Data-Driven Iterative Learning Model Predictive Control With Self-Modified Prior Knowledge.

Area of Science:

• Control Engineering
• Artificial Intelligence
• Manufacturing Systems

Background:

• Iterative learning model predictive control (ILMPC) is effective for batch manufacturing but struggles with varying operational conditions.
• Existing ILMPC methods require identical trials, limiting their use when production requirements change.

Purpose of the Study:

• To develop an adaptable ILMPC strategy that can handle unconformable prior information and trial-varying conditions.
• To enhance the flexibility of ILMPC for digitized batch manufacturing processes.

Main Methods:

• A data-driven self-modification scheme is embedded into ILMPC to adapt historical data to current trial conditions.
• An adaptive deep neural network (DNN) imitates control actions and generates reference signals for iterative learning.
• A tube control framework is employed for 2-D optimization to ensure time-domain bounded stability, mitigating DNN approximation errors.

Main Results:

• The proposed ILMPC method demonstrates superior adaptability to significant changes in operating reference and duration.
• Theoretical validation confirms the iteration-domain bounded convergence of the developed ILMPC system.
• Simulations on a nonlinear injection molding process show improved tracking performance and disturbance rejection.

Conclusions:

• The novel ILMPC approach effectively addresses limitations of traditional methods in dynamic manufacturing environments.
• The integration of DNNs and a self-modification scheme enhances ILMPC's robustness and flexibility.
• The method ensures both time-domain stability and iteration-domain convergence for reliable control.