Fault diagnosis method and system based on cost-sensitive active learning and double-flow network

Abstract

The application provides a fault diagnosis method and system based on cost-sensitive active learning and a double-flow network, which couples asymmetric misclassification cost and marginal uncertainty to construct query scores, so that the active learning sample selection process can preferentially identify and screen high-risk boundary samples that will lead to serious consequences once misjudged, thereby significantly improving the utilization efficiency of labeling resources and effectively reducing the fault miss rate under the same labeling budget. Meanwhile, by constructing a special double-flow network containing a global state space modeling branch and a local convolution mixed branch, the global dependence features and local impact features in the time-frequency graph are extracted and fused respectively, which greatly enhances the recognition ability of early weak faults and overcomes the defects of insufficient feature extraction of traditional single network. In addition, by introducing the confidence and uncertainty quantitative output and temperature scaling probability calibration mechanism, the reliability and engineering decision support value of the diagnosis result are significantly improved.

Images