Deep representation learning for process variation management in laser powder bed fusion

Laser Powder Bed Fusion (LPBF) is an additive manufacturing process where laser power is applied to fuse the spread powder and fabricate industrial parts in a layer by layer fashion. Despite its great promise in fabrication flexibility, print quality has long been a major barrier for its widespread implementation. Traditional offline post-manufacturing inspections to detect the defects in finished products are expensive and time-consuming and thus cannot be applied in real-time monitoring and control. In-situ monitoring methods by relying on the in-process sensor data, on the other hand, can provide viable alternatives to aid with the online detection of anomalies during the process. Given the crucial importance of melt pool characteristics to the quality of final products, this paper provides a framework to process the melt pool images by a configuration of Convolutional Auto-Encoder (CAE) neural networks. The network's corresponding bottleneck layer learns a deep yet low-dimensional representation from melt pools while preserving the spatial correlation and complex features intrinsic in the images. As opposed to the manual annotation of data by X-ray imaging or destructive tests, an agglomerative clustering algorithm is applied to these representations to automatically extract the anomalies and annotate the data accordingly. A control charting scheme based on Hotelling's T² and S² statistics is then developed to monitor the process's stability by keeping track of the learned representations and residuals obtained from the reconstruction of original images. Testing the proposed methodology on the collected data from an experimental build demonstrates that the method can extract a set of complex features that are inextricable otherwise by using hand-crafted feature engineering methods. Moreover, through extensive numerical studies, it is shown that the proposed feature extraction and statistical process monitoring scheme is capable of detecting the anomalies in real-time with accuracy and F₁ score of about 95% and 82%, respectively.