Abstract
In mechanical systems, unplanned equipment failure often leads to costly downtime, productivity losses and reduced reliability. Triplex pumps, widely used in industrial applications, are particularly vulnerable to mechanical faults such as seal leakage, blocked inlet and worn bearings. These faults, if undetected, can cause severe damage to system components. Recent advancements in machine learning (ML) have opened new opportunities for predictive maintenance, enabling early fault detection through pattern recognition in operational data. Motivated by the need for more proactive maintenance strategies, this research explores the application of ML techniques for fault classification in triplex pumps. An efficient predictive maintenance framework was developed to detect and classify mechanical faults using flow and pressure data from the pump system. Two learning models, artificial neural networks (ANNs) and random forest (RF), were implemented and evaluated for performance in multi-label classification. The machine learning process followed a complete lifecycle including pre-processing, feature engineering, model training and performance evaluation. Four ANN models were designed with varying architectures and training configurations, while two RF models were constructed using different set of features. Advanced signal processing techniques such as fast fourier transform (FFT), wavelet decomposition, statistical analysis and anomaly detection were employed to enhance the model’s input features. The models were trained and tested on Google Colaboratory using Python-based libraries and frameworks. ANN Model 4 achieved a test accuracy of 72.92% after incorporating extensive feature engineering. In contrast, the random forest (RF) model 1 outperformed all others, achieving a test accuracy of 83.33% and a precision of 96%, highlighting its superior performance in classifying pump faults. The evaluation metrics including accuracy, precision, recall and F1 score confirmed the robustness of the RF Model 1 for predictive maintenance applications. The model demonstrated high fault detection performance while maintaining low computational cost and generalisation ability. The findings established that ensemble learning methods, particularly random forests, offer a highly accurate and scalable solution for fault diagnosis in pump systems. The proposed predictive maintenance framework can be integrated into existing monitoring systems to enhance operational reliability, reduce maintenance costs and support data-driven maintenance planning across various industries. Future
x
developments may include real-time deployment, expansion of fault types and integration with digital twin platforms to further improve predictive capabilities.