Comparison of Physical and LSTM Neural Network Model of a High-Pressure Valve Used in the Steel Industry

This study presents models of an electro-hydraulic valve derived from physical principles and neural network techniques. Input-output models are constructed using experimental data from a hydraulic press machine in a steel manufacturing plant and as a plant of a closed loop system. The models are candidates for digital twins in the steel manufacturing plant. The physical model is derived based on fluid mechanics, fluid dynamics, and electronic principles. Two LSTM neural network models denoted NN1 and NN2 are employed for modeling the valve. The parameter estimation for each neural network model is conducted using distinct training datasets. This work compares the validation results of the models in the time and frequency domain. Both the physical model and NNs capture the main behavior of the valve. However, NNs have lower mean square error (MSE) compared to the physical model. NN2, trained the model using different operating conditions, is capable of modeling non-linearities (seen at high frequencies) and leakage effects of the system, that are not captured by NN1 and physical models. The fault caused by leakage is seen in the MSE vs cycles for the physical and NN1 models.

I. Introduction

Research on hydraulic valves has been ongoing for a long time. The electrohydraulic servo valve appeared in the 1940s. Novel actuation systems for such valves were developed during the 1950s. In the early 2000s, academic research on hydraulic valves increased due to the availability of high-speed microprocessors, which motivated the design of novel hydraulic valves. There are monographs with physical models of a wide range of hydraulic valves [1] [2]. Two-stage pro-portional cartridge valves with a poppet-type construction are used in applications with fast response, large flow, and high-pressure requirements [3] [4]. Xu et al. modeled a directional valve with three stages from physical principles, where two proportional poppet-type cartridge valves were in the second stage [5]. Han et al. proposed an optimal design for a two-stage proportional cartridge valve with a poppet construction type [6]. This type of valve is controlled by proportional-integral-derivative (PID) controller with a feedback loop, using signals continuous in amplitude. In contrast, other types of valves are controlled by a model predictive control (MPC) based controller using pulse width modulation signals in [7].