Condition Monitoring of a Reciprocating Compressor

This work uses condition monitoring techniques to classify the health of reciprocating compressor valves using machine learning. Condition monitoring is a field of study that aims to improve maintenance scheduling by reducing machine downtime and utilizing the full life of components. Condition monitoring uses real-time data to analyze machine performance and determine the health level of key components. Previous maintenance scheduling determined when maintenance needed to be done by either allowing the machines to run until a component significantly hinders the output of the machine or on time-based intervals which wastes useful component life. Condition monitoring aims to determine the current and future health of components to schedule maintenance around the working hours of the machines, improving the overall efficiency of the facility. Two common signals that are used to monitor reciprocating compressor valves are the pressure and vibration data. From the pressure data, the efficiency of the valves can be calculated and are good indicators of component wear and leakage. The vibration data can be examined to determine the timing of valve impacts. Each wear condition affects the timing and number of impacts differently allowing for the vibration data to be a good indicator of valve degradation. Different machine learning algorithms which range from statistical classifiers to deep learning models are used on the features that are created from the pressure and vibration data. Classification algorithms are used which have a set number of options that the algorithm can choose from. They require examples from each valve health level to be able to learn how they impact the features. This is why the algorithms need to be generalizable to other compressors because facilities won’t allow for bad components to be put in their compressors because their output would be reduced. Previous work used 4 features from the pressure data and was able to diagnose outlet and inlet valves from one cylinder with 17 health options with a classification accuracy of 98%. Current work is being done to improve the feature selection for vibration data to increase the classification accuracy which is currently at 90.5%.