Bearing fault diagnosis based on visual symmetrized dot pattern and CNNs

This paper presents a new bearing fault diagnostic method based on symmetrized dot pattern (SDP) and convolutional neural networks (CNNs). Firstly, a time-domain vibration signal is directly transformed into a snowflake image in the polar coordinate to visualize fault by using SDP technique, and the sample library of visual SDP graphs of each running state is established. Then, shape difference features of SDP images are automatically extracted by the designed CNNs model to form a feature vector. Finally, the formed feature vector is used as the input to a Softmax classifier for recognizing the bearing fault state. Relative to the fault visualization of time-frequency analysis methods, the snowflake image of bearing vibration signal is directly acquireded by SDP technique without Fourier transforms, which is simpler with better performance. Experimental results show that the proposed method using SDP and CNNs can not only accurately recognize the bearing states, but also identify the relative position that fault occurred. The proposed method is more applicable for intelligent fault diagnosis of rolling bearing with 100% diagnosis accuracy.