Global wavelet-integrated residual frequency attention regularized network for hypersonic flight vehicle fault diagnosis with imbalanced data

In the context of long-term exposure to harsh operating conditions, the hypersonic flight vehicle (HFV) is susceptible to failure. Regrettably, the current fault diagnosis methods for HFV primarily rely on shallow learning approaches, which struggle to identify the faults encountered in HFV effectively. Moreover, the imbalanced data further hamper the accuracy of fault recognition. Hence, this study proposes a global wavelet-integrated residual frequency attention regularized network (GWRFARN) for fault diagnosis in HFV. Firstly, a global waveletintegrated residual network is devised to enhance the feature extraction capabilities of the network. Secondly, a multiscale frequency combined attention mechanism embedded with discrete cosine transform and dilated convolution. This mechanism is designed to obtain more comprehensive and discriminative deep features. Besides, a frequency loss function is designed to select the optimal frequency components of the features. Thirdly, a dynamic exponential loss function that gives more attention to hard-to-classify samples is designed to handle the effects of data imbalance in real scenarios. Finally, an improved label smoothing regularization is proposed to enhance the generalizability of the network with imbalanced data. This approach assigns more reasonable initial soft labels to each category according to the imbalanced distribution of the dataset. And it dynamically adjusts these soft labels based on the evolving training result. To assess the efficacy of the GWRFARN, comprehensive comparisons are conducted using HFV datasets under multiple imbalance ratios. The performance illustrates the superiority of the proposed method in recognizing HFV faults.