In-line monitoring of complex sensor data for zero-defect additive manufacturing in the electromobility industry

The rise of additive manufacturing across various industrial sectors has brought new challenges in qualifying novel processes and complex products. At the same time, it has created opportunities by providing extensive data for enhanced quality assurance. Different types of sensor data streams (e.g., images, videos, acoustic emissions, thermal measurements) are potentially available in-line and in-situ, while the part is being produced. This imposes the need to rethink statistical and machine learning methods to enable a fast and effective identification of anomalies and defects since their onset stage, making sense of big and complex sensor data. In this study, we specifically explore novel methods for in-line and in-situ monitoring of a laser powder bed fusion process (L-PBF), where acoustic data are collected to detect cracking events in the manufactured part. The process has its own acoustic signature in the time-frequency domain that exhibits natural variations along the different phases of the layerwise printing process. Low-dimensional learning methods are explored to effectively separate out-of-control events (i.e., cracks) from the underlying data variability, and to enhance the detection of real defects. A real case study is presented, which involves the use of additive manufacturing to fabricate high-performance cooling structures directly on pre-existing substrates for advanced power electronics performance. Substrates include a ceramic layer which is subject to crack formation because of the heating and cooling cycles imposed by the laser melting of the copper structure. Enabling a fast and accurate detection of such cracking events is of paramount importance to bring this novel additive manufacturing application in a real series production.