Lakshmi Narain College of Technology

Abstract Accurate measurement of vehicle speeds is central to traffic safety, enforcement, and intelligent transportation systems, yet many established methods depend on intrusive pavement sensors, costly radar or light detection and ranging (LiDAR), or vision systems sensitive to environmental conditions. This study evaluates a low-cost, nonintrusive alternative that estimates vehicle speed from roadside pass-by noise using a Class 1 sound-level meter with a free-field microphone. A total of 2,607 pass-by events were analyzed using a physics-guided Doppler ridge framework, refined through calibration and machine learning. On the test subset, calibrated acoustic features achieved mean absolute errors (MAE) of 8.29 km / h , root-mean squared error (RMSE) of 11.39 km / h , and R 2 as 0.47, and trucks exhibited the best performance with MAE of 5.44 km / h with 87.50% of cases within ± 10 km / h of actual speed. A portion of the error can be attributed to vehicle speed being measured when it was at some upstream location. Vehicles can accelerate or decelerate before reaching the closest distance from the microphone. These results demonstrate that acoustic Doppler-based features enable reliable, scalable, and nonintrusive speed estimation, offering a practical dual use for existing noise monitoring infrastructure. This approach provides primary low-cost speed screening in parallel with noise pollution assessment. More sophisticated state-of-the-art instruments may be triggered in real-time using this information for confirmatory measurement.

Partial shading reduces solar PV array performance by causing mismatch losses and multiple power peaks, making conventional MPPT methods ineffective in tracking the global maximum power point.*. This paper proposes an intelligent hybrid control framework that integrates LSTM neural networks with ACO to dynamically reconfigure PV arrays for enhanced power generation. The LSTM model forecasts short-term irradiance and shading patterns using time-series sensor data, while the ACO algorithm determines the optimal series–parallel configuration of PV modules under predicted conditions. A real-time switching matrix implements the selected configuration, and a closed-loop feedback mechanism continuously updates both the LSTM model and ACO pheromone trails based on actual system performance. MATLAB-based simulation results demonstrate high prediction accuracy (R2 = 0.992), low error metrics (MAPE = 2.54%), and improved power extraction under partial shading, validating the effectiveness of the proposed mixture approach.

There are numerous concerns related to the extensive use of cloud computing, including performance issues with communication- and computation-intensive services on virtual cloud resources. The present study explores the utilization of a Deep Neural Network (DNN) model to predict and optimize SaaS instance configurations on the basis of analyzing a dataset on Kaggle, which harbors Amazon AWS SaaS Sales information. After the data preparation process, such as filling in missing values, handling outliers, and normalizing features, to guarantee the validity of the results, the data is separated into training, validation, and test sets. The DNN model is then trained and tested using a number of different performance measures: accuracy, precision, recall, F-measure, and loss. The model is found to be not only excellent but also very accurate, reaching an accuracy percentage of 96.33%, and a strong balance between precision and recall is maintained. The learning curves suggest that overfitting was not a factor during the training process, which was smooth and stable, while the confusion matrix supports the model's reliability. The DNN has consistently outperformed other common methods, including KNN, LGBM, ARIMA, AdaPCA, and SV, in different datasets. To sum up, the insights obtained from this study point to the robustness and the capability of the model in processing SaaS sales data as well as in optimizing cloud resource configurations.

This study experimentally characterizes Myrtus communis leaves and integrates those data into a computational model to improve indirect solar drying performance. Sorption isotherms were measured by the static gravimetric method at 28, 40, and 50 degrees C; GAB and the medicinal-plant-tailored LESPAM models best described adsorption/ desorption (maximum R 2 up to 0.9989 and reduced chi 2 as 6.3 & times; 10 -4 ). Thin-layer drying experiments in an indirect solar dryer (collector heat flux 700 W/m 2 ; inlet air 0.8 m/s) were fitted with several kinetic models: the Midilli model performed best (R 2 = 0.9944; reduced chi 2 = 1.79 & times; 10 -4 ). Experimentally, leaves reached marketable dryness in about 6 h under Constantine (Algeria) conditions. COMSOL Multiphysics CFD simulations, driven by measured sorption and kinetic parameters, reproduced intra-chamber maxima of 319 K and air speeds up to 0.83 m/s, and captured humidity evolution showing negligible vapor flux after 6 h. Non-uniform airflow is identified as the main cause of drying heterogeneity, and design/operational changes (e.g., adjusted inlet velocity and tray layout) are recommended to improve uniformity. As result, ensuring uniform airflow distribution helps in minimizing the experimental effort and reduce parameter dependency, hence simplify the overall process for the benefit of general use in industrial applications. The combined experimental-computational approach enables realistic performance prediction. Proposed design modifications has been studied, which leads to optimize the airflow distribution, and as result, reducing the heterogeneity inside the solar dryers has been achieved.

The rise of fake news online across various platforms (including social media) has made it crucial to develop automated systems that can accurately detect and classify false information as it spreads through the internet. However, traditional methods of identifying false information often use pre-existing features or employ single deep learning models, which limits their ability to contextualise the data they are analyzing, their ability to use additional metadata associated with the data, and their overall ability to be robust to class imbalance in a classification problem. To address these limitations, this paper provides a hybrid contextual and metadata-aware framework for detecting multi-class fake news from the LIAR dataset. The goal of this study is to combine different pieces of information (statement, subject, and context) related to the speaker with their credibility (through the use of their metadata). The proposed methodology consists of using DistilBERT to extract contextual embeddings, extract metadata features through feature engineering, learn feature representations through deep learning, and finally using XGBoost for classification. This hybrid framework achieved 0.84 accuracy, precision, recall, and F1 score compared to the baseline two-stage classifier which achieved an accuracy of 0.54 and F1 score of 0.44. The novelty of this approach is that it incorporates a combination of semantic, contextual, and credibility features into an ensemble learning framework. This research suggests that there is increased robustness in classification with a greater ability to handle multi-class false information detection.