Inspection-Oriented Predictive Quality Modeling for MDF Manufacturing Using Industrial Process Data and Machine Learning

Continuous medium-density fiberboard (MDF) production presents a persistent quality-assurance problem: destructive laboratory tests return results too late to prevent off-specification material from accumulating before a corrective response is possible. This study develops an inspection-oriented predictive quality framework using industrial Distributed Control System (DCS) data and automated machine learning, treating the production line as an integrated nine-stage system in which upstream process disturbances propagate through coupled thermomechanical and chemical operations before becoming visible in final panel properties. Two quality targets were modeled across Ultralight (UL) and Standard Thin (STD) panels using 3365 production batches and 327 DCS process variables. The pipeline combined Random Forest imputation, Pearson collinearity filtering (|r|≥0.8), target-specific feature selection, and stacked ensemble regression via H2O AutoML. The Vertical Density Profile Index (VSC), a plant-reported scalar derived from X-ray density profiling, was predicted accurately in both product families (test RMSE: 1.39 and 1.78, index units for UL and STD respectively), reflecting its close coupling to drying stability, resin dosing, and thermal conditions. Internal Bond strength (IB) was harder to predict, especially for thin STD panels (test RMSE: 78.93 kPa vs. 27.41 kPa for UL), as core-layer bonding mechanisms are only indirectly observable through standard DCS instrumentation. Model-agnostic feature importance rankings were physically coherent across both product families, with dominant predictors concentrated in drying, resin application, forming, and hot pressing, consistent with the coupled-subsystem nature of MDF quality formation. The historical dataset was dominated by acceptable and over-quality IB production, which precluded conformity classification and sampling-reduction analysis; a prospective dataset with near-threshold observations is required for those evaluations. Within that scope, the framework provides continuous quality estimates, identifies deviations from the desired operating range, and supports inspection planning as a complement to formal laboratory testing.