Convolutional variational autoencoder for real-time monitoring of high-dimensional partially observed ordinal categorical data streams

In the current rapidly developing technological environment, the integration of artificial intelligence (AI) into quality technology marks a transformative shift, presenting significant opportunities and challenges for online process monitoring and diagnosis. This study introduces a novel convolutional variational autoencoder (CVAE)-based adaptive sampling strategy tailored for high-dimensional partially observed ordinal categorical data streams (HPODS) under resource constraints. The proposed method compresses HPODS into a lower-dimensional latent space and reconstructs outputs from sampled latent vector, enabling effective data reconstruction and real-time monitoring. An adaptive sampling procedure is further developed to dynamically allocate data collection resources, balancing exploration and exploitation. Extensive simulations and a case study on leather defect detection validate the effectiveness of the approach.