Abstract: The Received Signal Strength Indication (RSSI) exhibits significant fluctuations and instability in the complex electromagnetic propagation environment indoors. Its characteristic distribution is easily influenced by factors such as multipath propagation and spatial structure differences, resulting in the RSSI signal having the characteristics of high dimensionality, sparsity, and difficulty in stable representation. This, in turn, restricts the indoor positioning accuracy and robustness based on RSSI. To address the problem of insufficient representation ability of RSSI signal characteristics in the high-dimensional space, this paper proposes an AutoEncoder-Spectral Convolution (AESC) model. This model first uses an autoencoder to perform unsupervised nonlinear compression on the original RSSI signal, reducing the feature dimension while suppressing redundancy and noise interference, and obtaining a compact and stable latent representation. Subsequently, a spectral convolution module is introduced at the latent feature layer to enhance the feature expression ability through discrete Fourier transform and learnable frequency-domain mapping, to depict the potential structural correlations in the feature dimension, thereby improving the model's discriminative performance. Experimental results based on the UJIIndoorLoc dataset show that AESC achieves a classification accuracy of over 97% in the building-floor multi-classification positioning task, and outperforms the comparison methods in terms of precision, recall rate, and F1 score. Further ablation experiments verify the effectiveness of the spectral enhancement module in improving the model performance, indicating that the proposed feature compression and enhancement collaborative strategy can effectively enhance the stability and robustness of RSSI fingerprint positioning.