Establishing Convolutional Neural Network Kalman Recurrent Variational Autoencoder Using Infrared Imaging for Process Monitoring: An Application in Spinning Disk Processes

The measurements for thermal changes are commonly mounted only at the fixed location in the operating process. In this article, to comprehensively understand the thermal inhomogeneity and the sampling variances, the nondestructive evaluation of infrared measurement is used to record real-time thermal changes. With the thermal image data, a data-driven deep dynamic latent variable model named the convolutional neural network-Kalman recurrent variational autoencoder (CNN-KRVAE) is proposed. It is constructed based on image data instead of sensor-measured data. The convolutional neural network is used to extract the important spatial thermal distribution features in the image data. Furthermore, the dynamic characteristics of the thermal information are represented by the Markov state-space relationship between the latent variables in the latent space. Moreover, the probabilistic model, variational autoencoder (VAE), is embedded in the model to describe the uncertainty in the process. Consequently, the probability density estimates instead of the point estimates are established to form the distribution-based monitoring indices in the latent and residual spaces. To verify the modeling and monitoring performance of the proposed model, CNN-KRVAE is used for an experiment of the spinning disk process.