August 2023 Volume 5

FORGING RESEARCH

Conclusions The present study showed that it is possible to use Industry 4.0 technology (and more specifically neural networks) to predict both complete fill/incomplete fill and monitor parts for surface defects in a forging environment. These outputs can be used to better monitor, automate, and improve the forging process. It is likely that other quality attributes and defects can be captured; this study is only a representative example of what can be accomplished. Furthermore, it does not appear that there is a single solution to setting up a neural network system for a forging process. Based on the power and flexibility that a neural network offers, it should be possible for companies to adapt and use a neural network strategy and set-up that best fits their specific needs. Specifically, the study revealed three key points: 1. Discrete and time-series data provided similar accuracy for the axisymmetric geometry considered. While additional confirmation is warranted, this suggests that companies will be able to use an approach that best meets their needs and available resources. 2. The training data mix is critical for accurate ANN predictions and should contain sufficient proportions of good and defective part conditions. 3. Vision based ANNs can use archival, non-specific geometry data for training purposes and facilitate system training. These are suited for detecting surface defects that are not directly linked to the deformation phase of the process or are generated on a casual basis. Some challenges remain before widespread neural network implementation can be realized at the forge plant level. The present study shows that Industry 4.0 will require more attention to e-data collection and analysis in the forging process. Vision hardware implementation can be in-situ or dedicated station. The latter facilitates and enables simpler techniques/algorithms to get good images for processing though it appears likely that in situ monitoring will be commonplace as image algorithm capabilities develop. For further information regarding the project or to learn more about participating in the technology deployment phase, please contact Prof. Joseph Domblesky via email at joseph.domblesky@marquette. edu or by telephone at 414-288-7832. References 1. https://www.mckinsey.com/capabilities/operations/our-insights/ covid-19-an-inflection-point-for-industry-40 2. https://openai.com/blog/chatgpt 3. Hochreiter, S., & Schmidhuber, J. (1997). Long Short-Term Memory. Neural Computation, 9(8), 1735–1780. https://doi. org/10.1162/neco.1997.9.8.1735 Acknowledgment This research is sponsored by the Defense Logistics Agency Information Operations, J68, Research & Development, Ft. Belvoir, VA, and by DLA-Troop Support, Philadelphia, PA.

Table 3. Confusion matrix showing ANN predicted and actual part fill/ incomplete fill for the forged aluminum part using load-stroke curves.

The results from the vision-based systems were found to be effective in detecting various surface defects under different conditions though the team is continuing to test the system to define its resolution and confirm its robustness. One example is shown below in Figure 7 using plasticine material which had a number of small and large scratches introduced on the surface. The actual piece, a cylindrical billet, is shown on the left and the processed image is shown on the right where the lighter regions indicate surface imperfections. It can be seen that the ANN algorithm was successful in defecting the defects and was not impacted by the surface curvature.

Figure 7. Comparison of actual workpiece surface (left) and digitized surface showing defects as light colored lines (right).

FIA MAGAZINE | AUGUST 2023 71