Early fault detection of space fluid circulating pumps is critical for preventing catastrophic on-orbit failures, yet the gradual and subtle nature of pump degradation makes early anomalies difficult to distinguish from normal variability. Existing reconstruction-based methods respond to instantaneous signal error rather than cumulative statistical drift, while high-dimensional distance-based methods suffer from covariance ill-conditioning. This paper proposes a multi-scale LSTM-PCA Mahalanobis framework that represents pump state using local mean and standard deviation features at three temporal scales and predicts their evolution with an LSTM. The prediction error is projected into a PCA-reduced space where a well-conditioned Mahalanobis distance serves as the anomaly score, and MC Dropout averaging combined with EWMA smoothing yields robust detection. Experiments on a real-world space pump dataset show that the proposed method achieves an AUC of 0.9496, outperforming four baseline methods, and detects anomalies 4.5 months before the labeled onset and 24 months before pump replacement.

Early fault detection,Space fluid circulating pump,Multi-scale feature extraction,LSTM,Mahalanobis distance