Fractional-Order Dynamics Learning and Control via Data-Driven Approaches: Taking Soft Manipulator as an Example.

Area of Science:

• Control Engineering
• Applied Mathematics
• Machine Learning

Background:

• Fractional-order calculus enables modeling complex dynamics but presents challenges in system identification and control.
• Accurate modeling and stable control are crucial for systems exhibiting memory and nonlocality.

Purpose of the Study:

• To present a unified data-driven framework to address challenges in modeling and controlling fractional-order systems.
• To introduce novel deep learning architectures and control strategies for enhanced system performance.

Main Methods:

• Developed a fractional-order deep Lagrangian network (fPLCS-DeLaN) integrating physical priors and self-attention mechanisms for learning system dynamics.
• Proposed a hybrid network-based disturbance observer (T2F-CRNN) combining CNN, recurrence, and fuzzy inference for robust uncertainty estimation.
• Designed a fractional-order controller with finite-time convergence, input saturation compensation, and sliding mode constraints.

Main Results:

• fPLCS-DeLaN achieved modeling errors at least one order of magnitude lower with a minimal increase in computational time.
• The proposed fractional-order controller significantly reduced transient (23.1%) and steady-state (87.6%) tracking errors.
• Experiments on a soft manipulator platform validated the framework's superior model learning and tracking performance.

Conclusions:

• The proposed data-driven framework effectively tackles the complexities of fractional-order systems.
• The innovations in deep learning and control design lead to substantial improvements in accuracy, robustness, and performance.
• This unified approach offers a promising direction for advanced control applications involving fractional-order dynamics.