Multi-label domain adversarial reinforcement learning for unsupervised compound fault recognition

Abstract

Compound fault composed of coinstantaneous multiple faults frequently causes the failure of a manufacturing system, which greatly reduces the reliability. When measuring the compound fault, two difficulties generally exist: (1) the complex correlation between different single faults, and (2) collected target samples without labels. To accomplish the cross-domain unsupervised compound fault recognition, this study proposes a multi-label domain adversarial reinforcement learning (ML-DARL) framework that implements two multi-label deep reinforcement learning (ML-DRL) models with adversarial domain adaptation. First, a source ML-DRL model is adopted to train a source feature network (SFN) and a policy network by using a dataset with labels (source domain). Then, a discriminator and a target ML-DRL model that includes a target feature network (TFN) are jointly trained with adversarial adaptation by simultaneously using the dataset without labels (target domain) and the source domain. Specifically, two outputs of TFN and SFN are regarded as fake and real components, respectively. Notably, the reward function in the target ML-DRL model is related inversely to the output of the discriminator for the fake component. Finally, a cross-speed case and a cross-location case are executed to verify the adaptation ability of the proposed method on cross-domain unsupervised compound fault recognition.