Detection of weak transient signals based on unsupervised learning for bearing fault diagnosis

Transient impulse contains abundant information of bearings status. When fault occurs, it is activated and would recur periodically or quasi-periodically. Its period can indicate where defects lie in. However, transient impulse is easily swallowed by background noise or interferences in part or in whole, especially at early stage of fault. This problem brings hard obstacles into faults detection. Considering that transient impulses are periodical or quasi-periodical and vibration signal has local similarity, the single transient impulse can be seen as one of shift-invariant features. In view of this, this paper derives adaptive and non-linear signal decomposition formulas and further proposes adaptive and unsupervised feature learning method by using convolutional restricted Boltzmann machine model. With respecting local waveform structures, this method can automatically capture shift-invariant patterns hidden in original signal and decompose the original signal into several sub-components at the cost of minimizing reconstruction error. Among these sub-components, the fault-related information, i.e., transient impulses signal, could be extracted likely. It provides a promising idea for intelligent signal processing by using unsupervised learning. Afterwards, Maximizing kurtosis is applied to select optimally latent fault component. Two real bearing experiments validate this method is effective and reliable in extraction of weak transient impulses.