Classification of bearing data based on deep belief networks

It is difficult for linear methods to analyze and classify the bearing signals. This research designs a classifier based on deep belief networks to classify the data which is collected from bearings. We use stacked Restricted Boltzmann Machines (RBM) to pre-train the networks and avoid the local minimal. The first three layers of the networks are non-linear and the last two layers are linear. The reason is that non-linear methods are capable of fitting the complex relationship of the data. When it comes to the training process, we pay much attention to the training times. Training too many times could lead to over-fitting while less training would cause under-fitting. In the proposed model, we choose the performance of all data as the criterion of training times. Upon getting the best performance, we stop the training and record the parameters. After training, we use back propagation (BP) algorithm for fine-tuning the weights of the networks. In the experiments, five types of bearing data including four fault data and one normal data are used and they can be classified by the proposed model.