Improving Reciprocating Compressor Valve Condition Monitoring via an Accelerated Lifetime Tester and Deep Learning

The goal of this research is to improve the current condition monitoring method for diagnosing reciprocating compressor valves. Leakage, seat wear, and spring degradation are three common valve faults that are seeded into all the valves of a Dresser-Rand ESH-1 dual-acting reciprocating compressor. The method converts raw vibration data into the time-frequency domain and determines if the Continuous Wavelet Transformation or the Smoothed Pseudo Wigner-Ville Spectrum can achieve a higher classification accuracy. The Continuous Wavelet Transformation can vary its resolution through two parameters while the Smoothed Pseudo Wigner-Ville Spectrum can vary its resolution through smoothing window sizes. Both time-frequency transforms are optimized and compared to determine which one can provide the highest classification accuracy using a linear discriminant classifier. A portion of the transformed data, the region of interest, is converted into a grey-scale image and a black-and-white image where distribution statistics are calculated and used as features for classifying valve health using a linear discriminant classifier. The region of interest is also examined, and the boundaries are tested to find the optimal region for each valve independently. This creates a methodology that can be used on all the valves in a Dresser-Rand ESH-1 reciprocating compressor and achieve a high classification accuracy. The classification accuracy can be improved further using deep learning which creates features automatically. This allows for a more complex feature space but requires more training data. The work includes transfer learning from GoogLeNet which is used for image classification allowing for faster training times but is limited by the input size. This takes longer to train than a statistical classifier but creates a larger separation between the classes. A second deep learning model is created which is a convolutional neural network with a depth of four. It is created from scratch and uses all the valve vibrations and outputs the valve that has a fault as well as the fault type. The convolutional neural network allows for one algorithm per compressor as opposed to one algorithm per valve. Both deep learning methods require data augmentation to increase the amount of data. Synthetic data is created from existing data which looks realistic to increase the classification accuracy and reduce overfitting which allows the algorithm to generalize to other compressors more easily. Wearing down poppets naturally takes a long time, so a poppet accelerator is designed and created to wear down poppets in a natural way but faster than a compressor. This system is used to understand better wear patterns in poppet valves as well as relate wear to the original feature space. The worn poppets are put in the compressor at progressively worn stages and plotted using the created algorithms.