Papers by Keyword: Machine Learning

Abstract: Air bending is a critical operation in the metalworking industry, where dimensional accuracy and process efficiency are essential to ensure product quality and economic viability. This work proposes an AI-driven design and optimization strategy which couples artificial intelligence, specifically artificial neural networks, with a quasi-random search algorithm for the metamodeling and optimization of the air bending process. An extensive simulation database was generated by varying geometrical, material, and process parameters, and neural-network-based metamodels were trained to predict the maximum punch force, maximum thickness reduction, and final bending angle, achieving high predictive accuracy with R² values exceeding 0.96. The metamodel was subsequently used to optimize process configurations by simultaneously minimizing the maximum punch force and the maximum thickness reduction while ensuring the target bending angle, leading on average to reductions of 46.7% in maximum force and 31.5% in thickness reduction compared to non-optimized cases. The results demonstrate that artificial intelligence provides an efficient and effective tool for the design and optimization of the bending process, significantly accelerating parameter selection while improving process quality and reducing manufacturing costs.

Abstract: The quality and dimensional accuracy of sheet metal components are strongly influenced by various sources of uncertainty, including variations in material properties, tool geometry and process parameters. Determining the specific source responsible for deviations in bending outcomes is usually costly and time-consuming, especially in industrial settings where numerous factors interact. In this study, a machine learning framework that can detect and quantify the impact of uncertainties in both air and bottom bending processes is presented. A dataset comprising forming results such as bending angles, final thickness and measured deviations, is used to train two neural networks metamodels (one for each process) that link input uncertainties to process outcomes. The predictive performance of these models was evaluated using different metrics achieving high predictive accuracy, with coefficients of determination close to 1 for most uncertainty sources in air bending and values above 0.95 for the majority of parameters in bottom bending. These results demonstrate the capability of the methodology to reliably identify dominant sources of uncertainty and support robust process optimization.

Abstract: This study proposes an automated framework for online cutting tool wear classification in CNC turning using low-cost optical equipment and Convolutional Neural Networks (CNNs). Longitudinal turning experiments were performed on CK45 medium carbon steel using a HAAS TL1 lathe under dry machining conditions. Tool wear evolution was monitored via a lathe-mounted digital microscope, with images classified into three distinct stages: Low (Vb<160 μm), Medium (160≤Vb≤200 μm), and Critical (Vb>200 μm). A shallow CNN architecture, consisting of three convolutional blocks and a Softmax output layer, was developed to balance model complexity with computational efficiency for potential edge deployment. To enhance robustness against positional changes, data augmentation techniques including random translations and rotations were applied. The results demonstrate good performance, with the model achieving 94.7% accuracy and a weighted F1-score of 95.4% on the testing subset. While the model showed exceptional performance in identifying Low and Medium wear, data scarcity in the Critical wear class remained a limiting factor for recall. Overall, the study confirms that shallow CNNs can accurately capture spatial hierarchies for image-based wear assessment.

Abstract: In modern precision manufacturing, optimizing complex processes like turn-milling is crucial for reducing production costs and ensuring high surface integrity. In this study the application of artificial intelligence, specifically machine learning (ML), for modeling turn-milling processes is investigated. The complexity of machining operations and the multitude of influencing input parameters often lead to time-consuming setups, particularly in single-part or small series manufacturing. Traditional process monitoring methods frequently fall short due to system complexity, prompting the exploration of ML for process optimization and automation. Focusing on orthogonal turn-milling, experimental data was collected to address regression problems such as tool wear and surface roughness, as well as tool condition classification. Three regression models - linear, polynomial, and support vector regression (SVR) - and four classification models - logistic regression, neural networks, support vector machines (SVM), and decision trees - were trained and validated using k-fold cross-validation. For regression models, root mean square error (RMSE) was used as the performance evaluation metric, while accuracy and F1-score were employed for classification problems. The results indicate that ML algorithms provide enhanced flexibility and accuracy compared to traditional statistical techniques, offering potential reductions in time and costs in process setups. By optimizing parameters iteratively, ML models demonstrate higher precision, reducing the need for extensive empirical research and the associated experimental costs. The developed models can be adapted to various manufacturing processes with minimal code adjustments, broadening their applicability and efficiency.

Abstract: The Guided Material Flow (GMF) process is an advanced variant of the Samanta process designed for the net shape cold extrusion of gears. The GMF process employs a modified die geometry to control material flow and significantly reduce maximum tool loads, effectively overcoming traditional process limitations. Key advantages include enhanced tooth tip strength and a reduction in face end deformations, which are characteristic defects in the conventional Samanta process. Minimising these deformations reduces the requirement for subsequent machining and enhances overall material efficiency. A numerical dataset was generated to train and validate data driven surrogate models, facilitating rapid process analysis without the computational cost of continuous Finite Element Analysis (FEA). The models developed in this paper enable the precise prediction of critical process outputs, including maximum punch force, die filling behaviour, material utilisation and strain hardening at the tooth tip. This paper details the numerical data acquisition, the specific training and validation methodologies of the machine learning models and demonstrates their capability to accurately predict complex process outcomes when varying the geometry of the die active surface in the GMF process.

Abstract: In Laser Powder Bed Fusion (LPBF), ensuring reproducible part quality remains a ma-jor challenge despite the availability of high-resolution in-situ monitoring systems, such as ExposureOptical Tomography (EOT). While EOT provides detailed layer-wise optical information, most exist-ing approaches focus on single-layer analysis or real-time process control and do not exploit the fullvolumetric information contained in the acquired data.This work presents a modular framework for the volumetric reconstruction and post-process anal-ysis of EOT image data. Sequential EOT images are processed using volumetric component segmen-tation (VCS) and fused into a three-dimensional OT Image Cube, forming a central volumetric datastructure called LPBF Cube. Each voxel encodes spatial and radiometric information and can option-ally be augmented with additional process metadata.Based on this representation, high-resolution two-dimensional slices are rendered along arbitraryorientations using profile-based slicing strategies for planar, cylindrical, and complex geometries.These slices enable intuitive, part-level inspection of laser exposure history and spatial process vari-ations. The framework is validated using geometric and radiometric analyzes, demonstrating goodagreement with nominal CAD geometry and a clear correlation between EOT-derived emission val-ues, laser energy input, and local cross-sectional area.The proposed approach extends the use of EOT data beyond layer-wise monitoring toward com-prehensive, volumetric part inspection and provides a practical basis for geometry-aware quality as-sessment in LPBF, particularly for prototyping and post-build evaluation.

Abstract: A machine learning (ML) framework was developed for the prediction of surface topography obtained with maskless grayscale laser lithography based on the spatial distribution of the applied laser energy dose, or virtual photomask. Artificial neural networks (ANNs) were employed, with the virtual photomask and its radial averages selected as input variables and the surface elevation selected as the output variable. Training of ANNs was carried out with data acquired from the production of models comprising a wide range of representative geometries. Hyperparameter optimization was performed by assessing the accuracy of trained ANNs, with the final configuration comprising a single hidden layer with 15 neurons and a Sigmoid activation function. The trained ANN was then employed within an iterative optimization algorithm to determine the best virtual photomask for the production of new objects by updating the virtual photomask based on the predicted error, thus automatically compensating for proximity effects and sharp dose transitions. The developed approach achieved a reduction in average build error from 2.8 µm to 1.3-1.5 µm compared to standard experimental approaches in a single build, improving not only accuracy but also greatly reducing time requirements for optimization of the process.

Abstract: The impregnation represents a crucial phase in liquid composite molding (LCM) processes. Researchers over the years have used various approaches for monitoring, based on smart weave, pressure sensors, dielectric. Among the LCM processes, the vacuum bag allows the use of visual systems for detecting the resin flow front. The integration of monitoring systems with controllers for automated management of process parameters leads to an improvement in the characteristics of the final manufactured component. In the present work, an AI-based system integrated with the control of a resin preheating system allows for improvement of the impregnation stage. A machine learning approach, based on the You Only Look Once (YOLO) algorithm, has been integrated with the visual monitoring system to detect and dynamically track the resin flow front in real time. The flow front position has been compared with the theoretical one, evaluated by using the Darcy’s law and based on the mismatch the controller suggests a proper in-time regulation of microwave power. The implemented system is capable of processing images through an AI-based algorithm and extracting the kinematic data of the flow front and integrating the information from the thermocouples and the visual system to control the microwave power.

Abstract: Software that can damage an information system asset is considered a malware, such information systems have been rendered to several destructive attacks mainly due to the emergence of the Internet. Conventional Antimalware software is not effective at eliminating malware due to its many evasion techniques such as polymorphism and code obfuscation. Antimalware software is ineffectual and defenseless against zero-day attacks as it can only eliminate malware for which it has signatures. K Nearest Neighbor, Decision Tree and Support Vector Machine are some of the leading classifiers that has successfully detect and classify Malware but optimal accuracy of detection has not been achieved, in addition, false positives and false negatives persists because the hyperparameters of these classifiers were not optimized and noise was not filtered out of the datasets using feature selection technique. The aim of this research is to develop an optimized malware detection and classification framework employing Principal Components Analysis to mitigate the curse of dimensionality while utilizing optimal hyperparameters of chosen classifiers to boost accuracy of malware detection and classification as well as reduction of false positives and false negatives. This research employed K Nearest Neighbor, Decision Tree, and Support Vector Machine to detect and classify malware with CICMalmem dataset to train the model. Grid search optimization was combined with K-fold cross-validation to optimize the hyperparameters of the selected classifiers in order to boost the model's performance and achieve high detection accuracy as well as low false positives and low false negatives. Machine learning performance metrics such as the F1 Score, Precision, Recall, and Confusion Matrix were used to evaluate the Research Model. K Nearest Neighbor generated Zero False Positives while KNN, Decision Tree and Support Vector Machine achieved Accuracy of 99%, 98.64, and 100% respectively.

Abstract: This paper examines the integration of Artificial Intelligence (AI) and Machine Learning (ML) tools in language learning for novice learners, with a focus on grammar correction, vocabulary acquisition, speech recognition, and enhancing motivation for continued learning. A survey to justify the selection of AI-based tools such as Large Language Models (LLMs), reinforcement learning, adaptive learning systems, and automatic speech recognition (ASR) was performed. Results indicated that LLMs significantly improved grammar accuracy, while reinforcement learning-based vocabulary tools enhance retention. Transformer-based architectures outperform traditional models in contextual language understanding and speech recognition, reducing word error rates. Additionally, AI-driven chatbots and gamification platforms improved engagement. Despite these advancements, challenges such as AI over-reliance, lack of deep personalization, and accent bias in ASR models remain. The research follows a comparative and mixed-methods research design approach with machine learning-based model evaluation, using structured and unstructured datasets. Model performance will be assessed through accuracy, precision, and recall metrics.