Abstract:Device RUL prediction is important in improving device reliability, safety, and reducing maintenance costs. By discovering the health status and potential faults of device in advance, RUL prediction helps reduce the risk of sudden failure, extend device life, improve work efficiency, and ensure normal operation of tasks. However, in the face of the increasing complexity of device and the increasing dimensionality of collected sensor data, traditional methods and certain deep learning methods have limitations in handling feature relationships, long time series data and mining important sensor data. In order to improve the prediction accuracy, a hybrid model based on MWCEN combined with Informer - MWCEN-Informer is proposed. MWCEN encodes the device time series data with dynamic work condition coding algorithm, fully extracts the feature information by one-dimensional multi-scale hybrid convolution of the device sensor information, the effective features are enhanced using the multi-branch channel attention mechanism, and the enhanced sensor data are input into Informer for analysing the correlation of the device sensor timing data to achieve more accurate device RUL prediction. Validation is carried out with a generic turbofan engine data set based on C-MAPSS, and the results show that the model reduces the RMSE by an average of 5.5% and the S-Score by an average of 4.7% on the four subsets, which effectively improves the RUL prediction accuracy of the device under complex operating conditions and complex faults.
