MMD-DCGAN: A Robust Approach for Missing Data Imputation in Manufacturing Processes

Data quality is a critical factor in enhancing the intelligence of process industries. Incomplete or corrupted data can significantly affect the performance of industrial soft-sensing models. The paper proposes an improved Deep Convolutional Generative Adversarial Network (DCGAN) model, enhanced with Lipschitz continuity and Maximum Mean Discrepancy (MMD), for the imputation of missing data. The proposed MMD-DCGAN model addresses common issues in traditional GANs, such as mode collapse, unstable training, and high computational costs. The model is validated using both synthetic benchmark data and real-world wastewater treatment data, demonstrating its superior performance in generating accurate and consistent data. The experimental results indicate that the MMD-DCGAN model reduces the imputation error by 10% for benchmark data and 15% for real-world data, proving its effectiveness and practical application potential in industrial scenarios.