Designing
Modern Data & AI
Systems

The detection and classification of defects in cotton yarn are crucial in maintaining the quality of textile production. This is hardly getting attention in the literature due to the complexities and non-homogeneous features appearing in the cotton yarn. This study explores the application of transfer learning techniques in convolutional neural networks (CNNs) to classify yarn defects, including neps, thick and thin places, hairiness, and snarls, as well as identifying non-defective yarns. A dataset of 1,250 images was divided into five classes to evaluate three CNN models: ResNet-50, VGG-16, and Inception-v3. Inception-v3 achieved the highest validation accuracy at 98.8%, followed closely by VGG-16 with 98%, while ResNet-50 reached 77.2%. Inception-v3 and VGG-16 were successful in detecting complex yarn defects. The study further emphasizes the capability of CNNs to automate yarn defect identification by decreasing the processing time, by allowing CNN models to integrate with GPUs.

Predicting dimensional shrinkage in knitted fabrics remains a complex challenge due to the non-linear interplay of material and process variables. This study introduces an Artificial Neural Network (ANN) model designed to estimate shrinkage in 100% cotton and polyester-elastane single-jersey knits. Built using TensorFlow-Keras with a feed-forward backpropagation architecture, the model integrates twenty-three critical inputs, including yarn count, stitch density, tightness factor, and machine settings. By capturing data from the knitting, pre-setting, and finishing stages, the ANN provides a comprehensive framework that surpasses conventional predictive models in both accuracy and scalability. The research demonstrates that the trained ANN can rapidly forecast shrinkage using known parameters, enabling proactive quality control during production planning. This digital approach significantly reduces the need for resource-intensive physical sampling and post-compacting tests. Validated against independent samples, the model showed high correlation coefficients and minimal error rates. This integration of machine learning into textile manufacturing offers a robust solution for enhancing productivity and dimensional stability across various fabric structures.

Financial fraud accounts for an estimated 5% loss in annual corporate revenue, severely undermining stakeholder trust and capital market stability. Undetected anomalous journal entries are a primary driver of these losses, yet traditional auditing methods and "black-box" machine learning models often fall short. Manual oversight is increasingly unscalable, while opaque algorithms lack the explainability required for regulatory compliance and forensic validation. To address these critical gaps, this research introduces an automated, explainable multi-agent AI system specifically engineered for journal-entry fraud detection. By leveraging a collaborative framework of specialized agents, the system achieves an impressive 90.48% overall accuracy. Beyond mere identification, the framework prioritizes transparency, providing auditors with clear reasoning for flagged transactions. This approach not only mitigates direct financial loss but also strengthens investor confidence and ensures sustainable corporate performance through proactive, high-precision financial oversight and robust regulatory alignment.