Abstract:High-precision control for nonlinear systems is impeded by the low-fidelity dynamical model and external disturbance. Especially, the intricate coupling between internal uncertainty and external disturbance is usually difficult to be modeled explicitly. Here we show an effective and convergent algorithm enabling accurate estimation of the coupled disturbance via combining control and learning philosophies. Specifically, by resorting to Chebyshev series expansion, the coupled disturbance is firstly decomposed into an unknown parameter matrix and two known structures depending on system state and external disturbance respectively. A Regularized Least Squares (RLS) algorithm is subsequently formalized to learn the parameter matrix by using historical time-series data. Finally, a higher-order disturbance observer (HODO) is developed to achieve a high-precision estimation of the coupled disturbance by utilizing the learned portion. The efficiency of the proposed algorithm is evaluated through extensive simulations. We believe this work can offer a new option to merge learning schemes into the control framework for addressing existing intractable control problems.
What problem does this paper attempt to address?
### Problems the paper attempts to solve
The paper aims to solve the problem of high - precision control in nonlinear systems, especially when the system is affected by the combined action of internal uncertainties and external disturbances. Specifically, the paper focuses on how to effectively estimate these coupled disturbances, thereby improving the control performance of the system. Traditional disturbance observer methods usually assume that the derivative of the disturbance is bounded, which has limitations in practical applications because this assumption requires that the system state is a priori stable, and a large derivative bound will lead to a large estimation error. In addition, although existing data - driven methods can handle time - varying external disturbances, they require a large amount of offline training and may not be able to achieve zero - error estimation in some cases.
### Main contributions of the paper
1. **Principle of variable separation**: A theorem (Theorem 1) is established to decompose the coupled disturbance into an unknown parameter matrix, a matrix related to the system state, and a vector related to the external disturbance, and the residual can be arbitrarily small.
2. **Supervised learning framework**: Based on the analytical assumption of external disturbances, a corollary (Corollary 1) is further developed, so that the unknown parameter matrix can be learned through supervised learning. The learning objective is formulated as a regularized least - squares (RLS) problem with a closed - form solution.
3. **High - order disturbance observer**: Using the learned knowledge, a high - order disturbance observer (HODO) is designed, which can achieve zero - error estimation of the coupled disturbance (Theorem 2).
### Method overview
1. **Decomposition of coupled disturbances**:
- Using Chebyshev series expansion, the coupled disturbance is decomposed into an unknown parameter matrix, a matrix related to the system state, and a vector related to the external disturbance.
- Through Theorem 1 and Corollary 1, the effectiveness and feasibility of this decomposition are proved.
2. **Learning of parameter matrix**:
- Construct a training data set, and use historical data to learn the unknown parameter matrix through the regularized least - squares (RLS) algorithm.
- Provide a closed - form solution of the learning objective to ensure the efficiency and stability of the learning process.
3. **Online estimation**:
- Design a high - order disturbance observer (HODO), and use the learned parameter matrix to estimate the coupled disturbance online.
- By designing an appropriate observer gain matrix, ensure that the estimation error converges exponentially to zero.
### Evaluation results
1. **Learning performance**:
- Through experiments on three nonlinear functions, the effectiveness of the supervised learning strategy is verified.
- Analyze the influence of measurement noise and parameter selection on learning performance, and the results show that appropriate parameter selection can achieve good learning results.
2. **Estimation performance**:
- In a second - order Newtonian system, by simulating coupled disturbances, the estimation performance of the proposed HODO is verified.
- The results show that, compared with traditional methods, the proposed HODO can estimate the coupled disturbance more accurately and significantly improve the tracking performance of the system.
### Summary
The paper proposes an effective method for estimating coupled disturbances by combining data - driven learning and control theory techniques. This method not only avoids the bounded - derivative assumption in traditional methods but also can achieve zero - error estimation without a large amount of offline training. The experimental results verify the effectiveness and superiority of this method.