A survey on multimodal learning for industrial diagnostics: a data dimensionality perspective

Abstract Industrial diagnostics, as an essential methodology for quality control, production optimization, and operational safety assurance, has garnered global research efforts to enhance diagnostic efficacy. With advancements in hardware and deep learning technologies, multimodal learning has emerged as a transformative approach in this domain. Within multimodal frameworks, data inherently assumes a pivotal role, where inter-modal characteristics fundamentally determine diagnostic strategies. This study adopts an innovative data dimensionality perspective to investigate multimodal learning applications in industrial diagnostic tasks systematically. Through comprehensive surveys of cutting-edge research across task requirements, data characteristics, and fusion methodologies, we delineate the utilization patterns of heterogeneous multimodal data in diverse diagnostic scenarios. Furthermore, we systematically categorize feature extraction and fusion strategies based on the types of data modality and their intrinsic properties. The study ultimately identifies three core challenges and proposes actionable research directions to address these challenges. By elucidating technological advancements through data characteristics, this work provides a holistic understanding of state-of-the-art developments and practical guidelines for applied researchers seeking to implement multimodal solutions in real-world industrial settings.