关键词: 算法, 神经网络, 数据填充, 预测, 长序列缺失

Abstract:

Chemical processes are characterized by highly complex operating conditions, including elevated temperatures, high pressures, and severe corrosive environments. These harsh conditions frequently induce sensor malfunctions and signal anomalies, leading to missing values in long-term industrial time series data. Conventional statistical imputation techniques and machine learning approaches often fail to simultaneously capture global temporal dependencies and localized dynamic features, thus exhibiting limited effectiveness in such scenarios. To address this challenge, this study proposes a novel Deep Adaptive Transformer-enhanced Multi-scale Iterative Generative Adversarial Network (DAT-MIGAN) for missing data imputation. The framework integrates the generative capability of MIGAN with the sequence modeling strengths of DAT, thereby overcoming the limitations of conventional GAN-based methods in representing both short- and long-term dependencies. By introducing multi-scale attention mechanisms within the latent space, the proposed method establishes a global–local collaborative imputation strategy that enables more accurate reconstruction of missing segments in long-sequence data. Extensive experiments conducted on the Tennessee Eastman (TE) benchmark dataset and real-world Chemical process plant data verify the robustness and effectiveness of DAT-MIGAN. Results demonstrate that the proposed approach substantially outperforms existing methods in handling complex missing patterns, offering a reliable solution for accurate long-sequence data imputation in the Chemical process industry.

Key words: algorithm, neural networks, data imputation, prediction, long-sequence missing