A Deep Belief Network Based Approach for Bearing Fault Diagnosis

Effective fault diagnosis techniques are crucial for normal and safe machinery operation. The development of data acquisition techniques allows for massive volumes of data to be collected and used for fault diagnostics and prognostics. The main challenge that faces existing methods is the dependency on extracting features manually. A modified deep belief network (MDBN) is proposed for the purposes of bearing fault classification. The proposed method can address the challenge between machinery big data and intelligent diagnosis by extracting the features automatically, with only applying a simple signal processing technique. Two more goals of the proposed method are to increase the speed of learning and to prevent overfitting. To increase the learning speed, momentum is added to the original deep belief network (DBN). To prevent the model from overfitting the training data, weight decay and sparsity of the hidden units are both brought into the proposed MDBN. The proposed MDBN based fault diagnosis proved to be more effective when compared to DBN.