In ultra-wideband wireless communication systems, the accurate identification of non-line-of-sight and line-of-sight environments is crucial for enhancing positioning accuracy and system performance. To this end, a feature extraction method based on an attention mechanism structure is proposed to enhance the sparse autoencoder. The SAE learns stable and effective feature representations through sparse coding, while the attention mechanism highlights crucial information via dynamic weight allocation. Their integration enables more precise feature dimensionality reduction. A semi-supervised classification strategy combining Gaussian Mixture Models and self-training Support Vector Machines is employed for LOS/NLOS state recognition. The GMM models the generative probability distribution parameters of the extracted feature space, while the self-training SVM enhances classification accuracy and robustness in marginal regions. Experimental results demonstrate that, when comparing the performance of different dimensionality reduction methods combined with LOS/NLOS classifiers in the NLOS recognition task, the proposed SAE-ATTENTION feature extraction strategy coupled with the GMM-SVM classification model exhibits significant advantages in recognition accuracy and reliability. In static and dynamic experiments, this approach achieved NLOS recognition accuracies of 95.20% and 93.05% respectively, effectively enhanced the model’s generalisation capability under conditions of limited labelled samples.