A hybrid Swin–DeiT transformer framework for automated welding defect detection in radiographic images M․Siva Ramkumar, S․K․Prasanna Lakshmi, Kiruthika Balakrishnan, Narmatha C, Rajendran T, Mohammad Arif, M. Sivaramkrishnan, D. Kavitha Computers and Electrical Engineering, 2026 • Swin Transformer backbone integrated with DeiT classification head. • Image-level split with 224 × 224 patches prevents data leakage. • SWRD dataset (3675 X-ray images, 7 defect classes) utilized. • Achieved 99.55% binary and 98.36% multi-class accuracy. • Ablation, cross-validation, and complexity analysis validate model. . Weld defect detection involves identifying issues such as pores, lack of fusion, and cracks in welded joints to ensure structures are safe and strong. In industries, using image classification to automatically detect these problems helps to reduce human errors, improve quality checks, and increase accuracy. It is necessary to improve safety, reliability, and productivity in automatic welding inspection systems. This research aims to build a smart and reliable system for detecting weld defects using image classification. This study uses deep learning models called Swin Transformer and DeiT (Data-efficient Image Transformer). The research uses the SWRD (Seam Weld for Defect Detection) dataset, which has images of major weld defect types. The aim is to obtain the best results in both multi-class and binary classification. To make the images clearer and better for training, preprocessing steps like image denoising and image enhancement are applied. Tests were done using the SWRD dataset for both multi-class and binary tasks. In binary classification, the model achieved 99.55% accuracy, 99.43% precision, 99.29% recall, and a 99.38% F1-score. In multi-class classification, it reached 98.36% accuracy, 98.4% precision, 98.3% recall, and a 98.3% F1-score. These results demonstrate the efficacy of the developed method for identifying common weld defects.
Chemical reaction and heat source effects on oscillatory suction in MHD flow through permeable media with Soret effect Lawanya T., Pragya Pandey, Sangeetha S., Kavitha D. Aircraft Engineering and Aerospace Technology, 2025 Purpose The current investigation is concerned with the Soret effect along with chemical reaction and radiation on flow of an electrically conductive, viscous fluid through a perpendicular plate, which is porous with oscillatory suction. The aim of this study is to investigate the effects of first-order temperature and chemical reaction and the transverse magnetic field characteristics. The closed form of solutions are obtained using the governing equations for concentration, energy and momentum. The perturbation technique was applied to find the result for the velocity field, temperature profiles and concentration distributions. Furthermore, the impact of various nondimensional parameters on fluid flow variables on the temperature field, velocity field and concentration dispersal was analyzed and the results were depicted graphically. Moreover, the skin friction and the rate of mass transfer (local Sherwood number) were analyzed using tables. In this work, an unsteady 2D flow of a laminar, viscid (Newtonian), electrically conducting fluid across a semi-infinite perpendicular permeable plate under motion in its plane (x-axis) embedded in a constant permeable structure was investigated. Design/methodology/approach In this work, an unstable 2D flow of a laminar, viscid (Newtonian), electrically conducting fluid across a semi-limitless perpendicular permeable plate under motion in its plane (x-axis) embedded in a constant permeable structure was investigated. The medium is considered to be under a transverse magnetic field with concentrated buoyancy effects. Furthermore, it is considered that no voltage is supplied, which indicates that there is no electrical field. The fluid properties are considered to be uniform. The concentration of the imparting species is considered as C′w at the plate; the concentration of the specimens away from the wall, C′8, is considered to be limitlessly less. The first-order chemical reaction is considered to be seen in the flow. Due to the semi-limitless plane surface considerations, the flow parameters are the functions of y′ and the time t′ only. The oscillatory suction velocity of the fluid at the plate normal to it is v′; initially, the plate relocates with the oscillatory velocity u′, in the direction of x that is in its plane. The pressure gradient is toward the x-axis. Findings The analytical solutions were obtained using the above analytical method for a few values of the governing parameters, such as the magnetic parameter (M), the permeability parameter (K), Schmidt number (Sc), chemical reaction parameter (Kr), Grashoff number for the concentration (Gm), Radiation parameter (N), Prandtl number (Pr), Chemical reaction parameter (Kr), Grashof number for heat transfer (Gr) and Heat source parameter (s). The influence of M, K, Sc, Kr, Gm, N, Pr, Kr, Gr and s on the fluid velocity, temperature and the concentration over the semi-infinite porous plate was obtained. Furthermore, the numerical computation was carried out using MATLAB. Originality/value In this chapter, the analysis of a free convective flow of a viscid compact, electrically conductive fluid was discussed during its flow through a plate in permeable condition with oscillatory suction with first-order temperature and chemical reaction and the transverse magnetic field. The problem formulation and the results were discussed. The following chapter explain the Soret effect of mass transfer and radiation with heat source on magnetohydrodynamics oscillatory viscoelastic fluid in a channel filled with porous medium.
E-Super edge magic graceful labeling of graphical image of trisomy ultrascan Mathematics in Engineering Science and Aerospace, 2025
Modelling the heat transfer of nanofluid towards a radiating stretching sheet of varying thickness using thermal flux Pragya Pandey, Dhatchana Moorthy Kavitha, Thangaraju Lawany Archives of Thermodynamics, 2025 The present paper targets the flow of fluid with Fe3O4 particles as nanomaterial over a non-linear elongated sheet with changing width. The process holds vital importance in various industries like paper manufacturing, extrusion of dyes and filaments, atomic reactors and many more. Nanofluids depict special features which give them the potential to be also used in power engines, refrigerators, power plants as well as pharmaceutical processes. Hence, the presented model is designed to intensify the rate of heat transfer and to reduce energy wastage, and tailor for the optimal selection of parameters like conductivity as well as viscosity, which will improve the effectiveness of the heat transfer process. The main idea behind this investigation is to calculate the effect of electromagnetohydrodynamics, Biot number, Eckert number, radiation along with the absorption factor. In this paper, the flow is modelled by using Navier-Stokes equations which are customised to Prandtl boundary layer equations. The Adams-Bashforth predictor-corrector is used to obtain numerical solutions. The present study helps to potentially improve and achieve the desired quality of the stretching sheet. Moreover, a negligible amount of activation energy is required, finding an economical way to get suitable out-puts.
Empowering Learners and Educators: The Transformative Role of Conversational AI in Personalized Education and Student Support M. Ruba, G. Ashwin Prabhu, Jayashree Deka, D. Kavitha, S. Sathiya, S. Pavithra Enhancing Student Support and Learning Through Conversational AI, 2025 In today's evolving educational landscape, conversational AI is emerging as a powerful ally for both learners and educators. This chapter explores how tools like chatbots, voice assistants, and virtual tutors are making education more personalized, supportive, and accessible. By interacting in natural language, these AI systems offer real-time guidance, answer questions, and adapt to individual learning needs. For educators, they reduce workload by handling repetitive tasks and providing insights into student progress. The chapter also addresses important concerns around privacy, bias, and digital inclusion, and shares real-world examples of how conversational AI is creating more engaging and equitable learning environments.
Dynamic Path-Controllable Deep Unfolding Networks with Planet Optimization: A Scalable and Efficient Approach for Software Defect Detection Kevin N. Shah, Chandrasekhar Rao Katru, Umesh Kulkarni, Jay Gandhi, D. Kavitha, Ramya Maranan Proceedings of the 4th International Conference on Intelligent Computing Information and Control Systems Icoiics 2025, 2025 Software defect prediction (SDP) is an important building block in enhancing software reliability through the identification of faulty modules prior to their deployment. In spite of progress made in applying deep learning to SDP, current models are prone to static learning processes, higher computation costs, and poor scalability for application to intricate software environments. These are the reasons behind elevated error rates, increased processing times, and decreased responsiveness in dynamic defect detection applications. Mitigating these challenges, a new framework, PathOpt-DUN, is introduced, which combines a Dynamic Path-Controllable Deep Unfolding Network (DPDUN) with Planet Optimization (PO) for optimal and precise defect detection. The approach utilizes software defect repositories under NASA's Metrics Data Program (MDP), with primary attributes including Lines of Code (LOC) and Halstead complexity measures. Preprocessing is applied via Global-Local Depth Normalization (GLDN) to provide consistent scaling and equally distributed features. The DPDUN model adaptively modulates learning pathways during training in response to changing patterns of defects, whereas PO optimizes hyperparameters to achieve the best convergence rate and prediction accuracy. Such dynamic coupling successfully balances defect detection accuracy with computational complexity. Experimental results on NASAMDP datasets show that PathOpt-DUN reaches 99.9% accuracy, 0.1% error rate, 99.7% fault detection rate, and lower computational time compared with state-of-the-art baseline models. These findings validate the superiority of PathOpt-DUN in providing robust defect prediction with greater reliability, scalability, and lower computational overhead. By overcoming major limitations of traditional methods, PathOpt-DUN becomes a useful and versatile solution for enhancing software defect detection in large-scale and dynamic software systems.
Steam Players Count Forecasting with LSTM Tuned by Modified RSA Marija Markovic Blagojevic, D. Kavitha, Luka Jovanovic, Tamara Zivkovic, Miodrag Zivkovic, Nebojsa Bacanin 2025 IEEE Zooming Innovation in Consumer Technologies Conference Zinc 2025, 2025 Precise estimation of the number of Steam users is crucial for optimal resource allocation, well-informed strategic choices, and the capability to fully leverage a game's potential throughout its lifespan. From a commercial perspective, this aspect significantly influences revenue generation, player engagement, and long-term planning. From the players' perspective, it allows server stability by anticipating peak loads and avoiding potential congestion, more balanced player matchmaking and tracking game's popularity trends. This study explores the effectiveness of an LSTM model refined through an enhanced variation of the highly acclaimed reptile search optimization algorithm. The developed approach was subjected to benchmarking simulations alongside other robust optimization methods, with the experimental findings distinctly highlighting the superior efficiency of the proposed technique. Ultimately, the proposed approach achieved the lowest MSE of 0.030559 and highest R2 of 0.430862 among all methods evaluated side by side comparisons.
A Convolutional Neural Network Approach for Rice Leaf Disease Detection in India Using Deep Learning K. Kishore Kumar, S. Sasikala, G. Maheswari, Mihir Kumar B. Suthar, Pothumarthi Sridevi, D. Kavitha, T. Vengatesh Journal of Neonatal Surgery, 2025 Rice is a staple food crop in India, and its production is critical for food security. However, rice crops are susceptible to various diseases that can significantly reduce yield and quality. Early detection of these diseases is essential for effective management and mitigation. This paper proposes a Convolutional Neural Network (CNN)-based approach for the automated detection of rice leaf diseases using deep learning. The model is trained on a dataset of rice leaf images collected from different regions of India, encompassing healthy leaves and those affected by common diseases such as blast, brown spot, and bacterial leaf blight. The proposed CNN architecture achieves high accuracy in disease classification, demonstrating its potential as a tool for early disease detection in rice farming. The results highlight the effectiveness of deep learning in agricultural applications, particularly in resource-constrained settings like India.
Evolutionary Optimization of Dominating Set-Based Virtual Backbone Cluster Scheduling for Enhancing Energy Efficiency in Asymmetric Radio WSNs International Journal of Intelligent Systems and Applications in Engineering, 2024
Fekete szegö problem for some subclasses of multivalent non-bazilevič function using differential operator Journal of Computational Analysis and Applications, 2019
