1D-CNN Based Multi-label Aircraft Electrical Signal Classification Method

The identification of malfunctions within the electronic load systems of spacecraft is a crucial component of the spacecraft’s predictive maintenance and health monitoring framework. For the purpose of real-time anomaly detection in these systems, it is imperative to swiftly and precisely discern intricate electrical signals. In the field of fault diagnosis, various data-driven methods, which are mainly based on machine learning, have been utilized for the signal classification task. However, the traditional machine methods consisted of feature extraction algorithms and classification algorithms may extract unrepresentative features and take a long time to run. In this paper, a one-dimensional convolutional neural networks (1D-CNN) based multi-label spacecraft electrical signal classification method is proposed. First, during the preprocessing stage, the one-dimensional signals are processed by wavelet denoising and segmented into equal length. Then, a well-designed 1D-CNN model is trained to classify the 19-class electrical signal. An average classification of 99.69% is achieved, while the test time is much shorter than traditional methods. Furthermore, to insight into the internal operation and behavior of the proposed 1D-CNN model, a visualization method for 1D-CNNs is suggested to analysis the behavior of each filter in the network, and the visualization results show that the filters in different convolutional layers exhibit different behavior characteristics.