Deep RL Based Notch Filter Design Method for Complex Industrial Servo Systems
- PDF / 2,488,313 Bytes
- 10 Pages / 594.77 x 793.026 pts Page_size
- 12 Downloads / 163 Views
ISSN:1598-6446 eISSN:2005-4092 http://www.springer.com/12555
Deep RL Based Notch Filter Design Method for Complex Industrial Servo Systems Tae-Ho Oh, Ji-Seok Han, Young-Seok Kim, Dae-Young Yang, Sang-Hoon Lee, and Dong-Il “Dan” Cho* Abstract: This paper proposes a deep reinforcement learning (deep RL) method for simultaneously designing several notch filters in complex industrial servo systems. Notch filters are highly effective for suppressing resonances in motion control systems and are widely utilized in industry. However, severe limitations exist in complex servo systems because there are many vibration modes that are difficult to identify. In such cases, several notch filters must be used, but the task of tuning these filters involves lengthy empirical procedures by well-experienced engineers. To automate this tuning process, this paper proposes a novel design method that can design several notch filters simultaneously for the first time. In this method, a deep deterministic policy gradient (DDPG) algorithm with a vector stability margin as the reward function is utilized to find filter parameters in the frequency domain. The proposed method simultaneously finds a set of many parameters for several notch filters that are optimal with respect to stability. Using a real industrial servo system that has multiple resonances, it is demonstrated that the proposed method effectively finds the optimal parameters for several notch filters and successfully suppresses multiple resonances to provide desired performances. Keywords: Deep reinforcement learning, highly complex servo system, multiple notch filters, resonance suppression.
1.
INTRODUCTION
Controller tuning and filter design problems are one of the classical topics in control. In industrial servo systems, many types of filters such as low-pass filters and notch filters are widely used to compensate for the highfrequency noise and resonances [1–4]. There are various methods to tune controllers and filters. Traditionally, for PID controllers, the Ziegler-Nichols tuning method and its variants are widely used, which model the plant as a first-order low-pass filter with delay [5,6]. More advanced H∞ optimization algorithms combined with system identification methods have also been proposed and applied to flexible-link manipulator [7], wafer stage control [8], and small-scale helicopter system [9]. Due to the limitation of the model-based control technique, several data-driven approaches have been proposed [10, 11]. Iterative learning control has been proposed to tune a feedforward controller for tracking and compensating for noise using the data obtained from a few iterations on a real plant [12,13]. Iterative feedback tuning can
directly tune the controller parameters by calculating the gradients of the controllers using the data obtained from different references [14]. Evolutionary algorithms such as the genetic algorithm and particle swarm optimization algorithm have been researched to optimize controller parameters and applied to various systems [15, 16]. A
Data Loading...
