Centrifugal Pump Fault Detection with Hybrid Feature Pool and Deep Learning
Wasim Zaman, Muhammad Siddique, Jong-Myon Kim
Abstract
This research introduces an innovative hybrid technique for diagnosing faults in centrifugal pumps utilizing vibration sensor data. The initial signal, filtered through a low pass filter with a 4.6 kHz cutoff frequency, undergoes both Stockwell and continuous wavelet transformations concurrently. These transformations yield two sets of time-frequency scalograms as inputs for separate convolutional neural networks. These networks extract vital features from the scalograms, which are then amalgamated into a hybrid feature pool. Because the pool has many dimensions, principal component analysis is used to reduce the feature space to a manageable latent space while keeping the essential features from both sets of scalograms. Subsequently, a two-layer Artificial Neural Network is employed to categorize pump states based on this reduced-dimensional feature space. Impressively, our approach attains a classification accuracy of 95.60%, outperforming existing state-of-the-art methods. This shows how well our strategy works. It combines different transformation techniques, hybrid feature extraction, and dimensionality reduction to make a method for diagnosing centrifugal pump faults that is very effective and quick.