Hybrid imputation of missing values using KNN on MEWMA-based adaptive process control

Missing data, a common issue in production processes due to factors like sample contamination and equipment malfunctions, can lead to a decrease in the recognition accuracy of control charts, especially in cases of shifting. To address this, we introduce an online adaptive weighted imputation technique that combines the strengths of K-Nearest Neighbor (KNN) and Exponentially Weighted Moving Average (EWMA) imputations. It utilizes an adaptive weight matrix for weighting both methods and an adaptive covariance matrix to optimize for missing structures. When dealing with data fluctuation, we assign a higher weight to the KNN method for its sensitivity, while the EWMA method is preferred for stationary data. This approach does not require data stacking; thus, the imputation process for missing data is conducted online. Consequently, based on the online Multivariate EWMA (MEWMA) control chart, real-time process monitoring can be achieved. To optimize the use of available information, we also adjust the covariance matrix with a weight matrix to emphasize complete data. The proposed technique outperforms traditional methods in performance monitoring by avoiding false alarms and quickly detecting anomalies during process shifts.