FAULT DETECTION AND ISOLATION USING A NEURAL NETWORK

摘要

This specification describes systems and methods for detecting and isolating faults in telemetry data using a machine-learned model. According to one aspect of this specification, there is described computer implemented method of training one or more discriminator neural networks to identify anomalous telemetry data. The method comprises: generating, using one or more modification neural networks, a set of anomalous telemetry data; processing the set of anomalous telemetry data using the one or more discriminator neural networks to generate a candidate classification for each of one or more of the channels of telemetry data in the set of anomalous telemetry data; processing a further set of telemetry data corresponding to equipment functioning correctly using the one or more discriminator neural networks to generate a candidate classification for each of one or more of the channels of telemetry data in the further set of telemetry data, each classification indicative of whether the discriminator classifies the corresponding one or more channels of telemetry data input to the discriminator neural network as containing an anomaly; updating parameters of the one or more modification neural networks based on a comparison of the candidate classifications of the anomalous telemetry data to ground truth classifications of the anomalous telemetry data; and updating parameters of the one or more discriminator neural networks based on a comparison of the candidate classifications of the anomalous telemetry data to ground truth classifications of the anomalous telemetry data, and a comparison of the candidate classifications of the further set of telemetry data to ground truth classifications of the further set of telemetry data.