Self-Supervised Learning-Based Multiparameters Optimization Framework in Servo Systems

Servo control systems exhibit strong coupling among multiple control parameters, posing dual challenges of efficiency and stability during optimization. Traditional parameter tuning strategies based on systematic optimization often face difficulties, such as the need for precise modeling, low optimization efficiency, high dependence on data, and structural intrusiveness. This letter rethinks the problem of strongly coupled multiparameter optimization from a feature extraction perspective and proposes a physics-constrained self-supervised optimization framework. By employing a biphase hierarchical encoder and an independent encoder structure, the framework extracts key control parameter features from the system to enable effective optimization of servo control parameters. Furthermore, the method incorporates prior feature knowledge for network preraining, significantly reducing reliance on real data. Experiments conducted under both PI and ADRC control structures demonstrate that the proposed approach achieves rapid error convergence, offering high optimization efficiency while maintaining the structure’s nonintrusiveness.