Learning precedence relations for manufacturing operations sequencing using convolutional recurrent neural networks

Abstract

Cyber manufacturing as-a-service or platform-based manufacturing, which connects buyers and manufactured parts suppliers through an online marketplace, have recently emerged with the goal of democratizing access to manufacturing capabilities. This approach to sourcing manufactured parts places intense pressure on suppliers to efficiently and optimally plan for part manufacturing to reduce production costs and become more competitive. Sequencing of manufacturing operations is an important step in the process planning pipeline, which has historically relied on the knowledge of human experts. An automated approach to operations sequencing has long been sought but is urgently warranted today given the skilled labor shortage in the certain parts of the world. While researchers have proposed various algorithms for automating operations sequencing, an underlying assumption of these methods is that precedence relations (commonly referred to as precedence constraints) among manufacturing operations must be manually defined to preprocess inputs to operations sequencing algorithms. This assumption has significantly hampered the generalizability of existing operations sequencing algorithms. Considering this limitation, in this work we present a data-driven approach to learn precedence relations for machining operations instead of relying on human expertise. By embedding the precedence relations from successfully produced parts as latent recurrent vectors, it is demonstrated that the proposed 3D-convolutional recurrent neural network (3D-ConvRNN) model yields 97.6% precedence relation validation accuracy, outperforming a 3D-CNN binary classifier. Furthermore, the proposed model is used in case studies to assess simple sequences of machining operations for realistic parts and to automatically generate operations precedence graphs as inputs to operations sequencing algorithms. Our results suggest that a data-driven approach to learning precedence relations can be beneficial for automating operations sequencing by augmenting or replacing manually defined precedence constraints.