Non-Destructive Damage Detection Using Laser Doppler Vibrometry and Machine Learning Statistical Techniques

Due to its ability to manufacture a single, complex part that either (1) could not have been built with traditional manufacturing or (2) would require the assembly of several sub-components, the use of Metal Additive Manufacturing (MAM) has increased over 300% in just the last 5 years(1,2). Even with the advancements in the technology of Metal Additive Manufacturing, the final as-built mechanical properties continue to possess variability that make it difficult for designers and engineers to utilize them effectively, especially on mission critical parts. The final material properties of an MAM part are process and operator-dependent and the MAM process is more complex than traditional metal manufacturing techniques. The ability to non-destructively validate the quality of an MAM part is critical as the utilization of metal these parts increases. This paper describes ongoing research that focuses on techniques that can predict if a given part meets customer requirements and is free from hidden defects, flaws, non-visible geometric fluctuations, and variations in mechanical properties that are critical to the design/analysis process and mission certification. The techniques use Laser Doppler Vibrometry to detect the frequency response spectrum for a given part then process that data with non-p-value statistical techniques (Machine Learning) to develop models that can be "trained" to detect a variety of quality issues. Machine Learning models work best when data from many samples with different unacceptable end-states are available and used to train and develop the models. These end-states could be the result of a variety of unacceptable variations in the MAM process, quality of the precursor powdered metal, quality of the millions of micro-welds that make up the MAM build process, size and number of inclusions and micro flaws in the material as well as residual stresses in the part. In this study, our goals were to (1) use a large enough number of parts to ensure the validity of the machine learning model and (2) focus on geometric defects and the effects of residual stresses. For these initial investigations, we utilized Off-The-Shelf parts to increase statistical reliability, improve our ability to train our Machine Learning Model and refine our experimental protocol while keeping costs within the budget for the research. In the second phase of this research, we will use MAM parts.