Government Engineering College, Thrissur

Alzheimer's Disease (AD) is a progressive and irreversible neurodegenerative disorder that poses a significant global health burden. Early and accurate prediction of the conversion from Mild Cognitive Impairment (MCI) to AD is critical for timely intervention and personalized treatment planning. In this study, we introduce AlzFusionFormer, a novel transformer-based deep learning framework specifically designed to forecast the progression of MCI into either MCI non-converted (MCInc) or MCI converted (MCIc). Unlike traditional architectures, AlzFusionFormer employs a multi-branch transformer fusion architecture that simultaneously learns from multi-modal data-including neuropsychological assessments, volumetric features of Regions of Interest (ROIs), and 3D structural MRI scans. The framework is trained and validated using the ADNI-1 and ADNI-2 datasets, achieving validation accuracies of 81.62% and 88.67%, and prediction accuracies of 79.76% and 86.89%, respectively. Importantly, we integrate SHapley Additive exPlanations (SHAP) and self-attention map visualizations to provide interpretable AI insights, revealing the relative importance of imaging and cognitive features in conversion risk. This interpretability supports clinically meaningful decision-making, enhancing trust and transparency in model predictions. Our results demonstrate that AlzFusionFormer significantly outperforms existing uni-modal and multi-modal baselines, showcasing the potential of transformer-based architectures in the early detection of neurodegenerative progression through context-aware, modality-sensitive representation learning.

This article presents a three-phase, three-wire (3P-3W) renewable-based hybrid charging infrastructure that includes a photovoltaic (PV) system, wind-powered self-excited induction generator (SEIG), storage battery, sources to provide power to small consumer loads as well as incorporating AC & DC charging stations. The generated PV power is employed locally to increase the self-consumption rate, whereas the power generated from the wind is utilized to feed AC loads and electric vehicles (EVs) connected at the point of common injection (PCI). The harmonics introduced by the charging stations are suppressed using a modified filtering generalized integrator (MFGI) based control technique. The system is equipped with ancillary services, such as maintaining the power quality (PQ) of the isolated system frequent switching of EV loads and small consumer loads, undisrupted power to loads, and reactive power compensation. Validation of the proposed hybrid system is presented through a performance evaluation of the presented technique with an enhanced phase-locked loop (EPLL) and Notch filter technique. The results are plotted using MATLAB/Simulink and verified with license hybrid optimization of multiple energy resources (HOMER) version 1.2.7 under different operating circumstances. Despite the elevated total harmonic distortion (THD) of 24.49% in the load current, the MFGI control effectively mitigates the supply current THD to 4.0%, which effectively complies with the IEEE-519 standard.

The design of advanced driver-assistance systems (ADAS) is essential to improve the safety and autonomy of rear wheel driven four-wheel vehicle in harsh conditions. This work introduces the design and development of a steering automation system for Lane Keep Assistance (LKA) in an rear wheel driven four-wheel vehicle with a parallel steering system. The system utilizes an ArduCam module to take real time images of the ground in front, and these are processed via machine learning techniques on a Raspberry Pi in order to identify lane edges with great precision. The corrective steering maneuvers are carried out by a motorized steering actuator based on the visual data after processing, and an encoder that is built into the actuator constantly tracks the steering angle and position. This closed-loop feedback affords accurate, real-time corrections to ensure lane discipline without driver intervention. Extensive calculations for steering effort, torque, and gear design confirm the system's mechanical viability. Combining low-cost vision sensing, embedded machine learning, and encoder-based feedback control, this project presents a viable and flexible solution for Lane Keep Assistance in rear wheel driven fourwheel vehicle, paving the way for safer, more autonomous off-road driving.

Accurate lesion detection and analysis in medical imaging are vital for early diagnosis and effective treatment planning across a range of diseases. Despite advancements in artificial intelligence, existing frameworks often exhibit limited robustness and struggle to handle diverse lesion characteristics such as varying shapes, sizes, and textures. To address these gaps, this study introduces MedScope-LDx (Medical Scope for Lesion Detection and Analysis), a comprehensive framework leveraging advanced models and optimization techniques to deliver a seamless, high-precision workflow for automated lesion analysis. The workflow begins with robust preprocessing techniques, including Contrast-Limited Adaptive Histogram Equalization (CLAHE), Gaussian Filtering, and pixel intensity normalization, to enhance image quality and standardize the inputs. These steps ensure that the medical images are artifact-free and optimized for subsequent feature extraction. Feature extraction is achieved through Hybrid Attention DenseNet (HADN), which integrates spatial and channel attention mechanisms to prioritize lesion-relevant features, ensuring the precise and efficient representation of complex lesion characteristics. Segmentation is performed using the Segment Anchoring Model (SAM), a transformer-based architecture that tokenizes features into patches and applies self-attention mechanisms to accurately delineate lesion boundaries. For lesion detection, the Mamba Vision framework combines convolutional layers, Mamba Vision Mixers, and Transformer blocks to extract global and local patterns, thereby enabling an accurate lesion categorization. The Black-Winged Kite Algorithm (BKA), a novel nature-inspired optimization algorithm, is employed to fine-tune the parameters of preprocessing, feature extraction, segmentation, and detection, ensuring optimal performance and computational efficiency. Extensive evaluation of medical imaging datasets demonstrated the effectiveness of MedScope-LDx. On the lung CT dataset, the framework achieved an accuracy of 99.25 %, precision of 98.85 %, recall of 98.60 %, and F1 score of 98.90 %, with a Dice Similarity Coefficient (DSC) of 96.5 %. The kidney CT dataset delivered an accuracy of 99.6 %, precision of 99.4 %, recall of 99.2 %, F1 score of 99.3 %, and ((DSC) of 98.5 %. On the lung X-ray dataset, the framework achieved an accuracy of 98.80 %, demonstrating strong performance in this domain. Moreover, the inference time is significantly reduced to 55.8 ms for the lung dataset and 54.3 ms for the kidney dataset, highlighting the framework's computational efficiency. By addressing the challenges in lesion diversity and data variability, MedScope-LDx offers a scalable and clinically viable solution for early disease diagnosis and management, setting a new benchmark for automated medical image analysis.

Fires in Urban high-rise structures and industrial areas pose significant challenges to traditional firefighting methods. Traditional firefighting methods often struggle to address the challenges posed by height, accessibility and rapid response. In such a scenario innovative technologies become vital for effective and efficient methods. This project introduces an unmanned aerial vehicle designed to suppress fire on high-rise building by using drone technologies and robotics. The drone is equipped with a stereo camera which will detect fire and measure its coordinates with the help of algorithms fed on the companion computer raspberry pi. Upon receiving the coordinates, the drone will station itself at a predetermined distance from the fire. The drone will adjust itself in the vertical direction for proper ejection of water at the fire. The water will be ejected through a nozzle integrated with the drone, which is connected to the pump at the ground via hose. This drone solution provides a promising solution for fire in high-rise structures and industrial areas.