Thadomal Shahani Engineering College

Dried leaves are a highly regarded renewable resource and the primary source of cellulosic plant material. It is rumored that dried leaf fibers may enhance the strength of polymer laminates comparably to synthetic fibers. This study is distinctive as it used artificial neural network (ANN) methodology to investigate the influence of dried leaves fiber, alumina, copper, and silicon carbide reinforcement on the thermal, physical, and mechanical properties of epoxy, polylactic acid, and vinyl-ester polymers. The wet layup procedure supported by an ultrasonication bath was employed to fabricate these composites under ambient circumstances. The findings indicate that the dried leaves-alumina fillers enhanced the mechanical and thermal stability of all three polymers compared to the other samples. The Fourier-transform infrared (FTIR) spectra indicate that the fillers within the matrix form robust interfacial bonds, possibly due to the generation of novel hydroxyl functional groups. The thermogravimetric analysis indicated that the hybrid composites composed of dried leaves, silicon carbide filler, and polymer exhibited superior thermal stability. The findings were statistically significant at the 95

Urban traffic congestion is a challenging problem that arises because our cities grow and the speeding number of vehicles makes it inevitable to have efficient traffic flow prediction models. The classic machine learning methods, as well as the popular ARIMA model, have difficulties in targeting and capturing the complexity and irregularity of urban traffic. The focal point of cities currently is to counter the problems with traffic which have been existing for a long time. To solve this, deep learning models like RNN, LSTM, GRU, and Transformer-based systems are being used for their ability to model sequential dependencies. Another study that analyses and compares these models for traffic flow prediction using the METR-LA dataset where the data comes from 207 sensors and are repeated in various cycles was carried out. The findings of the experiments revealed that GRU was the best of all the models and had a lower RMSE (10.3398) and a strong R2 score (0.6369) as well. It implies that the model learns traffic patterns well. Followed by LSTM with RMSE (10.4586) and R2 (0.6394), RNN performed decently. The Transformer model which did not perform up to the mark with RMSE (11.6764) and R2 (0.5392) were the two key components that hinted the limitations of the model in this case in the urban traffic complexities. Further research should deal with hybrid systems and data fusion techniques to improve the accuracy of the prediction process, making ITS (Intelligent Transportation Systems) more efficient in managing congestion and optimizing urban mobility.

Consumer-grade smartwatches facilitate highly accessible physiological data collection during daily routines. They frequently struggle, though, to distinguish cognitive stress from physical exertion purely because of restrictive single-sensor configurations. Conversely, clinical-grade medical devices deliver outstanding accuracy but remain cost-prohibitive and physically cumbersome. Seeking an optimal balance across accuracy, affordability, and form factor, we designed a comprehensive software stack and edge computing framework deployed upon a highly reliable hardware platform. This system utilizes an ESP32 microcontroller integrated with five sensors: ECG, EMG, GSR, PPG, and Infrared Thermography. The node acquires continuous biological signals in real-time, applying digital filters to segment raw streams into discrete time windows for precise feature extraction. We evaluated four diverse machine learning models to accurately classify data into Rest, Mental Stress, and Physical Stress. While the XGBoost classifier demonstrated a marginal advantage in raw accuracy, the Random Forest algorithm provided an ideal compromise, delivering robust predictive stability while requiring virtually zero computational overhead. Consequently, we quantized this Random Forest model into a TinyML format and flashed it directly onto the ESP32 chip. The resulting wearable prototype successfully executes real-time stress classification autonomously, eliminating any ongoing dependency on remote cloud servers.

Diagnosing Attention-Deficit/Hyperactivity Disorder (ADHD) presents several challenges, especially due to its reliance on subjective judgment and observation of behavior. While these age-old practices remain in vogue, they don’t usually provide consistency or objectivity in early and correct identification. With increasing interest in physiological signals as digital biomarkers, this study examines the application of Heart Rate Variability data in supporting automated ADHD classification. The data used in this research integrated HRV based statistical features, cognitive performance scores, and psychological screening test scores of 81 subjects. After feature extraction and feature selection, 28 significant features were retained. Three deep learning architectures Recurrent Neural Network (RNN), Gated Recurrent Unit (GRU), and Long Short-Term Memory (LSTM) were used to compare their performance in identifying ADHD patterns from sequential physiological signals. Collectively, these results justify the use of ML-based classifiers in facilitating clinical decision-making in the ADHD diagnosis process. Moreover, given their compatibility with wearables, it could offer cost-effective scalable solutions for early detection and continuous monitoring of mental health.

Generation of retinal images and their classification persist to be difficult because of the limits in data availability, imbalance in classes and the inaccuracy in representing and preserving minute features of generated images. Traditional GAN models, though being able to generate retinal images up to an extent, still lacked feature consistency, accuracy in global reconstruction, visibility of minute details, matching the fidelity and diversity, making them unsuitable and unreliable for augmentation in medical imaging. To tackle these issues, this research proposes a unique Conditional Dual Generative Adversarial Network, called CoFi-RetinaNet, which uses a hierarchical approach by combining a two-stage generator and ensures generation of retinal images which are not only realistic but also consistent, achieved by the strategic integration of structural production with pathological detail refining. First, the dataset is pre-processed, which includes resizing, normalization and structured augmentation, followed by training the model on the balanced dataset. Evaluation is carried out using performance metrics which are FID, IS, MMD and MOS scores, which clearly indicate that CoFi-RetinaNet performs significantly better than traditional models, achieving a test accuracy of 99.7%, demonstrating its robustness and enhancement in diagnostic results. The CoFi-RetinaNet model sets new benchmark of medical image generation, improving data quality and the accurate classification of retinal diseases.