Papers by Keyword: Machine Learning

Abstract: Designing a model that utilizes previously reported experimental data on graphene and metal oxide nanoparticle-based hybrids and nanocomposites to predict the gas sensor response can be a promising approach for developing innovative and effective gas sensors. In this work, experimental data were extracted from published reviews and research articles to build a dataset for training various machine learning (ML) models. The compiled dataset focuses on the rGO-SnO₂ nanohybrid-based chemiresistive sensor and includes features such as gas concentration (ppm), operating temperature (°C), sensor response (%), response time (s), and recovery time (s). The sensor response and gas concentration were considered as target variables, one at a time. Several machine learning models, such as random forest regression (RFR), support vector regression (SVR), gradient boosting regression (GBR), and extreme gradient boosting regression (XGBR), were employed to predict target variables. Prediction accuracy was evaluated using the coefficient of determination (R² score), root mean squared error (RMSE), and mean absolute error (MAE). Among all the models, the XGBR ML model achieved the best performance, with a maximum R² score (0.93) and minimum RMSE (0.52) and MAE (0.23) values when predicting gas concentration and a highest R² score of 0.99 with RMSE and MAE values of 7.97 and 5.92 when predicting sensor response as the target variable. This study demonstrates the application of machine learning for the rational design of rGO-SnO₂ nanohybrid-based NO₂ gas sensors, supporting their potential use in various applications such as indoor and outdoor monitoring and industrial gas leakage detection.

Abstract: Maintaining safe pipeline conditions is crucial to ensure sustainable and reliable transportation for energy and water. Pipelines are generally laid underground due to larger transport capacity, rapid construction speed, space restriction, and safety precautions. Nevertheless, they are prone to failures due to mechanical problems, extreme operation, and aggressive surrounding environmental conditions. The usage of machine learning methods to predict buried pipeline failures has risen recently due to its effectiveness in addressing the aforementioned problems. This paper reviews making predictions on different buried pipeline failures by adopting machine learning approaches, particularly artificial neural networks (ANN) and hybrid methods. It highlights the detail of the machine learning algorithms as well as the parameters that were used in the predictive models with concise elaboration. Findings show that the ANN method gives accurate failure prediction, while the hybrid method enhances the prediction accuracy. Nevertheless, there is no single absolute algorithm that can work best to solve all pipeline failures. Finding the most suitable machine learning algorithm for a specific pipeline failure will be a challenge to overcome. This review is expected to give more comprehension to industry players related to machine learning methods as a potential tool to solve various buried pipeline problems. Further, this review may prompt other interested researchers to further discover machine learning potentials and ways to increase its effectiveness.

Abstract: The increasing complexity of X-ray test data and the demand for precise, large-scale evaluations have exposed the limitations of manual flaw detection methods in Non-Destructive Testing (NDT). Traditional manual approaches, while widely used, are labor-intensive and prone to human error, often leading to inconsistencies and inefficiencies. An Artificial Intelligent driven (AI-driven) system has been developed to automate the analysis and evaluation of X-ray test data, improving both accuracy and efficiency. The system employs advanced image segmentation and classification algorithms, aligning with ISO standards to detect defects in welded structures. By reducing reliance on manual interpretation, the system enhances the reliability and speed of the evaluation process, while also allowing for expert oversight through manual corrections when necessary. Integration with a Laboratory Information Management System (LIMS) ensures streamlined data handling and traceability, minimizing human error in data recording. As the AI model processes more data, it continuously improves, adapting to evolving defect patterns and maintaining high performance. This paper details the system’s AI architecture, the methodology employed for X-ray image analysis, and the performance results from industrial applications, demonstrating how this technology addresses key challenges in NDT processes.

Abstract: Machine learning (ML) algorithms can improve and innovate the design of new, eco-friendly composite materials. Therefore, this study aims to forecast tensile strength for polyvinyl alcohol composite reinforced by crystalline nanocellulose (CNC) through ML regression algorithms. Moreover, 107 datapoints from the literature were used to train and test ML models. However, this dataset had missing values for the input variables considered, so an Iterative Imputation with an Extra Tree (ET) Regressor model as estimator was performed, which reached a determination coefficient of 0.88. This study implemented five ML algorithms to predict tensile strength: Adaptive Boosting, Extreme Gradient Boosting, Random Forest, ET, and Gradient Boosting (GB). Additionally, a hyperparameter optimization was carried out using the Random Search optimization technique, obtaining that the GB optimized model had the highest precision with a determination coefficient of 0.97. Moreover, it was determined that PVA hydrolysis degree, CNC percentage, and CNC diameter were the most important variables for the GB-optimized model through SHAP analysis.

Abstract: Aluminium based MMNCs have gained significant traction across various industries due to their superior stiffness, strength-to-weight ratio, and enhanced mechanical and tribological properties. Despite extensive research in this field, the application of ML techniques to predict the properties of these materials remains limited. Present work aims to predict the wear rate of A-MMNCs based on their chemical compositions. The nanocomposites were fabricated using ultrasonic assisted stir casting method and studied their wear results. Classification models achieved an accuracy of 0.92 with SVM, 0.95 with KNN, and 0.97 with ANN. Additionally, prediction models for wear rate yielded R² values of 0.8876 with linear regression and 0.9165 with ANN, with minimal MAE for the ANN model. Genetic algorithms were employed to optimize wear test parameters.

Abstract: This research uses a multi-domain technique to give a thorough analysis of mechanical gears health evaluation that includes time, frequency, and time-frequency signal analysis. The research seeks to discover patterns indicative of healthy, partially damaged, or fully damaged gear states using a variety of graphical representations, including time and frequency plots, the Short-Time Fourier Transform (STFT), and scalograms which are visual representations of the wavelet transform of a signal. Advanced machine learning models are used to improve diagnostic accuracy when manual identification of these trends becomes difficult. The goal is to achieve a validation accuracy greater than 70% a threshold selected based on prior studies indicating that this level ensures reliable fault detection for industrial applications while balancing computational constraints. The reliability and effectiveness of gear monitoring systems can be increased by integrating contemporary signal processing and machine learning approaches, as demonstrated by this research, which also advances the identification of gear faults. Among the conclusions are the outcomes of tests done to identify gear problems in which authors were able to train a model with more than 72% accuracy and able to propose Artificial Intelligence model for classification of faults in gears.

Abstract: There are a set of few among the 7billion+ people in the world with either hearing or speech impairment, their only means of communication is through the use of sign language. It is one of the most reliable methods of communicating with special needs people. A form of communication using visual patterns to express emotions and ideas helps bridge the gap for deaf individuals. However, when interacting with those who rely on spoken language, a communication barrier often arises. Currently, human interpreters are used to facilitate conversations between these groups, but this solution can be both costly and inconvenient. The necessity for developing a technology to aid the interpretation of sign language to the deaf community and to foster idea-sharing amongst all humans cannot be overemphasized. Much research has been carried out to acknowledge sign language using technology for most global languages. In this project, deep learning techniques were applied to develop a system for recognizing hand gestures in American Sign Language. A dataset was created using both two-dimensional and three-dimensional images of American gestures. To detect landmarks in these images, the MediaPipe framework was utilized. Additionally, Long Short-Term Memory (LSTM) networks were employed to improve gesture recognition by leveraging the temporal dependencies of hand movements. My work specifically focuses on recognizing contexts in sign language communication, enhancing the system's ability to understand not just individual gestures but also their meanings in different scenarios.

Abstract: The detonation cell size is an important parameter for evaluating gas detonations and developing techniques for explosion damage mitigation. The resource-intensive experiments on gas detonation are carried out under limited test conditions, often leading to interpolating, or extrapolating from available datasets. Existing detonation cell width prediction models using Machine Learning models either use chemical-based input parameter computed using Cantera or CHEMKIN II packages. Parameters defined by chemical mechanisms introduce both, model errors and parameter uncertainties that could affect the prediction accuracy. In the present paper, experimental gaseous detonation data for hydrogen-oxygen and hydrogen-air mixtures are statistically analyzed to establish features for detonation cell widths prediction using machine learning models. Machine learning models are trained and validated using experimental data available from literature and internal tests. SHapley Additive exPlanations method is used for feature impacts analysis on model predictions. For non-diluted mixtures, detonation cell widths prediction based on initial mixture composition, pressure, and temperature are made for hydrogen-air mixtures (mean absolute error of 0.02 mm) and for stoichiometric hydrogen-oxygen mixtures (mean absolute error of 0.16 mm) at an averaged 99% accuracy. The models’ performance is validated against most recent models and new datasets and the experimentally reported detonation cell width measurements uncertainties.

Abstract: Financial fraud remains a persistent challenge across various domains, particularly in public-sector financial operations, threatening the integrity and transparency of financial statements while eroding public trust. This highlights the need for continued advancement of fraud detection mechanisms to keep up with the ever-evolving fraud tactics. ML algorithms have proven to be one of the most successful methods for analysing large financial datasets to detect fraudulent patterns. This paper reviews the application of ML to detect fraud in financial transactions using ML-based algorithms, namely K-means, Support Vector Machine, Decision Trees, Naive Bayes, and Deep Learning, in fraud detection, analysing their use cases and effectiveness as reported in the literature. Additionally, the study experimentally compares the performance of a Convolutional Neural Network (CNN) model against a Logistic Regression model, with the CNN achieving an impressive 90% accuracy, outperforming Logistic Regression in fraud detection. The paper further investigates the financial features and indicators most relevant to fraud detection and explores the challenges and opportunities posed by large volumes of financial transactions. By addressing these areas, the study aims to provide insights into enhancing fraud detection mechanisms and strengthening the security and integrity of financial transactions in today's digital ecosystem, including government institutions.

Abstract: Accurate prediction of steel microstructure is critical for ensuring desirable mechanical properties in industrial applications. This research integrates metallurgical transformation models into a convolutional neural network (CNN) for the classification and quantitative analysis of steel microstructures, including ferrite, pearlite, bainite, and martensite. The model utilizes image-based grain structure recognition in combination with explicit mathematical relations, such as the Hall-Petch equation for yield strength, the Koistinen-Marburger equation for martensitic transformation, and the Avrami equation for ferrite and pearlite phase fraction prediction. By implementing these relations within a Python-based machine learning framework, the network not only classifies steel phases but also estimates grain size, transformation kinetics, and mechanical properties. The developed approach achieves an accuracy of over 90% in microstructure classification and enables real-time prediction of metallurgical characteristics from microstructure images. This research provides a new avenue for computational material science by integrating data-driven neural networks with fundamental physical models.