Research on elevator door anomaly detection method based on multisource spatial-temporal information fusion

Purpose Elevator door system failures are a leading cause of elevator malfunctions, impacting safety and operational efficiency. Existing anomaly detection methods often overlook the relative positional relationships among time-series data sequences. This study aims to propose a novel spatial-temporal information fusion approach to accurately identify abnormal states in elevator door systems. Design/methodology/approach This paper develops an integrated spatial-temporal fusion framework for anomaly detection. First, operational time-series data are structured into a node graph using a K-Nearest Neighbor (KNN) based method to model complex inter-sequence interactions. Subsequently, a Graph Convolutional Network (GCN) is employed to extract local dependencies and global spatial information, while a Bidirectional Long Short-Term Memory network (BiLSTM) captures temporal evolutionary characteristics. A multi-source information-driven feature fusion mechanism is then designed to enhance model robustness. The proposed method is experimentally validated using real-world elevator door operating data, including vibration and video-derived signals. Findings The experimental results demonstrate that the proposed method effectively identifies abnormal elevator door states, achieving a high detection accuracy of 96.27%. This confirms the framework's effectiveness and reliability in practical scenarios. Originality/value This research presents a novel KNN-based graph construction method that captures dynamic dependencies between time-series sequences based on their relative positions. Furthermore, it develops an integrated framework that concurrently models spatial structural relationships and temporal dynamics, overcoming the limitations of methods that treat these dimensions separately. Finally, it introduces a multi-source feature fusion mechanism that leverages the complementarity of information from different sources and dimensions, significantly enhancing the model's representation capability and robustness under complex operating conditions.