A Unified Approach to Screen Suspicious Machine Sets with Reduced Redundancy and Spuriousness in Multistage Manufacturing Processes

Abstract

Many modern manufacturing systems employ a multi-stage manufacturing process (MMP) that consists of multiple consecutive process stages with multiple machines aligned in each process stage. The defective rate of the MMP can be increased due to either or both individual machines or interactions between machines. In this research, a set of one machine that significantly increases the defective rate is defined as one- machine root cause machine set (MS), and a set of multiple machines exhibiting an interaction that significantly increases the defective rate is defined as multiple- machine root cause MS. Existing studies analyzed production log data and product quality data to screen a limited number of suspicious MSs to reduce the number of machines to be diagnosed for root cause identification. However, this screening often includes spurious MS and redundant MS because defective products caused by root cause MSs also pass through the MSs other than root cause MSs in MMP. A spurious MS is an MS that does not significantly increase the defective rate, but it can be screened as a suspicious MS when it frequently operates with root cause MSs in a production log dataset. A redundant MS is a union of a root cause and a spurious MS, and it is screened as a suspicious MS when the effect of the root cause MS on defective rate and interaction between the machines in the spurious MS are not distinguished. The aim of this research is to develop methods for suspicious MS screening that reduce redundancy and spuriousness. This is achieved through three research objectives. The first and second objectives are to reduce redundancy by screening suspicious MSs with suspiciousness measures that distinguish the effects of a root cause MS and a redundant MS on the defective rate. The full-way effect is defined as the effect on defective rate owing to the highest order of interaction between machines in an MS. Measuring the suspiciousness of the MS according to the full-way effect does not overrate the suspiciousness of the redundant MS. The method to calculate suspiciousness measures based on the multi-item association rule and feature attributions on the prediction model is developed in the first and second research objectives, respectively. The third objective is to develop the suspicious MS group seleciton method that mitigates spuriousness by screening MSs by defective coverage diversification. The defective coverage of an MS is a set of defective products that pass through the MS. The diversification of defective coverage of MSs in the suspicious MS group, as well as MSs’ suspiciousness maximization, are considered to reduce the chance of spurious MS being screened. This research provides the following contributions. First, it proposes unified methods for screening both one-machine MS and multiple-machine MS. The suspicious MSs are screeend based on the full-way effects; additional tasks for reducing redundancy are not required. Also, this research resolves the spuriousness for both one- and multiple-machine MSs. Second, the application of the proposed methods is expected to reduce the time and costs associated with diagnosing redundant and spurious MSs. Finally, the proposed methods are applicable to various MMPs because the methods do not require specific assumptions about the model or data requirements.