Development and Optimization of Parameter Recommendation Model for Automotive Resistance Spot Welding

Technical Summary: Development and Optimization of Parameter Recommendation Model for Automotive Resistance Spot Welding

Problem Statement
In automotive manufacturing, the design of Resistance Spot Welding (RSW) parameters is critical for joint quality and production efficiency. Traditional methods, such as weldability-lobe testing and Design of Experiments (DoE), are time-consuming and struggle to adapt to the increasing variety of sheet-stack combinations found in modern lightweight vehicles. While data-driven approaches (e.g., neural networks, ensemble learning) have improved quality prediction, they face significant limitations in parameter recommendation (the inverse problem). Purely data-driven models suffer from a non-uniqueness problem: for a given sheet-stack configuration and target nugget size, multiple valid welding schedules (e.g., high-current/short-time vs. low-current/long-time) exist. A model trained solely on statistical correlations may learn an averaged or inconsistent solution, leading to poor generalization when applied to new vehicle models or different production environments. Furthermore, existing models often lack physical interpretability, making it difficult to control specific process outcomes like cycle time or defect mitigation.

Methodology
The authors propose a physics-constrained ensemble learning framework to generate welding parameters (electrode force, preheating time/current, cooling time, welding time, and welding current) based on sheet attributes and a target nugget size.

1. Baseline Ensemble Model: A heterogeneous ensemble learning model was constructed using six base learners: Gaussian Process Regression (GPR), Random Forest Regression (RFR), Multilayer Perceptron (MLP), Support Vector Regression (SVR), LightGBM, and XGBoost. These models map input features (sheet thickness, strength, coating, stack-up, target nugget size) to the six output parameters. A weighted aggregation strategy, determined by five-fold cross-validation errors, fuses the outputs to maximize accuracy.
2. Physics-Informed Constraint (Heating Intensity): To address the non-uniqueness issue, the authors introduced heating intensity ($u$) as an explicit physical constraint. Defined as the ratio of the squared average welding current to the total welding time ($u = I^2 / T$), this variable quantifies the temporal concentration of heat input.
   • High $u$ corresponds to "hard" schedules (high current, short time).
   • Low $u$ corresponds to "soft" schedules (low current, long time).
3. Multi-Loss Optimization: The model employs a multi-loss objective function ($L = L_{para} + \lambda L_{HI}$) that simultaneously minimizes:
   • $L_{para}$: The prediction error of the welding parameters.
   • $L_{HI}$: The deviation between the heating intensity calculated from the generated parameters and a target heating intensity derived from measured data.
   • An adaptive uncertainty-based weighting strategy (inspired by Kendall et al.) was used to dynamically balance these two loss terms during training, preventing one objective from dominating the other.

Key Contributions

• Resolution of Non-Uniqueness: By incorporating heating intensity as an explicit constraint, the framework compresses the multi-solution space, ensuring that the generated parameters are unique and physically consistent for a given input and target strategy.
• Enhanced Cross-Vehicle Generalization: The model demonstrates superior transferability between different vehicle models (e.g., from a sedan to an SUV) compared to purely data-driven baselines. The physical constraint forces the model to learn transferable process relationships rather than vehicle-specific statistical patterns.
• Controllable Parameter Generation: The framework allows engineers to explicitly select "hard" or "soft" welding schedules by adjusting the target heating intensity input, offering a practical tool for optimizing cycle times or mitigating defects like expulsion.

Experimental Results
The study utilized a dataset comprising 450 unique two- or three-sheet stack-up configurations derived from production records of two vehicle models: a sedan (Vehicle 1) and a sport utility vehicle (Vehicle 2). While Vehicle 1 contained 265 configurations and Vehicle 2 contained 385 configurations, the total set of unique configurations formed from these materials was 450, with approximately 200 configurations common to both models. The dataset covered various steel grades (CR01, CR04, DP590, DP780, PHS1300T).

• Same-Vehicle Performance: The physics-constrained model improved prediction accuracy across all six parameters compared to the optimized baseline, with an average gain of approximately 1.35 percentage points. The most significant improvement was in preheating time prediction (95.3% to 97.5%).
• Cross-Vehicle Performance: The physics-constrained model significantly outperformed the baseline in transfer scenarios. While the baseline accuracy dropped to as low as 64.3% for preheating current in cross-vehicle tests, the constrained model maintained accuracies between 79.9% and 95.8%. The overall mean cross-vehicle accuracy improved from ~85.5% (baseline) to ~90.0%, with a maximum gain of 15.6% for specific parameters.
• Ablation Study: Removing the heating intensity loss term ($L_{HI}$) resulted in a performance drop of 0.1–15.6 percentage points in cross-vehicle tests, confirming that the physical constraint is the primary driver of generalization capability.
• Sensitivity Analysis: Experiments varying the target heating intensity confirmed that the model could systematically generate different current-time combinations (e.g., reducing time from 300ms to 232ms while increasing current from 8.9kA to 9.4kA) while maintaining the target nugget size (~5.0 mm), validating the controllability of the generated schedules.

Significance and Claims
The paper claims that this work establishes a pathway for the intelligent management of the entire RSW design and manufacturing process. By moving beyond purely data-driven regression to a mechanism-data fusion approach, the study addresses the critical challenge of generating reliable, unique, and interpretable welding parameters for diverse automotive applications. The authors assert that their framework enhances the engineering applicability of ensemble learning by ensuring physical consistency and enabling the controllable generation of process schedules, thereby reducing development cycles and improving manufacturing responsiveness in the face of increasing material complexity.