Software-defined self-learning control system for industrial robots by using reinforcement learning.

Area of Science:

• Manufacturing Automation
• Artificial Intelligence in Industry
• Control Systems Engineering

Background:

• Modern manufacturing demands adaptable systems for changing production needs.
• Current software-based controls have limitations in real-time response and handling unexpected equipment states.
• Skilled human operators are often required, posing a bottleneck.

Purpose of the Study:

• To propose a novel self-learning control system for manufacturing equipment.
• To enable existing equipment to adapt to new tasks and anomalous conditions via software updates.
• To improve system flexibility and reduce reliance on human operators.

Main Methods:

• Integration of anomaly detection and reinforcement learning (RL) algorithms.
• Utilizing a virtual environment to train RL models on various anomalous equipment statuses.
• Developing an anomaly detection algorithm to trigger pre-trained control models for specific statuses.

Main Results:

• The proposed system successfully adapted existing equipment to new tasks and states through software updates.
• Anomaly detection and control model switching occurred within 1.5 seconds.
• Validation was performed on a SCARA robot under simulated overcurrent conditions without extra sensors.

Conclusions:

• The self-learning control system effectively enhances manufacturing flexibility and responsiveness.
• The integration of anomaly detection and RL offers a viable solution for autonomous equipment adaptation.
• This approach reduces the need for specialized sensors and human intervention in dynamic manufacturing environments.