A Scalable Soft Perception Framework for Cross-Individual Fault Diagnosis Toward Multisource Heterogeneous Signals With Dynamic Dimensions

Classic deep learning (DL)-based fault diagnosis methods focus on homogeneous tasks with the same data distribution and sample shape between the training set and the test set, which is not applicable to multisource heterogeneous signals from multiple machine-level devices. Individual differences in devices, as well as inconsistencies between input features and dimensions, make classical diagnostic frameworks based on a single static model no longer applicable. This article proposes a scalable soft perception framework (SSPF) to address cross-individual fault diagnosis considering multisource heterogeneous signals (CIFD-MSHSs). A flexible collaborative neural network (FCNNet) modeling paradigm using multiple DL models is proposed to endow a universal perception of dynamic dimensional heterogeneous data. In addition, the proposed method considers the posterior optimization of local networks, which has good scalability and growth potential to quickly adapt to new scenarios. The proposed method and model are performed on multiple machine-level motors on real production lines and obtained unseen device accuracy (UDA) of over 90% for nine heterogeneous generalization tasks.