Deep Learning Fault Diagnosis Based on Model Updation in Case of Missing data

The sampling frequency of different sensor used to collect data may be different, which will result in a structure incomplete sample at a particular sampling point. It is a kind of data missing problem. Deep learning based fault diagnosis model may be inaccurate because there are fewer well-structured samples that can be used to train the DNN based fault diagnosis model. In this paper, the potential cross-correlation between missing variables and existing variables is used to obtain additional well-structured samples by establishing an interpolation model based on BP neural network. Using the new well-structured samples, an online update mechanism of the DNN fault diagnosis model is designed to update the parameters of DNN. It is effective to get more accurate fault diagnosis result since more structure incomplete samples is used in the training process. The experimental results show that the method proposed in this paper can effectively improve the accuracy of fault diagnosis in the case of missing data.