Deep Manifold Learning for Weak Signal Detection

Extraction of weak impulses is very essential for fault detection of rolling bearings. Since the weak transient impulse may be submerged in interferences or background noise, therefore, it still very challenging for weak signal enhancement. However, the reported weak signal enhancement techniques based on band-pass filtering or wavelet basis optimization to match the weak transient features, as a result, the performance, flexibility and adaptability are constrained. Aiming at the aforementioned shortcomings, deep manifold learning (DML) is proposed in this paper for weak signal detection. In the proposed method, a mapping model between the noisy waveform feature manifold (WFM) and the pure repetitive impulses is constructed and the high-dimensional WFM can be adaptively mined by deep compression. The mined signal obtained by encoding of the high-dimensional WFM is considered as the detected signal. The experimental results demonstrate that the proposed DML method is adaptive, flexible, and outperforms the conventional methods based on traditional manifold learning.