Automated Anomaly Detection of Laser-Based Additive Manufacturing Using Melt Pool Sparse Representation and Unsupervised Learning

Advanced thermal imaging is increasingly invested in direct energy deposition (DED) additive manufacturing (AM) to cope with information visibility of melt pool and tackle process inconsistency. However, there are key challenges regarding the feasibility of current image-guided monitoring methodologies in the DED process. First, high-resolution thermal images consist of millions of pixels captured by hundreds of frames lead to the curse of dimensionality in analysis. Second, the presence of various exogenous noise, ill-structured data, and significant cluster imbalance limit the capability of the current methodologies to perform real-time monitoring. The objective of this research is to advance the frontier of melt pool monitoring in DED process by designing an automated and unsupervised anomaly detection on high-dimensional thermal image data. Specifically, we develop a variational autoencoder to generate a low-dimensional representation of each input thermal image data. A Gaussian mixture model and K-Mean clustering are integrated with the generative model to split latent space into homogenous regions and detect anomalies. Experimental results show that the proposed methodology is highly effective in detecting defective melt pools with accuracy up to 94.52% and a false alarm rate of less than 2.1%.