Domain-Adversarial Imputation Transfer Network for Industrial System Working Condition Identification

Introduction: The working condition identification of an industrial system (WCIIS) is important for optimizing system operation in real time, thereby enhancing production efficiency. However, existing WCIIS methods face challenges such as cross-domain adaptation and data missing. Methods: To address these issues, we propose a novel domain adversarial transfer network (DAITN). Specifically, we propose a self-attention generation-adversarial multi-discriminator dual-imputation module (SGMDM) to address the problem of data missing in WCIIS. Furthermore, two asymmetric encoder networks are designed to learn hierarchical representations from both the source and target domains. The network parameters learned from the source domain are utilized to initialize the parameters trained by the target domain. Additionally, domain adversarial training with a loss function is employed to handle distribution drift between the source and target domains. Results: The performance of SGMDM is evaluated using the anode current signal (ACS), the Tennessee Eastman Process (TEP), and the Continuous Stirred Tank Heater (CSTH). The validation of the DAITN is conducted on ACS and TEP, and the accuracies on ACS and TEP are 99.17+0.69, 99.11±0.76, respectively. Discussion: The DAITN model works well because it combines random forests and a selfattention GAN to fill in loss data accurately, and a domain-adversarial network with two CNNs that share knowledge from a source to a target domain, helping the model generalize better across different datasets. Conclusion: Experimental results demonstrate that the proposed method effectively addresses cross-domain WCIIS challenges in the presence of data missing.