Sparse Robust Dynamic Feature Extraction using Bayesian Inference

Datasets of large-scale industrial processes are often high-dimensional and are characterized by outliers. Probabilistic latent variable models are effective for modeling such data complexities. However, the performance of such models is influenced by the number of latent variables and the adequacy of the noise model that describes the data complexities, such as outliers and skewness. This paper presents a probabilistic slow feature model that considers these two issues simultaneously. The latent space dimensionality is automatically obtained by modeling the emission matrix with a Laplace distribution, resulting in a sparse model. Further, the measurement noise is modeled with a skewed-t distribution to account for the outliers and asymmetry of the noise. The hierarchical representation of these two distributions is considered to obtain tractable solutions for the posterior distributions of the latent variables. The resulting model is estimated through variational Bayesian inference.