Improving the Performance of Storage Tank Fault Diagnosis by Removing Unwanted Components and Utilizing Wavelet-Based Features

This paper proposes a reliable fault diagnosis model for a spherical storage tank. The proposed method first used a blind source separation (BSS) technique to de-noise the input signals so that the signals acquired from a spherical tank under two types of conditions (i.e., normal and crack conditions) were easily distinguishable. BSS split the signals into different sources that provided information about the noise and useful components of the signals. Therefore, an unimpaired signal could be restored from the useful components. From the de-noised signals, wavelet-based fault features, i.e., the relative energy (REWPN) and entropy (EWPN) of a wavelet packet node, were extracted. Finally, these features were used to train one-against-all multiclass support vector machines (OAA MCSVMs), which classified the instances of normal and faulty states of the tank. The efficiency of the proposed fault diagnosis model was examined by visualizing the de-noised signals obtained from the BSS method and its classification performance. The proposed fault diagnostic model was also compared to existing techniques. Experimental results showed that the proposed method outperformed conventional techniques, yielding average classification accuracies of 97.25% and 98.48% for the two datasets used in this study.