Use of neural networks for quick and accurate auto-tuning of PID controller
Introduction
If industrial automation applications are considered, proportional, integral and derivative factor (PID) controllers are among the most diffused solutions, mainly due to their simple structure, which also appears very robust in most common process control situations [1].
In important industrial processes many hundreds of this kind of controllers can be installed and each one of these should undergo a specific tuning, in order to individuate the process dynamics for a suitable and effective control strategy.
Tuning the characteristic parameters of a PID controller is a procedure in most cases strongly dependent on the personal know-how and skill of the plant operator, although theoretical guide lines for tuning are available; therefore a lot of time should be spent for PID tuning and some experience is necessary, if whole plant capabilities are to be exploited according to a correct setting of PIDs [2].
Due to the above considerations, auto-tuning techniques, the tuning of a controller in a automatic way on demand of an operator or of an external trigger, are very interesting for an industrial environment, because they are time saving and allow to improve the efficaciousness of control parameter settings. Recent advances in PID auto-tuning procedures can be found in different industrial contexts, which significantly improved the process-control efficaciousness [3], [4], [5], [6].
It is well known that auto-tuning a controller typically requires to test the system of interest, then to automatically evaluate the controller parameters, according to different models of system and to different possible optimisation criteria [7], [8], [9], [10], [11].
Taking into account these aspects, auto-tuning could be considered an indirect measurement method of the PID parameters, so that a metrological approach could give a contribution to evaluate the efficaciousness of their setting. In particular, a comparison between the measurement uncertainty and the variability range for control-parameter setting that results from classical methods could help to validate new auto-tuning methodologies, in order to predict if the controller behaviour fits the control requirements.
Standard deviation of PID coefficients is influenced by many aspects, depending on the model approximation, on instrumentation measurement uncertainty, on environmental effects, on the need of using data-processing techniques able to shorten the time duration for auto-tuning.
In this paper metrological considerations will be used in order to improve auto-tuning methods of controllers with reference to high-performance automatic systems (high manufacturing rate, many controllers, etc.). In particular, actions will be considered able to reduce the effect on auto-tuning of environmental disturbances on instrumentation and of working-condition variation of the system to be controlled; the effect of using innovative data-processing techniques for significant reduction of the evaluation time will also be studied. Traditional and innovative approaches, based on neural-network analysis, will be considered and compared in order to select the most suitable methodology for real applications. Improving these aspects without affecting accuracy of parameter evaluation will appear very useful for practical purposes. A real situation will be considered not only to validate the proposed solutions with reference to an actual application, but also to experimentally consider in the proposed methodology the standard deviation contributions due to different aspects, difficult to model and to predict, and that only experimental activity and data allow to take into account. Theoretical and experimental activity will be carried on, also considering the possibility of getting procedures able to be easily and efficaciously implemented in a quality-assured industrial context.
Section snippets
The real application
The methodology will be discussed with reference to a real application concerning the automatic welding of edges of plastic bags for packaging. Welding temperature should be set at 150.0 °C, with a tolerance range of ±1.0°C.
The bag edges are welded by heating some pliers, that are depicted in Fig. 1, by means of a couple of cylindrical electrical resistors, shown in Fig. 2, inserted into the jaws of the pliers themselves. Resistors are fed by alternate current and the heating is modulated
Laboratory test bench
In order to carry on a preliminary set-up of the control procedure, a laboratory experimental device has been designed and realized, reproducing the real configuration in a simplified way.
The test bench is sketched in Fig. 4; the main components are summarized as in the following:
- •
a jaw
- •
two cartridge resistors (500 W at 230 V AC);
- •
two relays;
- •
two type-J thermocouples for pliers-temperature measurement;
- •
the Pac Controller MAX-4 by ELAU;
- •
a PROFIBUS module IB IL TEMP 2 by Phoenix Contact for temperature;
- •
a
The methodology
In order to compare the different auto-tuning procedures and to validate the shortened data-processing techniques to be implemented, research has been carried on according to the following steps:
- •
preliminary tests in order to experimentally evaluate the dynamic behaviour of the system and to define by comparison the most suitable auto-tuning techniques;
- •
design and implementation of innovative data-processing techniques to significantly reduce the auto-tuning duration time;
- •
evaluation of standard
Conclusions
In this paper metrological considerations have been used in order to improve auto-tuning methods of controllers with reference to high performance automatic systems (high manufacturing rate, many controllers, etc.).
In particular, actions have been considered able to reduce the effect on auto-tuning of environmental disturbances on instrumentation and of working condition variation of the system to be controlled; particular attention has been paid to the evaluation of accuracy of data-processing
References (13)
- et al.
Relay feedback auto-tuning of process controllers—a tutorial review
J Process Control
(2002) - et al.
A simple method of tuning PID controllers for integrator/dead-time processes
Comput Chem Eng
(2003) - et al.
Relay auto-tuning of PID controllers using iterative feedback tuning
Automatica
(2003) New tuning method for PID controller
ISA Trans
(2002)- et al.
Design of PID-type controllers using multiobjective genetic algorithms
ISA Trans
(2002) Automatic tuning of PID-regulators. report, Department of Control Engineering
(September 2002)
Cited by (44)
Tuning of extended state observer with neural network-based control performance assessment
2022, European Journal of ControlCitation Excerpt :Furthermore, neural networks can serve as a standalone adaptive controller [22,37], however, they usually come with no warranties about their control performance or robustness to noise and external disturbances, which limits their use in safety-critical applications. Instead of relying on a standalone neural network based controller, neural networks can be used to tune some well known control schemes, like PID [9] or ADRC [13,14,29]. This allows taking advantage of the innate robustness of these controllers, while improving their performance by a data-driven choice of the controller or observer gains.
A universal digital motion controller design for servo positioning mechanisms in industrial manufacturing
2020, Robotics and Computer-Integrated ManufacturingCitation Excerpt :However, PID has the potential problem of windup, and its performance would exhibit a degradation with the change of target reference or disturbance, even if some anti-windup methods are included. Intelligent methods, such as neural network or fuzzy logic, are also being integrated into controller design for performance improvement, see e.g., Giulio et al. [14–16]. In recent years, an emerging control technique related to PID, i.e, the fractional order control, is gaining increasing attentions from researchers, see e.g., Monje et al. [17–19].
Intelligent tuning method of PID parameters based on iterative learning control for atomic force microscopy
2018, MicronCitation Excerpt :In recent years, some intelligent methods were used to improve the PID parameters tuning process. Giulio D’Emilia proposed a theoretical-experimental approach based on neural networks to reduce the time of auto-tuning PID parameters, in the situations where the ratio of signal to noise is poor and the system model is difficult to obtain (D’Emilia et al., 2007). K Sinthipsomboon combined the fuzzy controller with the fuzzy self-tuning PID controller together to the servo electro-hydraulic system, thus the PID parameters were tuned by the fuzzy tuner (Sinthipsomboon et al., 2011).
Realtime performance analysis of different combinations of fuzzy-PID and bias controllers for a two degree of freedom electrohydraulic parallel manipulator
2015, Robotics and Computer-Integrated ManufacturingCitation Excerpt :Song and Liu [26] used self-tuning fuzzy–PID controller to control switched reluctance motor. Giulio et al. [27] used artificial neural network for tuning PID controller parameter. However none of the above studies had focused on controlling the motion simulation of a manipulator attached to more than one hydraulic cylinder.
Actor-Critic Learning of Variable Damping Injection for Quadrotor Attitude Robust Control
2023, 2023 20th International Conference on Electrical Engineering, Computing Science and Automatic Control, CCE 2023A Novel Entropy-Maximizing TD3-based Reinforcement Learning for Automatic PID Tuning
2023, Proceedings of the American Control Conference