Respiratory ailments, especially pneumonia, demand advanced diagnostic tools for timely and more accurate detection. Addressing this critical need, we introduce LungConVT-Net, an innovative architecture that blend the strengths of Vision Transformer (ViT) and Convolutional Neural Networks (CNN) to delineate among three crucial lung conditions viz., Viral Pneumonia, Bacterial Pneumonia, and COVID-19, as well as normal lung manifestations. The proposed model leverages depthwise separable convolutions, optimizing computational efficiency without sacrificing spatial filtering. Additionally, we integrate the proposed Dynamic Hierarchical Multi-Head Attention Convolution (DH-MHAC) and Adaptive Multi-Granular Multi-Head Attention (AMG-MHA) modules. These modules bridge the self-attention mechanisms with convolutions and utilize non-overlapping patches, culminating in enhanced feature extraction, respectively. A strategically incorporated Multi-Layer Perceptron (MLP) block within the AMG-MHA refines the model's prowess in understanding intricate data patterns. The Gradient Connection Enhancers (GCE) capture both long-range and short-range feature dependencies, addressing potential challenges in gradient descent and promoting training stability. Experimental evaluations, spanning from bi-class to complex quad-class combinations, reveal the model's performance compared to the state-of-the-art models. The results unveil AUC scores consistently surpassing 99% in most bi-class scenarios and show strong performance in complex multi-class settings, with AUC scores exceeding 99% for Pneumonia, COVID-19, and Normal categories. Moreover, in the quad-class combination, our model achieves an AUC score of 98.19%, highlighting LungConVT-Net's effectiveness in advancing respiratory disease diagnostics. The complete implementation code is publicly available: Codebase github.
Reliable, explainable prediction of milling surface roughness (Ra) in Inconel 625 enables tighter control of a notoriously difficult-to-machine superalloy. This study introduces an end-to-end, sensor-fusion and explainable deep-learning pipeline that integrates cutting-force and tri-axial vibration measurements. Raw signals are first smoothed via Savitzky-Golay filtering and then decomposed using ICEEMDAN to isolate physically meaningful oscillatory modes. A Lyapunovbased sensitive-mode criterion retains only the most informative components, while Sequential Feature Selection further reduces redundancy and mitigates overfitting. The refined feature set drives bidirectional recurrent regressors-Bi-LSTM and Bi-GRU-capable of capturing temporal dependencies in forward and backward directions for accurate Ra estimation. Model transparency is ensured through SHAP-based attribution, which links predictions to specific force-vibration features and modal scales, clarifying how multiscale dynamics influence surface finish. Taken together, this transparent, high-fidelity framework supports process optimization, tool-path tuning, and adaptive control in Inconel 625 milling. Using a CCRD dataset, the approach delivers high accuracy across tool variants. For cutting tool T1, the best XAI-Bi-LSTM model achieved R2 = 95.72 %, RMSE = 0.027 mu m, |R95%| = 1.85% and MAE = 0.023 mu m; forT2, R2 = 90.78 %, RMSE = 0.038 mu m, |R95%| = 1.90% and MAE = 0.034 mu m. SHAP analysis highlights depth of cut, feed, and entropy/spectral-center features as dominant contributors, aligning with known machining physics. The dataset showed maximum Ra = 0.62 mu m, with T2 producing higher roughness than T1. The results indicate that the proposed interpretable pipeline maintains strong predictive performance while exposing process-relevant factors that can guide parameter selection and monitoring.
Emotion identification is important for human–computer interaction, and its applications include medical and customer service provision. In the past, emotion recognition systems have been based on one modality, such as text, image, audio, or video, each with advantages. However, these single-mode approaches are often less accurate due to their failure to fully capture the complexity of human emotions. Several limitations characterize the existing methods, such as isolated analysis per modality, loss of contextual information, and sub-optimal performance. Usually, these approaches cannot accurately detect emotions in real-life situations where multiple channels exist through which people express themselves in terms of experiencing feelings. To overcome this situation,a multi-modal emotional recognition using a cross-modal fusion (MER-CMF) system is proposed that makes different data sources work together to understand emotional states more completely and holistically. The fusion process synergizes the strengths of each modality, thus enabling the system to capture subtle nuances in emotional expressions. Consequently, the MER-CMF model enhances emotion recognition accuracy significantly compared to existing methods. Our experimental results show that it outperforms unimodal baselines in terms of accuracy and F1 score, achieving an impressive accuracy of 98.32
This paper presents a discrete delta-domain fractional-order PID (FOPID) controller, tuned using dynamic Particle Swarm Optimization (dPSO), for precise DC motor speed control. The proposed approach directly discretizes the FOPID controller in the delta domain, ensuring improved numerical stability and continuous-time-like performance even at fast sampling rates. The controller is optimized for time-domain criteria and implemented on an Atmega328P microcontroller, with comprehensive validation via both simulation and hardware-in-the-loop (HIL) experiments. Compared with the conventional z-domain FOPID and PID controllers, the delta-domain design achieved up to 60% reduction in overshoot and 40% improvement in settling time, while maintaining steady-state error below 1%. These results confirm the practical viability of advanced digital fractional-order control for real-time industrial applications.
In this communication, a single-layered (resistive ink-dielectric-metal based) wide-band frequency selective surface microwave absorber using a unique pattern of resistive ink is presented analytically and experimentally for the reduction of radar cross section (RCS) in stealth technology. As compared to multilayer designs, single layer designs are more easy to deploy in practical applications like stealth technology. Further in contrast to other options, resistive ink has numerous benefits, including consistent wideband absorption, uniform resistance distribution, low cost, and material flexibility. In this work four-fold symmetric topology based on resistive ink is imprinted using screen printing technology on a copper-backed FR-4 dielectric substrate to yield polarization insensitive response. Absorptivity over 90 % has achieved over the frequency range of 17.2–26.5 GHz, associated to a fractional bandwidth of 42.5 % for normal incidence of electromagnetic (EM) wave with two absorption peaks at 19.4 GHz and 22.9 GHz covering the K U and K bands. In the entire band, more than 20 dB monostatic RCS reduction is obtained with a maximum RCS reduction of 38 dB at 22.9 GHz. Current distribution on the upper and lower surface, along with the induced electric fields at absorption peaks, and equivalent circuits have presented to examine the absorption mechanism. The proposed design is evaluated for transverse electric (TE) and transverse magnetic (TM) waves under oblique incidence, and for varying polarization angles in TE and TM mode, demonstrating wide angular stability. A prototype has been fabricated, and the measured outcomes are compared and validated against the simulation results. The novelty of the proposed absorber lies in its distinctive single layer λ 1 /11 ( λ 1 corresponding to the lower absorbing frequency) thin topology made from resistive ink, which shows polarization insensitivity and wide angular stability. All the above-mentioned attributes along with 18 % fractional bandwidth of 20 dB RCS reduction make it commercially appropriate for RCS reduction in stealth applications.
