Satellite Interpretable Anomaly Detection With Expert Experience-Based Algorithm Unfolding and Conditional Canonical Correlation Analysis

Artificial intelligence techniques have been extensively employed in anomaly detection tasks for massive systems and equipment across numerous industries, achieving notable success. Nevertheless, classical machine learning methods typically possess a simplistic design, which occasionally fails to satisfy the detection demands of minor anomalies in certain complex tasks. Conversely, the interpretability of deep learning methods is often insufficient to convince domain experts and operators. Consequently, achieving a balance between detection accuracy and interpretability remains a critical and often conflicting challenge. This article proposes an expert experience-based algorithm unfolding (EAU) network and a conditional canonical correlation analysis (CCCA) theory for anomaly detection of spacecraft under multiple operating conditions, aiming to ensure detection accuracy while enhancing interpretability. First, the EAU network integrates the experiential knowledge with the Lasso regression model and employs sparse coding to iteratively expand it layer by layer (LbL), facilitating deep feature extraction from the original telemetry data. Second, the CCCA method formulates the residual vector on the premise that the correlation between the regularization components of the input and output sets will change markedly before and after the anomaly appears. It subsequently compares the Hotelling T² statistic of each sample with the detection threshold, which was constructed based on the kernel density estimation (KDE) method, to discover the evolution of the anomaly. Finally, multigroup comparisons on two simulations and two real satellite-telemetry datasets verify the superior overall performance of the proposed method and provide guidance for selecting anomaly detection approaches.