Learning data from continuous multistage manufacturing processes: focusing on sequential nature of the processes, lead-time, and insufficient labels

Abstract

Manufacturers have implemented continuous multistage manufacturing processes (MMPs), in which raw materials are processed through multiple stages without any interruptions, based on their efficiency and flexibility compared to discrete MMPs. While numerous data-driven soft sensors effectively estimate key variables by reflecting the inherent characteristics of target processes, applying these approaches to continuous MMPs presents several challenges: first, obtaining intermediate output labels is often impossible or very difficult; second, determining lead times between stages is also challenging; third, the scarcity of final product labels complicates model training. To address these three challenges and accurately reflect the inherent characteristics of continuous MMPs, we propose three models, each tackling a specific challenge: (1) MMP Net, an FFNN-based model that accurately represents the control mechanisms in continuous manufacturing industries. The MMP Net sequentially injects special input elements into the hidden layers of an FFNN. As a single model, MMP Net does not require intermediate output information. The sequence of input features in continuous MMPs can be intuitively considered, and the dynamics across multiple stages systematically represented. (2) Multistage Net, novel machine-learning model designed for continuous MMPs of liquid products. In this model, several interstage blocks are organized in a hierarchical structure within a multistage module. The interstage block is proposed to capture the sequential dependency between the previous and current stages and concurrently explores the lead-time relationship. From the interconnected interstage blocks, the multistage module can learn the sequential nature of MMPs across all stages even in the absence of intermediate output labels. (3) Multistage MCN, a self-supervised learning approach focused on the insufficiency of product and lead-time labels in continuous MMPs. The architecture incorporates two key components: the innerstage and interstage blocks. To learn the sequential nature of MMPs and the random lead-time relationships without relying on intermediate output and lead-time labels, the architecture incorporates two key components: the innerstage and interstage blocks. To overcome the scarcity of final output labels, we have introduced self-supervised learning through two innovative pretext tasks: masked patch reconstruction and contrastive learning. Validation studies on real-world datasets demonstrate the superior performance of the proposed models compared to existing models. We believe this paper will significantly contribute to the advancement of machine-learning technologies, specifically tailored to address industrial problems.