LSTM 오토인코더를 이용한 라디에이터 고장 진단 및 설비보전 사례연구

초록

The radiator is essential equipment to release heat losses during the operation of the facility. In this study, the direction of facility conservation of the radiator was suggested using the LSTM autoencoder model, one of the deep learning algorithms. Vibration signals were acquired during the test while the damage of the radiator was simulated by random vibration test on the vibration bench. Data preprocessing was performed by extracting statistical characteristics of time series vibration signals data. To diagnose the operations of the radiator, the reconstruction error(MSE: Mean Square Error) of LSTM autoencoder model is configured as the fault level index. The specific MSE of the training data regular distribution was set as a threshold. If the MSE of the validation data exceeds the threshold, it was diagnosed as a fault status. In the result of the experiment with changing the threshold level and optimizing parameters, optimal precision-recall tradeoff relationship was checked. In addition, Area under cover was 0.99 in ROC curve depending on the threshold level. So, anomaly detection was performed successfully. We can see that the data area without a label could be diagnosed accurately through the learned deep learning model.