Deep Neural Network Pipelines for Multivariate Time Series Classification in Smart Manufacturing

In smart manufacturing, multivariate time series (MTS) data from interconnected sensors and actuators have been collected to model the product quality. However, high dimensional MTS data associated with complex functional structures has posed significant challenges for classical machine learning and statistical learning methods. As alternatives, deep neural networks (DNN) with highly non-linear structures and various data augmentation, pre-processing, and tuning techniques have been investigated for MTS data modeling. The recent transformation of step-by-step offline data analytics to the fast end-to-end computation pipelines motivated us to investigate DNN pipelines for MTS classification problems. However, execution of all the candidate DNN pipelines is computationally expensive, which calls for an effective approach. Thus, the adaptive top-N linear generative-discriminative (AT-LinGD) method is proposed as a learning to rank method that learns top-N ranked pipelines by iteratively exploring a small subset of all possible pipelines. It generates latent variables to describe pipelines and explore them to update the exploration set in a sequential manner. Thus, the adaptively generated latent variables enable the efficient and accurate ranking of the top-N pipelines with limited execution. A real case study of aerosol® jet printing process demonstrates the merits of the AT-LinGD model.