Unknown working condition fault diagnosis of rotate machine without training sample based on local fault semantic attribute

Data-driven fault diagnosis techniques for rotating machinery have exhibited highly promising results. However, these methods heavily rely on sufficient faulty data and presuppose that the source (model training) and target domains (model diagnosis) share a matching data distribution. In practical industrial settings, acquiring target domain data can be quite challenging, and the distribution between the source and target domains is expected to differ due to various working condition of mechanical equipment. In order to surmount these challenges and address state monitoring under unknown working conditions, this paper presents a novel fault diagnosis method designed for rotating machinery in the absence of target domain data. Firstly, this method involves constructing local fault state semantic attributes using source samples from limited known working conditions of the rotating equipment. Secondly, a dual embedding module is employed to map the relation between fault features and fault state semantic attributes in a high-dimensional embedding space. Thirdly, an improved loss function is designed to optimize the dual embedding module by balancing inter-class and out-of-class distances of fault samples. Finally, to prevent overfitting result from limited known working conditions, samples of an additional third working condition are introduced during the training of the dual embedding module. Experimental evaluations conducted on two bearing datasets and an automobile transmission fault dataset from First Auto Work demonstrate the effectiveness of the proposed method in accurately identifying faults under unknown working condition. The fault diagnosis recognition accuracy under unknown working conditions exceeds 99.2%, 96.1% and 88.2%, and the proposed approach can effectively address the issue of diagnosing with no target domain data in engineering problems.