The "INDIAai" National AI Portal of India offers weekly articles that showcase the AI research contributions of universities across the country. Each article provides an in-depth report on the work of a specific university, allowing researchers and students to present concise explanations of their research. 

This week, the portal highlights the top AI research contributions from Prasad V. Potluri Siddhartha Institute of Technology, Vijayawada.

Prasad V. Potluri Siddhartha Institute of Technology (PVPSIT), established in 1998, is a leading self-financed institution known for its academic excellence and discipline. Sponsored by the Siddhartha Academy of General and Technical Education, which operates 18 educational institutions, PVPSIT offers a range of undergraduate (B.Tech) programs in fields such as Computer Science Engineering, Artificial Intelligence, Data Science, and others, along with postgraduate (M.Tech) programs in ECE, ME, and an MBA.

The college is autonomous, AICTE-approved, permanently affiliated with JNTUK, and accredited by NAAC with an A+ grade. All undergraduate programs are accredited by the NBA, and the institution holds ISO 9001-2015 certification, UGC 2f/12B status, and an 'A' grade from the Government of Andhra Pradesh. PVPSIT is also ranked in the 101-150 band in the NIRF Innovation stream.

The college emphasizes research, innovation, and entrepreneurship, offering a robust placement cell with a highest package of 44 lakhs and MOUs with 60 reputed industries. Its curriculum is industry-aligned, featuring a Choice Based Credit System from PVP20 regulations, allowing students to pursue honors and minors. The alumni network actively contributes to the institution, providing insights into cutting-edge technologies and career opportunities.

Military object detection in defense using multi-level capsule networks

Authors:

• B. Janakiramaiah, Prasad V. Potluri Siddhartha Institute of Technology, Vijayawada, Andhra Pradesh, India 
• G. Kalyani, Velagapudi Ramakrishna Siddhartha Engineering College, Vijayawada, Andhra Pradesh, India 
• A. Karuna, University College of Engineering Kakinada, Jawaharlal Nehru Technological University Kakinada, Kakinada, Andhra Pradesh, India 
• L. V. Narasimha Prasad, Institute of Aeronautical Engineering, Hyderabad, Telangana, India 
• M. Krishna, Sir C R Reddy College of Engineering, Eluru, Andhra Pradesh, India 

Automatic target detection is crucial in military operations, relying on the ability to recognize military objects from captured images. Traditionally, convolutional neural networks (CNNs) have been employed for object recognition; however, CNNs face challenges related to location invariance and performance dependency on large training datasets. In military contexts, where available training data is often limited due to operational and security constraints, these limitations can significantly hinder CNN performance.

To address these challenges, this research introduces a novel neural network architecture based on capsule networks (CapsNet), specifically a multi-level CapsNet framework designed for efficient military object recognition with small training datasets. The framework was validated using a dataset of military objects collected from the Internet, consisting of five distinct military objects alongside similar civilian counterparts. Experimental results demonstrate that the proposed multi-level CapsNet framework achieves high recognition accuracy, surpassing the performance of conventional algorithms such as support vector machines and transfer learning-based CNNs. This approach offers a promising solution for enhancing military object recognition in scenarios with limited training data.

Intelligent system for leaf disease detection using capsule networks for horticulture

Authors:

• B. Janakiramaiah, Prasad V. Potluri Siddhartha Institute of Technology, Vijayawada, Andhra Pradesh, India 
• G. Kalyani, Velagapudi Ramakrishna Siddhartha Engineering College, Vijayawada, Andhra Pradesh, India 
• L. V. Narasimha Prasad, Institute of Aeronautical Engineering, Hyderabad, Telangana, India 
• A. Karuna, University College of Engineering Kakinada(A) Jawaharlal Nehru Technological University Kakinada, Kakinada, Andhra Pradesh, India 
• M. Krishna, Sir C R Reddy College of Engineering, Eluru, Andhra Pradesh, India 

Horticulture plays a vital role in the Indian economy, with mangoes being a significant crop. However, mango production is often affected by diseases such as anthracnose and powdery mildew, which are commonly caused by bacterial and fungal infections. These diseases can severely impact both the quality and quantity of mangoes. In recent years, deep learning architectures, particularly Convolutional Neural Networks (CNNs), have shown promise in detecting and classifying plant diseases. However, CNNs have limitations, particularly in rotational or spatial invariance.

To overcome these challenges, this research introduces a variant of Capsule Network (CapsNet) called Multi-level CapsNet, designed to accurately classify mango leaf diseases. The proposed architecture was validated on a dataset of mango leaves, including both healthy and infected samples, collected in a natural environment. The experimental results demonstrate that the Multi-level CapsNet model achieved a remarkable accuracy of 98.5% in classifying mango leaf diseases, outperforming other classification algorithms such as Support Vector Machines (SVM) and traditional CNNs. This work highlights the superior precision and robustness of the Multi-level CapsNet model in addressing the challenges of disease detection in horticultural crops.

A Novel Tunicate Swarm Algorithm for Optimal Integration of Renewable Distribution Generation in Electrical Distribution Networks Considering Extreme Load Growth 

Authors:

• Hemanth Sai Madupu, Department of Electrical and Electronics Engineering, Prasad V. Potluri Siddhartha Institute of Technology, Vijayawada, Andhra Pradesh, India 
• Padmanabha Raju Chinda, Department of Electrical and Electronics Engineering, Prasad V. Potluri Siddhartha Institute of Technology, Vijayawada, Andhra Pradesh, India 
• Sri Kumar Kotni, Department of Electrical and Electronics Engineering, University College of Engineering, JNT University, Kakinada, Andhra Pradesh, India 

In recent years, the integration of renewable-based distributed generation (DG) into electrical distribution systems has seen significant growth. DG is increasingly favored over conventional power systems for meeting power demands at load centers due to its efficiency and technical benefits, such as reduced distribution losses, improved voltage profiles, and enhanced voltage stability. Traditional load modeling in electrical distribution networks often relies on constant power (CP) load models, which are not realistic. To address this, realistic load modeling is achieved by representing an optimal mix of various load models, particularly in radial distribution systems where high R/X ratios make real power losses a significant concern.

This research introduces a tunicate swarm algorithm (TSA) for optimizing DG installation, considering various load models and load growth, with the goal of improving voltage profiles, reducing network losses, and enhancing voltage stability indices. The study focuses on minimizing real power losses to determine the optimal DG location and size. The proposed algorithm was tested on the IEEE 33-bus distribution system, with results demonstrating that TSA outperforms existing optimization algorithms from recent literature. The findings underscore the effectiveness of DG integration and the impact of different load models on the distribution system under load growth conditions.

Lightweight CNN Models for Product Defect Detection with Edge Computing in Manufacturing Industries

Authors:

• B. Janakiramaiah, Prasad V. Potluri Siddhartha Institute of Technology, Vijayawada, Andhra Pradesh, India 
• K Sai Sudheer, Tejas Networks, Bengaluru, Karnataka, India 
• L. V. Narasimha Prasad, Institute of Aeronautical Engineering, Hyderabad, Telangana, India 
• M. Krishna, Sir C R Reddy College of Engineering, Eluru, Andhra Pradesh, India 

Product defect detection is a critical aspect of quality control in the manufacturing industry. Traditionally, human visual inspection has been the standard method, but it is often labor-intensive, error-prone, and inconsistent. To address these challenges, deep learning-based Convolutional Neural Networks (CNNs) have been widely used to fully automate defect detection systems. However, the limited computing capabilities of real-time edge devices commonly used in manufacturing settings present challenges in running complex CNN models.

In response to this, the research introduces lightweight CNN models optimized for deployment on edge devices, striking a balance between defect detection accuracy, model training time, memory consumption, computing efficiency, and speed. The study applies transfer learning to lightweight CNN models for defect detection across fabric, surface, and casting datasets. These models were deployed on the NVIDIA Jetson Nano-kit edge device, demonstrating improved detection speed and robust performance metrics, including accuracy, sensitivity, specificity, and F1 measure, within the manufacturing industry context. This work highlights the effectiveness of lightweight CNN models in enabling efficient, real-time defect detection on resource-constrained edge devices.

An Interpretable Approach with Explainable AI for Heart Stroke Prediction

Authors:

• P Naga Srinivasu, Department of Computer Science and Engineering, Prasad V Potluri Siddhartha Institute of Technology, Vijayawada, India 
• Uddagiri Sirisha, Department of Computer Science and Engineering, Prasad V Potluri Siddhartha Institute of Technology, Vijayawada, India 
• Kotte Sandeep, Department of Information Technology, Dhanekula Institute of Engineering & Technology, Vijayawada, India 
• S. Phani Praveen, Department of Computer Science and Engineering, Prasad V Potluri Siddhartha Institute of Technology, Vijayawada, India 
• Lakshmana Phaneendra Maguluri, Department of Computer Science and Engineering, Koneru Lakshmaiah Education Foundation, Vaddeswaram, Guntur 522302, India 
• Thulasi Bikku, Computer Science and Engineering, Amrita School of Computing Amaravati, Amrita Vishwa Vidyapeetham, Amaravati, India 

Heart strokes present a major global health challenge, necessitating predictive models to aid in early detection and prevention. While numerous studies have employed machine learning (ML) and deep learning (DL) techniques for heart stroke prediction, many have struggled with the interpretability of these models in clinical settings, hindering their adoption by healthcare professionals. This research introduces an innovative, interpretable approach to heart stroke prediction using explainable AI techniques, designed to bridge this gap. Key contributions include a well-crafted model that incorporates resampling, data leakage prevention, and feature selection, all aimed at enhancing the model’s clarity and utility for clinicians.

Utilizing the Stroke Prediction Dataset, which encompasses 11 attributes, the research achieved a remarkable 95% accuracy with an Artificial Neural Network (ANN) model. To ensure the model's comprehensibility, interpretability measures like permutation importance and LIME were employed. Permutation importance offers global insights into feature significance, while LIME provides local, instance-specific explanations. Together, these methods offer a robust framework for understanding both the overall model performance and individual predictions. This comprehensive, explainable approach holds significant promise for improving heart stroke prediction in healthcare.