and adjusting early.

  Sign language is a vital communication mean for hearing-impaired individuals, combining manual gestures with non-manual signs like facial expressions and body movements, often requiring both hands and sequential actions. Recently, an automatic Sign Language Recognition (SLR) has gained increasing attention, with Machine Learning and Deep Learning systems achieving competitive performance. While convolutional neural network has been widely employed owing to their effectiveness in image-based recognition tasks, existing methods, however, often struggle with efficiency, adaptability, and real-time deployment. This paper proposes an Internet of Things-Integrated Deep Learning Model for Real-Time SLR to enhance the communication among individuals with hearing-impairment and non-signers. The framework employs IoT-based wearable sensors for capturing hand and finger movements, followed by Sobel filtering for noise reduction. MobileNetV3 is applied for lightweight feature extraction, while a Variational AutoEncoder enables robust sign detection. To further improve performance, an Improved Sparrow Search Algorithm is introduced for hyperparameter tuning, constituting the novelty of this work. Experimental results show that the proposed framework achieves an outstanding accuracy of 99.05% when compared to state-of-the-art systems, validating its robustness and effectiveness for real-time SLR applications

This study proposes an Intelligent Tutoring System (ITS) to enhance hand-knitting skills among Home Economics students through AI-driven personalized learning, addressing the limitations of traditional generic methods. The system integrates computer vision, adaptive algorithms, and interactive tutorials to provide real-time feedback and track progress. A study involving 60 students (30 control, 30 experimental) showed the ITS group achieved significantly higher post-test scores, confirming improved proficiency and engagement. Results reveal that the IT IS effectively accelerates skill acquisition and deepens understanding compared to conventional instruction.

Recently, irrigation management has been considered one of the most significant areas of research in smart vertical farming. Hence, it is essential to optimize freshwater usage for smart vertical farming management due to the lack of freshwater sources. It is observed that the soil moisture level and temperature data need to be appropriately examined and analyzed to predict the water irrigation level in a smart farming platform. Hence, in this work, the Internet of Things (IoT) sensors have been utilized to collect and monitor the soil moisture level, ambient temperature level, and humidity level data effectively. Besides, the collected sensor information has been analyzed and predicted to recognize the appropriate utilization of the optimum level of freshwater using Grey Wolf optimizer integrated recurrent network models. Therefore, this approach successfully analyzes the sensors' data and predicts the required level of irrigation based on motor ON and OFF conditions. The generated data from the sensor has been evaluated using the Keras model using the python language, and the performance is assessed based on the accuracy ratio. This model obtained a maximum of (0.995%) accuracy in forecasting the optimum irrigation level. The proposed system will utilize less voltage to minimize the power consumption ratio up to 35% in the irrigation process with 99.5% accuracy in forecasting the optimum irrigation level.

Health reconnaissance frameworks are currently a more significant issue and examination subject. A few applications, like military, home consideration, medical clinic, athletic preparation, and the crisis control framework, have been laid out for wellbeing observation research. Competitors' lives require a lot of activity and exercise for wellness and wellbeing. The capacity to screen the imperative indications of the competitor that mirror the physical and physiological state of the individual, particularly during an apprenticeship, is fundamental both for the competitor and for the mentor to keep away from overtraining, wounds, and sickness or to change the power and time as per the information estimated — wearable checking gadgets associated with remote correspondence advances. In the model, utilizing remote innovations implies that devices utilized by competitors discuss information with other remote hubs progressively and make a small correspondence organization. The utilization of remote sensor correspondence and the need to impart between sensors has prompted the formation of wireless sensor networks (WSN) and wireless body area networks (WBANs). This paper presented a wireless sensor network-based athlete health monitoring (WSN-AHM) method and concentrated on their growth phases. Since it is a remote and versatile wellbeing reconnaissance arrangement, it can give medical care specialist organizations a valuable remote checking device to diminish the expense of their administrations. WSNs and their correspondence advancements and principles can be utilized in these reconnaissance applications, accentuating wearing exercises through the entire and relative show of realities on well-known correspondence conventions.

Floods are among the most devastating natural disasters, causing widespread damage to infrastructure, homes, and human lives. Rapid assessment of flood severity is critical for effective disaster response and resource allocation. This study explores several deep learning approaches for flood water level classification using UAV imagery. A curated dataset of 2,000 UAV images from diverse regions, including India, the United States, and Brazil, was developed and augmented to improve generalization. Multiple architectures were evaluated, including pre-trained CNNs, ResNet50v2, MobileNetv2, Vision Transformers, and Swin Transformers, with and without the Convolutional Block Attention Module (CBAM) and adaptive learning strategies. Experimental results reveal that integrating Vision Transformers with CBAM achieves the highest classification accuracy of 90.6%, while a hybrid CNN–Vision Transformer model further improves performance to 92.3%. These findings highlight the potential of attention-based hybrid models for precise flood severity mapping. The proposed framework can aid rescue teams and disaster management authorities by prioritizing high-risk areas, enabling faster response and optimized allocation of resources during emergency operations.

In recent years, EEG based recognition and characterization of brain states has received much interest due to the advances in deep learning and machine learning methods. The non-invasive and highly inexpensive activity of EEG presents a patient with details concerning the activity and the conditions of the brain. The synthesis of artificial intelligence (AI) models (convolutional neural networks (CNNs), long short-term memory networks (LSTMs), and collaborative knowledge options has been explored in a series of studies that recognize the mental state accurately in a large number of cases. The literature focuses on introducing strong, explainable models as well as on multimodal data to boost classification accurateness and reliability. The results are a 1D CNN and a LSTM network were trained separately and in a hybrid, architecture (CNN-LSTM) to classify the EEG signals. The models were appraised using accurateness, accuracy, recollection, F1-score, and confusion matrix analysis.

Forecasting the energy consumption of heating, ventilation, and air conditioning (HVAC) chillers is vital for enhancing building efficiency, reducing operating costs, and supporting sustainability goals. However, the task remains challenging due to nonlinear system dynamics, strong dependence on weather conditions, and the scarcity of high-quality real-world datasets. In this work, we employ the Chiller Energy Data from Kaggle, which contains 13,561 cleaned records collected between August 2019 and June 2020, incorporating ten operational and meteorological features. Six baseline models, namely the Evolutionary Attention-based Long Short-Term Memory (EALSTM), Bidirectional LSTM (BILSTM), standard LSTM, Gated Recurrent Unit (GRU), Temporal Convolutional Network (TCN), and Artificial Neural Network (ANN), are first benchmarked to assess their forecasting capability. To further improve predictive accuracy, we integrate EALSTM with ten meta-heuristic optimization algorithms, focusing on the Greylag Goose Optimization Algorithm (GGO) and comparing it with alternatives such as Harris Hawks Optimization (HHO), Artificial Physics Optimization (APO), Simulated Annealing Optimization (SAO), Grey Wolf Optimizer (GWO), and others. The optimized GGO+EALSTM framework achieves state-of-the-art performance with a mean squared error of 6.83×10−6 and an R2 value of 0.98, reflecting a 96% reduction in error relative to simple feedforward models and significant improvements over other recurrent networks and optimizer-enhanced variants. The main contributions of this study include a structured benchmarking of neural architectures for chiller forecasting, the first systematic comparison of ten meta-heuristic optimizers applied to deep learning in this domain, and a visualization-based error analysis that strengthens interpretability and supports practical deployment. These results establish optimization-enhanced EALSTM as a robust and generalizable framework for HVAC energy forecasting, paving the way toward more efficient, reliable, and sustainable building energy management.

This study addresses the challenge of smart-home energy forecasting across multiple appliances under varying temperature and seasonal regimes, aiming to improve demand planning and household energy efficiency. The analysis leverages a 100,000-row dataset from Kaggle, encompassing appliance type, time of consumption, outdoor temperature, season, and household size. The study benchmarks several recurrent neural network models, including Long Short-Term Memory (LSTM), Bidirectional LSTM (BiLSTM), Gated Recurrent Unit (GRU), and Bidirectional RNN (BiRNN), as well as a feedforward Artificial Neural Network (ANN). A novel enhancement, the Evolutionary Attention-based LSTM (EALSTM), is introduced, and its hyperparameters are optimized using the Greylag Goose Optimization (GGO) algorithm. The performance of GGO-optimized EALSTM is compared to other metaheuristics, such as Differential Evolution (DE), Genetic Algorithm (GA), Quantum-Inspired Optimization (QIO), JAYA, Bat Algorithm (BA), and Stochastic Fractal Search (SFS). The results indicate that GGO-optimized EALSTM outperforms all other models, achieving superior accuracy across multiple metrics, including MSE, RMSE, MAE, r, R2 , RRMSE, NSE, and WI. Key contributions of the paper include (i) the establishment of an appliance- and season-aware forecasting benchmark, (ii) a comprehensive optimizer comparison for EALSTM using GGO, and (iii) the provision of actionable visual analytics to enhance the understanding of energy demand patterns and model errors.

This paper addresses the challenge of predicting and analyzing electricity consumption patterns in Tetouan, Morocco, using time-series data. The dataset consists of 52,416 observations with 9 features, collected from the SCADA system of electricity consumption across three zones. The primary goal is to enhance forecasting accuracy and optimize prediction models through machine learning (ML) algorithms, including both timeseries models and advanced optimization techniques. We compare the performance of several baseline ML models, such as BiLSTM and Continuous Time Stochastic Modelling (CTSM), with their optimized versions, utilizing optimization algorithms like Greylag Goose Optimization (GGO), Bat Algorithm (BA), and Whale Optimization Algorithm (WOA). The results show that the optimized CTSM model, using GGO, achieved substantial improvements, including the lowest Mean Squared Error (MSE) of 7.09E-07 and the highest R² of 0.990, demonstrating superior accuracy and stability. The contributions of this work include (i) benchmarking various ML models for time-series forecasting, (ii) introducing the use of optimized CTSM with meta-heuristics, and (iii) evaluating model performance using a comprehensive set of statistical metrics.

The complex nature, non-linear dynamics, and inherent volatility of stock markets make it difficult to provide accurate predictions. Recent developments in the area have shown the efficiency of some machine learning methodologies in predicting financial stock prices. However, emerging markets, such as Iraq, face additional challenges due to the lack of fundamental data needed to support predictive analysis. In this study, we present a novel framework that focuses on overcoming this issue and predicting the next-day closing prices of the Iraq Stock Exchange (ISX) main index, using only available historical closing prices to engineer 12 technical indicators. The goal is to compensate for the lack of important Open, High, and Low prices data while improving prediction accuracy. We used four machine-learning algorithms in the form of Random Forest (RF), Support Vector Machine (SVM), Artificial Neural Network (ANN), and K-Nearest Neighbor (KNN), which were optimized using grid search hyperparameter tuning technique. The performance of the models was evaluated using Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), and Coefficient of Determination (R²). The comparison analysis resulted in the SVM with the linear kernel yielding the best performance (RMSE = 16.25, MAPE = 1.15, R² = 0.989), followed closely by the ANN (RMSE = 18.25), RF (RMSE = 26.76), then KNN (RMSE = 55.77). The current study introduces two main contributions: (1) the feasibility of using engineered features to achieve reliable predictions in markets with incomplete data, and (2) the critical role of using hyperparameter optimization to enhance models accuracy. The framework we propose provides a practical model for predicting stock prices in resource-constrained emerging markets.

Retinopathy of prematurity (ROP) remains the leading cause of blindness in children. The detection and treatment of this disease mainly depend on subjective evaluation of the features of retinal blood vessels. This method is not only time-consuming but also prone to errors. The increasing number of such cases demands an urgent need for automated models to improve the accuracy and efficiency of diagnosis and treatment. This paper presents a method for early detection of ROP using the Swin Transformer, a hierarchical vision transformer architecture. This work focuses solely on the screening stages for ROP, as documented between 2015 and 2020, based on a dataset composed of 3720 retinal images from preterm infants, kindly made available by the Al-Amal Eye Center located in Baghdad, Iraq. The proposed model achieved a classification accuracy of 98.67% on a clinical ROP dataset. The results highlight the importance of the most recent in-depth learning methods in enhancing early detection techniques, ultimately leading to improved clinical outcomes for at-risk infants.