Low-Cost Functional Testing Based on Data Imputation Integrating Fault Tree Analysis and XGBoost

Abstract

Functional testing is essential for ensuring the quality of electronic products. As system complexity increases, the cost of functional testing—particularly during the motherboard testing stage—has risen significantly. Designing an efficient testing strategy is therefore key to reducing overall testing costs. Although data imputation methods can enhance the effectiveness of strategies by addressing missing data, current approaches do not adequately account for the impact of correlation information between system modules on the input information of the data imputation model. This oversight results in suboptimal imputation performance, making it challenging to further reduce testing costs. To overcome this limitation, we propose a data imputation method that integrates fault tree analysis (FTA) with eXtreme gradient boosting (XGBoost), effectively combining system fault mode analysis with data-driven modeling. First, faults in nonbus data transmission are additionally considered, and an enhanced fault tree for motherboard testing is constructed. Next, correlations among system modules are quantitatively represented as system event associations within the fault tree, based on which the attributes requiring imputation are determined. Then, a top–down quantitative analysis method driven by both mechanisms and data is introduced to infer the states of basic events from intermediate event states. This mapping from intermediate events to basic events reduces the proportion of missing data. Based on this, missing values are imputed using the tree-based XGBoost model. Experiments conducted on transformed real-world manufacturing data demonstrate that effective testing strategies can be dynamically developed under varying conditions using the proposed method. Testing costs are reduced by up to 5.64% in high-yield scenarios, 6.66% in low-yield scenarios, and 7.07% during long-term yield fluctuations. Furthermore, the defect level is reduced by as much as 51.02%.