A Graph Neural Network Model for Elevator Anomaly Detection Based on CAN Bus Data

As elevator usage time gradually increases, elevator faults become more frequent. Constrained by the lack of communication protocol disclosure from elevator manufacturers, current elevator anomaly detection typically necessitates additional sensors to acquire the dataset. However, this approach is marred by increased hardware costs, impacts on the elevator's original system, and unclear anomaly correlation. In this paper, we propose an elevator dataset scheme based on CAN bus data, eliminating the need for additional sensors while obtaining various critical elevator signals. For anomaly detection, we have i suggested a Graph Neural Networks which labels normal data as running up or down single-layer, multi-layer and stationary while all fault data are represented with one label. The approach detects anomalies by analyzing the Gini index under the operating state of elevators. It does not rely on expert knowledge nor limited to specific finite anomalies. Experiments conducted on an elevator in Shenzhen, China, demonstrated the model achieving 98.4% accuracy when a certain amount of anomalies was included in the training set. The model exhibited the ability to recognize unknown types of anomalies not present in the training set. Maintaining a high accuracy rate for both serious and minor anomalies, the model may detect subtle anomalies in the elevator before they manifest as serious faults, contributing to the predictive maintenance of the elevator.