Open Set Recognition of Intelligent Bearing Fault Diagnosis Based on Hybrid Neural Network and Improved OPENMAX

Rolling bearings, as critical mechanical components in modern industry, can cause significant economic losses and personnel injuries when they fail. Therefore, intelligent fault diagnosis systems are among the most important systems in various machinery. Traditional deep learning diagnostic methods often fail to correctly identify unknown categories, mistakenly recognizing them as one of the known categories, which brings substantial potential risks. To achieve the open set recognition task in intelligent bearing fault diagnosis, this paper incorporates residual block structures and Squeeze-and-Excitation attention mechanisms into the convolutional neural network, constructing a CNN-Resnet-SE hybrid neural network. This is then combined with an improved OPENMAX. The specific overall process is as follows: first, the original vibration signals are processed using fast Fourier transform, then the features are extracted by the CNN-Resnet-SE network, and finally, decisions are made through the improved OPENMAX. And experiments were designed on two public datasets, the Case Western Reserve University dataset (CWRU) and the Paderborn University dataset (PU), to further validate the superiority of the proposed model.