Experimental modeling of a hydraulic load sensing pump using neural networks
Xu, Xin Ping
The traditional method of modeling dynamics of a nonlinear hydraulic system is to develop mathematical describing equations based on "laws of nature". The development of these describing equations for a physical hydraulic system often requires that engineering intuition and a priori knowledge of the system be combined with mathematical properties of the equations. In addition, the problems of accurately measuring or defining physical parameters or coefficients have frequently restricted the interpretation of modeling results to specific operating points with very limited modeling accuracy. An alternative modeling approach is to establish or approximate mathematical relationships of a dynamic system based on observed input-output data. Neural networks have been a class of very attractive mathematical model structures that can be used to establish these mathematical relationships, because of their proven capabilities of approximating many nonlinear functions. The overall objective of the research in this study is to develop a neural network simulation package to assist designers in the simulation and configuration of hydraulic circuits. The specific objective of this thesis is to explore the capabilities of neural networks to approximate the nonlinear dynamics of a particular hydraulic component using an experimental approach. In this thesis, the use of partially recurrent neural networks with the conjugate gradient training algorithm to model a hydraulic load sensing pump was investigated. A simulation study was first conducted using "noise-free" data to examine the modeling errors in order to provide a clear insight into the mechanism of the modeling error accumulation over the transient state with a recurrent type of model structure. The established concepts and approach were then applied to experimentally model a load sensing pump. An experimental system was designed and constructed with particular attention paid to the design and generation of sufficiently rich input signals, and to the selection of an appropriate sampling rate. The data obtained on the testing of the load sensing pump dynamics are used in the training and testing of the neural models. The analysis and discussion showed that the training accuracy and the error accumulation were the two most critical factors in examining and interpreting the overall modeling accuracy. It has been established through the work presented in thesis that a partially recurrent neural network is capable of approximating the dynamics of a hydraulic load sensing pump with very satisfactory accuracy. The major contributions of this study are as follows: (1) the study identified the modeling error accumulation problem to be a major cause for the deterioration of the dynamic modeling accuracy, and suggested a means to effectively reduce the error growth in order to improve the modeling accuracy; (2) the applicability of the neural network approach to modeling a "real-world" hydraulic component using actual data was investigated, and practical constraints imposed by the actual hydraulic experimental testing facilities (not revealed by theoretical or simulation studies) were studied. The experimental implementation successfully established, from a practical point of view, the feasibility of the neural network approach to modeling a real hydraulic component, and suggested that the hydraulic data quality in terms of data precision and data distribution significantly affected final model accuracy.