Problem
Color judgment in dyeing runs on two things: a CCM (Computer Color Matching) instrument and the human eye. The problem was that the two often disagreed. Fabric the CCM called identical could look different to an inspector. Because the error values the CCM measured could not be trusted, final judgment ended up resting on human eyes — inefficient, and above all inconsistent. A failed judgment meant writing a correction recipe and dyeing again.
Data collected
Equipment, quality and volume data were gathered from the dyeing equipment PLC, the CCM inspection equipment PLC and keyed-in volume information. Thirteen variables were used in the analysis, including dyeing run length, equipment temperature process value (TRD_TEMP_PV), equipment speed (TRD_SPEED1) and dyeing color difference DE. The dataset came to roughly 31.8 million records and 157MB.
Modules applied
An XGBoost regression model was built to predict dyeing color difference. As a boosting method it trains and predicts quickly and carries overfitting regularization, which keeps it relatively stable against the spread of real shop-floor data. Alongside the prediction model, variable importance analysis established how much each condition acts on color difference.
Result
Color difference in the dyeing process was confirmed to be predictable in advance. Being able to predict quality mid-process from equipment operating conditions and dyeing batch conditions provides a basis for managing quality on data rather than waiting for inspection results. The variable importance analysis gives the material for explaining why a color defect occurred.
When the instrument and the eye disagree
In dyeing, color decides pass and fail. But it was being measured two ways, and the two often disagreed.
A CCM (Computer Color Matching) instrument measures color numerically. Yet fabric the CCM called identical could look different to a person — not once, but repeatedly.
So the floor treated CCM readings as reference and left final judgment to the eye. A first inspector looks; difficult cases go to a second judgment; even passed product gets a final color judgment from the customer.
That structure carries two costs. Human time goes into judgment continuously, and the standard shifts from person to person.
Capturing the moment color is decided
The dyeing process runs like this. Dye and auxiliaries are sent to the dyeing machine in the standard quantity set by the recipe; water and fabric go in; temperature and time are controlled to put color into the fabric.
In other words, color is a result produced by equipment operating conditions. How the temperature rose and how long it held, what speed the equipment ran at, how large the batch was — these decide the final color.
If so, those conditions should be able to predict color difference. Thirteen variables were drawn from the dyeing equipment PLC, the CCM inspection PLC and volume information: dyeing run length, equipment temperature process value, equipment speed, and the outcome value, dyeing color difference DE. Some 31.8 million records were gathered.
Why XGBoost
XGBoost regression was chosen for color difference prediction.
Boosting weights the errors made by weak models and passes them to the next model to correct. It keeps working at what it got wrong, which raises accuracy.
XGBoost adds parallel training and overfitting regularization to that. Shop-floor data is often concentrated around particular conditions, and regularization holds back a model that only fits the training data.
What mattered more than prediction
The analysis confirmed that dyeing color difference can be predicted in advance — that the outcome can be estimated from in-process conditions alone, before the fabric reaches the CCM.
But what mattered more in this case was variable importance analysis. It gives a number for how much each condition acts on color difference.
That means there is now material for answering why a color defect occurred. A visual judgment reports the result and nothing about the cause. You need the cause to fix the recipe and avoid the same defect on the next lot.
What remains
This is not a story about AI replacing the human eye. A person is still there at final judgment.
What changed is the basis of the judgment. Once something explained only by feel starts being explained by data, that standard survives a change of staff.