Durative Monitoring of Sulfur Hexafluoride Characteristic Gases under Hydrogen Interference Using a Time2Vec-Encoded CNN–Transformer–LSTM Model Based on a Heterogeneous Gas Sensor Array

Gas-insulated switchgear (GIS) systems extensively employ sulfur hexafluoride (SF 6 ) as an insulating medium and are widely deployed in modern power systems. Under partial discharge (PD) conditions, SF 6 decomposes to generate hazardous byproducts such as H 2 S, SO 2, CO, and a certain amount of H 2 . To mitigate the cross-sensitivity interference among gas sensors when detecting mixed gases, a heterogeneous gas sensor array was designed, integrating three distinct sensor types: metal oxide semiconductor (MOS) sensors, an electrochemical sensor, and a Pd–Au alloy hydrogen sensor. A novel detection framework incorporating a Time2Vec-encoded CNN–Transformer–LSTM deep learning model was proposed for the qualitative identification and quantitative prediction of tetra-component gas mixtures in the SF 6 background. The experimental data set was collected over two consecutive days, where the data from Day 1 were augmented to improve the model’s generalization performance. Among the three data augmentation strategies evaluated, Gaussian random noise injection yielded superior results in both classification and regression tasks. This approach achieved a classification accuracy of 97.0% and an average F1-score of 97.3%. For concentration estimation, the proposed model attained an average R 2 value of 97.6%, with the RMSE for H 2 S, SO 2, CO, and H 2 recorded at 0.251, 0.415, 3.023, and 5.701 ppm, respectively. In addition, comparative evaluations with four classical machine learning models─SVM, RF, KNN, and MLP─substantiated the superior accuracy and robustness of the proposed model. Ultimately, the contribution of the Pd–Au alloy hydrogen sensor to the overall performance of the heterogeneous sensor array was comprehensively evaluated. Experimental findings substantiated the sensor’s exceptional selectivity for H 2 and its pivotal role in effectively mitigating cross-sensitivity effects among the other sensors. The integration of a heterogeneous sensor array with the proposed framework exhibits a strong potential for accurate online monitoring of SF 6 decomposition products in GIS systems.