An end-to-end framework for fault detection and diagnosis electric rudder servo system-based under imbalanced data condition

Abstract

The status of the electric rudder servo system status directly affects the whole controlled system. We develop an innovative intelligent test platform for the electric rudder servo system and propose a new end-to-end method called K-clusters Synthetic Minority Over-sampling Technique-Multiscale Convolution Neural Attention Network for fault detection and diagnosis. Among them, the new method called K-clusters Synthetic Minority Over-sampling Technique is applied to address imbalances between and within classes by supplementing samples for the electric rudder servo system. Furthermore, to prevent the loss of sensor information due to the initial position, we propose a 2D-Multiscale Convolution Neural Attention Network to extract the 1D-Multisensor information. Compared to other models, the K-clusters Synthetic Minority Over-sampling Technique-Multiscale Convolution Neural Attention Network approach achieves state-of-the-art results, with an average G_mean of 0.99,646 and an average F2 Score of 0.99,375 in 10-fold validation. Finally, the total program execution time for all experimental models is calculated, which demonstrates the high efficiency of the 2D K-clusters Synthetic Minority Over-sampling Technique-Multiscale Convolution Neural Attention Network approach.