MRD-Nets: Multi-Scale Residual Networks With Dilated Convolutions for Classification and Clustering Analysis of Spacecraft Electrical Signal

The fault detection of spacecraft electronic load systems is a crucial part of the spacecraft prognostics and health management system. To detect the abnormal state of spacecraft electronic load systems, complex electrical signals should be processed rapidly and accurately. For the fault detection of spacecraft electronic load systems, a robust unsupervised clustering analysis method and an accurate supervised classification method are of great importance. However, the traditional machine learning methods have poor performance when processing high-dimensional signal data because of the lack of ability to extract complex features from the signals. Therefore, neural-network-based deep learning (DL) models which can extract features from signals automatically are more suitable in this situation. In this paper, a novel convolutional neural network (CNN) module, the multi-branch residual module with dilated convolutions (MRD module), is proposed to extract multi-scale features from the electrical signal. Then, a well-designed CNN model named MRD-CNN is presented for the supervised classification task of signal. Furthermore, for the unsupervised clustering task, the clustering variational autoencoder with MRD modules (MRD-CluVAE) is proposed. The MRD-CluVAE can extract high-quality features from signal data and output the clustering results directly. To evaluate the performance of the proposed models, comparisons among the proposed models and other baseline algorithms are carried out. The experimental results show that the MRD-CNN model achieves higher classification performance and stability, while the MRD-CluVAE has a higher clustering accuracy than other algorithms. These methods can be utilized to resolve the classification and recognition problems of spacecraft electronic load signals.