Evolving Through Steps: Double-Memory-Driven Anti-Forgetting Framework for Incremental Zero-Shot Fault Diagnosis

Abstract

Zero-shot fault diagnosis (ZSFD) can identify unseen faults by predicting fault attributes labeled by human experts. We recognize the need for ZSFD to handle continuous changes in practical industrial processes, that is, the model’s ability to update for newly collected fault categories and attributes while avoiding forgetting the diagnosis ability learned before. Therefore, the incremental ZSFD (IZSFD) paradigm is proposed, which incorporates category increment and attribute increment tasks for both conventional and generalized ZSFD (GZSFD) paradigms. For the category increment, the number of categories continuously increases due to new categories being collected or recognized. For the attribute increment, the number of attributes continuously increases as experts deepen their understanding of each category. To achieve IZSFD, we present a double-memory driven anti-forgetting framework (DM-AF) that aims to learn new fault categories and attributes. DM-AF accumulates knowledge from two perspectives: features and attribute prototypes. The feature memory is established through a generative model that employs designed anti-forgetting training strategies, addressing the accumulation of generation errors over multiple learning stages. The attribute prototype memory is established through the diagnosis model, and the proposed memory-driven prototype update strategy allows the update of the attribute prototype memory without requiring the storage of samples. The effectiveness of the proposed method is verified by a real hydraulic system and the Tennessee-Eastman benchmark.