Method for robot spray coating thickness characteristics prediction using a GA-WOA bi-level self-tuned two-stage physically feasible range-constrained conditional adversarial regression GAN

Robotic spray painting is a strongly experience-dependent manufacturing process in which coating quality is governed by multiple nonlinearly coupled process parameters, while experimental data are costly to acquire. To enable pre-spray coating thickness characteristics prediction under small-sample conditions, this paper proposes GAWOA-BiTS-PFRC-CRGAN, a Physically Feasible Range-Constrained conditional adversarial regression framework for robotic water-based paint spraying. The method formulates coating thickness characteristics prediction as a conditional dual-output regression problem and adopts a generator-discriminator architecture to learn the nonlinear mapping from process parameters to film-quality indicators (center-point mean film thickness and cross-sectional thickness range). To improve trainability and mitigate oscillations typical of adversarial regression, a two-stage training strategy is introduced, consisting of supervised pre-training followed by adversarial joint training. Physically feasible range-constrained mapping is embedded in the forward inference to enforce engineering-feasible output ranges. Furthermore, a GA-WOA bi-level online weight self-tuning mechanism is developed to adaptively optimize multi-loss weights using validation performance as fitness, reducing reliance on manual hyperparameter tuning and improving training stability. Experiments on an ABB robot spraying platform demonstrate that the proposed method achieves consistently lower prediction errors than representative baselines (MLP, SVR, RF, RBF, ANFIS, and SimpleGAN), and ablation/robustness studies verify the contributions of key modules to prediction accuracy, training stability, and robustness under the investigated experimental conditions.