비정상 신호를 가진 연료 펌프의 상태 진단에서 Conv1D 모델의 시간 간격 별 예측 정확성 비교

초록

This study deals with the condition detection of centrifugal pump supplying liquid to marine engines. The purpose is to detect the anomaly and the failure modes of the pump by applying the vibration signals from the simulated failures on the experimental bed. Since the pump rarely experiences faults as well as used to degradation experiments, the vibration signal was gathered from an experimental bed simulating the bearings and lubricant failures. Considering the non stationarity of variable speed, the convolution is applied to extract the feature of time series rather than the frequency feature. The advantage of CNN implicitly extracting features from a non-stationary signal is used to extract the features applied to Conv1D. After learning the features, a multi-layer perceptron (MLP) was connected to failure classification to identify the operation state. Finally, it is suggested that the time series-based feature extraction can be applied to the condition monitoring of a centrifugal pump with variable speed and non-stationarity

키워드