University Visvesvaraya College of Engineering

Multivariate Time Series (MTS) forecasting plays a pivotal role in domains such as finance, healthcare, energy management, climate prediction, and IoT systems, where timely and accurate forecasts are critical for decision-making. Traditional sequence modeling approaches, including RNNs, LSTMs, CNNs, and TCNs, often face challenges in capturing long-range dependencies, multivariate interactions, and efficient parallel computation, limiting their scalability and accuracy. Transformer models, leveraging self-attention mechanisms, parallelism, and flexible sequence modeling, have recently emerged as a powerful alternative. This survey aims to provide a comprehensive, structured overview of Transformer-based approaches for MTS forecasting, including early models, efficient attention mechanisms, decomposition-based architectures, patch-based methods, hybrid models, and large-scale pretrained foundation models.We present a comprehensive taxonomy of existing approaches, datasets, and evaluation metrics, highlighting their comparative strengths and limitations. Additionally, we discuss key challenges such as computational complexity, interpretability, data sparsity, and generalization, and outline future research directions, including spatiotemporal modeling, self-supervised learning, physics-informed designs, explainable architectures, and federated learning. The survey consolidates existing knowledge and provides guidance for researchers and practitioners to select, adapt, and innovate Transformer-based models for effective MTS forecasting.

This study reports the design, fabrication, and experimental investigation of 3D-printed honeycomb structures subjected to drop-weight impact testing. Hexagonal honeycomb structures were designed using CAD software and fabricated using Fused Deposition Modeling (FDM) with PLA (Polylactic Acid) material. The primary objectives were to analyze energy absorption capacity, deformation characteristics, and failure mechanisms under controlled dynamic impact loading. Drop-weight impact tests were conducted at three distinct heights (1.0 m, 1.5 m, and 2.0 m), with a striker mass of 63.6 kg, systematically evaluating different impact energy levels. Force-displacement behavior and post-impact failure modes were documented to assess structural response. Results demonstrated that the honeycomb structure absorbed 624 J, 936 J, and 1248 J of energy at 1.0 m, 1.5 m, and 2.0 m drop heights, respectively, with corresponding displacements of 28–32 mm, 33–39 mm, and 45–50 mm. Specific energy absorption (SEA) values increased from 9.81 J/kg to 19.62 J/kg with increasing impact height. Progressive cell wall buckling and densification were identified as primary deformation mechanisms. This work contributes to understanding the performance of polymer-based cellular structures under dynamic impact loading and provides a foundation for optimization in crashworthy applications in aerospace, automotive, and protective equipment sectors.

The rise in Electric Vehicle (EV) usage has significantly increased the need for high-power fast charging systems that must meet rigorous performance standards. This paper examines the performance of 50 kW grid-connected EV charger. The design of the charger utilizes a dual stage Vienna rectifier at the front end, combined with a bi-directional DC-DC stage, to achieve a high-power factor, minimize total harmonic distortion (THD), and maintain stable operation of the DC bus. Vienna rectifiers are commonly utilized in high-power electric vehicle chargers because of their excellent efficiency greater than 94% and nearly unity power factor. An analysis of 50 kW charger that employs a Vienna rectifier focuses on its output ripple, input signal distortion, and power input factor. Under standard operating conditions, the performance remains equable. However, once the battery charge status surpasses 80%, the notable decline in performance occurs. In this scenario, both ripple and THD increase, and the power factor strays from unity, potentially harming the battery State of Health (SOH) during constant-voltage charging. Based on research results, this paper quantified the implication of current ripple on conversion efficiency in Electric vehicle charger through experimental verification and the results communicate that the current ripples have important influence on EV chargers.

This paper describes a Post-Quantum Cryptography (PQC) framework that employs Multi-Key Homomorphic Encryption (MKHE) and blockchain to secure data sharing and computation in IoT based e-healthcare systems. The system provides privacy-preserving analytics on encrypted electronic medical records while being quantum resistant. The methodology comprises algorithmic formulations for secure data storage, sharing, and homomorphic computing, which have been validated through extensive simulations with Python-based cryptographic tools. Key performance measures such as system reaction time, end-to-end delay, calculation time, and energy consumption were studied for various IoT device counts and security settings. Compared to previous models (SEF-IoMT and $\mathrm{E}^{2} \text{SCEC}^{2}$), the suggested framework reduced reaction time by up to 28%, energy consumption by 23%, and end-to-end latency by 31%. These results demonstrate that the suggested technique provides good post-quantum security while being scalable and efficient in resource-constrained IoT environments.

The rising integration of IoT devices into various applications has amplified security risks, particularly in resource-limited setups with constrained computing capacity and storage. Standard intrusion detection approaches often prove inadequate for these contexts owing to their intensive resource needs. This work proposes a host-based intrusion detection system customized for IoT infrastructures, harnessing compact and adaptable machine learning strategies. The developed system examines diverse classifier groups, encompassing tree-oriented models like Decision Tree and Random Forest, boosting-oriented techniques such as LightGBM and XGBoost, support vector machines including LinearSVC, and linear-based classifiers encompassing Logistic Regression, Stochastic Gradient Descent, Passive-Aggressive methods, and Perceptron algorithms. To identify the optimal detection algorithms suited to varied IoT application environments, a multi-factor decision framework integrating TOPSIS, VIKOR, MOORA, and WASPAS ranking methodologies is utilized. Outcomes from trials reveal that linear classifiers and streamlined tree models deliver strong detection performance alongside brief training periods and lowered processing demands.The solution achieves a top accuracy rate of 98.7%, with training times under two seconds for the majority of algorithms, yielding roughly 15-20 times enhancement in efficiency relative to typical ensemble-driven intrusion detection systems.