This research provides reliable technical support for transformer condition-based maintenance and full life-cycle management, and has significant practical value for ensuring the stable operation of industrial power systems, such as those in textile manufacturing. However, traditional state evaluation methods have limitations in extracting complex signal features, easily leading to inaccurate assessments and resource waste, making it difficult to meet the needs of intelligent upgrading and online monitoring of new power systems. This study aims to utilize the deep feature learning advantages of stacked autoencoders to solve the problem of insufficient diagnostic accuracy of traditional methods under complex operating conditions, and to construct an efficient and accurate transformer state evaluation system. By collecting multi-source operating data such as transformer temperature, vibration, electrical parameters, and oil chromatography, a stacked autoencoder model was designed after preprocessing. A training strategy combining hierarchical pre-training and overall fine-tuning was adopted, and a multi-dimensional evaluation index system was established. Experimental results show that the model can effectively extract deep features of complex signals, significantly separate heterogeneous states, and achieve an overall state recognition accuracy of over 95.7%, outperforming traditional methods such as frequency response methods and support vector machines, and exhibiting stronger anti-interference capabilities. This research provides reliable technical support for transformer condition-based maintenance and full life-cycle management, and has significant practical value for promoting smart grid construction and reducing operation and maintenance costs.

power transformer, condition assessment, stacked autoencoder, deep learning, fault diagnosis