Fault diagnosis of wind turbine bearing using a multi-scale convolutional neural network with bidirectional long short term memory and weighted majority voting for multi-sensors

In order to solve the problems of insufficient extrapolation of intelligent models for the fault diagnosis of bearings in real wind turbines, this study has developed a multi-scale convolutional neural network with bidirectional long short term memory (MSCNN-BiLSTM) model for improving the generalization abilities under complex working and testing environments. A weighted majority voting rule has been proposed to fuse the information from multi-sensors for improving the extrapolation of multisensory diagnosis. The superiority of the MSCNN-BiLSTM model is examined through experimental data. The results indicate that the MSCNN-BiLSTM model has 97.12% mean F1 score, which is higher than existing advanced methods. Real wind turbine dataset and an experimental dataset are used to demonstrate the effectiveness of the weighted majority voting rule for multisensory diagnosis. The results present that the diagnosis result of the MSCNN-BiLSTM model with weighted majority voting rule is higher respectively 1.32% and 5.7% than the model with traditional majority voting or fusion of multisensory information in feature-level.