With rapid development of internet of things ( IOT) and information technology, mobile location-based service ( LBS) is gaining more and more research focus in recent years. It also stimulates the development of indoor localization technology. Owing to the advantage of its pervasive deployment, WiFi fingerprint-based indoor localization has drawn much attention to academia. However, the fluctuation of WiFi signals and other interference always influence the localization performance. In this paper, a hierarchical deep learning indoor localization framework (HDLIL) is proposed to solve the problem of mobile device localization in indoor environments and predict the specific position. To capture and learn the reliable fingerprint features, a feature extraction module based on variational autoencoder (VAE) is introduced to characterize the latent representation of the training data. Also, to train the localization model, we feed the reconstructed training fingerprints as well as corresponding labels to a 3-layer deep neural network, of which the output of current layer is set as the input for the next layer, followed by a location output module based on the concatenated softmax classification. In the localization phase, the localization fingerprints ( testing fingerprints) is set as the input, the HDLIL model is invoked and the output of final layer is the predicted location of the mobile device. In addition, to evaluate the localization performance, we conducted the experiment in a real indoor scene, and several influence factors are discussed. The result indicates that the proposed HDLIL model can attain a superior localization performance.