Feedforward Compensation Analysis of Piezoelectric Actuators Using Artificial Neural Networks with Conventional PID Controller and Single-Neuron PID Based on Hebb Learning Rules

Abstract

This paper presents a deep analysis of different feed-forward (FF) techniques combined with two different proportional-integral-derivative (PID) control to guide a real piezoelectric actuator (PEA). These devices are well known for a non-linear effect called “hysteresis” which generates an undesirable performance during the device operation. First, the PEA was analysed under real experiments to determine the response with different frequencies and voltages. Secondly, a voltage and frequency inputs were chosen and a study of different control approaches was performed using a conventional PID in close-loop, adding a linear compensation and a FF with the same PID and an artificial neural network (ANN). Finally, the best result was contrasted with an adaptive PID which used a single neuron (SNPID) combined with Hebbs rule to update its parameters. Results were analysed in terms of guidance, error and control signal whereas the performance was evaluated with the integral of the absolute error (IAE). Experiments showed that the FF-ANN compensation combined with an SNPID was the most efficient.

Advances in Tracking Control for Piezoelectric Actuators Using Fuzzy Logic and Hammerstein-Wiener Compensation

“…In comparison with recent works that were related to a similar control architecture and device; the results presented in [24] reached an IAE of 0.049, which means that this research could improve 3% of the value; as PEAs are for high-precision, then any enhancement in tracking is completely accepted. Additionally, the authors of [49] provided a similar FLC combined with FF for the PEA with lower displacement range, but the error is higher than the one reached in this research.…”

“…The triangle wave is a complex source as a reference, since it consists of high-frequency harmonics that can raise the tracking difficulty, according to [46]. Thus, a triangle wave was used to obtain the hysteresis graph and as a reference for guidance, a configuration as in [24] where the period chosen was 2 s with a maximum amplitude of 150 V, since, at the utmost driving voltage, the non-linearity analysed has its maximum reflection. Figure 2 is the PK4FYC2 hysteresis graph for two triangle cycles, where the first ascending curve has its initial point that begins at the origin or also called Initial point and ends at the Upper target point.…”

“…Figure 2 is the PK4FYC2 hysteresis graph for two triangle cycles, where the first ascending curve has its initial point that begins at the origin or also called Initial point and ends at the Upper target point. Provided that the amplitude and frequency of the input do not change [24], this demeanour only occurs for the first rising curve, since the following cycles converge to a single hysteresis curve, where the lower converging point and the upper target point will be permanent. According to this analysis, the conversion from voltage to displacement for reference design cannot be done by multiplying the maximum displacement divided by the maximum voltage which implies a straight line between the initial point and the upper target point.…”

“…Feedforward-feedback controllers can increase the control accuracy of PEAs through the advantages that both techniques can provide separately [23]; furthermore, this strategy can produce multiple structure combinations. In [24], the hysteresis was compensated with a linear and artificial neural network (ANN) combined with a conventional proportional-integral-derivative (PID) and a neural type; results shown that the combination of ANN with the neural PID has been the most precise one, since the integral of the absolute error (IAE) was reduced to 0.049. The authors of [25] described the FF compensation by a Prandtl-Ishlinskii hysteresis model merged with sliding mode control (SMC) as feedback; results unveiled an error less than 1%.…”

Modeling the prediction of hydrogen production by co‐gasification of plastic and rubber wastes using machine learning algorithms

“…Changes in the synaptic weight from one node to another are dependent on the learning rule. Based on Hebb's learning rule, there will be an increase in the synaptic weight between an input and output neurons if there a large signal in the input neurons results in a corresponding large signal in the output neurons 47 . Understanding the influence of the synaptic weight on the model output could help in optimizing the performance of the 1‐LMLP neural network.…”