Self-Learning Disturbance Observer (SLDO)

• Self-Learning Disturbance Observer (SLDO) is an adaptive framework that combines traditional observer theory with learning mechanisms to estimate and compensate for unknown, time-varying disturbances.
• It employs techniques like Chebyshev polynomial decomposition, neuro-fuzzy systems, and experience replay to address complex nonlinear dynamics and improve robustness.
• SLDO guarantees exponential convergence and enhanced performance over conventional disturbance observers, proving effective in robotics, AUVs, and other precision control applications.

A Self-Learning Disturbance Observer (SLDO) is a class of observers for dynamical systems whose core purpose is real-time estimation and compensation of disturbances—particularly those that are unknown, time-varying, or coupled with system states and exogenous signals—by combining observer theory with explicit learning mechanisms. SLDOs systematically integrate model-based estimation with data-driven adaptation (e.g., regularized regression, adaptive neuro-fuzzy systems, reinforcement learning, or experience replay) to achieve greater robustness, accuracy, and convergence guarantees compared to conventional disturbance observers (DOBs), especially in the context of high-precision robotics and nonlinear control under uncertainty (Jia et al., 2024, Kayacan et al., 2021, Kayacan, 2021, Wang et al., 2019, Li et al., 2020, Kayacan et al., 2021).

1. Formal Class of Systems and Problem Setting

SLDOs are applicable to a wide range of plant models, including general control-affine, nonlinear, and uncertain systems with both matched and mismatched disturbances. A general system model is:

$\dot{x} = f_x(x) + f_u(x)u + \Delta(x, d(t))$

where $\Delta(x, d(t))$ is an analytic disturbance term coupling the system state $x$ and exogenous disturbance $d(t)$. The observer’s objective is precise estimation of $\Delta(x,d(t))$ to enable high-fidelity feedback or feedforward compensation (Jia et al., 2024).

In advanced variants, the disturbance is modeled as the output of an exosystem:

$$\dot{\varepsilon}_T = S \varepsilon_T, \quad d(t) = D \varepsilon_T$$

where $S$ (unknown) and $D$ are constant matrices, permitting the generation and learning of disturbances with arbitrary unknown temporal structure such as periodic signals (Li et al., 2020).

2. Structural Principles and Architectures

SLDOs universally depart from classical DOBs by introducing a self-learning module that progressively dominates disturbance estimation, enabling unbiased and robust operation under general, time-varying uncertainty.

2.1. Chebyshev Polynomial Decomposition and HODO

The SLDO framework in (Jia et al., 2024) achieves universal function approximation of coupled disturbances by expressing $\Delta(x, d(t))$ via Chebyshev series:

$\Delta(x, d(t)) \approx \Theta\, B(x)\, \xi(t)$

with $\Delta(x, d(t))$0 an unknown parameter matrix, $\Delta(x, d(t))$1 the state-dependent Chebyshev basis, and $\Delta(x, d(t))$2 the time-dependent Chebyshev vector. This transforms coupled disturbance learning into an off-line regularized least-squares regression for $\Delta(x, d(t))$3, and enables the design of a polynomial higher-order disturbance observer (HODO) that guarantees exponential convergence:

$\Delta(x, d(t))$4

$\Delta(x, d(t))$5

The observer gain $\Delta(x, d(t))$6 is chosen to render the error dynamics exponentially stable via pole placement or time-varying least-squares observer synthesis (Jia et al., 2024).

2.2. Observer–Neuro-Fuzzy Hybrid Structures

A major class of SLDO architectures employs a parallel configuration: (i) a conventional DOB path (e.g., BNDO) providing an initial estimate and (ii) a learning-enhanced path, typically a neuro-fuzzy system, adapting to the residuals and learning dynamics unmodeled by the nominal observer (Kayacan et al., 2021, Kayacan, 2021, Kayacan et al., 2021).

• Conventional path: Realizes a basic estimation cascade such as $\Delta(x, d(t))$7, which alone is only bias-free for constant disturbances.
• Learning path: Employs an interval type-2 neuro-fuzzy system (T2NFS) or a more general NFS, driven by the temporal derivatives of $\Delta(x, d(t))$8, to adaptively model disturbance components that vary in time or depend on unmodeled dynamics. Adaptation proceeds via sliding-mode learning laws, e.g.,

$\Delta(x, d(t))$9

where $x$0 are rule outputs and $x$1 balances upper/lower memberships (Kayacan et al., 2021).

• Feedback error learning (FEL) loop: The difference between conventional and NFS output is used as the sliding variable for adaptation, progressively driving the observer output to rely exclusively on the learned path.

2.3. Experience Replay and Adaptive Regression

The approach in (Li et al., 2020) introduces experience replay for online parameter adaptation. Past regressor data is stored and leveraged in a “concurrent learning”-style batch update, ensuring exponential parameter convergence, which is otherwise limited to slow asymptotic rates in online gradient-based estimators.

2.4. Deep Learning/RNN Approaches

In fully data-driven RL-based SLDOs, RNN (specifically, stacked GRU) structures encode past state and control history into a latent disturbance representation. The observer and controller are trained jointly using actor-critic RL, allowing the controller to “anticipate” future disturbance trends (Wang et al., 2019).

3. Learning, Estimation, and Adaptation Laws

SLDO core methodologies integrate both off-line and online learning.

3.1. Off-line Regularized Least Squares

For Chebyshev-based SLDOs, offline system identification uses a regularized least-squares objective:

$x$2

with a closed-form solution (see [(Jia et al., 2024), Eq. 3]). Model order $x$3 and regularization $x$4 are tuned via cross-validation.

3.2. Online Adaptive Learning

For neuro-fuzzy architectures, adaptation is explicit:

• Neuro-fuzzy rule adaptation is performed via sliding-mode laws, with learning rates chosen to dominate worst-case time derivatives of disturbances, guaranteeing finite-time escape from the initial bias and convergence to the true disturbance.
• Experience replay update augments gradient adaptation by incorporating stored historic regressors and outputs. A rank condition on the data suffices for parameter identifiability and exponential convergence (Li et al., 2020).

4. Stability, Convergence, and Robustness Guarantees

Rigorous convergence proofs are ubiquitous in SLDO literature.

• Lyapunov-based proofs establish uniform exponential (or finite-time) convergence of estimation error:

$x$5

leading to $x$6 exponentially for disturbance estimation error.

• SMC-style learning laws guarantee that, provided the learning rate exceeds the maximal disturbance derivative, the estimation error converges to zero in finite time under bounded, time-varying $x$7 (Kayacan et al., 2021, Kayacan, 2021, Li et al., 2020).
• Robustness to measurement noise and unmodeled dynamics is enhanced in type-2 fuzzy and replay-based estimators, as evidenced by reduced MSE and superior performance in challenging disturbance scenarios.

5. Performance and Comparative Assessment

Simulation across varied platforms (robotic manipulators, chaotic oscillators, position regulation of AUVs) empirically confirms SLDO superiority.

Method	Tracking RMSE	Disturbance Estimation MAE	Convergence Time
Baseline (PD)	0.50	–	–
Standard DOB	0.15	0.10	5 s
SLDO (HODO)	0.02	0.005	2 s

• For example, in (Jia et al., 2024), SLDO achieved one order of magnitude lower estimation MAE and tracking RMSE than standard DOBs.
• Neuro-fuzzy SLDOs demonstrated an RMS tracking error reduction of 60% over BNDOs in strongly time-varying $x$8 (Kayacan et al., 2021).
• SLDO robustness to both matched and mismatched, time-varying disturbances is consistently better than DOB and even sophisticated nonlinear or integral SMC schemes (Kayacan, 2021, Kayacan et al., 2021).

6. Implementation Guidelines and Practical Considerations

SLDO implementation involves both offline (regression/data collection) and online (real-time adaptation) phases:

• Data requirements: Offline learning requires $x$9–$d(t)$0 samples for functions over moderate-dimensional spaces (Jia et al., 2024).
• Polynomial order and regularization: Typically $d(t)$1–$d(t)$2 suffices; higher $d(t)$3 increases computational cost ($d(t)$4 for $d(t)$5 basis functions).
• Observer gain selection: For observer dynamics, eigenvalues of the error system are placed in $d(t)$6 to $d(t)$7 regime for a suitable speed/robustness tradeoff.
• Complexity: Both HODO and neuro-fuzzy pathways are computationally tractable for real-time deployment, with example implementations reporting sub-millisecond update times for $d(t)$8 fuzzy rules (Kayacan et al., 2021).
• Chattering: Mitigated via boundary layers or continuous approximations of the sign function in SMC adaptation laws.
• Robustness to residual modeling error: If Chebyshev expansion error is non-negligible, low-gain integral or $d(t)$9 correction can be added.

7. Extensions, Limitations, and Outlook

SLDO methodologies extend naturally to multi-input multi-output systems, higher-order plants, and can be combined with model-based or deep-learning controllers. Reinforcement learning-based self-learning DOB architectures (e.g., DOB-Net) demonstrate the feasibility of fully data-driven SLDOs, albeit at the expense of sample efficiency and the need for careful domain adaptation (Wang et al., 2019).

Limitations include:

• Dependence on the richness and diversity of collected data for regression- or replay-based methods.
• The need for boundedness in disturbance derivatives to guarantee adaptation-convergence.
• Increased computational complexity as system dimension or disturbance representation complexity grow.

Prospective research explores embedded, real-time flight deployment, meta-learning for rapid adaptation to unseen disturbance classes, and integration with $\Delta(x,d(t))$0 robustification techniques (Jia et al., 2024).

Bibliography

• "Learning-based Observer for Coupled Disturbance" (Jia et al., 2024)
• "A Self-Learning Disturbance Observer for Nonlinear Systems in Feedback-Error Learning Scheme" (Kayacan et al., 2021)
• "Sliding Mode Control for Systems with Mismatched Time-Varying Uncertainties via a Self-Learning Disturbance Observer" (Kayacan, 2021)
• "DOB-Net: Actively Rejecting Unknown Excessive Time-Varying Disturbances" (Wang et al., 2019)
• "Adaptive Finite-time Disturbance Rejection for Nonlinear Systems using an Experience-Replay based Disturbance Observer" (Li et al., 2020)
• "Feedback Linearization Control for Systems with Mismatched Uncertainties via Disturbance Observers" (Kayacan et al., 2021)