Anomaly detection methods using melt pool imaging in laser powder bed fusion

Abstract

In recent years, much interest is invested in metal additive manufacturing (AM) for its flexibility in fabricating a wide range of complex geometries infeasible for conventional subtractive manufacturing. The implementation of metal AM is restricted for its propensity for process defects, diminishing the quality and consistency of manufactured components. Accordingly, it is critical to detect when these faults appear, ensuring informed process knowledge is maintained, and mitigating the material and sunk-time cost of failed prints. Therefore, research efforts have been directed to metal AM monitoring methods and the detection methodologies that use this information. While complex machine learning approaches can leverage monitoring data fully to detect anomalies within a metal AM print, these methods have limited transferability and require large quantities of training data, slowing its deployment to a metal AM system. Although these approaches are accurate, these systems lean on extensive training datasets to function; this is a lengthy collection process in metal AM and may not be feasible when considering difficulties in labeling unique faults. Furthermore, these approaches rely on obscure features within the monitored data and unclear decision metrics, limiting generalizability of these methods to geometries and faults contained in the training dataset. Last, many approaches do not incorporate the dynamic behavior of the system, instead evaluating signals in isolation without context. To address these concerns, the work in this thesis aims to apply simple unsupervised detection approaches, enabling minimal training lead time to deployment in a laser powder bed fusion (LPBF) metal AM system. First, implementation of coaxial melt pool monitoring is discussed. An implementation of a thermal camera to capture the temperature of the melt pool is presented, allowing for more complete process knowledge in the experimental LPBF testbed setup. A comparison of the temperature to intensity measurements highlighted improvements to the sensing approach. With respect to applying simple detection methods, the process is compressed to a time series consisting of a characteristic feature of the melt pool. This compression highlights a physical, interpretable aspect of the melt pool while enabling online detection through data reduction. Two methods are then applied to represent this time-dependent characteristic: the spectrogram of the data that tracks the frequency evolution against time, and an autoregressive (AR) model that performs predictions based on the history of the melt pool time series. The nominal behavior of this feature is learned by leveraging the time-series evolution. Statistical tests are applied to both methods to characterize the statistics of the nominal response, declaring faults when responses do not fall within the expected nominal statistical distribution. Several statistical tests are assessed and evaluated on a dataset containing both naturally occurring and injected faults. Through the simplicity of the feature compression, time-evolution modeling, and the statistical tests, only minimal training data is necessary, equivalent to a single test print before detection algorithm deployment. Lastly, a study on selecting an optimal compression feature is presented, along with a investigation into several simple deep learning models under the same training dataset limitation to compare the efficacy of the presented detection algorithms.