Search

Big data plays an important role in the fourth industrial revolution, which requires engineersand computers to fully utilize data to make smart decisions to optimize industrial processes. Inthe additive manufacturing (AM) industry, laser powder bed fusion (LPBF) and direct metallaser solidification (DMLS) have been receiving increasing interest in research because of theiroutstanding performance in producing mechanical parts with ultra-high precision and variablegeometries. However, due to the thermal and mechanical complexity of these processes, printingfailures are often encountered, resulting in defective parts and even destructive damage to theprinting platform. For example, heating anomalies can result in thermal and mechanical stresson the building part and eventually lead to physical problems such as keyholing and lack offusion. Many of the aforementioned process errors occur during the layer-to-layer printing process,which makes in-situ process monitoring and quality control extremely important. Although in-situsensors are extensively developed to investigate and record information from the real-time printingprocess, the lack of efficient in-situ defect detection techniques specialized for AM processesmakes real-time process monitoring and data analysis extremely difficult. Therefore, to helpprocess engineers analyze sensor information and efficiently filter monitoring data for transport and storage, machine learning and data processing algorithms are often implemented. These algorithmsintegrate the functionality of automated data processing, transferring, and analytics. In particular,sensor data often takes the form of images, and thus, a prominent approach to conducting imageanalytics is through the use of convolutional neural networks (CNN). Nevertheless, the industrialutilization of machine learning methods often encounters problems such as limited and biasedtraining datasets. Hence, simulation methods, such as the finite-element method (FEM), are usedto augment and improve the training of the deep learning process monitoring algorithm.Motivated by the above considerations, this dissertation presents the use of machine learningtechniques in process monitoring, data analytics, and data transfer for additive manufacturingprocesses. The background, motivation, and organization of this dissertation are first presentedin the Introduction chapter. Then, the use of FEM to model and replicate in-situ sensordata is presented, followed by the use of machine learning techniques to conduct real-timeprocess monitoring trained from a mixture of experimental and replicated sensor image data. Inparticular, a cross-validation algorithm is developed through the exploitation of different sensoradvantages and is integrated into the machine learning-assisted process monitoring algorithm.Next, an application of machine learning (ML) to non-image sensor data is presented as a neuralnetwork model that is developed to estimate in-situ powder thickness to account for recoater arminteractions. Subsequently, an integrated AM smart manufacturing framework is proposed whichconnects the different manufacturing hierarchies, particularly at the local machine, factory, andcloud level. Finally, in addition to the AM industry, the use of machine learning, specificallyneural networks, in model predictive control (MPC) for dynamic nonlinear processes is reviewedand discussed.

Additional file 1: Real-time quantitative PCR primer sequences used in this study. PCR, polymerase chain reaction; bp, base pair.

Additional file 1. PRISMA checklist and flow diagram