Intelligent Lot‐Level Acceptance Sampling via Active Learning, XGBoost, and Wilson Confidence Intervals

ABSTRACT Traditional acceptance‐sampling plans rely on fixed sample sizes and random sampling principles. They cannot directly incorporate lot‐to‐lot quality differences and statistical uncertainty into the decision‐making process. This study proposes a novel acceptance‐sampling framework that integrates active learning (AL), XGBoost‐based defective‐probability estimation, and an early decision mechanism based on the Wilson confidence interval to make lot‐based quality acceptance/rejection decisions in the laminate flooring production process more efficient and statistically reliable. In the proposed method, a small initial sample is collected for each lot; these samples are labeled within the tolerance rules, and an XGBoost‐based classifier is trained. The trained model estimates the probability of defective products and prioritizes which products to inspect using uncertainty sampling. The Wilson confidence interval for the defective rate is calculated from the sample set updated at each iteration, and decisions to early accept, early reject, or continue inspection for the lot are made based on the lower and upper limits of this confidence interval. Within this framework, three Wilson policies representing different decision levels (Conservative Wilson Policy: CWP, Moderately Conservative Wilson Policy: MCWP, and Adaptive Wilson Policy: AWP) have been defined. Experimental results on a dataset of 3000 products, comprising 2000 training and 1000 validation items, obtained from laminate flooring production, demonstrate that the proposed method can significantly reduce inspection burden without compromising reliability. A 100% accuracy was achieved across all policies, with no false lot acceptances or rejections. The consistent performance results obtained on training and independent validation lots validate the generalization capability of the proposed approach.