Abstract: To address the urgent need for a multifunctional portable terminal used in health monitoring of spacecraft electromechanical products, a 2.5 kg integrated hardware–software device for fault diagnosis and degradation prediction was developed. The hardware features a 16-channel data acquisition box with a maximum sampling rate of 156 kHz, capable of interfacing with multiple types of sensors, including vibration, temperature, flow rate, rotational speed, and torque. The embedded software integrates four individual neural network models–BPNN, PSO-BPNN, GA-BPNN, and FNN–along with a multi-neural-network fusion (MNN) model. It also incorporates two degradation prediction algorithms (linear regression and Gaussian process regression) and a vibration analysis module, enabling a comprehensive workflow encompassing data acquisition, storage, diagnosis, and prediction. Six-fold cross-validation results showed that the MNN model significantly improved fault classification accuracy and model stability compared with those of single neural network models. For a spacecraft fluid loop pump with 36 samples under six-fold cross-validation, all test set diagnoses were correct, yielding a mean squared error (MSE) of only 0.010 3. Fatigue spalling faults on the pump bearing raceway, along with inner- and outer-ring spalling faults of the B7004C bearing, were accurately identified by the device. This research provides reliable modeling algorithms and integrated software–hardware support for health management of electromechanical products.