Energy consumption intelligent modeling and prediction for additive manufacturing via multisource fusion and inter-layer consistency

Abstract

A novel deep network termed 3DPECP-Net is proposed to address an important, interesting, yet challenging problem in intelligent manufacturing: Accurately predicting energy consumption (EC) in the additive manufac-turing (AM) process, commonly known as 3D printing (3DP). This task is of vital importance in fulfilling the low energy criteria under the policies to tackle energy shortage and climate change. According to the intelligent EC modeling in 3DP, the multisource data consists in three parts including pixel-source, motion-source, and processing-source. Compared with existing solutions, which either only harness the geometric characteristics of computer-aided design (CAD) models or simply integrate the processing parameters into the prediction, the proposed 3DPECP-Net is able to seamlessly fuse relevant data in 3DP, taking full advantage of the complementary information of diverse datasources for more accurate prediction. Specifically, the multisource fusion frame (MSFF) stacks multiple transformer blocks to catch global spatial information, and convolutional blocks to catch local feature details. Furthermore, we propose a continual attention memory network (CAMN) to ably extract and store the inter-layer consistency, enforcing to efficiently learn an exhaustive representation for 3DP EC prediction. Extensive experiments are conducted on an in-house dataset and the experimental results demonstrate that the proposed 3DPECP-Net achieves a superior prediction performance than state-of-the-art methods.