This dissertation presents a unified methodology of prognostics evaluation for multiple product and process manufacturing system. Under the proposed framework, several prognostics tasks, including failure predictability evaluation, component Remaining Useful Life (RUL) estimation, and root cause analysis, are systematically addressed. While the methods introduced in this work are applicable for general manufacturing system with multiple operational conditions, it places a particular emphasis on the semiconductor environment and addresses some unique challenges in wafer fabrication equipment prognostics. All methods developed in this work are tested with use cases from the fabrication facilities in GLOBALFOUNDRIES.
A Partial Least Square (PLS) Regression-based method is developed for failure predictability evaluation, RUL estimation and fault diagnosis. A group of so-called Generalized Remaining Useful Life (GRUL) curves, which are defined to reflect various degradation patterns, are combined as multivariate output of PLS regression model. The PLS model selects the variations from the inputs that have the best correlation with the output, based on which an optimal RUL target curve is obtained to represent the degradation pattern of the equipment. The predictability of the failure mode can be evaluated by comparing the RUL estimation provided by the model and the lead time requirement of actual maintenance practice. Furthermore, a variable blocking contribution strategy is proposed to enable a user to hierarchically drill down to the input variables and identify the root cause of the failure.
As today's equipment is constantly operating under different regimes for various products' specification while its condition keeps degrading, it is critical to monitor the condition consistently such that the schedule of operation and maintenance can be optimized. The PLS modeling strategy is further extended to address this issue with separate models built for individual operational regimes. Assuming that data from each process spread widely over the whole degradation process, individual PLS models are built to capture the impact of different processes on the equipment. Given a tool is running on a particular process, the RUL for the current process as well as the equivalent RUL for other processes can be obtained. With this information, a recommendation can be made on future production such that equipment availability is maximized and maintenance can be scheduled in a timely manner.