A CNN-GRU Fusion Mathematical Model for Positioning Jump Correction in Integrated Navigation Systems Authors: Mingyang Deng, Guangjiao Chen Published: 2026 ABSTRACT Positioning jumps are a primary cause of trajectory discontinuities in urban canyon environments, severely hindering the widespread adoption of autonomous vehicles. This paper proposes a CNN-GRU fusion-based method for correcting such positioning jumps. A complete 15-dimensional error-state extended Kalman filter (EKF) framework is established to analyze the jump generation mechanisms from three perspectives—pseudorange distortion, inertial drift, and filter gain divergence—thereby justifying the use of inertial measurement unit (IMU) time-series data for anomaly prediction. An end-to-end mapping model is further constructed, in which a one-dimensional convolutional neural network (CNN) extracts cross-channel spatial features from multi-axis inertial data, while a gated recurrent unit (GRU) captures long-term temporal error evolution. A hysteresis navigation quality factor and a piecewise Huber loss function are incorporated to enable hierarchical adaptive optimization. Experimental results demonstrate that the proposed method reduces the positioning root mean square error (RMSE) from 1.24 m to 0.70 m, achieves a jump suppression rate of 43.5%, and maintains a single-frame inference latency of 11.8 ms, meeting the real-time requirements for future autonomous driving localization. 1. Introduction High-level autonomous driving systems demand continuous, high-precision, and robust lane-level localization to underpin environmental perception, path planning, and vehicle motion control [1]. The GNSS/INS integrated navigation system, which synergistically combines the long-term absolute accuracy of GNSS with the high-rate continuous output of inertial measurement units (IMUs), has become the de facto positioning solution for mass-produced intelligent vehicles. However, in typical urban canyons, under viaducts, and in dense foliage, line-of-sight (LOS) satellite signals are severely attenuated, and multipath reflections substantially distort pseudorange observations [2]. These effects induce instantaneous meter-level positioning jumps and step-like trajectory discontinuities, which in turn cause oscillations in path planning and instability in lateral control, thereby posing a serious threat to autonomous driving safety [3]. Conventional approaches to mitigate localization errors fall into three main categories. The first category comprises fixed-gain Kalman filters and their adaptive variants, which adjust the noise covariance matrix based on residual statistics [4,5]. Nevertheless, such methods offer only limited improvement for slowly varying noise and exhibit pronounced lag when responding to abrupt non-line-of-sight (NLOS) outliers, failing to eliminate step-type jumps fundamentally. The second category leverages additional sensors such as LiDAR and cameras in multi-sensor fusion frameworks; while these schemes enhance robustness, their high hardware cost and computational burden impede wide deployment in low-cost vehicle terminals [6,7]. The third category, driven by recent advances in deep learning, employs temporal neural networks for navigation error prediction and compensation [8,9,10]. Convolutional neural networks (CNNs) are particularly effective at extracting spatial coupling features from multi-channel sensor data, whereas recurrent neural networks (RNNs), including gated recurrent units (GRUs), excel at modeling long-term temporal dependencies [11]. Despite these efforts, existing CNN-GRU integrated navigation models are pre-dominantly designed for error compensation under complete GNSS outages, with little systematic attention paid to the localization jump problem caused by partial signal degradation [12]. Moreover, most deep learning-based navigation models lack rigorous mathematical derivation and interpretability [13]. In addition, the common practice of using a single loss function cannot adaptively discriminate between severe abnormal outliers and mild trajectory jitters, often resulting in either insufficient jump suppression or excessive smoothing distortion [14]. This paper proposes a hierarchical adaptive model fusing a 1D CNN with a modified GRU for positioning jump correction. Theoretically, GNSS NLOS errors propagate through the EKF and leave traces in the error-state evolution. Section 2.3 shows that the posterior estimation error depends on both GNSS and IMU noises, implying that IMU-state consistency encodes predictive information for anomalies. This motivates a deep architecture to extract IMU spatial features and map them to EKF corrections. The main contributions are as follows. First, we derive a 15-D error-state EKF to analyze three mechanisms of NLOS-induced jumps, establishing the theoretical basis for IMU prediction. Second, we propose a modified GRU with adaptive gated scaling for enhanced long-term memory, and construct an interpretable spatio-temporal fusion framework to capture IMU spatial features and error temporal evolution. Third, we introduce a hysteresis quality factor for adaptive hierarchical discrimination and design a piecewise Huber loss to balance robustness and fidelity by differentially handling outliers and jitters. Fourth, we perform vehicle experiments and ablation studies on urban canyon data, with MATLAB analysis, to validate accuracy, jump suppression, and re-al-time performance. The remainder is organized as follows. Section 2 presents GNSS/INS fundamentals and jump generation. Section 3 details the proposed 1D-CNN/mGRU model. Section 4 describes experiments and results. Section 5 concludes and discusses future work. 2. Navigation Fundamentals and Localization Jump Mechanism The research demonstrates that IMU time-series data can provide a motion baseline, propagate GNSS anomalies into detectable deviations, and enable data-driven mapping of these deviations. Therefore, leveraging the mechanism of predicting EKF localization errors from IMU time-series data, this paper proposes a correction model architecture. The network is specifically designed to: (a) extract spatial features from multi-channel IMU data via a 1D CNN; (b) model the temporal evolution of localization errors using a modified GRU; and (c) perform hierarchical adaptive correction under the guidance of a hysteresis quality factor. 2.1. 15-Dimensional Error-State EKF Model The loosely coupled GNSS/INS integrated navigation architecture is adopted, and a 15-dimensional error state vector is constructed to comprehensively describe position error, velocity error, attitude misalignment angle, gyroscope drift, and accelerometer bias. The state vector is defined as [15]:(1)x=δpT,δvT,ϕT,bgT,baTT∈R15 where δp∈R3 and δv∈R3 denote the three-dimensional position error and velocity error in the navigation coordinate system, respectively; ϕ∈R3 represents the attitude misalignment angle vector; bg∈R3 and ba∈R3 are the drift vector of the gyroscope and the bias vector of the accelerometer, respectively. (1) State Prediction Equation The discrete state prediction process of EKF is expressed as [15]:(2)x^k|k−1=Fkx^k−1|k−1+wk,wk∼N0,Qk(3)Pk|k−1=FkPk−1|k−1FkT+Qk where Fk is the time-varying state transition matrix; wk denotes the system process noise vector obeying Gaussian distribution; Qk represents the process noise covariance matrix; Pk|k−1 is the prior state estimation covariance matrix. (2) GNSS Observation Equation The GNSS position and velocity observation equations are constructed as follows [15]:(4)zk=Hkxk+vk,vk∼N0,Rk where zk is the GNSS observation vector; Hk denotes the observation mapping matrix; vk represents the Gaussian observation noise vector; Rk is the observation noise covariance matrix. (3) Filter Update Equation The Kalman gain calculation, posterior state estimation, and covariance update formulas are given by [15,16]:(5)Kk=Pk|k−1HkTHkPk|k−1HkT+Rk−1(6)x^K|k=x^k|k−1+Kkzk−Hkx^k|k−1(7)Pk|k=I−KkHkPk|k−1 Traditional EKF adopts a fixed Rk matrix, which cannot adapt to the sharp degradation of GNSS signal quality in NLOS scenarios. The mismatch of noise statistical characteristics leads to excessive Kalman gain, amplifies abnormal observation errors, and eventually causes severe localization jumps. 2.2. Definition and Generation Mechanism of Localization Jump (1) Quantitative Criterion of Localization Jump To accurately identify abnormal localization jumps, the 2-norm of the inter-frame planar position deviation is defined as the jump amplitude evaluation index [16]:(8)Δdk=‖P^k|k−P^k−1|k−1‖2 Combined with vehicle kinematic physical constraints, the maximum feasible single-frame displacement satisfies the following condition:(9)Δdk≤vmaxΔt+12amaxΔt2 where the maximum vehicle speed vmax=15 km/h, the maximum acceleration amax=3 m/s2, and the system sampling interval Δt=0.1 s. If the actual inter-frame displacement fails to satisfy Equations (8) and (9), the current epoch is judged as a localization jump outlier without physical motion support. (2) Three Mathematical Causes of Localization Jump (a) Multipath pseudorange distortion: The ideal geometric pseudorange observation model of GNSS satellites is [16]:(10)ρi=ri+cδtu−δts+Ii+Ti+ϵρ where ri is the geometric distance between the satellite and the receiver; c denotes the speed of light; δtu and δts represent the receiver clock error and satellite clock error, respectively; Ii and Ti are the ionospheric delay and tropospheric delay; ϵρ denotes random noise. In NLOS scenarios, reflection interference introduces an extra deviation Δρbias, forming distorted pseudorange ρ^i=ρi+Δρbias and directly causing meter-level coordinate mutation. (b) Discontinuous inertial drift compensation: Under long-term GNSS outage conditions, the accelerometer bias continuously accumulates position errors through double integration:(11)Δpdrift=∬Cbnbadt2 where Cbn is the direction cosine matrix from the body coordinate system to the navigation coordinate system. When satellite signals recover suddenly, the filter compensates for all accumulated drift errors at one time, resulting in discontinuous trajectory steps. (c) Kalman gain divergence: The fixed observation noise matrix cannot adapt to the sharp increase in residuals in NLOS scenarios. The excessive filter gain makes distorted GNSS observations dominate the state update, further amplifying localization oscillation and jump distortion. Although GNSS NLOS errors are not directly observable from IMU measurements, the resulting localization jumps are manifested as discrepancies between IMU-predicted states and GNSS observations. Specifically, the innovation vector y˜k=zk−Hk x^k|k−1 becomes statistically abnormal during NLOS periods. Since this innovation anomaly correlates with the IMU temporal sequence (through the prediction step), the IMU data implicitly encode the information needed to predict jumps. This theoretical insight motivates our end-to-end learning framework: the network learns the mapping from IMU patterns to correction vectors without explicit GNSS NLOS detection. 2.3. Theoretical Basis for IMU-Based Prediction of Localization Jumps A critical question arises: since GNSS NLOS errors originate from satellite signal blockage, how can a model relying primarily on IMU measurements predict the resulting localization jumps? This subsection establishes the theoretical foundation for the feasibility of IMU-based jump prediction from three complementary perspectives. (1) Kinematic Constraint Transitivity Under NLOS conditions, GNSS pseudorange observations contain large positive biases, causing the EKF to generate position estimates that violate the vehicle’s physical motion constraints. Specifically, the true vehicle motion must satisfy [15,16,17]:(12)‖Pk−Pk−1‖2≤vmax⋅Δt+12amaxΔt2 where vmax and amax and are the maximum feasible velocity and acceleration given the vehicle’s dynamic envelope. When GNSS observations are corrupted by NLOS errors, the EKF output violates this inequality, producing a “jump” pattern in the position trajectory. Crucially, the IMU measurements, which are insensitive to GNSS signal quality, provide a continuous and physically consistent representation of the vehicle’s actual motion. The discrepancy between the EKF-estimated motion (contaminated by NLOS) and the IMU-indicated motion (physically plausible) thus serves as a detectable signature of GNSS anomalies. (2) Error-State Temporal Propagation To formalize the relationship between IMU measurements and EKF output errors, we analyze the error-state propagation mechanism. From the EKF prediction and update equations, the posterior state estimation error can be rigorously derived as [16,17]:(13)x˜k|k=I−KkHkFkx˜k−1|k−1+I−KkHkwk−1−Kkvk where wk−1∼N(0,Qk−1) denotes the IMU process noise (primarily gyroscope drift and accelerometer bias), and vk∼N(0,Rk) denotes the GNSS observation noise, which becomes abnormally large under NLOS conditions. This equation reveals a crucial property: the GNSS observation noise vk does not directly appear in the IMU measurements, but its effect propagates through the Kalman gain Kk and contaminates the entire state vector. In other words, the NLOS-induced error leaves detectable “traces” in the state evolution pattern—traces that manifest as deviations from the expected IMU-state consistency. The IMU measurements Uk=[ωk,ak] provide a reference baseline for what the state evolution should be in the absence of GNSS corruption. By observing the temporal sequence of IMU data within a sliding window Uk, the network learns to identify these deviation patterns and predict the corresponding correction δ^k. This is mathematically feasible because the mapping from IMU features to EKF errors is a continuous function (albeit highly nonlinear) that can be approximated by the proposed 1D-CNN/mGRU architecture. (3) Data-Driven Functional Approximation It is important to clarify that the network does not attempt to “infer” the physical NLOS conditions from IMU data that a physically underdetermined inverse problem. Instead, the network learns a data-driven approximation of the complex nonlinear relationship [17,18]:(14)δk=FUk;θ+ϵk where F is an unknown but deterministic function mapping the recent history of IMU measurements to the EKF output error, and ϵk represents the irreducible approximation residual. During training, the network observes a large number of examples {Uk,δktrue} collected across diverse urban NLOS scenarios. Through supervised learning, it identifies the statistical regularities between IMU-derived motion features and EKF error patterns—essentially learning a surrogate model for the complex interaction between vehicle dynamics and GNSS error propagation. This approach is conceptually aligned with residual learning [17,18], where the network learns to predict the residual between the corrupted estimate and the ground truth. 2.4. Predictability Analysis of Localization Jump Urban localization jumps are not purely random noise, and the spatio-temporal characteristics of IMU data can support the prediction and correction of localization anomalies. The theoretical basis is summarized as follows. First, inertial navigation errors satisfy Lipschitz continuity, and the drift evolution is smooth in a short-time window, ensuring traceable error variation. Second, satellite signal quality degrades gradually in urban occlusion scenarios, providing an effective early warning window for jump prediction. Third, vehicle motion states such as turning and bumping are strongly coupled with multipath interference intensity, which can be fully characterized by six-axis IMU observations. Fourth, urban occlusion scenarios exhibit spatial repeatability, enabling the network to learn typical temporal laws of localization jumps. 3. Proposed 1D-Cnn and Modified GRU Hierarchical Correction Model 3.1. Temporal Regression Task Modeling The single-epoch six-axis IMU observation vector is constructed as uk=[ωkT,akT]T∈R6, where ωk and ak denote angular velocity and specific force observations, respectively. A sliding time window with length T=50 is adopted to construct the network input matrix [19]:(15)Uk=Uk−T+1,Uk−T+2,…,Uk∈RT×6 The sliding window length is set to T=50 corresponding to 50 × 0.1 s = 5 s of historical IMU data at the 10 Hz sampling rate. This value is selected through a systematic ablation study over T∈10,20,30,50,80,100 on the validation set. The results show that T=50 achieves the optimal balance between prediction accuracy and computational latency: shorter windows T≤30 fail to capture sufficient temporal context for low-frequency navigation error evolution, while longer windows T≥80 introduce redundant information and increased inference latency without commensurate accuracy gains. The end-to-end nonlinear mapping relationship of the proposed model is defined as [19]:(16)δ^k=MθUk;Θ where δ^k is the predicted localization correction vector; Mθ⋅ denotes the mapping function of the proposed network; Θ represents all trainable parameters of the network. The optimal parameter set is obtained by minimizing the loss function [19]:(17)Θ*=arg minΘ1N∑Lk=1NMθUk;Θ,δktrue where δktrue is the real localization error solved by high-precision real-time kinematic (RTK) ground truth; N denotes the total number of samples; L⋅ is the custom loss function. 3.2. 1D-CNN Spatial Feature Extraction IMU measurements comprise six-axis synchronized signals with inherent cross-channel coupling. 1D-CNN is adopted over 2D-CNN or Transformer for two reasons: (i) the temporal dimension of IMU data exhibits local stationarity, making the local receptive field of 1D-CNN physically interpretable as capturing short-term motion patterns; (ii) the computational complexity of 1D-CNN is O(T·C·K), significantly lower than 2D-CNN O(T·C·H·W) and attention-based O(T2·C), which is critical for onboard real-time deployment. A two-layer 1D-CNN is adopted to extract cross-channel spatial coupling features of multi-sensor IMU data. For the input temporal matrix U∈RT×C (C=6), the convolution calculation is expressed as [20]:(18)Zi,f=σ∑c=1C∑l=0L−1Ui+l,c⋅Wl,c,f+bf where the activation function σ⋅=ReLU; Wl,c,f and bf denote the convolution kernel weight and bias, respectively. The first convolution layer adopts a kernel length of 5 and 64 feature channels, while the second convolution layer adopts a kernel length of 3 and 128 feature channels. Batch normalization and max-pooling operations are embedded to eliminate redundant features and enhance spatial feature representation capability. The 1D-CNN architecture is designed with two convolutional layers. The first layer adopts a kernel length of 5 and 64 feature channels, while the second layer adopts a kernel length of 3 and 128 feature channels. These hyperparameters are selected via a grid search on the validation set, with kernel lengths searched over (3,5,7) and channel numbers searched over (32,64,128,256). The optimal configuration—kernel lengths (5,3) and channels (64,128)—achieves the best trade-off between feature extraction capability and computational efficiency. The larger kernel in the first layer captures short-term motion patterns, while the smaller kernel in the second layer extracts finer spatial features from the intermediate representations. Batch normalization and max-pooling operations are embedded after each convolutional layer to eliminate redundant features and enhance spatial feature representation capability. 3.3. Modified GRU Temporal Memory Unit Traditional GRU lacks sufficient long-term temporal feature retention capability for low-frequency, slowly varying navigation errors. In this paper, an adaptive scaling coefficient is introduced to optimize the update gate structure, improving the long-sequence error modeling performance. The complete formulas of the modified GRU are presented as follows. Standard GRU suffers from rapid information decay when processing low-frequency navigation errors. The update gate ut determines the retention of historical states. However, under slowly varying GNSS degradation, the error evolves as a low-frequency perturbation, yet the standard update gate remains overly sensitive to instantaneous inputs, leading to premature state refresh. To address this, we introduce a scaling coefficient γ = 0.8 to compress the update gate activation, yielding slower state evolution and enhanced long-term memory retention. Reset gate:(19)rt=σWrzt+Urht−1+br where γ is the scaling coefficient, set to 0.8. γ=0.8 is selected based on grid search. The optimal γ lies in [0.75, 0.85], corresponding to the typical NLOS error correlation time constant of 5–10 s. This alignment indicates that the scaling coefficient effectively matches the temporal characteristics of urban GNSS degradation. The scaling coefficient γ=0.8 is introduced to suppress the update gate, thereby enhancing the long-term memory retention of the GRU. The value is determined through a grid search over γ∈0.5,0.6,0.7,0.8,0.9,1.0 on the validation set, with γ=0.8 yielding the minimum validation RMSE. Theoretically, when γ<1, the effective memory time constant is τ≈1/1−γ=5 time steps, which aligns with the typical correlation time scale of IMU drift errors (approximately 0.5 s at 10 Hz).(20)ut=σγ⋅Wuzt+Uuht−1+bu Candidate hidden state:(21)h˜t=tanhWhzt+Uhrt⊙ht−1+bh Output hidden state:(22)ht=1−ut⊙ht−1+ut⊙h˜t where ⊙ denotes the Hadamard product; zt is the spatial feature vector extracted by the CNN. Two layers of modified GRU are stacked with 64 and 32 hidden units, respectively, balancing temporal fitting accuracy and embedded computational efficiency. 3.4. Hysteresis Hierarchical Correction Strategy To adaptively distinguish strong and weak interference working conditions, a dimensionless navigation quality factor is constructed by integrating jump amplitude, position dilution of precision (PDOP), and the number of valid satellites:(23)Qk=αQ˜k+(1−α)Qk-1 where α∈(0,1) is the hysteresis smoothing factor that controls the temporal persistence of the quality evaluation, preventing frequent mode switching due to transient fluctuations. Qk-1 is the quality factor from the previous time step. Qk denotes the hysteresis navigation quality evaluation factor at time step k, which integrates both instantaneous and historical navigation quality indicators; Q˜k is the instantaneous quality indicator computed from current-frame measurements, defined as:(24)Qk=ω1ΔdkΔdth+ω2PDOPkPDOPth+ω3Nth−Nsat,kNth where the weight coefficients ω1=0.4,ω2=0.3,ω3=0.3; ΔdK being the single-frame displacement, the threshold parameters Δdth=1.0 m, PDOPth=2.0, and Nth=6. To avoid frequent working condition switching and oscillation, a hysteresis switching mechanism is designed. If Qk≤0.3 for five consecutive frames, the working condition is judged as stable and superior, and the basic correction strategy is adopted. If Qk≥0.7 for two consecutive frames, the working condition is judged as a severe NLOS condition, and the adaptive attenuation correction coefficient is introduced:(25)ηk=exp−λ⋅ΔdkΔdth The final correction output of the model is δ^k=McoarseUk⋅ηk+δkINS. For the transition interval 0.3