DASA-Fuse: A Novel Lightweight Multisensor Fusion Framework for Fault Diagnosis of Hydraulic Components

Hydraulic systems are complex nonlinear systems with concealed fault characteristics, making accurate fault diagnosis challenging. To address these issues, this article develops a lightweight multisensor fusion framework, DASA-Fuse, for hydraulic component fault diagnosis. The framework consists of two main components. First, a novel time-series representation method, density-aware partitioned recurrence plot (DAPRP), is proposed. By introducing a density-aware symbolic modeling mechanism, DAPRP reformulates conventional recurrence analysis, enabling robust characterization of nonlinear dynamic behaviors under noise interference and nonuniform data distributions. Based on DAPRP, a multichannel DAPRP (MDAPRP) is further constructed to achieve unified fusion of heterogeneous multisource sensor signals. Second, a task-specific attention mechanism, grouped spatial cross-attention (GSCA), is proposed to model directional structural dependencies in recurrence-based representations. GSCA is further embedded into the MobileViT architecture to construct a lightweight network, namely structure-aware MobileViT (SAMViT), which enhances the extraction of fused structural features without significantly increasing computational overhead. Experimental results on both a self-made hydraulic test rig dataset and a public dataset demonstrate that DASA-Fuse achieves average diagnostic accuracies of 96.69% and 97.46%, respectively. Furthermore, compared with existing state-of-the-art methods, the proposed approach exhibits favorable diagnostic performance and cross-dataset applicability.