A satellite fault diagnosis method based on time-series data feature cases

Abstract

Satellite telemetry data are characterized by sensor heterogeneity, high dimensionality, strong temporal correlations, and limited fault samples, posing significant challenges to traditional fault diagnosis methods that rely heavily on expert knowledge, manual labeling, and handcrafted rules. To address these issues, this paper proposes a temporal feature case-matching fault diagnosis algorithm named DAT-CLSTM (Dual Attention Temporal-CNN-LSTM). Unlike conventional Siamese-network-based approaches, the proposed method employs a CNN-LSTM hybrid network as the core feature extractor and incorporates a dual-attention mechanism that jointly captures temporal-step and sensor-feature dependencies. An improved similarity measurement model is constructed to achieve efficient fault diagnosis through temporal feature matching between historical fault cases and target samples. Comparative experiments with multiple baseline algorithms demonstrate that DAT-CLSTM achieves superior diagnostic performance and exhibits strong generalization capability in small-sample and multi-task scenarios. The proposed method provides an effective technical reference for intelligent on-orbit fault diagnosis of satellites.