Detection of Process Variation in a Cold Forging Process through Smart Manufacturing
Author(s): Vignesh Selvaraj, Andrew Glaeser, Kangsan Lee, Namjeong Lee, Yunseob Hwang, Sooyoung Lee, Seungchul Lee, and Sangkee Min
Publication: International Symposium on Precision Engineering and Sustainable Manufacturing (PRESM)
Citation: Vignesh Selvaraj, Andrew Glaeser, Kangsan Lee, Namjeong Lee, Yunseob Hwang, Sooyoung Lee, Seungchul Lee, and Sangkee Min, “Detection of Process Variation in a Cold Forging Process through Smart Manufacturing,” International Symposium on Precision Engineering and Sustainable Manufacturing (PRESM), Online symposium, Korea, November 15 – 18, 2020.
Defective products in the manufacturing process induce a large loss annually. Employees on the manufacturing inspect the products in order to detect defectives but it is impossible to inspect every product and it costs too much. With current progress of electronics and communication technology, new and effective methods can be used which is so-called Smart Manufacturing. Thus, we installed a remote monitoring system on a cold heading machine that produces automotive fasteners. Prediction of tool wear is the eventual goal of this study, but we first examined the monitoring system and analyzed the characteristics of acquired data by AI methods. In this paper, we suggest using accelerometer on the purpose of management of a cold forging machine and testify monitoring system what we installed with various data driven analyses in order to see the feasibility. We conducted a few preprocessing methods on 3-axis vibration data such as wavelet transform, temperature calibration, and extracting features with randomly generated functions, which results in useful data having various forms. Furthermore, we investigated part classification, observation of signal changing, and detecting die deterioration using conventional machine learning algorithms such as CNN. The designed CNN model achieved part classification accuracy as high as 87% when looking at data for 1 month. Binary classification with data from the beginning and end of the day was also conducted to see the variation of signal over time within a day, which achieved a classification accuracy of 80% meaning the data is varying. Moreover, we did the same thing for 3-day data and plotted a graph that is similar to the Flank tool wear curve. With the temperature calibrated data, we trained a model to do binary classification for this dataset. We propose how to plot a tool wear graph using predicted accuracy.