Centrifugal Pump Health Condition Identification Based on Novel Multi-filter Processed Scalograms and CNN

This paper proposes a fault diagnosis method for centrifugal pumps (CP) based on multi-filter processed scalograms (MFS) and convolutional neural networks (CNN). Deep learning (DL) based autonomous Health-sensitive features extraction from continuous wavelet transform (CWT) scalograms are popular adoption for the health diagnosis of centrifugal pumps. However, vibration signals (VS) acquired from the centrifugal pump consist of fault-related impulses and unwanted macrostructural noise which can affect the autonomous Health-sensitive features extraction capabilities of the deep learning models. To overcome this concern, novel multi-filter processed scalograms are introduced. The new multi-filter processed scalograms enhance the fault-related color intensity variations and remove the unwanted noise from the scalograms using Gaussian and Laplacian image filters. The proposed techniques identified the ongoing health condition of the centrifugal pump by extracting fault-related information from the multi-filter processed scalograms and classifying them into their respective classes using convolutional neural networks. The proposed method resulted in higher classification accuracy as compared to the existing method when it was applied to a real-world centrifugal pump vibration signals dataset.