An ART2 (Adaptive Resonance Theory) neural network was trained to detect failures in a liquid rocket engine (LRE) from noisy normal steady-state sensor data.Power spectra of successive windows of individual sensor data were presented to an ART2 neural network to learn.The trained network was then tested to verify its effectiveness of failure detection and failure onset detection.Sensor data were collected from ground tests RS61 of the variable thrust LRE.The test results show that detected by the neural network corresponds with that the failure onset determined by experts from their post-test analyses.