Detecting Low Level Shifts in Noisy Process Data (2025-US-30MP-2587)

In a production environment, control charts are indispensable for differentiating common cause and special cause variation. With established process, the major focus of control chart reviewers is identifying and responding to large deviations from the expected process performance, along with a minor focus on drift toward the upper or lower control limits. 

Small periodic shifts in the process are easy to miss, and even when observed, are often ignored because they only represent a small amount of directional noise. Identifying and investigating these shifts provide an opportunity to identify accidental manufacturing experiments and input changes that have not caused a problem yet.

The toolkit for finding these shifts has traditionally centered on the use of cumulative sum (CUSUM) and exponentially weighted moving average (EWMA) charts. Both methods require some level of parameter tuning to “correctly” identify shift boundaries. Updates to the JMP CUSUM and EWMA platforms have made iterative interactions with these methods much simpler. In addition to the traditional methods, a machine learning method using a fused lasso approach is also effective for quickly identifying shift boundaries.

In this paper, CUSUM, EWMA, and fused lasso are demonstrated and compared for their ability to detect small shifts and ignore spurious patterns in noisy process data.