This study provides insights into deep learning-based feature extraction and classification recognition algorithms for electronic messages, particularly the application of autoencoders and their variants in unsupervised learning. By employing an unsupervised greedy hierarchical training algorithm, deep learning architectures such as deep belief networks (DBNs) are efficiently processed and optimised, thereby overcoming the challenges inherent in the optimisation process of deep structures. The research highlights the central role of autoencoders in electronic information processing, where the architecture maps data from the sample space to the representative feature space through encoding and decoding processes, and extracts deep features of the data by minimising the reconstruction error. In order to address the potential overfitting problem of deep learning models, this study employs regularisation techniques including weight attenuation and bias attenuation, as well as the introduction of noise as a regularisation mechanism in the model to enhance the generalisation ability and robustness of the model. The study also involves sophisticated time–frequency representation techniques such as power spectral density analysis and s-algorithms for feature extraction of electronic signals, demonstrating the high adaptability and effectiveness of deep learning in processing electronic information. The performance of the deep learning algorithms is evaluated by conducting experiments on seven different datasets, including methods such as Gaussian kernel mapping, RDFM, and DMFA. The experimental results show significant differences in the performance of different algorithms on different datasets, emphasising the importance of choosing the appropriate algorithm based on the characteristics of the dataset. In addition, the experiments demonstrate the effectiveness of deep learning algorithms, especially stacked self-encoders, in the task of classifying electronic information. This study not only reveals the potential of deep learning in the field of electronic information processing but also provides a theoretical and experimental basis for future research, especially in improving the algorithms’ ability to generalise, to handle high-dimensional data, and to utilise finite features when the sample size is insufficient. By continuously optimising deep learning models and algorithms, more breakthroughs are expected in the field of electronic information feature extraction and classification and recognition.