Robust System Fault Classification & Detection System (RobustDetect)

NASA Glenn Research Center (GRC) is actively seeking improvements to fault detection and diagnosis, prognosis, fault recovery, and mitigation strategies to achieve greater system autonomy and resilience. For this, American GNC Corporation (AGNC) is proposing the Robust System Fault Classification & Detection System (RobustDetect) based on the observation that a classifier can only be as good as the quality of input data provided to it. The traditional neural network classification approach for fault diagnostics entails training the model on a variety of healthy and faulty operational states of a system to be monitored and deploying the model for real-time inference of system states given input sensor data streams. This assumes that the sensor data itself is an accurate measurement (of similar quality to the sensor data in the training set). However, transducers themselves can experience a myriad of faults (noise, drift, bias, etc.) that can be either intermittent or sustained. Using inherently incorrect data to make a conclusion about the state of a system being monitored is flawed as it will yield false positives/negatives regardless of classifier quality. The RobustDetect system addresses this problem by providing a sensor data validation algorithm that can be applied on sensor data to assign greater weight to sensors that are more likely to be healthy before being fed to a downstream classifier. The scheme involves a learning stage where a set of neural networks are trained to learn different relationships and correlations among the available sensors. Changes to these learned correlations can then be leveraged to determine if a given sensor is experiencing a fault itself and to either generate a virtual reading for that sensor (as a temporary self-healing strategy) or to assign less weight to that sensor within the classification pipeline. The result is a new approach to fault classification that is robust to sensors with degraded or faulty measurements.