Self-Data-Driven Intelligent Fault Diagnosis of Gearboxes With Sparse Group Lasso

The vibration spectrum of a gearbox generally involves meshing frequency factors and a series of harmonics due to the time-varying meshing stiffness. When a fault occurs in a gear of the system, a group of sidebands will appear around the meshing frequency of the gear and its harmonics. Therefore, it is crucial to detect the gearbox faults by monitoring the appearance of group sidebands in the vibration spectrum. To detect the group sidebands automatically from the vibration spectrum, this paper develops a self-data-driven spectrum sparsity method based on sparse group lasso. The gearbox fault diagnosis problem is described as a group penalized regression problem. A new group number-oriented lasso (GNoL) algorithm is proposed to select the group sidebands of the fault gears according to the change in the spectrum. To determine a suitable group number of sidebands, the GNoL algorithm optimizes the penalty coefficient automatically driven by an expected group number. An experimental degradation test of a gearbox system and an industrial case study of heavy truck gearboxes are conducted to verify the proposed method. The results demonstrate that it is effective in locating the fault gears automatically driven by the self-data of a monitoring gearbox system.