Applications of smart manufacturing in industrial cold forging
File(s)
Date
2024-03-27Author
Glaeser, Andrew
Advisor(s)
Min, Sangkee
Metadata
Show full item recordAbstract
A recent paradigm shift in manufacturing has seen many companies move toward smart manufacturing, where intelligent decisions are made based on high volume data analysis with goals of reducing/eliminating down time, improving operating cost, and achieving factory autonomy. With the emergence of Internet of Things (IOT), large scale data collection is possible which provides an opportunity to gain insights into many processes. Advancements in computing have led to the development and application of many analysis tools, such as deep learning and machine learning models, which allow for advanced pattern recognition and can be used in many diverse ways. Application of smart manufacturing has not been exhaustively researched, and one area with potential to benefit from smart manufacturing is industrial cold forging. Cold forging is the process of forming metal into a desired geometry through use of dies and punches. Due to the extreme loads placed on cold forging tools, tool failure is inevitable. Detecting tool failure is crucial to product integrity, as even subtle tool failures can cause product defects that are reason for reject. Some cold forging processes produce parts at speeds up to several parts per second. In these processes, early tool failure detection is imperative in reducing cost of scrap. In this study, the feasibility of applying smart manufacturing in industrial cold forging was investigated. First, a process monitoring system was designed and installed in a cold forging factory. The collected vibration data was used to classify variation on the machine resulting from the different part numbers that were produced by it. After transforming the data using wavelet transform, a Convolutional Neural Network (CNN) was trained and achieved testing accuracy of 98%. Additional data was used to evaluate the performance of the model; with novel data the CNN model achieved 71% accuracy. Based on these observations, it was concluded that the vibration data contained meaningful features that allow for variation detection. In order to investigate if the proposed method could classify conditions resulting in defective products, controlled experiments were conducted where 6 commonly encountered failure conditions were intentionally induced on a cold forging machine. In one analysis, all failure conditions were combined into one class, and a CNN was trained to classify between good conditions and failure conditions. In another analysis, a CNN model was trained to classify between failure conditions. The highest observed accuracy was 98% and 95%, respectively. The results indicate that the proposed method for classification has high potential to detect failures in real-time applications, but further research is required to determine how the model will perform with novel data and to bring the system to real-time capability. Additional research was conducted on the feasibility of using an Acoustic Emission (AE) sensor to enhance the existing method for purposes of failure prediction. A mock-up assembly was constructed to simulate a simplified cold forging process where both an AE sensor and an accelerometer were mounted. Data was taken with two different tool conditions, cracked and uncracked. A CNN model trained with AE sensor data achieved 98% accuracy when classifying the two conditions, while a CNN model trained with acceleration data achieved 91% accuracy. It was further discovered that the accelerometer data is more sensitive to noise after artificial white noise was added to each signal. A final test showed that the model trained with AE signal has decreased performance if the sensors are move to a location away from the event. Based on these observations, an AE sensor may be able to provide insightful data if located close to the event. Application of an AE sensor will be explored further in additional research.
Subject
Mechanical Engineering
Permanent Link
http://digital.library.wisc.edu/1793/85097Type
Thesis