Sparse classification based on dictionary learning for planet bearing fault identification

Abstract

Planet bearing vibrations feature high complexity due to the intricate kinematics and multiple modulation effects. This leads to difficulty in planet bearing fault identification. In order to overcome this difficulty, a sparse classification framework based on dictionary learning is proposed. It operates directly on raw signals and is free from steps involved in conventional pattern identification such as feature design which requires prior expertise. First, a raw signal matrix is generated by partitioning the raw signal into segments, where each segment in all signal states has the same number of data points, and the length of the segment should guarantee that at least two adjacent fault impulses with the maximum interval can occur. Then, a dictionary initialized with the training sample set is learnt from the signal matrix, based on which the sparse representation is carried out afterwards. A dictionary learnt over signals under a certain state is best suited for signal reconstruction under the same state only but cannot recover signals well under other states. Inspired by this fact, sparse classification can be accomplished by comparing signal recovery errors over dictionaries under different states. The proposed method is validated using the experimental data of a planetary gearbox. Localized faults on the outer race, roller element and inner race of planet bearings are all identified successfully.