Elevator Fault Diagnosis Based on a Graph Attention Recurrent Network

Abstract

Elevator fault diagnosis is critical for ensuring operational safety and reliability in modern vertical transportation systems. Traditional approaches, which rely on time- and frequency-domain signal analysis, often struggle with the issues such as noise sensitivity, inadequate feature extraction, and limited adaptability to complex scenarios. To address these challenges, this paper proposes a Graph Attention Recurrent Network (GARN) which integrates graph-structured signal representation with spatiotemporal feature learning. The GARN employs a limited penetrable visibility graph to transform raw vibration signals into noise-robust graph topologies, preserving critical patterns while suppressing high-frequency noise through controlled edge penetration. An adaptive attention mechanism dynamically fuses triaxial features to prioritize the most relevant information for fault diagnosis. The GARN combines a graph convolutional network to extract spatial correlations and a gated recurrent unit to capture temporal fault progression, enabling holistic and accurate fault classification. Experimental results based on real-world elevator datasets demonstrate the superior performance of the GARN, showcasing its strong noise resistance, adaptability to complex fault conditions, and ability to provide reliable and timely fault diagnosis, making it a robust solution for modern elevator systems.