A Transformer-based Approach for aAnomaly Detection in Wire eElectrical Discharge

Hammed, Waleed; Al-Rubaye, Ameer H.; Bashar, Bashar S.; Imran, Merzah Kareem; Rzooki, Mustafa Ghanim; Hashesh, Ali Mohammed

doi:10.30486/mjee.2022.696524

Original Article

A Transformer-based Approach for aAnomaly Detection in Wire eElectrical Discharge

Authors

Medical technical college, Al-Farahidi University, Baghdad, Iraq
Department of Petroleum Engineering, Al-Kitab University, Altun Kupri, Iraq
Department of Medical instruments engineering techniques, Al-Farahidi University, Baghdad,10021, Iraq
Building and Construction Engineering Technology Department, AL-Mustaqbal University College, Hillah 51001, Iraq
Medical Device Engineering, Ashur University College, Baghdad, Iraq
Al-Hadi University College, Baghdad,10011, Iraq

Abstract

Although theoretical models of manufacturing processes are useful for understanding physical events, it can be challenging to apply them in real-world industrial settings. When huge data are accessible, artificial intelligence approaches in the context of Industry 4.0 can offer effective answers to real production challenges. Deep learning is increasingly being used in the realm of artificial intelligence to address a variety of issues relating to information and communication technology, but it is still limited or perhaps nonexistent in the industrial sector. In this study, wire electrical discharge machining—a sophisticated machining technique primarily used for computer hardware components—is applied to effectively forecast unforeseen occurrences. By identifying hidden patterns in process signals, anomalies, such as changes in the thickness of a machined item, may be efficiently anticipated before they occur. In this study, a model for anomaly detection in the sequence of thickness change in the machined component based on transformers is suggested. Our method is able to achieve 94.32 % and 94.16 % accuracy in Z 135 and Z 15 datasets, respectively. Also, it forecasts the abnormalities inside the sequence 1.1 seconds in advance, according to our tests on a dataset that has been introduced.