Abstract: The extreme environment of underground coal mines, characterized by narrow spaces, coal dust, and uneven illumination, renders Global Navigation Satellite System (GNSS) signals ineffective. Furthermore, wireless positioning technologies such as Ultra-Wideband (UWB) are susceptible to multipath effects and obstructions from metallic structures. Consequently, achieving accurate and stable positioning for mobile equipment like roadheaders has become a critical challenge for intelligent mining. To address the degradation of positioning accuracy caused by susceptibility to interference in visual feature matching and the error accumulation of Inertial Measurement Units (IMU) in such complex underground environments, a visual-inertial integrated positioning method fusing motion priors and a Long Short-Term Memory (LSTM) network is proposed. In the visual front-end, a Region of Interest (ROI) generation mechanism based on IMU motion priors is designed. IMU pre-integration information is utilized to predict camera motion and establish a local ROI to constrain the feature search range. A two-layer probabilistic model is then introduced to jointly evaluate candidate features from the dimensions of appearance similarity and motion consistency, thereby eliminating mismatched point pairs. For back-end fusion, an error compensation framework cascading an Extended Kalman Filter (EKF) and an LSTM network is constructed. The front-end EKF performs preliminary IMU state estimation and tightly-coupled fusion with visual observations. Subsequently, the back-end LSTM network conducts deep temporal modeling of the nonlinear errors present in the EKF residuals, predicting the position error compensation, which is then fed back to the EKF front-end. Simultaneously, a three-level closed-loop optimization mechanism is designed to achieve deep synergy between back-end compensation and front-end correction through adaptive covariance resetting, sliding window smoothing, and online bias calibration. Experimental results on the EuRoC public dataset demonstrate that the proposed method achieves absolute trajectory errors of 0.130 and 0.251 m on the MH_01 and V1_01 sequences, respectively, which is comparable in accuracy to the advanced algorithm Air-SLAM. On a self-collected dataset simulating an underground coal mine environment, the positioning error under normal roadway conditions is 0.705 m, representing a 5.4% improvement over Air-SLAM. Under conditions simulating dust and intense light interference, the positioning error is 0.820 m, a 12.8% improvement over Air-SLAM. During periods of severe visual degradation, the LSTM module reduces the average positioning error from 0.508 m to 0.319 m, an optimization margin of 37.2%. Results show that the proposed method can effectively suppress feature mismatches and IMU cumulative drift in complex underground environments, significantly enhancing the accuracy and robustness of the positioning system under extreme conditions, thereby providing technical support for the long-distance autonomous navigation of roadheading equipment.