| Title |
A Lightweight Deep Learning Model based on Image Encoding for Failure Classification of Motor Mechanical Facilities |
| Authors |
안동주(Dongju An) ; 신재광(Jaegwang Shin) ; 이수안(Suan Lee) |
| DOI |
https://doi.org/10.5573/ieie.2022.59.5.57 |
| Keywords |
Timeseries classification; Image encoding; Deep learning; Lightweight model |
| Abstract |
The failure of mechanical facilities used in industrial sites accounts for a significant portion of bearings, rotators, belts, and axes. When facilities fail due to mechanical or electrical causes or performance degrades, vibration is commonly generated and current or the like shows abnormal movement. It is essential to easily detect and predict failures that occur unspecified in this situation. Therefore, in this paper, a lightweight deep learning model was proposed using a method of encoding time series data generated by a sensor attached to a mechanical facility into an image. Three methods were used for image encoding, and a CNN-based deep learning classification model was created for each method. The CNN model was created through experiments with the most accurate model with small parameters. When analyzing the accuracy of the CNN model and experimenting with which encoding methods are more efficient and suitable for learning, it was confirmed that the GASF method was generally more accurate among the three image encoding methods. It is expected that various sensor data used in the industry can be used for various applications using the lightweight deep learning model based on image encoding proposed in this paper. |