An autonomous recognition framework based on reinforced adversarial open set algorithm for compound fault of mechanical equipment

• Published in: Mechanical systems and signal processing Vol. 219; p. 111596
• Main Authors: Wang, Zisheng, Xuan, Jianping, Shi, Tielin
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.10.2024
• Subjects: Autonomous compound fault recognition; Deep adversarial neural network; Deep reinforcement learning; Open set recognition; Autonomous compound fault recognition; Deep reinforcement learning; Open set recognition; Deep adversarial neural network
• ISSN: 0888-3270; 1096-1216
• DOI: 10.1016/j.ymssp.2024.111596

How to automatically recognize compound fault of mechanical equipment based on data-driven algorithms, has always been a research focus in modern intelligent manufacturing. Nonetheless, owing to the intricate nature of intelligent models in comprehending unknown knowledge, existing methods for compound fault recognition necessitate a training dataset comprising enough known compound fault samples. In order to address this issue, this paper first proposes an autonomous recognition framework based on reinforced adversarial open set (RAOS) algorithm, which can accurately recognize compound faults by utilizing the label information from single fault samples. Firstly, single fault and compound fault are represented as the known class and unknown class, respectively. Subsequently, a feature sub-network and a policy sub-network with adversarial learning are devised to classify and align the features of samples of two classes. Additionally, a deep reinforcement learning (DRL) model is formulated following a question-and-answer paradigm. Ultimately, the DRL model undertakes a sequence decision-making task to optimize the feature sub-network and the policy sub-network. Through a laboratory experiment and an engineering application, the results robustly validate the effectiveness of the proposed ROSA framework in accurately recognizing compound fault samples, even when only single fault samples are known in the training dataset.