<p>Unsupervised domain adaptation (UDA) has been widely applied to multisensor signal analysis in industrial scenarios. Time-series data generated in industrial environments are typically characterized by high dimensionality and large-scale, which impose significant requirements on computational efficiency and inference performance of learning models. However, most existing multivariate time-series domain adaptation methods primarily focus on aligning the marginal distributions of target variables, while largely overlooking the potential yet sparse dependency patterns among variables. To address this limitation, this paper proposes a novel multivariate time-series unsupervised domain adaptation framework based on structure-dependency consistent representation learning, termed CSIDA. Specifically, CSIDA first performs temporal block pairing to capture signal-specific fragments and their contextual dependencies. Subsequently, a dual convolutional attention mechanism is employed to learn cross-variable feature representations that preserve consistent underlying physical information. Finally, a maximum mean discrepancy (MMD) loss is introduced to further reduce the distribution gap between the source and target domains. Extensive experiments conducted on three publicly available real-world datasets and one internal industrial dataset demonstrate that CSIDA enables large-scale model training and deployment for practical industrial monitoring scenarios on parallel computing and high-performance computing platforms.</p>