Distinguishing cause from consequence is a big challenge for many machine learning techniques. We built a tool, using outlier detection techniques and Bayesian Networks to combine domain expertise and data that helps to identify problematic behavior in ASML lithography machines and automatically suggests potential root causes of the problem.
While our machine learning techniques excel in separating the signal from the noise in large amounts of sensor data, distinguishing correlation and causation remains a massive challenge. The latter is just the challenge we are facing when we use data-driven methodology for diagnostics. An ASML lithography machine has an enormous complexity, so when an issue arises, we need to find techniques that can both handle and use these large amounts of sensors to identify problem behavior, but also understand the workings of our machine to be able to separate cause from consequence.
In this work, we combine unsupervised techniques to identify problem behavior in an ASML lithography machine and combine this with Bayesian Networks to identify the underlying root cause. We choose for unsupervised techniques, due to nature of both the problem behavior and our ASML machines: the population is small in comparison to the dimensionality of the problem, and chances are big that behavior arises that we have not witnessed before. The Bayesian Network is created by combining ASML domain knowledge with data, in order to correctly suggest potential root causes, explainable by domain expertise.
We demonstrate our created tool that incorporates domain knowledge and interprets sensor data to come to a diagnosis. We show the challenges we had building it, and the python tools that were considered and used.