Fault diagnosis in hydraulic systems via multi-channel multi-modal fusion

Hydraulic systems are inherently complex and nonlinear, often prone to subtle, concurrent faults. These characteristics pose challenges for fault diagnosis, especially when time-domain data are limited. This paper studies multi-fault diagnosis problem in hydraulic systems, and proposes a novel multi-fault diagnostic framework, namely, multi-channel multi-modal attention fault diagnosis network (MC-MM-AFDN), to improve the diganosis accuracy with limited data. MC-MM-AFDN employs a parallel multi-channel architecture to extract and fuse sensor data features across different sampling frequencies. Specifically, each channel in MC-MM-AFDN adopts a dual-branch structure, with one processes temporal data using temporal convolutional network blocks, while the other converts such data into 2D images via gram angle sum fields for spatial features extraction using 2D convolutional neural network blocks. Features from both branches are then fused via a multi-head cross-attention mechanism for complementary spatiotemporal information integration. To further improve multi-channel fusion efficiency, fusion weights for each channel are also optimized using an improved snow ablation optimizer. Experiments on public datasets are conducted to validate the proposed method. Results show that the MC-MM-AFDN achieves fault diagnosis accuracy exceeding 99.55% on hydraulic system datasets, maintaining robust performance even with limited sample sizes and under noisy conditions.