| Title |
A Study on the Practical Implementation Technique of Motor Gear Fault Diagnosis Using a Mix-Up Algorithm and Auto Encoder |
| Authors |
최의진(Eui-Jin Choi) ; 한지훈(Ji-Hoon Han) ; 송승민(Seung-Min Song) ; 홍선기(Sun-Ki Hong) |
| DOI |
https://doi.org/10.5370/KIEE.2024.73.1.87 |
| Keywords |
Deep Learning; Motor fault diagnosis; Auto-encoder; Anomaly detection; IoT; Moving average; Mix-up |
| Abstract |
In this paper, a practical data acquisition method and fault diagnosis method for applying motor fault diagnosis in the field are studied. Since it is very difficult to acquire fault data in the field, unsupervised learning methods that can be trained using only normal state data are mainly used. However, unsupervised learning methods are very vulnerable to disturbances and are difficult to express the fault level. Disturbances in the field include electrical noise, which can cause data acquisition devices to fail due to low power quality in the field and affect the measured signal, and mechanical noise, such as external vibrations, which directly affect the operation of the electric motor. To minimize the impact of electrical noise, it is common to use hardware filters. Unlike previous studies that use theoretical cutoff frequency setting methods, this paper proposes a method to find the appropriate cutoff frequency using LRP(Layer-wise Relevance Propagation) analysis, one of the XAI(eXplainable A.I) techniques, for practical data acquisition.
Fault diagnosis in the field requires robust unsupervised learning algorithms that can ignore the presence of mechanical noise in the signal, such as transient shocks. This is solved by post-processing the output of the auto-encoder with a moving average filter. To represent the fault level a data generation technique, the mix-up algorithm, is used. A method is proposed to threshold the auto-encoder multiple times with data generated by the mix-up algorithm. This proposed method shows the availability of motor fault diagnosis considering the on-site data. |