Condition Monitoring Systems for Axial Piston Pumps: Mobile Applications
Abstract
Condition monitoring of hydraulic systems has become more available and inexpensive to implement. However, much of the research on this topic has been done on stationary hydraulic systems without the jump to mobile machines. This lack of research on condition monitoring of hydraulic systems on mobile equipment is addressed in this work. The objective of this work is to develop a novel process of implementing an affordable condition monitoring system for axial piston pumps on a mobile machine, a mini excavator in this work. The intent was to find a minimum number of sensors required to accurately predict a faulty pump. First, an expert understanding of the different components on an axial piston pump and how those components interact with one another was discussed. The valve plate was selected as a case study for condition monitoring because valve plates are a critical component that are known for a high percentage of failures in axial piston pumps. Several valve plates with various degrees of natural wear and artificially generated damage were obtained, and an optical profilometer was used to quantify the level of wear and damage. A stationary test-rig was developed to determine if the faulty pumps could be detected under a controlled environment, to test several different machine learning algorithms, and to perform a sensor reduction to find the minimum number of required sensors necessary to detect the faulty pumps. The results from this investigation showed that only the pump outlet pressure, drain pressure, speed, and displacement are sufficient to detect the faulty pump conditions, and the K-Nearest Neighbor (KNN) machine learning algorithms proved to be the least computationally expensive and most accurate algorithms that were investigated. Fault detectability accuracies of 100% were achievable. Next, instrumentation of a mini excavator was shown to begin the next phase of the research, which is to implement a similar process that was done on the stationary test-rig but on a mobile machine. Three duty cycle were developed for the excavator: controlled, digging, and different operator. The controlled duty cycle eliminated the need of an operator and the variability inherent in mobile machines. The digging cycle was a realistic cycle where an operator dug into a lose pile of soil. The different operator cycle is the same as the digging cycle but with another operator. The sensors found to be the most useful were the same as those determined on the stationary test-rig, and the best algorithm was the Fine KNN for both the controlled and digging cycles. The controlled cycle could see fault detectability accuracies of 100%, while the digging cycle only saw accuracies of 93.6%. Finally, a cross-compatibility between a model trained under one cycle and using data from another cycle as an input into the model. This study showed that a model trained under the controlled duty cycle does not give reliable and accurate fault detectability for data run in a digging cycle, below 60% accuracies. This work concluded by recommending a diagnostic function for mobile machines to perform a preprogrammed operation to reliably and accurately detect pump faults.
Degree
Ph.D.
Advisors
Vacca, Purdue University.
Subject Area
Artificial intelligence|Mathematics
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.