Unsupervised feature learning using locality-preserved auto-encoder with complexity-invariant distance for intelligent fault diagnosis of machinery

Unsupervised feature learning using locality-preserved auto-encoder with complexity-invariant distance for intelligent fault diagnosis of machinery

연구 분야: Artificial Intelligence

학회: Applied Intelligence

Unsupervised feature learning (UFL) has been recognized as a promising feature extractor in machinery fault diagnosis, where the auto-encoder is a very popular UFL framework. For the auto-encoder methods, however, it is still a great challenge to learn discriminative features from complex signals in an unsupervised manner. In this paper, a new UFL method named locality-preserved auto-encoder (LPAE) is proposed by explicitly designing a locality-preserved penalty term. Concretely, the penalty term constrains local geometry of samples in the original space to be well preserved in the reconstruction space, enabling more discriminative features to be learned accordingly. To better formulate this term, the complexity-invariant distance (CID) is employed to measure similarity between two mechanical signals so as to construct a reliable neighbor graph. On a rolling bearing dataset, experimental results verify that the proposed LPAE can learn sufficiently discriminative features from complex vibration signals collected from varying operating conditions, and achieves a remarkable and superior diagnosis performance over the existing advanced UFL methods. Moreover, the effectiveness of CID has been adequately validated by comparing with several other distance measurement methods. The proposed LPAE can be applied to the feature extraction stage of machinery fault diagnosis, which provides a potential solution for engineers to realize unsupervised learning of discriminative features.