Electro-hydraulic SBW fault diagnosis method based on novel 1DCNN-LSTM with attention mechanisms and transfer learning

Abstract

The Electro-hydraulic Steer-by-Wire (EH-SBW) is the future trend of commercial vehicle steering systems due to their dual actuator redundancy. The accurate fault diagnosis of the EH-SBW is an important basis for ensuring the safety of commercial vehicles. However, due to the strong coupling between dual actuators, rapid load changes and difficult sampling lead to the difficulty of existing techniques to meet the fault diagnosis of EH-SBW. Therefore, we propose the EH-SBW fault diagnosis method based on novel 1DCNN-LSTM with attention mechanisms and transfer learning, which contains a fault type diagnosis neural network (FTDNN) model and a fault degree estimation neural network (FDENN) model. The steering feature extractor (SFE) in the FDENN model fuses the driver’s long time-domain driving trend signals and the short time-domain signals from the steering system to extract the feature map, so that the feature map have more valid information in the case of fast load changes. Based on the result of the FTDNN and the feature map, the scaling attention layer in the FDENN amplifies the features that are favourable to the fault degree estimation accuracy and suppresses the expression of unfavourable features, which reduces the impact of strong coupling relationships on the fault degree estimation accuracy. The dynamic sample allocation strategy performs transfer learning of the SFE and scaling attention layer through a small number of samples, which enables the FDENN model to effectively estimate the degree of failure of the hydraulic mechanism. The experimental results show that the diagnostic accuracy of the proposed method is 94% when the electric mechanism fails. The diagnostic accuracy of the proposed method is 92% when the hydraulic mechanism fails.