Compound fault diagnosis of hydraulic system based on sample screening and joint analysis with pressure sensors pairs

Hydraulic systems play a crucial role in industrial control technology. This article proposes a compound fault diagnosis approach for hydraulic systems using a two-stage attention-multiscale convolutional neural network (CNN) framework paired with prediction interval estimation (PIE). Specifically, a model-agnostic PIE method constructs pseudo-label sets from non-stationary samples across various equipment failure modes. For each failure mode, the intersection of prediction interval sets from multiple components is taken to obtain conservative and credible samples. Sensors are then selected based on hydraulic circuit topology and fault impact analysis. Next, a dedicated network model is established for each component. Within the encoder module, dimensional segmental embedding (DSW) is applied to sensor data, followed by feature mapping using cross-time and cross-dimensional attention mechanisms. The extracted features are fused via a multiscale CNN. Experimental results show diagnosis accuracy for four components reaching 99.88 %, 97.44 %, 100 %, and 98.55 %, respectively. Furthermore, significant improvement is observed on the dataset optimized by PIE-based screening, demonstrating the method's effectiveness in identifying compound faults even when using only two pressure sensors with low fault-mode correlation scores. This approach enhances the reliability and robustness of hydraulic system condition monitoring.