Machine learning insight in quality control and maintenance management of drilling mud pumps

This study proposes a machine learning-based framework for predictive maintenance and quality control of drilling mud pumps, targeting critical operational challenges such as equipment failure, downtime, and high maintenance costs. The methodology involves using a dataset of 820 time-indexed operational records comprising historical maintenance logs, sensor readings, and operational parameters. Four supervised learning models were applied: Random Forest (RF), Multiclass Logistic Regression (MCLR), Feedforward Neural Network (FNN), and Long Short-Term Memory (LSTM). Although each record contains static measurements, the dataset is chronologically ordered by cumulative operating time and failure events, enabling the construction of temporal input sequences for sequence-based learning models. Data preprocessing included imputing missing values, normalization, and categorical encoding. Performance evaluation was conducted using stratified k-fold cross-validation across training, validation, and test sets to mitigate overfitting and preserve failure-class distributions. The results show that among the models, LSTM achieved the highest predictive accuracy (RMSE = 0.7273; R² = 0.9153), demonstrating superior capability in capturing temporal degradation patterns associated with drilling mud pump failures. These findings contribute to the development of intelligent maintenance systems in oilfield operations, offering a scalable, data-driven approach to improve pump performance and operational sustainability. Machine learning framework developed for predictive maintenance and quality control of drilling mud pumps. Comparative testing of classical, ensemble, deep learning, and temporal models using identical industrial datasets. Temporal learning (LSTM) shown to markedly improve prediction accuracy with time-ordered maintenance data. Key contributors to pump failure identified: working pressure, pumping rate, mud viscosity, mud density, and cumulative operating time. Scalable analytical framework proposed, adaptable to real-time monitoring and larger industrial data streams.