| Summary: | 碩士 === 國立高雄科技大學 === 電機工程系 === 107 === Collecting historical data from normal to fail is necessary to effectively build a failure prognosis model for diagnosing failures of a metal fastener forming process. However, it takes long time for collecting the failure progress of the forming die and is difficult to obtain and diagnose the gradual information of various failure modes of fastener forming process. The fastener forming industry is challenging how to stabilize production by reducing failure cost and time to repair.
This study develops a failure prognosis system with three characteristics dummy sample generation, feature auto-extraction, and failure modes diagnosis for a fastener forming process. In dummy sample generation, samples with gradual fail can be generated by mixing different proportions of failure samples using GAN (generative adversarial nets). The abnormal signals can be filtered and features are auto-extracted using autoencoder. With specified models built by neural network and random forest, various failure modes can be diagnosed. Finally, an online web process monitor is provided via MQTT protocol.
The research results indicate that the accuracies are up to 95% when diagnosing failure modes including abnormal length, core notch, cavity adhesion, and ill lubrication in the fastener forming process. The MAPEs are between 4.3% and 11.7% when estimating degrees of failure modes core notch and cavity adhesion. Therefore, this system is promising to monitor and estimate die states, and achieves the goal of estimating forming process failures.
|
|---|
