Automation of fault diagnosis for low-voltage equipment based on digital twins

Abstract

Addressing the engineering challenges of high physical node density, strong aggregation of electrical fault characteristics, and low efficiency of traditional non-networked manual inspection in modern Industrial Internet of Things (IIoT) and smart buildings, this paper proposes and implements a highly efficient automated fault diagnosis computing architecture based on high-fidelity digital twins and cross-modal fusion features. A multi-dimensional virtual state machine space for complex physical entities is utilized, and an improved dynamic graph connectivity network is employed to extract the high-order spatial connection physical dependencies between devices. Furthermore, a topology fusion-based lightweight YOLO feature is innovatively implemented to synchronously capture abnormal distortions in the device's appearance, completely severing the hidden propagation path of cascading faults between information silos. By introducing a hardwareaware adaptive hybrid precision prediction strategy, the massive parameter model is successfully and losslessly physicalscaled down to Jetson edge computing micronodes, achieving a closed-loop control between cloud-based macro-parameter optimization and edge-based low-latency inference. Large-scale stress experiments on a multi-dimensional test dataset of up to 150,000 units demonstrate that this architecture, while effectively filtering out multi-source importance interruptions, not only significantly increases the overall system diagnostic rate to 97.8% under the condition that a single inference operation only takes 18.2 milliseconds, but also achieves an F1 score as high as 0.968. This achievement reconstructs the interlocking mechanism of low-voltage monitoring from the underlying logic, providing a technical reference for the management of asset digital health loop systems.