Fault diagnosis for centrifugal pumps using deep learning and softmax regression

Fault diagnosis of centrifugal pumps is critical to lower its operating and maintenance costs. Due to the non-stationary and non-linear characteristics of vibration signals of centrifugal pumps, a large number of approaches for feature extraction and fault classification have been developed. However, these traditional methods spend too much time extracting features, reducing feature dimension and fusing different features. To resolve the issue, this paper presents an effective unsupervised self-learning method to achieve the fault diagnosis of centrifugal pumps, which uses deep learning method to adaptively extract fault features from non-stationary vibration signals and softmax regression model is used to identify possible failure modes automatically. In particular, the stacked denoising autoencoder (SDA) of deep learning models is selected to learn effective feature representations and we improved fault pattern classification robustness by corrupting the input data. The effectiveness and feasibility of the proposed method are validated by experiments in this paper.