Adaptive Convergent Visibility Graph Network: An interpretable method for intelligent rolling bearing diagnosis

Abstract

In the domain of mechanical equipment maintenance, the necessity for efficient and accurate fault diagnosis is critical. Traditional Graph Neural Network (GNN) methods, which employ time-series data for fault diagnosis, have proven effective but are far from perfect. Their common pitfall lies in mapping time-series data into graph data, often leading to loss of crucial temporal information and computational inefficiencies. These limitations could result in suboptimal diagnosis, potentially compromising the longevity and performance of mechanical systems. Driven by the need to improve on these limitations, we introduce ACVGN, a novel end-to-end intelligent diagnostic framework that unites the strengths of the Adaptive Convergent Visibility Graph (ACVG) algorithm and an enhanced DiffPool model. The cornerstone of our approach, the ACVG algorithm, adeptly transforms time-series data into graph format, thereby preserving both local and global dynamics from the original data. This rich representation is then processed by our improved DiffPool model, a powerful GNN model purposefully designed for high-accuracy classification tasks. The effectiveness of the proposed ACVGN framework is substantiated by its performance on the widely-used rolling bearing dataset, where it outshines existing methods in terms of both mapping efficiency and fault diagnosis accuracy. These promising results not only reinforce the effectiveness of our proposed method but also highlight the potential for its wider applicability in other scenarios involving time-series data analysis and graph-based machine learning tasks. In conclusion, our study advances the development of more intelligent, efficient, and precise diagnostic tools for mechanical devices, ensuring more effective fault detection and diagnosis, and thereby potentially improving the lifespan and functionality of mechanical systems.