Global and Local Feature Extraction Using a Convolutional Autoencoder and Neural Networks for Diagnosing Centrifugal Pump Mechanical Faults

Centrifugal pumps are important types of electro-mechanical machines used for fluid and energy conveyance. Mechanical faults in centrifugal pumps lead to abnormal impacts in the vibration signal of the system. Those impacts induce nonstationarity in vibration signals and hence complex time-frequency domain signal analysis techniques are required to investigate the mechanical fault features of centrifugal pumps. In this paper, an end-to-end pipeline for diagnosing faults in centrifugal pumps is proposed. To create a two-dimensional representation of the transients that appear in the vibration signals due to centrifugal pump operating conditions, first, a 1/3-binary tree fast kurtogram is computed. Next, a convolutional autoencoder and convolutional neural network are trained to autonomously extract global and local features from the kurtograms. Then, global, and local features are merged to form a joined feature vector that contains different visual features that are extracted using convolutional deep architectures using their specific loss functions during the training. Finally, this feature vector is propagated to a shallow-structured artificial neural network to accomplish fault identification. The proposed framework has been validated by the dataset collected from a real industrial centrifugal pump test rig. The results obtained during the series of experimental trials demonstrated that the introduced method achieved high classification accuracies when diagnosing faults based on signals collected under 3.0 and 4.0 bars of pressure.