Transfer learning-based cross-environment UWB NLOS identification and error mitigation

High-precision indoor positioning remains sensitive to non-line-of-sight (NLOS) propagation and multipath interference. This study investigates cross-environment adaptation of Encoder-only Transformer (ET) models for raw channel impulse response (CIR) sequences. The adaptation configuration combines optional Adapter insertion after self-attention with selective updates of existing transformer modules and is selected separately for NLOS classification and ranging-error regression under limited target-domain supervision. TPE and hyperband are used for configuration search and early pruning. Experiments on a public dataset collected in four indoor environments show that, under the evaluated settings, ET obtains an average classification accuracy of 84.94% and an average mean absolute ranging error of 28.23 cm, compared with 82.51% and 44.70 cm for CNN. ET-light uses approximately 10.92% of the parameters of ET and yields an average accuracy of 83.38% and an average MAE of 31.50 cm, illustrating the observed trade-off between predictive performance and model size.