Robust Deep Learning Algorithms for Selective Laser Melting Monitoring
The objective of this research is to enable real-time in-situ monitoring for the Selective Laser Melting (SLM) process, by providing diagnostic feedback from monitoring that can be used to automate and adjust SLM system parameter settings. The ultimate goal is to improve SLM product quality and manufacturing productivity. We propose a deep learning approach to monitoring that takes in-situ videos as input data fed to convolutional neural networks (CNNs). We describe the entire monitoring framework, including running SLM experiments, collecting SLM video data, image processing for ex-situ generated height maps, generating labels for in-situ data from ex-situ measurement, and training CNNs with labeled in-situ video data. Experimental results show that our approach successfully recognizes the desired SLM process metrics (e.g. size, continuity) from in-situ video data.
In order to train effective CNNs, besides collecting extensive SLM video data, we also need to label it. We have automated the process of generating labels from ex-situ measurements of the corresponding finished SLM experimental output. The ex-situ measurements provide high-precision height maps for the product surface, to which we apply our proposed image processing algorithm to calculate process quality metrics as labels.
However, our proposed automated labeling approach requires high-precision height maps, which are generated from an expensive Structured Light Microscope. It might not be readily available to other researchers and institutions, or not enough machine time may be available to label all experiments for which there is video. Thus there might not be enough labeled data to train effective CNNs. This research also combines semi-supervised learning with our original approach to address this problem. Semi-supervised learning method enables other researchers to address the problem without requiring a huge amount of labeled data.
In addition, in practice, another issue is label noise. Even though the data labels were generated using high-precision height maps, the labels are not perfect and might still contain incorrect labels, known as ``noisy'' labels. We propose novel approaches to improve neural networks' performance when they are trained under label noise. The proposed approaches can be easily combined with other existing approaches that address the label noise problem to further improve the prediction accuracy, with very few additional hyperparameters that need to be tuned. Experimental results demonstrate that our approaches can significantly improve CNN models' prediction accuracy when training neural networks with noisy labels.