Abstract: Objectives: Inertial measurement unit (IMU) data interruption can significantly degrade the positioning and attitude accuracy of integrated navigation systems. Methods: Considering that IMU observations are high-frequency time-series signals with strong temporal correlations between adjacent epochs, a bidirectional long short-term memory (BiLSTM)-based time-series prediction method is proposed for IMU data recovery. The proposed method exploits both forward and backward temporal dependencies to reconstruct missing IMU observations within interrupted segments. Two representative road sections are randomly selected to simulate IMU interruption scenarios, and the proposed method is compared with the autoregressive integrated moving average (ARIMA) model and the support vector machine (SVM) model. Results: Experimental results show that, compared with traditional methods, the proposed approach reduces the mean absolute error (MAE) by approximately 30%–47% and improves the coefficient of determination (R²) to 0.84, demonstrating higher prediction accuracy in IMU data recovery. Integrated navigation results further verify that the proposed method accelerates error convergence after interruption, suppresses error accumulation to a certain extent, and improves positioning and attitude accuracy under short-term IMU outage conditions. Conclusions: The proposed BiLSTM-based method effectively repairs interrupted IMU data and enhances the accuracy of post-processed integrated navigation solutions. However, because it relies on complete time-series information, the method is only applicable to post-processing applications and is not suitable for real-time online navigation services.