Real‐time data‐driven PID controller for multivariable process employing deep neural network

Abstract

The complex industrial processes exhibiting nonstationary and multivariable with time‐varying dynamics result in low accuracy. Also, stability compensation is difficult to be obtained by a conventional PID controller. Hence, a deep learning‐based data‐driven PID controller is designed for unmodeled dynamics compensation for complex industrial processes. In this research work, a nonlinear PID controller is designed with a deep neural network (DNN) model from unmodeled dynamics of the complex industrial processes. To validate the performance, results from stability compensation and convergence of the model parameters for closed‐loop systems were obtained. When tested on a real‐time twin tank system, it achieved an accurate output flowrate with 97.65% accuracy and 1.89% peak overshoot compared with conventional PID controller. Both simulated and experimental results validate that proposed controller has improved stability and uniform convergence of system variables. The proposed deep learning‐based PID controller was employed on a twin tank control system. This confirms the feasibility and practical application of a real‐time complex process.

Data-Driven Adaptive Steady-State-Integral-Derivative Controller Using Recursive Least Squares With Performance Conditions

“…The proposed control algorithm has the advantage of not needing any system parameters and dynamics for target value tracking. Related previous works are based on model-free adaptive control [3], [4], [5], [6], [7], [8], [9], [10], [11], data-driven control, and artificial intelligence control [16], [17], [18], [19], [20], [21]. In the previous related works, generally, the conventional proportional-integral-derivative control scheme was adopted, and the gain adaptation law was proposed using various methods.…”

“…Because the derivative of function of the Lyapunov candidate with respect to time t should always be negative for asymptotic stability, the derivative of the function J with respect to time t was derived and used for stability analysis. Using (16), the derivative of J with respect to time t can be derived as follows [10], [17].…”

“…• Sinusoidal target velocity (simulation) Figs. [14][15][16][17][18][19][20][21][22] show the simulation-based performance evaluation results when a sinusoidal target velocity was applied as the target state.…”

“…Therefore, various studies using data-driven and learning methods have been conducted to improve system control performance. Jeyaraj and Nadar [16] designed a deep learning-based data-driven PID controller for unmodeled dynamics compensation of complex industrial processes. The authors investigated specific modeling of data-driven proportionalintegral-derivative for complex industrial processes with consideration of online PID parameter tuning.…”

Robust adaptive control of nonlinear dynamic systems using hybrid sliding mode regressive neural learning technique

“…Researchers in Liu et al (2021b) suggested the improved distributed differential evolution (DDE)-based meta-heuristic method for training the NN-based control model applied to uncertain nonlinear plants. Jeyaraj and Nadar (2022) proposed an online adaptive PID control approach applied to nonlinear MIMO plants based on NN learning method. The principal disadvantage of these NN controllers relates to the enormous computing-time that hampers the facilitation of these schemes in practice.…”