In underground garage environments, traditional RSSI-based fingerprint positioning is compromised by fluctuating radiation, multipath effects, and interference, leading to feature distortion and positioning errors. This paper proposes a dual-source sensing indoor positioning method that integrates environmental radiation perception with signal analysis to enhance system robustness. In the offline phase, a bidirectional fusion model combining BiLSTM and BiGRU is employed to capture both long-term and short-term radiation effects on RSSI. By leveraging multi-head self-attention, the model constructs an adaptive fingerprint database that accommodates varying radiation conditions. During the online matching phase, an exponential power normalization technique is used to map RSSI signals to a unified scale, mitigating hardware-related interference. An AP-aware clustering algorithm is introduced to select RPs based on AP signal quality and suppress matching deviations through density estimation. Experimental results in underground garages demonstrate the method′s strong performance. Under known radiation conditions, it achieves an average positioning accuracy improvement of 11.05%~25.38% over baseline methods. Under unknown radiation conditions, the BGLA-constructed fingerprint database enables it to outperform comparative approaches by 27.55%~35.71% in average positioning accuracy.