TL;DR
This paper introduces a novel Robust Fuzzy-Neural Sliding-Mode Controller (RFSC) that automatically adapts its fuzzy-neural network structure to control unknown nonlinear systems. By combining self-structuring fuzzy-neural networks (SFNNs) with sliding-mode and robust control, the RFSC achieves accurate tracking performance even in the presence of disturbances, eliminating the need for manual tuning of the network architecture.
What Problem Does This Paper Address?
Designing controllers for complex, unknown nonlinear systems is a significant challenge. Traditional fuzzy-neural network (FNN) based adaptive control methods often require a predefined network structure, which is difficult to determine optimally. This trial-and-error process can lead to either computationally expensive systems with too many fuzzy rules or poor performance with too few. Furthermore, these systems can struggle with external disturbances and parameter uncertainties. This paper addresses the need for an adaptive control system that can automatically learn and adapt its structure and parameters to achieve robust tracking performance for unknown nonlinear systems, even when faced with disturbances.
Methods
The core of the proposed RFSC system is a Self-Structuring Fuzzy-Neural Network (SFNN) identifier. This SFNN operates in two phases:
- Structure-Learning Phase: This is the novel aspect. Instead of a fixed structure, the SFNN dynamically adjusts its architecture through three mechanisms:
- Growing Membership Functions: New membership functions are added when input data falls outside the range of existing ones, ensuring better coverage of the input space.
- Splitting Fuzzy Rules: Existing fuzzy rules are split when their associated weights experience large variations, allowing for finer approximation of complex dynamics.
- Pruning Fuzzy Rules: Redundant or insignificant fuzzy rules are removed to reduce computational load and prevent the network from growing excessively.
- Parameter-Learning Phase: Once the structure is adapted, the parameters of the membership functions and fuzzy rules are fine-tuned using a gradient descent algorithm derived from a Lyapunov function. This phase aims to accurately estimate the unknown system dynamics.
The SFNN acts as the computational controller, estimating the system's lumped dynamics. A robust controller is then designed to guarantee L2 tracking performance and attenuate disturbances, ensuring stability even with approximation errors from the SFNN. The overall control law is a combination of these two components.
Datasets
Not discussed in the paper. The paper focuses on applying the proposed RFSC to simulated nonlinear dynamic systems.
Key Insights and Ideas
- Automated Network Design: The most significant insight is the development of a structure-learning algorithm for fuzzy-neural networks that eliminates the need for manual, trial-and-error tuning. This greatly simplifies the design process and makes adaptive control more practical.
- Robustness through Hybridization: Combining the adaptive learning capabilities of SFNNs with the guaranteed stability of sliding-mode control and the disturbance attenuation of robust control creates a powerful and resilient control system.
- Dynamic Adaptation: The ability of the SFNN to grow, split, and prune rules in real-time allows it to adapt to changing system dynamics or operating conditions, leading to more efficient and effective control.
- Balancing Complexity and Performance: The structure-learning mechanisms directly address the trade-off between the number of fuzzy rules and the desired control performance, aiming for a concise yet effective network structure.
Key Results
The proposed RFSC system was applied to three nonlinear dynamic systems: two chaotic systems (Duffing's equation and Chua's circuit) and a general nonlinear system. In all cases, the simulation results demonstrated:
- Favorable Tracking Performance: The RFSC achieved accurate tracking of desired trajectories for the system states, even in the presence of external disturbances.
- Effective System Identification: The SFNN identifier successfully estimated the unknown system dynamics and the lumped disturbance.
- Concise Network Structure: The structure-adaptation algorithm resulted in a relatively small number of fuzzy rules (e.g., 5 rules for the Duffing system, 5-7 for Chua's circuit, and 5 for the general nonlinear system), indicating efficient learning.
- Robustness to Disturbances: The system maintained good tracking performance when disturbances were introduced, showcasing the effectiveness of the robust control component.
Limitations and Future Work
- Parameter Tuning: While the network structure is adapted automatically, the paper mentions that the parameters for the RFSC (e.g., learning rates, thresholds) are selected through "some trails" to achieve the best performance. This suggests that some manual tuning might still be required for optimal performance.
- Computational Load: Although the pruning mechanism aims to reduce computational load, the paper doesn't provide explicit comparisons of computational complexity against other methods.
- Real-world Implementation: The paper focuses on simulation results. Further validation on real-world hardware would be beneficial to assess practical implementation challenges.
The authors suggest that the proposed RFSC scheme can be applied to other unknown nonlinear dynamical systems, implying a broad applicability and potential for further research in various control applications.
Summary updated