A deep patch network with spatiotemporal meta-parameter learning for soft sensor modeling of industrial processes

Abstract

Soft sensing techniques have been widely employed in industrial processes to predict hardly measurable quality variables using easily measurable process variables. Effective modeling strategies for representing spatiotemporal features are highly desired in soft sensing applications for dynamic industrial processes. Deep networks, such as graph convolutional network and gated recurrent unit, are frequently employed to capture the spatial or temporal characteristics of sequential data. Nevertheless, effectively modeling and utilizing spatiotemporal heterogeneity remains a challenging problem. Thereby, a patch network with spatiotemporal meta-parameter learning (PMPL) is proposed in this paper, which presents an uniform framework of modelingspatiotemporal heterogeneous feature representations for soft sensing. In PMPL, a patching strategy is utilized to segment times series into sub-series-level patches, allowing the retention of local semantic information while capturing long-term dependencies. Next, a novel spatiotemporal meta-parameter learning approach is proposed to generate specific parameters for a spatial encoder, temporal encoder and spatiotemporal decoder through establishing corresponding embedding layers. In this way, the spatial, temporal and spatiotemporal dependencies within the process data can be adaptively captured and comprehensively refined from a uniform perspective. Extensive experiments are conducted based on two real production processes, which illustrate thefeasibility and efficacy of the proposed model.