The detection and analysis of power transformer oil is an effective method for power transformer fault diagnosis, and it is of great significance to quickly identify the oil sample of power transformer oil for power transformer fault diagnosis. The detection technology of conventional power transformer oil is mainly gas chromatography, which is complicated and not suitable for on-line detection, and it can't find the fault hidden danger of transformer in time. A method for transformer fault diagnosis of laser-induced fluorescence spectroscopy (LIF) technology and extreme learning machine (ELM) algorithm is proposed. Four oil samples were collected: thermal fault oil, electrical fault oil, local moisture-affected oil and crude oil. The spectral data of different oil samples are obtained by using a laser generator to emit fluorescence. The spectral data are processed by MSC and SNV preprocessing algorithms to prevent noise and other factors. Subsequently, the use of KPCA and PCA dimension reduction, the main components are taken 5, KPCA processing shows that MSC pretreatment of the cumulative contribution rate of the highest, 99%, MSC pre-processed PCA model cumulative contribution rate is still more than 95%, Original-KPCA and Origin-PCA model cumulative contribution rate of less than 65%, you can find that the use of pretreatment model, cumulative contribution rate has increased. Finally, the data after the two dimensions are regression fit by ELM. Experiments show that KPCA and PCA are two kinds of dimensional reduction methods. The KPCA algorithm performs best, the processing time is short, and the reliability and efficiency of the model are improved. In the same KPCA dimension reduction mode, the fitting excellence R² of the MSC-ELM model is 0.999 41, the mean square error MSE is 0.074%, and the SNV-ELM fit is 0.999 08, the mean square error MSE is 0.129%, The Original-ELM fitting excellence R² is 0.996 95, the mean square error MSE is 0.399%, and the comparison can be found that MSC is better than SNV after processing. The MSC-KPCA-ELM model performs best, the prediction value is closer to the real value. The mean square root error is the smallest. The results show that the MSC-KPCA-ELM model, combined with laser-induced fluorescence spectroscopy technology, is more suitable for the rapid diagnosis of whether or not the power transformer has failed, which type of fault is accurately determined, and the operation safety of power equipment is guaranteed.