A Bearing Fault Diagnosis Method Based on Branch Convolution Neural Network

Rolling bearings are one of the most important components of rotating machinery, to ensure stability, reliability, and safety of its operation, intelligent fault diagnosis plays a vital role. In the industry, the condition of class imbalance and few fault samples exists widely in the fault diagnosis field. In order to diagnose the state of bearings accurately under the kind of conditions, this paper proposes a one-dimensional Branch Convolutional Neural Network with Wide first-layer kernel (BWCNN) to reduce the impact of data imbalance and few fault samples. First, the normalization was applied to the original time-domain data to attenuate the effect of distribution differences of different features on the model and improve the convergence speed of the model. Then, using a two-branch network constructed by a one-dimensional convolutional network with a first-layer wide convolutional kernel, the model parameters was iteratively updated during the training phase of the network with resampling strategy and a cumulative learning strategy, making it possible for the model to learn generic patterns for all classes of samples and fault patterns for a few classes. Finally, the trained model was used for fault diagnosis. The proposed method's performance was evaluated using two widely used experimental datasets. On the selected datasets, the average macro F1 score reaches 99.5% when the healthy samples on the training set divided by the sum of all fault samples on the training set equals 20, and the number of samples for each class of faults on the training set is 10.