Wavelet Gaussian process for multi-scale monitoring of imaging data

Advanced sensing brings a plethora of imaging data to handle process variations in additive manufacturing (AM). Realizing the full potential of sensor-based statistical process control (SPC) methods greatly depends on the delineation of defect characteristics and anomalous patterns. However, traditional methods primarily focus on analyzing and monitoring AM images in a single scale. It is common that these images contain hidden patterns pertinent to process deviations that are not accessible in a single-scale view. This paper presents a new wavelet Gaussian process (WGP) framework for multi-scale analysis of variance (ANOVA) and image-guided SPC of AM. First, AM imaging data are examined through multi-scale wavelet lenses to reveal hidden defects in different spatial-frequency resolutions. Second, WGP models are developed to analyze different sources of variation in wavelet coefficients. Specifically, we separate the contribution of standard profile, layerwise variation, and process shifts. We further design generalized likelihood ratio (GLR) formulation and multi-scale hypothesis testing approaches to fuse test statistics from different wavelet sub-bands for AM quality monitoring. The proposed WGP approach is evaluated and validated through simulation and real-world case studies. Experimental results demonstrated the effectiveness of the proposed WGP method for multi-scale analysis and monitoring of imaging data.