This paper presents a 10-transistor (10T) SRAM-based compute-in-memory (CIM) architecture that enhances noise tolerance, reliability, and energy efficiency for accelerating binary deep neural networks (BNNs). The proposed single-ended 10T cell utilizes feedback cutting across cross-coupled inverters to enhance writability while maintaining read stability. Embedded XNOR logic enables efficient in-memory dot-product operations by reducing data movement and increasing throughput. The simulation is carried out in 45 nm CMOS technology using Cadence Virtuoso at 0.9 V. The architecture delivers static noise margins (SNM) between 140 mV and 220 mV, RSNM between 150 mV and 195 mV, and SVNM of 290 mV over Vdd levels of 750-950 mV. The write-tripping voltage (WTV) and write tripping current (WTI) are 560 mV and -187.71 & micro;A, respectively, with an SINM of 185 & micro;A. Compared to the most recent 8T,10T XNOR-CIM architectures, a TensorFlow-based inference study on MNIST and CIFAR-10 datasets with Gaussian noise (sigma = 0.3) shows negligible accuracy loss (6.79% and 1.10%, respectively), showing computational robustness. Furthermore, a 500-sample Monte Carlo analysis establishes stable flip-time distributions in Q/QB nodes, ensuring constant performance under varying conditions. The proposed 10T XNOR-SRAM CIM provides a scalable and stable foundation for energy-efficient in-memory computing, making it an appealing solution for edge AI and neural network inference systems.
This study proposes a Deep Learning (DL) framework based on a Multilayer Perceptron (MLP) to predict student performance in personalized English language learning environments. The dataset comprises academic features (test scores and previous assignment scores), behavioral features (engagement rate and time spent on the learning platform), and temporal statistical features, including rolling mean, standard deviation, and performance improvement rate. Data preprocessing techniques such as missing value imputation, normalization, and feature engineering are applied to ensure data quality and consistency. Recursive Feature Elimination (RFE) is employed to identify the most influential features, reducing training complexity while improving predictive performance. The MLP model is trained for 100 epochs and evaluated using standard regression metrics. Experimental results demonstrate excellent predictive accuracy, achieving a Mean Absolute Error (MAE) of 0.0937, Root Mean Square Error (RMSE) of 0.6542, and an R & sup2; score of 0.9916. Residual analysis indicates errors are centered around zero with no significant systematic bias, while close convergence between training and validation errors suggests strong generalization and minimal overfitting. Furthermore, actual-versus-predicted comparisons show a high degree of agreement, confirming the model's reliability. The proposed framework enables timely identification of at-risk learners and supports personalized educational interventions. The novelty of this research lies in integrating academic, behavioral, and temporal features within an MLP-based decision-support framework for student performance prediction in language education.
Hydroponic cultivation demands intelligent and adaptive regulation of nutrient chemistry and environmental parameters; however, existing IoT-based monitoring frameworks often struggle to accurately interpret irregular, noisy, and highly interdependent sensor streams, limiting autonomous decision-making and stable crop development. To address these challenges, this work proposes HydroFormer, a unified deep learning-driven hydroponic intelligence framework that integrates continuous-time dynamic modeling, graph-aware feature learning, transformer-based predictive analytics, and cooperative optimization for autonomous hydroponic management. The proposed framework introduces a Graph Neural Ordinary Differential Equation architecture capable of modeling nonlinear temporal interactions among pH, TDS, temperature, humidity, and water-level variables while preserving continuous system dynamics and physicochemical dependencies. To enhance representation quality, Echidna Optimization selectively refines latent features by eliminating redundant and weak system descriptors, thereby improving predictive robustness and computational efficiency. Furthermore, HydroFormer employs a cross-patch attention transformer mechanism to simultaneously perform hydroponic state classification, multivariate growth prediction, anomaly recognition, and intelligent actuator inference through joint short-term and long-range dependency learning. In addition, Mandarin Duck Cooperative Optimization adaptively tunes the network and training hyperparameters to achieve globally optimized learning behavior under complex IoT sensing conditions. Experimental analysis demonstrates that the proposed framework achieves superior hydroponic intelligence performance with 99.27% accuracy, 99.41% precision, 99.18% recall, and 99.29% F1-score, alongside highly reliable continuous predictions with a mean absolute error of 0.021, root mean square error of 0.042, and R2 value of 0.996.
Wireless Sensor Networks (WSNs) are widely used to monitor and collect data in environments with limited human access. Despite their potential, WSNs face persistent issues such as limited battery power, inefficient routing, and unreliable data caused by sensors or transmission faults. These challenges reduce the network lifetime and degrade the quality of the collected information. To overcome these limitations, this work introduces an integrated optimization framework that improves clustering, routing, and fault detection. The proposed approach adopts a density-driven clustering mechanism to form energy-balanced clusters and employs a Shrike Optimization Algorithm (SHOA) to identify suitable cluster heads by considering residual energy, node distribution, and trust levels. Routing paths are further refined using a Catch Fish Optimization Algorithm (CFOA), which dynamically selects energy-efficient and stable multihop paths. To ensure data reliability, a graph sample and aggregate attention network (GSAAN) was implemented to detect and isolate faulty data in real time. Simulation results demonstrated that the proposed framework consistently reduced energy consumption, extended network lifetime, and enhanced reliability compared with benchmark algorithms such as Mod PSO, CHBCO, IABC-C, HPO-WPBFT, and ASSO-SSO. Notably, it achieves a packet delivery ratio of 99%, fault detection accuracy of 98.7%, average delay of 155 ms, and network lifetime of up to 3400 rounds, demonstrating its effectiveness in critical WSN applications, including environmental monitoring and smart city deployments.
