Search

Accurate fault diagnosis of bogie bearings is beneficial for ensuring safe operation of railway trains. With the development of detection technology and optimization of maintenance strategies, increasingly detailed undetected or previously ignored faults are being found. However, most of the existing fault diagnosis methods are based on the closed-set hypothesis and fewer consider the occurrence of unknown faults in the future. The features of these unknown faults may overlap with the distribution of known faults, which inevitably affects the judgment of the model. This study proposes an open-set diagnosis method for unknown bogie bearing faults by a hybrid open score relation network (HOSRN). First, an auxiliary feature module is added based on the relation network framework, and the prototype center cross-entropy loss is introduced to enhance the feature extraction capability of the module. Second, a hybrid open score is constructed, based on the outputs of the modules, to simulate the distribution of known and unknown samples. Third, the open score of the training set is used to establish an adaptive threshold, which is used as the criterion for rejecting the unknown samples. Finally, the effectiveness of the proposed method is validated in three rolling bearing datasets, and the results show that the proposed method outperforms the comparison open-set diagnosis method.

Rotating machinery is an important component of modern electromechanical systems and its failure can result in significant economic losses. However, existing deep learning methods only consider the features within each sample, not the neighborhood relationships among samples; this results in poor performance when few labeled samples are available. To overcome this problem, we developed a meta-adaptive graph convolutional network (MAGCNet) to uncover the neighborhood relationships among samples and construct better features for the fault diagnosis of rotating machines when labeled samples are scarce. The wavelet-packet coefficient matrices of raw vibration data are extracted and defined as node features in a graph. To enhance the correlation properties of the few samples, an adjacency matrix is constructed by measuring the Euclidean distance between time- and frequency-domain characteristics and adding prior knowledge. The graph is divided into a series of subgraphs that are trained to optimize the initialization parameters of the adaptive graph convolution layers. The effectiveness of the proposed method was verified using datasets from the drivetrain diagnostics simulator (DDS) test rig and wind-turbine gearboxes.