Abstract Low cost Micro Electromechanical Systems (MEMS) sensors are widely deployed in embedded systems, the Internet of Things (IoT), and cyber physical environments; however, their raw output under nominally static conditions may exhibit substantial temporal, spectral, distributional, and... Low cost Micro Electromechanical Systems (MEMS) sensors are widely deployed in embedded systems, the Internet of Things (IoT), and cyber physical environments; however, their raw output under nominally static conditions may exhibit substantial temporal, spectral, distributional, and statistical complexity that cannot be captured by a single analytical representation. This study proposes and applies an integrated multidomain signal processing and unsupervised anomaly detection framework for the comprehensive characterization of multivariate static MEMS accelerometer data. The framework combines time domain statistics, Fast Fourier Transform (FFT) and Welch power spectral density (PSD) analysis, short time Fourier transform spectrograms, wavelet domain energy decomposition, autocorrelation and rolling root mean square (RMS) analysis, Pearson inter channel correlation, signal to noise ratio (SNR) estimation, Shannon and spectral entropy measures, rolling statistical change point detection, and two complementary unsupervised anomaly detection methods Isolation Forest and PCA reconstruction based detection. The framework is applied to a four channel dataset acquired from a stationary MPU-6050 accelerometer, comprising three raw acceleration axes and a derived percentage error channel, containing approximately 37,459 observations. The results reveal pronounced heterogeneity among the sensor channels: Sensors 1, 2, and 4 exhibit strongly non Gaussian, heavy tailed amplitude distributions with extreme skewness and excess kurtosis, whereas Sensor 3 displays substantially greater amplitude variability, strong temporal persistence, and concentrated spectral organization. Frequency domain analysis identifies a distinctive approximately 21 sample periodic component in Sensor 2, while time frequency and wavelet analyses expose localized transient events not captured by global spectral summaries. Dual entropy analysis demonstrates that amplitude domain and frequency domain complexity are distinct and independently informative properties. Change point detection identifies two statistical regime transitions, and both Isolation Forest and PCA reconstruction consistently identify 24 anomalous windows (anomaly rate H” 1.028%) in the absence of predefined labels. The integrated findings demonstrate that no single analytical domain is sufficient to characterize static MEMS sensor behavior, and that a multidomain framework is necessary to distinguish persistent nominal structure, transient anomalous deviations, and sustained statistical regime changes. The proposed framework provides a structured, reproducible, and label-free analytical foundation for sensor monitoring and anomaly detection in resource constrained and IoT-enabled systems. Read More Read Less

Abstract Handwritten digit recognition systems are often expected to perform reliably under imperfect input conditions, yet simple neural architectures are frequently evaluated only on clean benchmark data. This study systematically evaluates the robustness of a multilayer perceptron (MLP)... Handwritten digit recognition systems are often expected to perform reliably under imperfect input conditions, yet simple neural architectures are frequently evaluated only on clean benchmark data. This study systematically evaluates the robustness of a multilayer perceptron (MLP) trained on the MNIST handwritten digit dataset under four controlled perturbation families: additive Gaussian noise, salt and pepper impulse noise, contiguous pixel masking, and spatial translation. The trained MLP achieves a clean test accuracy of 92.75%. Perturbation analysis using accuracy, precision, recall, F1-score, confusion matrices, and accuracyversus- strength curves shows that the model is comparatively resilient to sparse and localized degradations, maintaining 86.40% accuracy under salt and pepper noise and 82.95% under pixel masking. Performance decreases more noticeably under Gaussian noise, with accuracy falling to 77.35%. In contrast, even small translations cause severe degradation, with accuracy dropping to 33.35% at a 2-pixel shift and below 5% at shifts of 4–5 pixels. Gradient-based saliency maps reveal that the network relies on localized stroke pixels at fixed spatial positions, explaining its pronounced vulnerability to translation. The findings demonstrate that MLP robustness is highly perturbation-specific and fundamentally constrained by the absence of spatial inductive biases such as weight sharing and local connectivity. The study provides practical guidance for deploying lightweight classifiers and highlights the importance of preprocessing or translation invariant architectures in real world handwriting recognition applications. Read More Read Less

Abstract The convergence of Internet of Things (IoT) technologies, artificial intelligence, and educational data analytics has generated considerable interest in personalised English language instruction within multicultural learning environments. However, the extent to which IoT-derived behavioural, linguistic, and engagement... The convergence of Internet of Things (IoT) technologies, artificial intelligence, and educational data analytics has generated considerable interest in personalised English language instruction within multicultural learning environments. However, the extent to which IoT-derived behavioural, linguistic, and engagement features can reliably predict overall learner performance remains insufficiently examined, particularly with respect to model transparency and pedagogical interpretability. This study conducts a comprehensive predictive analysis of Overall_Performance using the English Learning Interaction Dataset, comprising 15,000 learningsession records and 54 features collected from IoT-enabled smart classrooms at Chinese universities. Three ensemble tree-based regressors (Gradient Boosting, Random Forest, and Extra Trees) were trained to predict continuous performance outcomes, while Random Forest and Logistic Regression classifiers were applied to a dichotomised high-performance target (  75). Additionally, one-way analysis of variance (ANOVA) was employed to compare mean performance across six teaching strategies. Model interpretability was assessed using SHapley Additive exPlanations (SHAP), Partial Dependence Plots (PDP), and Individual Conditional Expectation (ICE) curves. Results revealed negligible predictive utility across all models: ensemble regressors yielded R² values at or below zero (best: Gradient Boosting, R² = - 0.0022; MAE  15.0; RMSE  17.3), and binary classifiers produced ROC-AUC values indistinguishable from chance (  0.50). One-way ANOVA detected no statistically significant difference in Overall_Performance among the six teaching strategies (F = 0.237, p = 0.946). Despite weak overall predictivity, SHAP and feature-importance analyses consistently identified Communication_Effectiveness, Pronunciation_Score, Assignment_Marks, Feedback_Score, and Grammar_Score as the relatively strongest predictors, while PDP/ICE plots revealed flat average marginal effects accompanied by substantial between learner heterogeneity. These findings indicate that the current tabular feature set is insufficient to predict overall performance with practical accuracy or to differentiate among pedagogical approaches. The study delivers a “Tabular Gradient Boosting” analysis and highlights the need for longitudinal modelling frameworks and alternative outcome metrics to realise the full potential of AI-driven personalised English education. Read More Read Less