Prediction of the relationship between yarn types and weaving defects detected in the confection process with artificial neural networks

This thesis aims to model and predict weaving defects detected during the garment production process based on the properties of the yarns that cause these defects, using Artificial Neural Networks (ANN) and Deep Learning (DNN) approaches. The main focus of the study is to analyze the influence of the physical and structural characteristics of yarns on woven fabric quality. Within the study, yarn count, yarn type, fabric weight (GSM), density, pile height, size, and produced meter were used as input parameters, while the output variable was the fabric quality class (GOOD, MEDIUM, POOR, REJECT) determined according to the number of defects per 100 meters. Input data were processed through normalization and numerical encoding techniques, and the output variable was modeled using one-hot encoding based on defect ratios. The initial ANN model consisted of a single hidden layer and was later expanded into a Deep Fully Connected Neural Network (DNN) structure with increased layer depth and neuron count. The Adam optimizer, ReLU and Softmax activation functions, and categorical cross-entropy loss were employed during training. Model performance was evaluated using accuracy, F1-score, and balanced accuracy metrics. Due to the class imbalance in the dataset, the model occasionally predicted only the majority class. When data balancing (oversampling) was applied, the model successfully predicted all classes with improved representativeness. The unbalanced dataset yielded 77.7% validation accuracy, while the balanced dataset achieved 74.79% test accuracy, a macro F1-score of 0.739, and balanced accuracy of 0.740. Particularly, the "REJECT" class achieved a recall of 0.95 and an F1-score of 0.83, indicating the model's effectiveness in identifying critical weaving defects at early production stages. The findings reveal that the relationships between weaving parameters and fabric quality are nonlinear, and that deep learning architectures can model these complex interactions more effectively than conventional ANNs. The developed model provides a predictive decision-support tool that can assist in yarn selection, minimize quality losses, and enable proactive process optimization in textile manufacturing.