Nonparametric EWMA monitoring of high-dimensional data streams using neural networks

With the rapid development of sensor technology, modern industrial systems generate large volumes of high-dimensional process data. Distinguishing normal and abnormal operating states from such data streams is an important problem in statistical process control. Yet traditional multivariate control charts often rely on specified distributional assumptions that may fail for skewed or heavy-tailed data. Existing machine learning-based charts, however, may lack the capacity to effectively capture complex nonlinear dependencies. In this paper, we propose a new nonparametric control chart that combines a neural network with an exponentially weighted moving average (EWMA) scheme for monitoring high-dimensional data streams. Numerical studies validate the effectiveness of the proposed chart in detecting diverse mean and covariance shifts. It remains robust and computationally efficient, even under data imbalance and high-dimensional settings (p = 1000). Furthermore, compared with representative benchmark control charts, the proposed chart generally achieves shorter or comparable detection delays. A case study, based on Backblaze hard disk drive data, further demonstrates its practical effectiveness for real-world fault detection.