A multi-sensor fusion-based transfer learning for the linear actuator fault diagnosis under varying operating conditions

Abstract Electro-mechanical actuator has gained increasing interest in the high-end applications and it can significantly impact the normal operation of the equipment when it breaks down. Therefore, the fault diagnosis of actuators has become a research hotspot. However, the current diagnosis method has been limited by the following weakness: (1) the use of a single sensor can be insufficient and perform poorly due to strong noise, (2) actuators normally experience various operating conditions, (3) collecting a large amount of labeled data under each condition can be difficult and time-consuming. To solve these problems, this paper proposes a fault diagnosis method for the linear actuator transmission components based on multi-sensor information fusion (MIF) and transfer learning techniques. Firstly, a fault diagnosis model based on MIF and parallel convolutional neural networks (PCNNs) is proposed, which takes the time-frequency maps generated by continuous wavelet transform as input. Group normalization and multi-sensor fusion techniques have been employed to accelerate network convergence and improve diagnostic accuracy, achieving high-precision fault diagnosis under single operating conditions. Then, by fine-tuning the MIF-PCNN model, a fault diagnosis model based on MIF and PCNN model transfer (MIF-PCNN-MT) has been established, addressing the issue of model performance degradation caused by small sample sizes and varying operating conditions. Finally, the proposed method has been validated using experimental datasets from actuator transmission systems. The experimental results indicate that the MIF-PCNN model and MIF-PCNN-MT model achieve average diagnostic accuracies of 93.58% and 93.45%, respectively, both demonstrating good diagnostic performance.