Machine learning iterative filtering algorithm for field defect detection in the process stage

Abstract

Domestic gas boilers, which are seasonal appliances, undergo total inspections in the process stage and only the products that have passed the final shipment inspection are released. However, defects in the field do occur even in the products that have passed total inspection. In order to detect field defects as much as possible in the process stage, this study proposes the machine learning iterative filtering (MLIF) algorithm, which iteratively filters predicted defect-free products. This algorithm applies a unique method that uses n learners formed through multiple random undersampling in the majority class of defect-free products. In addition, the sampling random number and the threshold for classification decisions are specified under the condition where the recall result of classification is 100%. As a result, the classification prediction performance of the minority class is improved because it is extremely rare that an actual defective product is incorrectly predicted as a defect-free product. Experimental results showed that the final classification performance of the test data of the MLIF algorithm was 87%, which is approximately 11%p higher than that of single learner (76%). Finally, a process management dashboard, which shows various information about processes, field data, and the predicted results of the MLIF algorithm, to facilitate process decision making is presented.