Missing-Data Filling Method Based on Improved Informer Model for Mechanical-Bearing Fault Diagnosis

Missing data, which are inevitable in real-time data monitoring and acquisition systems for mechanical bearings, will degrade the detection accuracy of the bearing working state. Incomplete observations in a target dataset with missing data are usually discarded, causing loss of some data features and inaccurate bearing-fault diagnosis results. To avoid this problem, we propose a long sequence time-series filling method named Di-Informer, which is based on an improved Informer model. After improving the weight and structure of the Informer model, we employed the model as the generator in the overall network structure and constructed a corresponding discriminator and loss function to form the final generative neural network Di-Informer, which can fill the missing data of a mechanical rolling-bearing monitoring system. The effectiveness of the Di-Informer model was verified through a series of training and testing processes on the bearing-life dataset of our laboratory research group and on publicly available data from the University of Case Western Reserve (CWRU).