Industrial Time Series Modeling With Causal Precursors and Separable Temporal Convolutions

Abstract

When applied to industrial processes, predictive models (PMs) easily fail to depict the overall complexity of a production plant. Multiple factors can interfere during the production process and the requirements in accuracy, safety and efficiency are very high. To fully achieve the potential of PMs, it is useful to integrate typical characteristics of the manufacturing process and build procedures that account for the domain peculiarities. Cause-effect relationships that entail the flow of an industrial process are rarely considered in PMs for manufacturing despite the relevance that these relationships have in practice. In this work we present a two-step procedure that uses a causal discovery method for time series, eventually validated by domain experts, together with a new neural network architecture named Separable Temporal Convolutional Network (S-TCN) to create a PM for industrial multivariate time series. Causal precursors exploit the sequentiality of the process flow, grouping process machines by their effective temporal activation. The S-TCN architecture is based on separable temporal convolution and enables the efficient forecast of more distant temporal connections than common temporal models. A numerical validation is presented on a large class of synergetic nonlinear stochastic processes. We apply the proposed procedure in an ultra-processed food plant collecting more than 100 days of active production. The presented procedure achieves better results when compared with state-of-the-art algorithms in both conditions.