The process of dyeing the fabric can cause textile waste due to typical coloring methods when wet. However, a professor at Wilson University at North Carolina State University found a way to solve this problem using machine learning.
Professor Warren Jasper discovered that the amount of color change from wet to dry is not uniform between different colors, and found that the nonlinear relationship makes it difficult to predict results based on a single color sample.
“The fabric is dyed while it's wet, but the target shade is dry and wearable. This means that if there is an error in the color, you can't know until the fabric is dry,” he said. “More than the fabric is dyed all the way while waiting for that drying to occur. That leads to a lot of waste because you can't catch the error until later in the process.”
To solve this problem, Jasper has developed five machine learning models, including neural networks designed to map this type of nonlinear relationship. Jasper trained the machine to predict the outcome using visual data from 763 fabric samples of various colors, both wet and dry.
Each machine learning model outperformed those who were not using AI in terms of accuracy, but neural networks outperformed everything else with median errors of 0.01 and 0.7 using Ciede2000, a standardized color difference equation. Other machine learning models show the Ciede2000 error ranges of 1.1 and 1.6, with a baseline of 13.8. Error values above 0.8-1 are generally considered external tolerance limits for the textile industry.
Jasper discovered that neural networks could significantly reduce waste caused by color errors. This is because fabric manufacturers can better predict the final result of dyeing before large amounts of fibers are accidentally colored. He outlined his findings in his paper, “Controlled Study on Machine Learning Applications for Predicting the Color of Dry Fabrics from Wet Samples: The Effects of Dye Concentration and Pressure,” published in the Journal. fiber.
Machine learning and AI are primarily tapped by textile industries in other sectors, recycling and cyclicality. Jasper said he hopes the study will lead to wider use of similar machine learning tools across the broader textile industry.
“We're a little behind the curves of textiles. The industry has begun to move further towards machine learning models, but it's very slow,” he said. “These types of models can provide powerful tools to reduce waste and increase continuous dyeing productivity, which accounts for more than 60% of dyed fabrics.”
