Development and Evaluation of Artificial Neural Network Techniques for Predictive Maintenance of a Reciprocating Compressor

This work focused on the development and evaluation of ANN-based techniques for the predictive maintenance of a reciprocating compressor. Predictive maintenance, PdM entails minimizing unplanned downtime while extending the life of the equipment through smart monitoring and analysis to predict impending failure. The predictive power of ANN has been exploited in this study to shift from reactive approaches towards maintenance to proactive ones, with improved operational efficiency costeffectively. Operations data over three consecutive months were retrieved and utilized to train the model. The model iterated through the steps of forward propagation, computation of the loss, and back-propagation to optimize the weights and biases using the training data. From this, it was observed that the ANN model was very effective in identifying the failure conditions at which compressors will fail, with an average of 98% on all predicted cases. The model's robustness is underlined by the high value of precision, recall, and F1-scores it returns; hence, it can be relied upon for failure prediction. Among those that were advisable in boosting the model were regularization, cross-validation, optimization of hyper-parameter, augmentation of data, and simplification of the model. The new direction that this study gives to failure prediction in compressors, especially the reciprocating ones, has performed impressively by showing how much machine learning and neural networks can contribute to industrial predictive maintenance