Helicopter transmission system anomaly detection in variable flight regimes with decoupling variational autoencoder

Condition monitoring of helicopter transmission system is the main focus of Health and Usage Monitoring System. Existing anomaly detection methods for transmission system ignore the challenges caused by variable flight regimes of helicopter, which greatly hinders their application in application scenarios. For the first time, we propose Decoupling Variational AutoEncoder (DVAE) as a pioneering solution for anomaly detection task under variable working conditions. It does not rely on any prior knowledge related to the task and greatly expands the application scenarios of anomaly monitoring tasks. With the purpose of decoupling original signals from regime information, we focus on three aspects: network mechanism, training strategy, and optimization theory, to extract domain-invariant latent features with regard to flight regimes. Specifically, domain shifting mechanism is built to train learnable transform index and implement cross-domain transformation adaptively. Domain adversarial regression strategy is proposed to alternately guide the decoupling process about regime information. Theoretical derivation guarantees that when algorithm converges, latent features would be independent of domain parameters. The effectiveness of DVAE is evaluated with three progressive experiments, including parts-level fault simulation, components-level fault simulation, and actual-level fault under practical scenarios of helicopter transmission system. Results show that it can timely and accurately detect actual failure of the helicopter transmission system under variable flight regimes and alert it before onboard Health and Usage Monitoring System.