Anomaly detection of asset data in power monitoring systems based on knowledge graph reasoning

Abstract Edge computing nodes face severe constraints on both computing and storage resources, and the dynamic correlation of asset data remains complex within power monitoring systems. To address this challenge, this study proposes an anomaly detection framework that is based on knowledge graph reasoning and tailored for asset data in such systems. The framework first constructs a lightweight embedding method for knowledge graphs, then introduces a time-sliding window mechanism that builds a sequence of dynamic subgraphs, thereby adapting to the dynamic evolution of the network topology. Following this, the approach combines relational attribute modulation with low-rank matrix decomposition, thereby reducing both the model’s global parameter count and its memory overhead. It then proposes a relation-aware lightweight graph attention network. Within this network, a differential projection decoupling mechanism separates the coupling effects between target nodes and relational features. In addition, the approach introduces a sparse attention truncation selection strategy to filter out low-frequency background noise. Based on real datasets collected from substations and distribution networks, the results show that the proposed model achieved an area under the receiver operating characteristic curve (AUC-ROC) of 0.982 on dataset A and an accuracy of 0.949 on dataset B. The low-rank matrix decomposition strategy reduced memory usage by 94.68 %, while the differential projection decoupling mechanism maintained a stable F1-score of 97.25 %. Furthermore, six months of continuous monitoring with the dynamic time-sliding window kept the F1-score stable at around 97 % and yielded an average alarm delay of only 0.8 s. These findings indicate that the proposed detection model effectively enhances the robustness of the power monitoring system against concept drift and network topology evolution, thereby providing technical support for lightweight intelligent monitoring and the safe deployment of edge devices for power assets.