Session: 01-13-01: Congress-Wide Symposium on NDE & SHM: Computational Nondestructive Evaluation and Structural Health Monitoring Count

Paper Number: 94486

94486 - Unsupervised Online Anomaly Detection of Metal Additive Manufacturing Processes via a Statistical Time-Frequency Domain Approach 

Additive manufacturing (AM) is a nascent field of manufacturing that relies on the addition rather than the removal of material to construct a desired article. One method of AM is laser powder bed fusion (LPBF) which uses metal powder and a high-powered laser to adhere layers of molten metal. As with many methods of AM, smaller imperfections during the construction process can lead to major repercussions in part quality in subsequent layers, often resulting in a failed build or severe structural faults. Consequently, determining a robust fault detection process is of great interest. Of the varied methods of fault detection, online defect detection shows promise for quickly assessing build errors. The ability to quickly determine if and when a fault has occurred in an AM build is critical for future work in the field. 

In this work, a novel framework for identifying process faults in-situ is proposed. This fault detection methodology is based on using real-time measurement of the melt pool thermals during the AM process, and consists of the following key components: (1) analysis of the time-frequency content via spectrogram, (2) development of a nominal model constructed from nominal operation data, and (3) a comparative metric quantifying the deviation of current signals from the expected response.

To accomplish objective (1), the thermal signal is converted to the frequency domain using the fast Fourier transform (FFT), giving the spectrogram response. The frequency domain is critical in examining certain patterns that occur as the laser returns to recurrent positions. Faults which disrupt the repetition of the signal can be observed in the spectrogram. By measuring the deviation of a test spectrogram from the expected nominal case, we can form a detection framework. A principal component analysis (PCA) is performed to obtain a healthy basis from known nominal data, forming target (2). A reduced quantity of spectrogram PCs are then combined with this baseline to form a reconstructed spectrogram. Lastly, goal (3) is accomplished by taking the root mean squared error (RMSE) as a classification metric. Intuitively, spectrograms similar to the baseline phase will exhibit low reconstruction errors, whereas anomalies with errant frequency responses will struggle given limited PCs. From a separate nominal dataset, we determine the statistics of the expected performance of the reconstructed model and determine a statistical threshold. This method notably requires no labeled data, as faulty responses will naturally deviate from nominal responses. Signals which fall outside of this statistical threshold are determined to be faulty, as the reconstruction metric is expected to be statistically unlikely to fall outside this confidence bound if the spectrogram responses are similar.

Utilizing this testing framework, a model with satisfactory accuracy was obtained. This metric was determined through true positive and false positive rates. Individual tolerances were determined for each unique raster pattern. The primary types of faults detected were over/under-melting faults, which resulted in larger and insufficient melt pools, respectively. This methodology can detect faults with good accuracy for varied scan geometry and shows good promise in scaling to different geometry. 

Presenting Author: Alvin Chen Rensselaer Polytechnic Institute

Presenting Author Biography: Alvin Chen is a second year PhD student at Rensselaer Polytechnic Institute. He is currently performing research on in-situ fault detection for metal additive manufacturing. His current areas of interest are: additive manufacturing, process control, and control systems.

Authors:

Alvin Chen Rensselaer Polytechnic Institute
Fotis Kopsaftpoulos Rensselaer Polytechnic Institute
Sandipan Mishra Rensselaer Polytechnic Institute