A neural network strategy for learning of nonlinearities toward feed-forward control of pressure-compensated hydraulic valves with a significant dead zone

A velocity feed-forward-based strategy is an effective means for controlling a heavy-duty hydraulic manipulator; in particular, a typical valve-controlled hydraulic manipulator, to compensate for valve dead-zone and other hydraulic valve nonlinearities. Based on our previous work on the adaptive learning of valve velocity feed-forwards, manually labelling and identifying the dead-zones and the other nonlinearities in the velocity feedforward curves of pressure-compensated hydraulic valves can be avoided. Nevertheless, it may take two to three minutes or more per actuator to identify a pressure-compensated valve's highly nonlinear velocity feed-forward in real-time with an adaptive approach, which should be reduced for realistic applications. In this paper, inspired by brain signal analysis technologies, we propose a new method based on deep convolutional neural networks comparing with the previous method to significantly reduce this online learning process with the strong nonlinearities of pressurecompensated hydraulic valves. We present simulation results to demonstrate the effectiveness of the deep learning-based learning method compared to the previous results with an adaptive control-based learning.