Fault Detection Using Human–Machine Co-Construct Intelligence in Semiconductor Manufacturing Processes

Fault detection (FD) utilizing “principal component-based
${k}$
-nearest neighbor rule” (PC-
${k}$
NN) has been previously studied. However, these studies do not explicitly account for the distribution of process variables in the manufacturing process. In addition, they do not incorporate the expert’s domain knowledge. To account for these issues, we introduced a new technique, FD using human machine co-construct intelligence (FD-HMCCI) that explicitly accounts for the distribution of process variables and integrates the expert’s knowledge in the principal subspace. In this technique, the expert knowledge is represented as expert envelopes, which are the range of values variables can take within which the expert believes that the process is acceptable. Similarly, the range of values of the variables within which the PC-
${k}$
NN classifies the process as acceptable are represented as
${k}$
NN envelopes. FD-HMCCI calibrates the parameters such that the aggregate score, which combines agreement (overlapping area between the expert and
${k}$
NN envelope), disagreement (the non-overlapping area) and tail risk (the conditional expectation of the variables’ distribution outside the
${k}$
NN envelope), is maximized. For demonstration, the technique is implemented to calibrate
${p}$
of PC-
${k}$
NN that is used for FD in Varian E500 implanter, operated in a semi-conductor foundry.