Abstract: The underground Received Signal Strength Indicator (RSSI) signal exhibits non-stationary characteristics such as sharp spikes, high-frequency jitter, and trend drift under the influence of multipath propagation, occlusion, and electromagnetic interference, resulting in large positioning errors. Existing positioning methods lack collaborative suppression of multi-source interference, and their feature extraction and multi-scale feature fusion are insufficient. To address these problems, an underground RSSI positioning method based on multi-stage denoising and a dual-branch temporal network was proposed. Multi-stage denoising suppressed spike interference, high-frequency jitter, and trend drift through outlier elimination and interpolation repair, adaptive Kalman filtering, and wavelet-domain adaptive gating, respectively, thereby producing a more stable RSSI sequence with preserved details. The dual-branch temporal network introduced the first-order difference as an auxiliary disturbance prior, extracted features in parallel through a trend branch and a disturbance branch, and adaptively fused them via a channel attention mechanism. A Bidirectional Long Short-Term Memory (Bi-LSTM) network was then used to capture contextual temporal dependencies, ensuring trajectory smoothness and continuity in complex dynamic environments. Test results showed that the RSSI signal became more stable after multi-stage denoising while preserving local dynamic features without excessive smoothing. The dual-branch temporal network achieved high accuracy, F₁-score, precision, and recall with fast convergence; in tests under different scenarios, both accuracy and F₁-score exceeded 85%, demonstrating good generalization. In continuous positioning tasks under dynamic environments, the average positioning error of the proposed method was only 0.12 m.