Accurate diagnosis of transformer faults can effectively improve the enduring reliability of power grid operation. Aiming at overcoming the problems of long time consumption and low diagnostic rate in the past diagnosis methods, this article designs a laser-induced fluorescence (LIF) detection system, which can be combined with a multi-scale one-dimensional convolution neural network (MS1DCNN) to diagnose transformer fault categories. The structural parameters of MS1DCNN are optimized using the improved wild horse optimizer (IWHO). Electrical fault oil, thermal fault oil, normal oil and locally damped oil are used as raw materials for the experiment. First, the LIF spectral data of the four kinds of oil samples are obtained, and the spectral data obtained are pretreated by standard normal variate (SNV) and multiple scattering correction (MSC), and the dimensions are reduced by linear discriminant analysis (LDA) and kernel principal component analysis (KPCA). Then the dimensionality reduced data are imported into the MS1DCNN algorithm for learning, and the parameters of MS1DCNN are optimized using the IWHO algorithm. Finally, the experiment shows that the efficiency and precision of LIF technology for raw data extraction are higher than for traditional methods; in comparison with the same type of algorithm, MSC has a better preprocessing effect, KPCA has a better dimensionality reduction effect, MS1DCNN has a better prediction effect, and IWHO has a better optimization effect. Compared to them, the MSC-KPCA-IWHO-MS1DCNN model has the best diagnostic ability, with a mean square error (MSE) of 4.9037 × 10-4, mean absolute error (MAE) of 0.0179, and goodness of fit (R2) of 0.9996. Transformer fault intelligent diagnosis is necessary for the sustained and stable operation of power networks.