Spacecraft fault prediction is difficult because flight telemetry is high-dimensional, non-stationary, and only sparsely labeled. This paper examines how artificial intelligence can be integrated into satellite systems engineering to support early warning and fault diagnosis without replacing deterministic fault detection, isolation, and recovery (FDIR) logic. The study compares Artificial Neural Networks (ANN), Support Vector Machines (SVM), Recurrent Neural Networks/Long Short-Term Memory (RNN/LSTM), and hybrid ANN-EKF models for representative AOCS, EPS, thermal, communication, and onboard-data-handling fault cases.

Because operational fault labels are limited, the revised methodology distinguishes between public spacecraft anomaly benchmarks such as SMAP/MSL, NASA Prognostics Center of Excellence degradation datasets, mission/event logs, and mission-representative fault-injection simulations. The dataset section now specifies source categories, labeling methods, telemetry variables, and sim-to-real differences. Performance is assessed using accuracy, precision, recall, F1-score, AUC/ROC, RMSE/MAE, execution time, false-alarm behavior, and interpretability through SHAP and LIME.

The results indicate that temporal models such as LSTM are most suitable for gradual degradation and drift-type anomalies, while SVM/PCA and rule-assisted classifiers remain useful when labeled data are small. Hybrid physics-informed AI is appropriate when subsystem dynamics are available. The paper further defines onboard feasibility conditions: training and validation are assumed to be ground-based, while onboard use is limited to lightweight inference subject to latency, memory, processor, power, and flight-software constraints. The contribution is therefore an AI-assisted monitoring architecture for satellite fault prediction, not a claim of fully autonomous flight qualification.