Quantitative analysis of transformer fault gas acetylene photothermal spectroscopy data based on CNN-WAM-LSTM

The detection of acetylene content in the fault gas of oil immersed transformers is the key to determining the type of internal faults and safe operation of the equipment. The photothermal spectroscopy method has a high sensitivity advantage in detecting fault gases, but there are problems such as information redundancy in the analysis process of photothermal spectroscopy data. This article proposes an improved convolutional neural network quantitative model that integrates wavelength attention mechanism and long short-term memory model, and designs and builds a photothermal spectroscopy platform based on Sagnac interference structure to collect C₂H₂ multi concentration samples. The dataset is augmented with data augmentation to enhance the robustness and generalization of the model. Optimize the model structure and network weights through experiments, complete model training, and verify prediction performance. Verify the contribution of wavelength attention and long short-term memory models through ablation experiments, and compare them with SVR, PLS, and XGBoost. The results indicate that the model has good prediction accuracy, outperforming traditional regression models in R², RMSE, MAE, and MAPE indicators, and is suitable for quantitative analysis of gas spectra in transformer faults.