Methodology for Integrated Failure-Cause Diagnosis with Bayesian Approach: Application to Semiconductor Manufacturing Equipment

Semiconductor Industry (SI) is facing the challenge of short product life cycles due to increasing diversity in customer demands. As a result, it has transformed into a high-mix low -volume production line that requires sustainable production capacities. However, significant increase in the unscheduled equipment breakdowns, limits its success. It is observed that in a high-mix low-volume production, product commonality is inversely proportional to failure occurrences and number of corrective actions in each failure. This provides evidence of misdiagnosis for equipment failures and causes. Moreover, equipment is believed to be the only source for product quality drifts that increase the unscheduled breakdowns and result in unstable production capacities. In this paper, we propose two defense lines against increasing unscheduled equipment breakdowns due to misdiagnosis. We argue that product quality drift can be traced to product itself, process and maintenance events, besides equipment. The Bayesian Belief Network (BBN) is proposed using symptoms, collected across drift sources, that improves equipment breakdown decisions by accurately identifying the source of product quality drift. The misdiagnosis of equipment failures and causes, if equipment is found as a source of drift, is another significant factor for increasing unscheduled equipment breakdowns. Existing failures and causes diagnosis approaches, in the SI, model equipment as a single unit and use fault detection and classification (FDC) sensor data. We also argue that these are the key reasons for the misdiagnosis because of neglected facts that production equipment is composed of multiple modules and FDC sensors undergo reliability issues in a high-mix low-volume production line. Therefore, to improve these misdiagnosis, another BBN is proposed that uses statistical information, collected from the equipment database, at the module level. These BBN models are evaluated in a thermal treatment (TT) workshop at the world reputed semiconductor manufacturer. The BBN model for the identification of the source of product quality drift (failure mode) demonstrates 97.8% prediction accuracy; whereas, module level BBNs for equipment failures and causes diagnosis are found 45.7% more accurate than equipment level BBN.