Root cause estimation of faults in production processes: a novel approach inspired by approximate Bayesian computation
针对仅能获取最终检测数据、批次生产且数据非高斯分布的多阶段大规模生产过程，提出一种融合近似贝叶斯计算与过程模型的方法，从正常与异常数据中计算各部件对故障的贡献分数，从而估计故障根因候选。
This paper presents a methodology for estimating root causes of faults in multistage mass production processes that have three properties: (1) only the final inspection data can be acquired, (2) hundreds of products are manufactured in a lot-wise manner, and (3) the acquired dataset does not always follow a Gaussian distribution. The proposed method consists of two components: (i) derive the distribution of part variables from the inspection dataset by fusing the approximate Bayesian computation (ABC) and the process model, and (ii) derive the root cause scores from the normal and abnormal datasets, which quantify how much each part contributes to the abnormal condition. The proposed method can estimate candidates of the fault causes, and numerical experiments are performed to explore the effectiveness and limitations of the method. Furthermore, the application to actual data of an internal camera module yields consistent results with design information given as domain knowledge.