Method Based on Stacking-Attention to Find Decomposition Indicators of Discharge Mechanism in C4F7N-CO2-O2 Gas

The correlation between the decomposition products and discharge faults in the electrical equipment with the environmentally friendly insulating gas mixture C4F7N/CO2/O2 remains inadequately revealed and the corresponding characterization methods also need to be determined. In this article, a feature fusion algorithm based on the stacking-attention mechanism is proposed. The ratios of decomposition products from C4F7N/CO2/O2 mixtures in various proportions under different discharge conditions are utilized as the dataset. Multiple feature extraction algorithms are employed as base learners to derive feature subsets from diverse dimensions. Subsequently, these feature subsets are fused using the attention mechanism as the meta-learner, and the contribution of each feature value is determined, thereby identifying the optimal feature subset. Experimental results demonstrate that the optimal feature values extracted by this method, when applied to classification algorithms such as support vector machine (SVM) and artificial neural network (ANN), effectively distinguish between defects including corona discharge, spark discharge, and suspended discharge in environmentally friendly gas-insulated equipment. Based on the decomposition characteristics, the ratios of CO/(C3F6+CF4) and CF4/C3F8 are proposed as diagnostic indicators for identifying discharge faults in engineering applications of C4F7N/CO2/O2 gas-insulated equipment. The research results of this article provide both theoretical and technical support for the operation and maintenance of gas-insulated electrical equipment.