Machine Condition Monitoring with Gaussian Mixture Model-Probabilistic Clustering for Pumps

Abstract

Pump systems play a critical role in various industries, and ensuring their reliability and timely maintenance is paramount. This research investigated the application of Gaussian Mixture Models (GMM) to condition monitoring and fault detection for pumps. The research begins by collecting and pre-processing extensive pump data from the Warri Refinery Petroleum Company, encompassing 374 samples of pre-processed vibration signals on various operating conditions and fault scenarios. The data were statistically analyzed, and then GMM, renowned for their ability to model complex data distributions, and K-means, a traditional clustering technique, was employed to cluster the dataset. The GMMs and K-means clustering were implemented by using suitable libraries on Python 3.0 software. The optimum hyper-parameters were determined using a grid search method. Then the clusters were created using both models, and their performance was investigated by calculating the silhouette and BIC scores. The obtained clusters were then assessed for their uniqueness to identify fault types and other pump conditions. Based on a hyper-parameter grid search, the optimum number of clusters was found to be 6 and a random state of 54. Comparative analyses revealed that GMMs outperform K-means having silhouette scores of 0.68 and 0.51, respectively. The application of GMMs showcases their potential for proactive maintenance, by identifying different anomalies such as those resulting from faulty sensors, and outboard and inboard faults. This study demonstrates the effectiveness of GMM-based clustering in accurately identifying different operational states and detecting anomalies within pump data. The application of GMMs provides a practical and effective means of enhancing pump system reliability and maintenance strategies.