Human-AI cooperative generative adversarial network (GAN) for quality predictions of small-batch product series

Abstract

This paper emphasizes the importance and the novel methodology of predicting product quality for smart manufacturing, particularly for small-batched power equipment productions with demand-specified variations in the energy sector. Accurate evaluation and parameter adjustments are crucial for achieving the highest-quality results. To predict final product quality, even with a small sample size of a specific transformer type (and unique design specification), we proposed a novel method using a generative adversarial network (GAN) for model training and fine-tuning. This approach is crucial in the context of the digital transformation of complex industrial machinery industries. This research was undertaken with a prominent power transformer manufacturer and its supply chain collaborators. To train and validate the model, data were gathered from actual systems, utilizing the expertise of the company’s personnel. The dataset included critical power transformer metrics, including core loss values, which are crucial for accurate predictions. GAN generated realistic, high-quality samples that enhanced the training process and enhanced the model’s generalization capabilities, ultimately resulting in more accurate predictions. The experimental findings indicate that the proposed approach offers manufacturers a powerful tool for predicting the quality of complex, high-value, and highly specialized industrial products, ultimately leading to a reduction in production costs. Furthermore, in comparison to the models employed in previous studies within this research series, which include AdaBoost, ARIMA-AdaBoost, and LSTM-AdaBoost, the GAN model has been enhanced to address quality prediction models for other small-batch transformer productions. In the present study, the generator and discriminator components of the GAN effectively generated usable data to enhance the limited dataset, thereby mitigating the challenges associated with a small sample size. Consequently, the methodology presented in this study has the potential for broad application across diverse industrial manufacturing sectors, thereby mitigating the constraints associated with predicting product quality and substantially enhancing the accuracy of the models.