A health-state assessment method for electromechanical actuators in reusable rockets based on channel-aware temporal memory

Abstract

Electromechanical actuators (EMAs) are critical components of thrust vector control in reusable launch vehicles, but health-state assessment is hindered by limited, heterogeneous, and noise-sensitive telemetry. This paper presents MCTM-Net, a channel-aware temporal memory network that preserves channel identity before feature fusion and combines local temporal extraction with selective temporal memory. The method is evaluated on the National Aeronautics and Space Administration Flyable Electromechanical Actuator (NASA FLEA) coupled-actuator testbed using telemetry from a fault-injected X actuator, a normal reference Y actuator, and a load Z actuator. Across three training seeds and three noise realizations per seed at −5 dB additive white Gaussian noise (AWGN), MCTM-Net achieves an accuracy of 98.37% ± 2.30% and a macro-averaged F1 score (macro-F1) of 95.60% ± 6.36%. Fixed-seed comparisons, four non-Gaussian disturbances, ablations, and channel analyses provide complementary evidence within the coupled-testbed protocol.