As a typical deep unsupervised learning model, autoencoder can automatically learn effective abstract features from unlabeled samples.In recent years, autoencoder has been widely used in target recognition, intrusion detection, fault diagnosis and many other fields.Thus, the theoretical basis, improved methods, application fields and research directions of autoencoder were described and summarized comprehensively.At first, the network structure, theoretical derivation and algorithm flow of traditional autoencoder were introduced and analyzed, and the difference between autoencoder and other unsupervised learning algorithms was compared.Then, common improved autoencoders were discussed, and their innovation, improvement methods and relative merits were analyzed.Next, the practical application status of autoencoder in target recognition, intrusion detection and other fields were introduced.At last, the existing problems of autoencoder were summarized, and the possible research directions were prospected.
