A Coarse-to-Fine Neural Network Framework for Multitask Fault Diagnosis Across Diverse Converter Types

A multi-task neural network (MTNN)-based model for joint fault diagnosis of multiple converter types (JFD-MCT) provides significant benefits by improving diagnostic performance, reducing model development costs, and promoting the advancement of portable and universal diagnostic tools. However, challenges remain in effectively extracting shared fault features, with susceptibility to negative transfer caused by interference factors. To address these issues, this paper introduces a coarse-to-fine multi-task neural network (CFMNN) for JFD-MCT. Through an in-depth analysis of fault correlations and diagnostic interference factors across various converter types, the mechanisms driving negative transfer are systematically uncovered. Guided by domain expertise, the learning tasks are hierarchically decomposed from coarse to fine, enabling CFMNN to iteratively extract shared fault features while mitigating interference. Additionally, a novel regularization technique and a task prompt network are integrated to further enhance the performance and robustness of CFMNN. The proposed method is rigorously validated through mathematical derivations and comparative experiments, demonstrating its superior fault diagnosis accuracy and generalization across diverse converter topologies and operating conditions, along with a substantial reduction in model construction costs.