Semantic Consistency Assisted Neuron Adding Network for Generalized Zero-Shot Diagnosis With Incremental Learning Capability

Generalized zero-shot diagnosis has received extensive attention recently in the field of network science and engineering. However, most of them may suffer from two drawbacks: (1) Destructing the collaboration between the semantically consistent features and distribution characteristics of seen/unseen faults, and (2) most of them are trained on predefined classes, neglecting the abundant information contained in the newly collected samples. Accordingly, this paper introduces the semantic consistency assisted neuron adding network with incremental learning (SCNAN-IL). First of all, a semantic consistency feature extractor is designed to synthesize the features that consistent with attributes, rendering the generalization of diagnosis knowledge from seen faults to unseen ones. Subsequently, seen and unknown faults are identified automatically through neuron adding network with joint updates of structure and parameters. After that, sample incremental learning strategy is carried out to distill the knowledge and quantify the sample effects from the real-time data adaptively, which facilitates the model performance by involving preferred samples and additional faulty information without complete model retraining. Finally, the effectiveness and superiority of SCNAN-IL are substantiated through Tennessee-Eastman process and ASHRAE RP-1043 centrifugal chiller by multi-granularity hierarchical attributes strategy.