Cloud‐Edge Collaborative Framework‐Based Fault Diagnosis Method for Power Metering Smart Laboratory Equipment Using Improved Transformer

With the continuous advancement of power system intelligence, the need for efficient fault diagnosis of equipment in power metering smart laboratories has become increasingly pressing. This paper presents a novel fault diagnosis approach for power metering smart laboratory equipment, which is based on a cloud‐edge collaborative framework and an improved transformer model. First, a cloud‐edge collaborative framework is established. In this framework, the edge layer is responsible for collecting data from power metering equipment, performing preliminary preprocessing and feature extraction. The network layer ensures the stable and secure transmission of data. The cloud layer receives the data for in‐depth analysis and fault diagnosis and collaborative work is realized through a dynamic task allocation mechanism. Second, the transformer model is enhanced by introducing a multimodal data fusion module to encode and fuse different types of data, adopting an improved self‐attention mechanism (SAM) to highlight key data and conducting a lightweight design to adapt to edge computing nodes. Based on this framework and model, fault diagnosis is carried out, including data collection and preprocessing, feature extraction at both the edge and cloud layers, fault diagnosis and classification using the improved transformer model and finally, the feedback of diagnosis results to optimize the model. This study introduces significant innovations to the transformer architecture and cloud‐edge framework specifically for power metering equipment fault diagnosis, including a multimodal data fusion module, an enhanced SAM and dynamic edge‐cloud collaboration. The proposed improvements demonstrate superior performance over conventional methods through specialized adaptations for electrical system diagnostics and optimized resource allocation between edge and cloud computing layers. Experimental results show that compared with traditional fault diagnosis methods such as decision trees, random forests, long short‐term memory and the original transformer, the proposed method has significant advantages in fault diagnosis accuracy (reaching 98.5%), real‐time performance (with an average diagnosis time of 0.2°s) and model stability. This provides a strong guarantee for the stable operation of power metering smart laboratories.