Model-Knowledge Mutual Reinforcement for Few-Shot Cellular Network Fault Diagnosis

Abstract

Data-driven cellular network fault diagnosis faces the great challenge of lacking sufficient high-quality labeled data from real networks. Also, how to exploit existing domain knowledge to aid the diagnosis in an efficient manner needs to be addressed. To obtain a high-performance diagnosis system based on a coarse dataset and knowledge base, we propose a Robust Belief Weighting Framework (RBWF) based on abductive learning and belief rule structure for the few-shot fault diagnosis. Our method is mainly innovative in three parts: our framework can obtain weights from insufficient knowledge base to effectively revise pseudo-labels to enhance classifier model performance. Meanwhile, our framework can make the coarse knowledge base more comprehensive by taking advantage of the revised pseudo-labels obtained from enhanced classifiers. To further improve performance, we also design Pattern Extraction (PET) to learn local information from unlabeled data and apply the Synthetic Minority Oversampling Technique (SMOTE) to generate new samples of the minority class. Experimental results illustrate the effectiveness and stability of the proposed framework using the cellular networks data. Furthermore, the PET and SMOTE enhance the diagnostic performance of the minority class.