Condition Monitoring of Reciprocating Compressors using Probabilistic Neural Network and Optimization with Genetic Algorithm

Demand for cost-effective, reliable operation and safety of machinery, especially reciprocating compressors, which are among the most expensive machines in maintenance, requires accurate troubleshooting and fault classification. Due to their advantages, data-driven methods are often preferred to physical modeling methods for fault detection. This research simulates the mathematical model of a two-stage reciprocating compressor and conventional faults for use as a monitored system. The artificial neural network used is the probabilistic neural network whose main task is classification. Classification classes include one healthy compressor class and seven defective compressor classes for eight classes. Classification with the probabilistic neural network was performed using time domain and envelope spectrum characteristics. Then, the selected features are optimized using a genetic algorithm before feeding into the probabilistic neural network. Classification with the probabilistic neural network using time-domain characteristics shows a poor classification percentage with 44% Correct accuracy. But classification with a probabilistic neural network and envelope spectrum features has a 95% correct classification accuracy. Also, optimizing the selection of statistical features of the time-domain and frequency envelope spectrum with a genetic algorithm brings 48 and 99% correct accuracy in classification, respectively.