In the area of transformer noise monitoring research, condition monitoring technology based on the combination of AI technology and modern signal processing methods gradually becomes a hot research topic. General pattern recognition methods often have higher requirements for signal acquisition and processing and are often restricted by the limited generalization ability of the models. In order to solve this problem, a sparse autoencoder-based transformer state recognition method is proposed. The sparse autoencoder (SAE) is trained by FFT spectral signals, and the intrinsic features of big data are adaptively refined into simple feature functions. Intelligent diagnosis of transformer conditions is realized by the expression of feature functions. The experimental results show that, compared with BP (backpropagation) neural network and SVM (support vector machine) classification algorithms, this method can quickly and effectively improve the accuracy of transformer state recognition, which is of great significance for industrial development.