Reinforcement learning-based real-time optimization of friction stir welding parameters for copper–aluminium dissimilar interfaces
Аннотация
Friction Stir Welding (FSW) of aluminium alloy dissimilar welding is essential to lightweight constructions in the automotive, aerospace and energy sector, yet stabilizing the quality of the joints in this context continues to be a challenge because of extreme thermal gradients, Intermetallic compound (IMC) formation and the dynamics of the process. Traditional offline or rule-based optimization methods are not flexible to material flow and heat production changes in real-time, which usually leads to defects and poor mechanical behavior. The paper was dedicated to the optimization of the main FSW process parameters, such as the tool shoulder-to-pin diameter ratio (SP), tool rotational speed (TRS), and tool traverse speed (TTS) with the help of a reinforcement learning (RL)-based real-time optimization algorithm to maximize the weld integrity and performance. The suggested methodology incorporated the in-situ sensor feedback, such as torque, axial force, interface temperature, and tool vibration in a digital process environment to allow the continuous state monitoring during welding. A model-free RL model based on Deep Q-Network (DQN) and Proximal Policy Optimization (PPO) to dynamically adjust SP, RS and TF in order to regulate heat input and material flow. An objective rewarding multi-objective was developed to reduce IMC thickness, thermal variability, and maximise tensile strength and joint effectiveness. The dissimilar aluminium alloy butt joints were experimentally validated to converge rapidly on policy and to be able to stay stable amidst process disturbances. The RL-optimized process demonstrated a 19.7% increase in tensile strength (142 ± 5.8 MPa to 170 ± 4.2 MPa), a 23.3% reduction in IMC thickness (8.6 ± 0.7 µm to 6.6 ± 0.5 µm), and a 17.2% improvement in electrical conductivity compared to the optimized static baseline. The study demonstrates that reinforcement learning enables intelligent and self-optimizing FSW through real-time adjustment of SP, TRS, and TTS, thereby offering a scalable route toward autonomous manufacturing of high-performance Cu–Al dissimilar joints. Reinforcement learning-enabled real-time optimization of Cu–Al friction stir welding for enhanced joint integrity and performance. • Introduces the first real-time reinforcement learning framework for optimizing Cu–Al friction stir welding parameters. • Integrates in-situ sensing with deep RL (DQN and PPO) for autonomous, closed-loop process control. • Achieves simultaneous reduction of intermetallic compound thickness and enhancement of tensile strength and joint efficiency. • Demonstrates superior process stability and defect suppression compared to conventional fixed-parameter FSW. • Provides a scalable pathway toward intelligent, self-optimizing dissimilar material joining for Industry 4.0 applications.