Global Academy of Technology

Air quality degradation has emerged as a significant urban challenge in recent decades, with fine particulate matter (PM2.5) recognized as one of the most hazardous pollutants. PM2.5 arises from both primary sources, including vehicular emissions, biomass burning, and industrial activities, and from secondary particle formation via atmospheric reactions involving precursor gases such as SO₂, NOx, NH₃, and VOCs. Exposure to these pollutants is closely associated with elevated risks of respiratory and cardiovascular diseases. This research applies machine learning (ML) algorithms to forecast PM2.5 concentrations using hourly data from Bengaluru, a major metropolitan area in southern India. Multiple models—Extra Trees, k-Nearest Neighbors, Random Forests, Support Vector Machines, CatBoost, and Decision Trees—were developed and evaluated using R2, RMSE, MAE, and MSE as performance metrics. The ExtraTreesRegressor demonstrated superior predictive performance, achieving an R2 value of 0.92 and the lowest error among the models tested. Hyperparameter optimization and Explainable AI techniques, including LIME and SHAP, were employed to improve model interpretability and robustness. The findings indicate that ensemble ML models can deliver reliable short-term forecasts of PM2.5 concentrations, thereby supporting air quality professionals and policymakers in issuing timely health advisories and implementing effective emission control measures.

This paper explores a low-voltage, thermally stable and high PSRR bandgap reference circuit made with an op-amp in GPDK 45nm CMOS technology. Accurate voltage references are crucial for the biomedical LDOs, like those used in ECG and blood pressure monitoring devices. If the reference isn't stable or precise, the whole system's performance decreases. Our work brings together CTAT and PTAT voltage behaviors in a compact op-amp circuit. The main idea is to improve the accuracy without raising the quiescent current. Simulation results show a reference voltage close to 1.2 V, with only minor changes across the supply range. The circuit's temperature coefficient comes in at 10.4 ppm/℃, line regulation stays between 2.8 and 3.6 %/V, and PSRR hits 44 dB at 100 kHz. Also, these tests cover a wide temperature range from -40 ℃ up to 125 ℃ and supply voltages between 1V to 3V. This circuit stands up well in tough thermal conditions, keeps it cool when temperatures changes, and tackles supply noise efficiently. It's a solid option for future biomedical LDO systems and other applications.

The increasing deployment of Internet of Things (IoT) and embedded electronic systems has created a growing need for efficient and reliable anomaly detection mechanisms capable of operating under strict computational and energy constraints. Traditional cloud-based monitoring approaches often introduce latency, communication overhead, and energy inefficiencies, limiting their suitability for real-time embedded applications. To address these challenges, this paper proposes an energy-efficient Tiny Machine Learning (TinyML) framework for real-time anomaly detection in embedded electronic devices. The proposed system integrates a lightweight convolutional neural network architecture with model optimization techniques, including pruning and quantization, enabling efficient deployment on microcontroller-class hardware. The framework processes sensor data locally through a pipeline consisting of data acquisition, preprocessing, feature extraction, and TinyML-based inference, thereby reducing dependency on remote cloud infrastructure. Experimental evaluation using NASA C-MAPSS turbofan engine degradation dataset demonstrates that the proposed approach achieves 95% detection accuracy, outperforming conventional machine learning models such as Support Vector Machines, Random Forest, and standard Convolutional Neural Network (CNN) architectures. Furthermore, the optimized TinyML implementation significantly reduces inference latency and energy consumption, making it suitable for low-power embedded environments. The results highlight the potential of TinyML-enabled intelligent monitoring systems to enhance the reliability and efficiency of next-generation embedded electronics and IoT devices.

Thanks to the accessibility of machine learning (ML) techniques, financial technologies are growing in almost every facet of finance, including asset management (robot advisors), payments (blockchain currencies), and peer-to-peer lending. High accuracy is sometimes attained by machine learning models at the price of inadequate explainability. Furthermore, high-risk AI systems dependent on ML should be “trustworthy” and adhere to a collection of obligatory standards, including sustainability and fairness, as per the proposed legislation. As of right now, no standardized measures exist that can facilitate a thorough monitoring of the reliability of AI implementations in the financial industry. To close the lag, we provide a collection of collected statistical techniques depending on the Lorenz Zonoid approach which may be utilized for evaluating and tracking the reliability of an AI application over time. In particular, the techniques will assess Explainability, Accuracy, Fairness and Sustainability.