Semantic Reconstruction of Multimodal Process Data With Dual Latent Space Constraints

Industrial processes often generate multimodal data with complex dynamics and distinct characteristics across multiple stages or conditions. Reconstructing the intrinsic semantic information from such data is essential for process monitoring and fault diagnosis. However, existing feature extraction methods often prioritize minimizing reconstruction error, which can overlook the importance of semantic interpretation and lead to limited accuracy and interpretability in the reconstructed results. To address this limitation, a novel semantic reconstruction framework for multimodal process data, driven by dual latent space constraints, is proposed. This approach utilizes a semantic consistency constraint and a multimodal characteristic constraint to extract latent space representations that effectively capture the intrinsic characteristics of the multimodal data. The core innovation of this framework lies in the integration of these dual constraints to obtain a comprehensive and interpretable representation of multimodal data. By jointly optimizing the dual latent space constraints and balancing reconstruction accuracy with interpretability, the proposed approach goes beyond simply minimizing the reconstruction error and focuses on learning expressive latent features that enable effective semantic interpretation. Experiments on three industrial benchmark datasets demonstrate the excellent performance of the proposed method, achieving an average accuracy of 96.73% and a maximum improvement of 13.34% compared to other methods.