A Novel Multimodal Process Monitoring Method Based on Improved Just-in-Time Learning-Assisted Stacked Quality-Related Autoencoder

The complex manufacturing processes are characterized by strong nonlinearity, high dynamics, and multimodality, which make it difficult to effectively monitor production safety and product quality. Therefore, there is a need to focus on the data characteristics of manufacturing processes and develop effective quality monitoring methods. In this context, this article proposes a novel improved just-in-time learning-assisted stacked quality-related autoencoder (IJITL-SQAE) for multimodal process monitoring. First, we construct the stacked quality-related autoencoder (SQAE) and quality-related autoencoder (QAE) basic models with interpretability for extracting deep-quality-related features. The proposed model takes into account the correlation and causality between variables through a nonfully connected form. Second, just-in-time learning (JITL) is improved to enhance the query efficiency and reliability of similar samples. The clustering algorithm is integrated in JITL to substantially reduce the query range. And multiple similarity metrics are fused based on distance and angle to improve the reliability of similar samples. Subsequently, similar samples are used for model parameter training to extract quality-related features, and monitoring statistics are constructed based on the extracted features to accomplish multimodal process quality-related fault detection. Finally, the effectiveness of this monitoring strategy is verified through a real manufacturing process, the hot strip mill process (HSMP). The results show that the proposed SQAE and QAE models outperform traditional methods. Meanwhile, the IJITL-SQAE multimodal monitoring strategy can effectively solve the multimodal quality monitoring problem.