Summary: | 碩士 === 國立高雄第一科技大學 === 電機工程研究所碩士班 === 106 === Die wear, e.g. changes of die dimensions and forming space, can be caused from accumulated impacts and frictions in the forming stage of producing metal fastener. It could reduce forming accuracy and increase production costs due to high defect rates. Because observing the die internal structure is inefficient, experienced experts are needed to adjust or replace the die by observing the fastener external quality. Besides, the poor production environment induces high employee turnover. The challenge of the fastener forming industry is how to automatically and intelligently estimate die state for achieving stable fastener production.
This study develops a die state estimation system for on-line monitoring die states of a fastener forming machine. In system architecture, the system can automatically separate a critical pressure interval from the sensing loads in forming, and extracts the features including the time and frequency domains by interval. Based on the pressure features, a model for estimating die state can be established by the deep neural network classifier and the autoencoder with Tensorflow. This neural architecture will be used to verify the die state definition. Finally, a web interface through the MQTT protocol is provided allowing users to remotely monitor the die states according to the estimation model.
In research results, taking 209,283 times of forming data of the key stage of a fastener machine as an example, the estimation accuracy of die state nears 95% after extracting 50 MB feature dataset from 1.15 GB raw data. In addition, this system can detect lubrication issue and the fastener length can be estimated correctly by using the corresponding features and models. Therefore, this system is promising to monitor and estimate die states, effectively.
Keywords: die state, pressure sensing feature, deep neural network, autoencoder.
|