Spacecraft exhibit intricate structural designs and operate in highly dynamic environments, where anomaly detection plays a pivotal role in monitoring spacecraft status and ensuring mission success. Traditional anomaly detection methods face challenges of low accuracy and poor adaptability when applied to non-stationary and noise-contaminated telemetry data. This paper proposes an intelligent anomaly detection framework integrating representation learning with dynamic threshold optimization to enhance detection accuracy and reliability. The methodology comprises three key components. First, a stacked autoencoder performs nonlinear dimensionality reduction on high-dimensional time-series data to extract low-dimensional intrinsic features while suppressing noise interference. Second, a Neural Circuit Policies model with bio-inspired sparse connectivity and adaptive time constant mechanisms is employed for temporal pattern modeling and prediction. Finally, a multi-objective optimization algorithm dynamically adjusts anomaly thresholds to balance precision and recall, effectively addressing the performance degradation of fixed thresholds under abrupt data distribution shifts. In the drift anomaly detection experiments on the simulation dataset, the proposed method achieved F1-score improvements of 65.1%, 50.5% and 8.8% compared to LSTM, Transformer and TFT methods, respectively. For the real-world satellite dataset, our approach demonstrated superior performance with F1-score enhancements of 53.0%, 51.0% and 41.0% over the same three baseline methods. Experimental results on two spacecraft datasets demonstrate that the proposed method significantly improves prediction accuracy compared to baseline methods. The proposed framework establishes a novel technical approach for on-orbit autonomous health management of spacecraft, with substantial implications for enhancing the safety and reliability of deep space exploration missions.