Achieving spatiotemporal coordinated control through the integration of material parameters based on CNN-LSTM-attention mechanisms for precise fault diagnosis in hydraulic systems during composite material forming processes

Abstract In response to the issue of low diagnostic accuracy in composite material hydraulic presses operating under harsh conditions, such as high temperatures and noise levels, This paper incorporates dynamic rheological parameters from composite material forming processes into a deep learning framework, thereby proposing an innovative dual-branch ‘spatiotemporal collaborative’ deep learning diagnostic model. This model innovatively integrates sensor signals with material dynamic process parameters. A convolutional neural network (CNN) extracts spatial correlations from ten-channel hydraulic signals, while an Long Short-Term Memory (LSTM) branch captures the temporal evolution of the signals and the real-time physical quantities of the composite material. These heterogeneous features are then fused via a cross-attention mechanism, which prioritises critical fault characteristics and achieves deep decoupling between process fluctuations and actual system failures. Experimental results demonstrate 99.1% accuracy in identifying three core failure modes: malfunctions of the hydraulic pump, cooler and valve. Further comparative experiments and ablation studies validate that incorporating dynamic process parameters reduces the model’s false positive rate (FPR) to 0.5%, proving its exceptional robustness and reliability in complex industrial forming environments. This advancement provides a high-precision solution for intelligent maintenance in automotive lightweight production lines.