Abstract Real-time monitoring for Additive Manufacturing (AM) processes can greatly benefit from spatial-temporal modeling using deep learning. However, existing, deep-learning approaches in AM are case-dependent, and therefore not robust to changes of control inputs and data types. As AM is dynamic and complex, this limitation leads to a lack of systematic, DL approaches for real-time monitoring of AM, which involves a large number of varying control parameters and monitoring data. To address the challenge, this paper introduces a novel approach for developing spatial-temporal models to monitor Laser Powder Bed Fusion (LPBF) processes using deep learning on real-time, monitoring data. First, we present a novel model for representing in-situ-monitoring and control data of LPBF at multiple scales. Second, from the model, we extract spatial-temporal relationships for in-situ monitoring of LPBF processes. Third, we present a spatial-temporal, modeling approach using the architecture of convolutional long short-term memory (LSTM) to monitor the spatial-temporal relationships and detect anomalies. A case study used convolutional LSTM Autoencoder on optical, melt-pool-monitoring data, one of the most widely adopted data types in in-situ monitoring of LPBF. The data was generated from an LPBF testbed called the Additive Manufacturing Metrology Testbed. The novel, learning approach enables spatial-temporal modeling of AM dynamics directly from real-time data for the monitoring of varying AM environments. This methodical approach provides the potential to fuse realtime data at multiple, spatial-temporal scales.

Abstract Additive manufacturing (AM) has brought positive opportunities with phenomenal changes to traditional manufacturing. Consistent efforts and novel studies into AM use have resolved critical issues in manufacturing and broadened technical boundaries. Build time estimation is one of the critical issues in AM that still needs attention. Accurate build time estimation is key for feasibility studies, preliminary design, and process/production planning. Recent studies have provided the possibility of neural network (NN)-based build time estimation. In particular, traditional artificial NN (ANN)- and convolutional NN (CNN)-based methods have been demonstrated. However, very little has been done on the performance comparison for build time estimation among the different types of NNs. This study is aimed at filling this gap by designing various NNs for build time estimation and comparing them. Two types of features are prepared as inputs for the NNs by processing three-dimensional (3D) models: (1) representative features (RFs) including dimensions, part volume, and support volume; and (2) the set of voxels generated from designating the cells occupied by the workpiece in a mesh grid. With the combination of NN types and input feature types, we design three NNs: (1) ANN with RFs; (2) ANN with voxels; and (3) CNN with voxels. To obtain large enough label data for reliable training, we consider simulation build time from commercial slicing applications rather than actual build time. The simulation build time is calculated based on a material extrusion process. To address various cases for input models, two design factors (scale and rotation) are considered by controlling the size and build orientation of 3D models. In computational experiments, we reveal that the CNN-based estimation is often more accurate than others. Furthermore, the design factors affect the performance of build time estimation. In particular, the CNN-based estimation is strongly influenced by changing the size of 3D models.

The study investigates the impact of build orientation policies on the production time in additive manufacturing (AM) for mass customisation business models. Two main orientation policies are considered: (1) Laying Policy (LP) that focuses on reducing the height of parts; and (2) Standing Policy (SP) that aims to minimise the projection base plane of parts to reduce the number of jobs. While LP minimises the build time per job since parts have low height, it could increase the total completion time as the number of parts increases. On the other hand, SP takes longer build time per job due to the high height of parts, where it could lead to a fewer number of jobs. Several numerical experiments have been conducted based on Stereolithography (SLA). The results show that, when the number of parts is experimentally about 40, SP could be more preferred than LP for minimising the completion time where the shape tendency of parts is likely to affect the extent of preference for the policies. When 40 parts with long and flat shape are considered, SP reduces the completion time by 15.7% over the default policy, the initial orientation of a part, while LP reduces by only 6.6%.