Inertial navigation compensation method for 5G signal missing condition

Abstract

The invention relates to the technical field of communication positioning, in particular to an inertial navigation compensation method for a 5G signal missing situation, which comprises the following steps: acquiring original inertial data including data of a triaxial accelerometer and a triaxial gyroscope through an inertial measurement unit, and calculating the data of the triaxial accelerometer and the triaxial gyroscope by an inertial navigation unit according to the data provided by the inertial measurement unit; the position, the speed and the posture of an object are calculated through integration, 5G data and data obtained through calculation of an inertial measurement unit are fused and optimized through Kalman filtering, all the optimized data and 5G position increment information serve as input and output of an AT-GRU algorithm respectively, and training prediction of the model is achieved. According to the method, the 5G signal is tracked and predicted and compensated in real time by using the characteristics in the inertial navigation data information, the continuity and robustness of the positioning process are ensured, the positioning precision in a complex scene is improved, and the problem that in the prior art, when a positioning target is in a relatively complex environment, the 5G positioning precision is reduced due to object shielding, and the positioning precision is influenced is solved. And even the 5G positioning signal is completely lost.

Description

Technical Field

[0001] This invention relates to the field of communication positioning technology, and in particular to an inertial navigation compensation method for situations where 5G signals are missing. Background Technology

[0002] With the development of modern society, obtaining high-precision location has become a growing need. For outdoor positioning, satellite positioning has become increasingly mature after decades of development. Currently, there are four major global satellite navigation systems: the US GPS system, China's BeiDou Navigation Satellite System, Russia's GLONASS system, and Europe's Galileo system, achieving centimeter-level positioning accuracy in ideal environments. However, for indoor scenarios, satellite positioning is unusable due to building obstructions and signal attenuation. Research on high-precision indoor positioning has become a hot topic, resulting in numerous positioning technologies and algorithms. Common indoor positioning technologies include UWB positioning, 5G positioning, pseudosatellite positioning, Wi-Fi fingerprint positioning, Bluetooth positioning, and inertial positioning. Each of these has its own advantages and disadvantages. UWB positioning offers high accuracy but has high deployment costs and its accuracy is affected by complex environments; 5G positioning has wide coverage and requires no additional equipment, but its accuracy is poor in areas with sparse base stations; fingerprint positioning requires maintenance; inertial positioning offers strong autonomy and high short-term accuracy, but suffers from cumulative errors and low long-term positioning accuracy.

[0003] To address the problems associated with single positioning methods, many researchers have considered multi-source fusion positioning. Some positioning methods are complementary; for example, 5G positioning is susceptible to environmental influences but does not have cumulative errors, while inertial positioning offers high accuracy in the short term but suffers from cumulative errors, leading to significant errors over long periods and requiring timely correction. Due to this complementary nature, numerous algorithms exist for the fusion of 5G and IMU (Inertial Measurement Unit). Commonly used algorithms include Kalman filtering-based fusion algorithms and their variants, such as extended Kalman filtering, unscented Kalman filtering, and particle filtering.

[0004] 5G positioning can be affected when the target is in a complex environment. For example, if the target is obstructed by an object, non-line-of-sight errors can occur between the direct path of the base station and the terminal. Multiple reflections before the 5G signal is received can also cause multipath effects. These factors can reduce 5G positioning accuracy and may even lead to a complete loss of the 5G positioning signal. Therefore, we propose an inertial navigation compensation method for situations where the 5G signal is missing. Summary of the Invention

[0005] Based on the technical problems existing in the background technology, this invention proposes an inertial navigation compensation method for 5G signal loss. It uses the features in the inertial navigation data to track and predict 5G signals in real time, ensuring the continuity and robustness of the positioning process, improving the positioning accuracy in complex scenarios, and solving the problem in the prior art that when the positioning target is in a relatively complex environment, object occlusion will reduce the 5G positioning accuracy, or even cause the 5G positioning signal to be completely lost.

[0006] This invention provides the following technical solution: an inertial navigation compensation method for 5G signal loss, comprising the following steps:

[0007] The inertial measurement unit (IMU) collects raw inertial data, including data from the three-axis accelerometer and the three-axis gyroscope. The inertial navigation unit (INS) uses the data provided by the IMU to calculate the position, velocity and attitude of the object through integration. The 5G data and the data calculated by the IMU are fused and optimized by Kalman filtering. The optimized data and the 5G position increment information are used as the input and output of the AT-GRU algorithm to achieve model training and prediction.

[0008] When the 5G positioning is good, it is in training mode. At this time, inertial navigation data and 5G positioning results are collected. The inertial navigation acceleration, angular velocity, and the calculated velocity and attitude angle are used as the model input, and the position increment of 5G in the n-frame at the corresponding time is used as the model output to train the network.

[0009] When 5G positioning results are missing due to environmental reasons, a trained model is used for prediction. The input is inertial navigation acceleration, angular velocity, and calculated velocity. The output is the predicted position increment at time k in the n-frame, which is then added to the position at time k-1 to obtain the predicted position at time k and participate in system fusion.

[0010] Preferably, the self-attention mechanism in the AT-GRU algorithm is used to enable the model to dynamically pay attention to all other elements in the sequence when processing each element, and to assign different weights according to the correlation between them.

[0011] Preferably, inertial navigation compensation uses O INS -ΔP 5G The model takes data from inertial navigation as input and outputs the 5G position increment in the station-centric coordinate system.

[0012]

[0013] The input and output of the prediction model are determined based on the principles of inertial navigation positioning and 5G positioning.

[0014]

[0015] Among them, λ, h represents latitude, longitude, and altitude, v N ,v E ,v D R represents the northward velocity, eastward velocity, and ground velocity. M,k-1 R represents the radius of the meridian calculated using the position at time k-1. M,k-1 / 2 This represents the meridian radius calculated using the latitude at the midpoint of the time interval. Position updates are primarily related to velocity. According to the mechanized velocity update algorithm, the velocity at time k is updated using the following expression: i represents the inertial coordinate system, e represents the geocentric coordinate system, n represents the navigation coordinate system, b represents the body coordinate system, and the value in the upper right corner of the variable represents the projection in that coordinate system. ω ie It is the Earth's rotational angular velocity vector, ω en It is the angular velocity vector of the n-frame relative to the e-frame, and C is the transformation matrix. It is a gravity / Gorbachev integral term. It is a specific force measured by the accelerometer. It is the angular velocity measured by the gyroscope;

[0016] According to the above formula, the position increment of 5G in the n-frame is only related to the inertial navigation acceleration, angular velocity, calculated velocity, and calculated attitude angle. Therefore, the input of the prediction model is the inertial navigation acceleration, angular velocity, calculated velocity, and calculated attitude angle, and the output is the position increment of 5G in the n-frame.

[0017]

[0018] The preferred GRU network structure of the AT-GRU algorithm operates as follows:

[0019] First, the reset gate r1 and update gate Z1 are calculated using the current input x1 and the previous hidden state h1, respectively determining whether to ignore past information and adjust the state update level. Then, candidate hidden states are generated using the historical information filtered by the reset gate and the current input. Finally, update the goalkeeper's previous state h1 and the candidate state. Weighted fusion is performed to output a new hidden state h2.

[0020] Preferably, the specific calculation formula for the GRU structure is as follows:

[0021] Z t =σ(W z ·[h t-1 x t ]+b z );

[0022] r t =σ(W r ·[h t-1 ,xt ]+b r );

[0023]

[0024] Where W is the weight matrix, b is the bias term, σ is the sigmoid function with an output value between 0 and 1, θ is the convolution, and Z is the update gate. t Control the hidden state h at the current moment t In the middle, the hidden state h in the previous time step t-1 With the current candidate hidden state The weight ratio, when Z t As Z approaches 1, the model relies more on historical information; when Z... t When the value is close to 0, the model focuses more on the current input and resets the gate r. t This determines the degree of influence of the previous hidden state on the current candidate hidden state. If r t When r approaches 0, the model resets historical information and generates a new state based solely on the current input; if r... t A value closer to 1 retains more historical context.

[0025] This invention provides an inertial navigation compensation method for 5G signal loss scenarios. For the first time in a fusion architecture, it applies an AT-GRU network with a self-attention mechanism to position increment prediction compensation in 5G signal loss situations, thereby achieving high-precision and robust continuous positioning capabilities even with signal loss. This not only improves the overall system performance but also overcomes the strong dependence of existing fusion methods on wireless signal continuity. Attached Figure Description

[0026] Figure 1 This is a diagram of the positioning structure of the present invention.

[0027] Figure 2 This is a plan view of the 5G positioning system of the present invention.

[0028] Figure 3 This is a flowchart of the inertial navigation compensation method based on the AT-GRU algorithm of the present invention.

[0029] Figure 4 This is the GRU network structure of the present invention.

[0030] Figure 5 This is a CDF diagram showing the positioning results of this invention.

[0031] Figure 6 This is a graph showing the relationship between the positioning error and time in this invention. Detailed Implementation

[0032] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] like Figure 1 As shown, this invention provides a technical solution: an inertial navigation compensation method for 5G signal loss, and derives the model's input and output based on the positioning principles of 5G positioning and IMU positioning. This method uses data from the inertial navigation system as input and the 5G position increment in the station-centric coordinate system as output. A mapping relationship between input and output is established through a training mode, and the position increment at a certain moment is predicted through a prediction mode, participating in system fusion to complete the inertial navigation compensation positioning work under 5G signal loss. Experimental verification shows that this method has certain advantages.

[0034] Related work in this approach includes a 5G and IMU fusion algorithm based on loosely coupled Kalman filtering, as well as a GRU prediction network and a self-attention mechanism. The Kalman filter algorithm, as an important optimal estimation theory, is widely used in various fields, particularly in integrated navigation systems, where it can achieve higher-precision positioning solutions. The adoption of GRU theory significantly reduces the number of parameters and improves training efficiency while retaining dependency modeling capabilities, making it particularly suitable for time series prediction tasks. The core idea of ​​the self-attention mechanism is to allow the model to dynamically pay attention to all other elements in the sequence when processing each element, and to assign different weights based on the correlation between them. In this invention, 5G stands for 5th Generation Mobile Communication Technology, IMU stands for Inertial Measurement Unit, AT stands for Self-Attention Mechanism, GRU stands for Gate Recurrent Unit, INS stands for Inertial Navigation System, TDOA stands for Time Difference of Arrival, and TOA stands for Time of Arrival.

[0035] like Figure 3 As shown, the specific steps are as follows:

[0036] The IMU (Inertial Measurement Unit) is responsible for collecting raw inertial data, including data from the three-axis accelerometer and three-axis gyroscope. The INS (Inertial Measurement System) uses the data provided by the IMU to calculate the object's position, velocity, and attitude (heading, pitch, roll) through integration. However, because integration errors accumulate over time, the accuracy of the INS gradually decreases. Therefore, the 5G data and the data calculated by the INS are fused and optimized using Kalman filtering to reduce errors. The optimized data and the 5G position increment information are used as the input and output of the AT-GRU algorithm, respectively, to train and predict the model.

[0037] Based on this architecture, the system can flexibly switch between two operating modes: training mode and prediction mode. When 5G positioning is good, it operates in training mode. In this mode, inertial navigation data and 5G positioning results are collected. The inertial navigation acceleration, angular velocity, and calculated velocity and attitude angle are used as model inputs, and the position increment of 5G at the corresponding time in the n-frame is used as the model output to train the network. When 5G positioning results are missing due to environmental factors, the trained model is used for prediction. The inputs are inertial navigation acceleration, angular velocity, and calculated velocity, and the output is the predicted position increment at time k in the n-frame. This is then added to the position at time k-1 to obtain the predicted position at time k, which is then used in system fusion.

[0038] The principle and specific workflow of the AT-GRU algorithm are as follows:

[0039] like Figure 4 The diagram shows the GRU cell structure. The main variables include the state information h from the previous time step. t-1 Input x t and current status information h t It contains only two gates: the vector Z corresponding to the update gate. t and the vector r of the corresponding reset gate t The update gate controls the influence of the previous state on the current state, while the reset gate controls the degree to which information from the previous state is ignored. A smaller reset gate value results in greater ignoring of information. σ represents the sigmoid function, and tanh is the activation function.

[0040] The specific operation process of the network is as follows: First, the reset gate r1 and update gate Z1 are calculated based on the current input x1 and the previous hidden state h1, which respectively determine whether to ignore past information and adjust the state update degree; then, candidate hidden states are generated using the historical information filtered by the reset gate and the current input. Finally, update the goalkeeper's previous state h1 and the candidate state. Weighted fusion is performed to output a new hidden state h2. The GRU network achieves dynamic control over "preserving history and updating the current" in time-series data through a gating mechanism. It can learn the temporal dependencies in the input sequence and output a temporal state representation that is meaningful for subsequent tasks.

[0041] The specific calculation formula for the GRU structure can be expressed as follows:

[0042] Z t =σ(W z ·[h t-1 ,x t ]+b z );

[0043] r t =σ(W r ·[h t-1 ,x t ]+b r );

[0044]

[0045]

[0046] Where W is the weight matrix, b is the bias term, σ is the sigmoid function with an output value between 0 and 1, and θ is the convolution. Update gate (Z) t Control the hidden state h at the current moment t In the middle, the hidden state h in the previous time step t-1 With the current candidate hidden state The weighting ratio. When Z t As Z approaches 1, the model relies more on historical information; when Z... t When the value is close to 0, the model focuses more on the current input. Reset gate (r) t Then it determines the degree of influence of the previous hidden state on the current candidate hidden state. If r t When r approaches 0, the model "resets" historical information and generates a new state based solely on the current input; if r is close to 0... t A value close to 1 retains more historical context. Through these two gating mechanisms, GRU can dynamically adjust the ratio of information retention to forgetting during training, thereby more efficiently capturing long-term dependencies in the sequence. GRU achieves dynamic control of information flow through update and reset gates, solving the gradient vanishing problem of traditional RNNs while simplifying the complex structure of LSTM. In time series processing, each forward propagation step of GRU involves the computation of reset gate, candidate hidden state, and update gate, ultimately generating the hidden state at the current time step.

[0047] Self-attention is a deep learning technique for processing sequential data. Its core idea is to allow the model to dynamically pay attention to all other elements in the sequence while processing each element, assigning different weights based on their correlation. Specifically, self-attention achieves this by calculating a "relevance score" (or attention score) between each element and other elements. These scores are obtained by representing each element as three vectors: query, key, and value, and then performing a dot product operation on these vectors. To make the scores more interpretable and avoid excessively large values, the dot product results are typically scaled (e.g., divided by the square root of the key vector's dimension). Then, the softmax function is used to convert these scores into a probability distribution, i.e., attention weights. Finally, the model performs a weighted summation of the value vectors based on these weights, resulting in a new representation for each element that incorporates information from other elements in the sequence, with different elements contributing differently to the current element. The formula is shown below:

[0048]

[0049] The hidden state update of the GRU network relies on a fixed time step, which may lead to insufficient information transmission. Furthermore, the GRU network only focuses on the current input and the hidden state from the previous time step at each step, making it difficult to directly capture dependencies between long-distance elements in the sequence. By incorporating a self-attention mechanism, the GRU network can enhance its global dependency modeling ability, enabling the model to more directly capture long-distance dependencies. Therefore, this invention incorporates a self-attention mechanism into the GRU network as a prediction network.

[0050] This invention, for the first time in a fusion architecture, applies an AT-GRU network with a self-attention mechanism to location increment prediction compensation in 5G signal loss scenarios, thereby achieving continuous positioning capabilities with high accuracy and robustness even in signal loss. This innovation not only improves the overall system performance but also overcomes the strong dependence of existing fusion methods on wireless signal continuity, demonstrating significant technological progress and innovation.

[0051] Example:

[0052] Data acquisition: In order to implement the method provided by this invention, inertial navigation data and 5G positioning results are acquired in two different environmental scenarios. The inertial navigation data mainly includes inertial acceleration, angular velocity, and calculated velocity and attitude angle. The 5G positioning results used are specifically the position increment of 5G in the corresponding coordinate system.

[0053] Scenario 1 is an open road environment with no objects significantly obstructing or reflecting the signal. Non-line-of-sight errors and multipath effects are minimal, resulting in high 5G positioning accuracy with few outliers. The test equipment is placed on a handcart, which is then pushed by the experimenter. Due to the slow speed, the 5G ranging device's sampling frequency is 1Hz to collect relevant data.

[0054] Scenario 2 is a simulated tunnel environment. Due to the narrowness of the space and the reflection of signals from the walls, there may be significant multipath effects, which would greatly affect the accuracy of 5G positioning and result in positioning anomalies. Because of the increased vehicle speed, the 5G equipment sampling frequency is set to 5Hz and it is installed on pillars on both sides of the tunnel.

[0055] Network training and prediction involve collecting relevant data at one point in the scene for network training, performing predictions over a continuous 5-second time interval, and comparing the error between the predicted and actual results. The root mean square error formula below is used to calculate the prediction error.

[0056]

[0057] In Scenario 2, the network is first trained using high-performing test data. Two seconds are taken as the 5G missing time, and different algorithms are used for prediction, which is then fused with the inertial navigation system. The prediction error is calculated in the same way as above.

[0058] like Figure 2 As shown, the site is approximately 250m long, with a pair of 5G base stations installed on pillars on both sides every 25m. The test equipment was placed on a test vehicle, and personnel drove along the centerline, keeping the horizontal coordinate in the 5G relative coordinate system stationary. The maximum speed of the test vehicle was 20km / h. Calibration was performed using total stations and laser rangefinders to provide accurate reference values ​​for the positioning results.

[0059] The derivation of the inertial navigation compensation principle first requires determining the input and output of the prediction model. Existing research commonly uses models such as O... INS -ΔP GNSS-INS Model and O INS -ΔP GNSS Model. O INS -ΔP GNSS-INS The model takes inertial navigation data as input and the error between the GNSS positioning position and the inertial navigation calculated position as output. INS -ΔP GNSS The model takes inertial navigation data as input and GNSS positioning position increment as output.

[0060] ΔP GNSS-INS =P GNSS -P INS ;

[0061]

[0062] Among them, O INS -ΔP GNSS The model error originates solely from GNSS, while O INS -ΔP GNSS-INS The model's errors originate from GNSS and INS, with the latter causing a larger error. Analogous to this model, this invention employs O... INS -ΔP 5G The model takes data from inertial navigation as input and outputs the 5G position increment in the station-centric coordinate system.

[0063]

[0064] The input and output of the prediction model are determined based on the principles of inertial navigation positioning and 5G positioning.

[0065]

[0066] Among them, λ, h represents latitude, longitude, and altitude, v N ,v E ,v D R represents the northward velocity, eastward velocity, and ground velocity. M,k-1 R represents the radius of the meridian calculated using the position at time k-1. M,k-1 / 2 This represents the meridian radius calculated using the latitude at the midpoint of the time interval. Position updates are primarily related to velocity. According to the mechanized velocity update algorithm, the velocity at time k is updated using the following expression: i represents the inertial coordinate system, e represents the geocentric coordinate system, n represents the navigation coordinate system, b represents the body coordinate system, and the value in the upper right corner of the variable represents the projection in that coordinate system. ω ie It is the Earth's rotational angular velocity vector, ω en It is the angular velocity vector of the n-frame relative to the e-frame, and C is the transformation matrix. It is a gravity / Gorbachev integral term. It is a specific force measured by the accelerometer. It is the angular velocity measured by the gyroscope.

[0067] According to the above formula, the position increment of 5G in the n-frame is only related to the inertial navigation acceleration, angular velocity, calculated velocity, and calculated attitude angle. Therefore, the input of the prediction model is the inertial navigation acceleration, angular velocity, calculated velocity, and calculated attitude angle, and the output is the position increment of 5G in the n-frame.

[0068]

[0069] Experimental results verify:

[0070] 1) Scene 1: Open road environment

[0071] The experimental results, comparing the proposed algorithm with other algorithms, are shown below:

[0072] Table 1 Comparison of prediction algorithm errors

[0073] Prediction Algorithm Name Prediction error (m) LSTM 0.31 GRU 0.25 AT-GRU 0.22

[0074] As shown in the table above, the proposed AT-GRU algorithm, as a prediction network, achieves a 29.03% improvement over the LSTM network and a 12% improvement over the GRU algorithm. Experimental results demonstrate that the proposed AT-GRU algorithm has better prediction accuracy than other networks, can provide better constraint information when 5G signals are missing, and suppresses the accumulation of errors in inertial navigation.

[0075] 2) Scenario 2: Simulated tunnel scenario, the positioning results are as follows Figure 5 and 6 As shown.

[0076] Table 2. Root Mean Square Error of Positioning

[0077]

[0078]

[0079] The positioning results above show that the proposed algorithm achieves a 42.1% improvement compared to pure inertial navigation estimation, a 31.25% improvement compared to LSTM-based prediction, and a 26.7% improvement compared to GRU prediction. This demonstrates that the proposed algorithm can provide constraints when 5G is missing, preventing positioning errors from accumulating over time. Combining the CDF plot and the error plot at each time step, it is evident that the proposed algorithm has stronger predictive capabilities than other prediction algorithms, thus resulting in better positioning accuracy after fusion.

[0080] In multi-source fusion positioning, algorithms for 5G and IMU fusion are quite common, but research on predictive compensation using inertial navigation for missing 5G signals is still lacking. Since there is a wealth of research on GNSS signal loss during GNSS and inertial navigation fusion, this invention draws on that research and combines it with 5G and IMU fusion algorithms for improvement.

[0081] For situations where GNSS signals are missing, some literature employs the following methods: GNSS / INS integrated navigation methods based on hybrid neural networks (CNN and GRU) and improved robust adaptive Kalman filtering (IRAKF); Long Short-Term Memory (LSTM) neural networks with integrated attention mechanism (AT-LSTM) to assist in error compensation of GNSS / INS integrated navigation systems; and GNSS pseudorange error prediction methods based on machine learning, which combine residual information from the inertial navigation system (INS) with other GNSS observation parameters, such as carrier-to-noise ratio (C / N0), satellite elevation angle, and pseudorange rate consistency, for model training.

[0082] Because the signal characteristics of 5G and inertial navigation data differ from those of GNSS scenarios, the applicability and advantages / disadvantages of the aforementioned methods for inertial navigation compensation in the absence of 5G signals need to be re-evaluated. Simultaneously, a balance needs to be struck between computational efficiency and prediction accuracy, ensuring the real-time performance of the algorithm on embedded devices, and improving generalization through diverse training data and online adaptive capabilities. Finally, experimental verification is required to test the system's robustness under different interruption durations and dynamic scenarios.

[0083] The prediction and compensation methods mentioned above are all implemented in GNSS scenarios, and have some drawbacks when 5G signals are missing. Firstly, model adaptability needs to be considered, mainly referring to the mismatch between model structure and data type. 5G data (such as TOA / TDOA measurements) is usually time series rather than image or spatial data. The spatial feature extraction capabilities of CNNs may be redundant in this scenario, leading to wasted computational resources and degraded model performance. Secondly, there is the issue of handling abrupt changes during signal recovery. During 5G signal recovery, there may be discontinuities between the pseudo-measurements generated by the model and the actual 5G observations, causing drastic fluctuations in the Kalman filter state updates, requiring additional smoothing transition mechanisms. Furthermore, for GRU networks, their hidden state updates depend on a fixed time step, which may lead to insufficient information transmission. Also, GRU networks only focus on the current input and the hidden state of the previous time step at each step, making it difficult to directly capture the dependencies between distant elements in the sequence.

[0084] The 5G signal itself presents a complex issue of multipath interference. In indoor environments, 5G signals are more susceptible to multipath effects, and signal interruptions may be accompanied by sudden noise or abnormal data. Therefore, system design must be tailored to handle such unforeseen circumstances. Consequently, when determining the model conditions for inertial navigation signal compensation, adaptability and computational efficiency must be comprehensively considered, and appropriate fusion models and algorithms must be selected. Simultaneously, the accuracy and generalization ability of the model must be improved to achieve more accurate high-precision fusion positioning of 5G and IMU.

[0085] Furthermore, most existing 5G+IMU fusion positioning methods typically employ extended Kalman filtering (EKF), unscented Kalman filtering (UKF), or particle filtering as fusion techniques. These methods all rely on the continuous availability of 5G positioning information, therefore their main drawbacks lie in:

[0086] 1) Unable to cope with prolonged 5G signal interruptions. Once the 5G reference signal is blocked, interfered with, or the scheduling is interrupted, the fusion system degrades to pure IMU navigation, and the IMU has inherent cumulative errors, resulting in a linear or exponential increase in positioning errors.

[0087] 2) No feedforward prediction mechanism is introduced for missing signal compensation. Traditional methods rely solely on IMU inertial navigation calculations during signal loss periods, without a model-assisted prediction mechanism to constrain the process. Consequently, they lack contextual information for modeling and cannot predict the position and state at future moments.

[0088] 3) The model structure is too simple and cannot adaptively adjust the signal confidence level. Most algorithms use a fixed weight fusion strategy, which cannot dynamically adjust the weight ratio of IMU and 5G in weak 5G or non-line-of-sight (NLOS) error scenarios.

[0089] The proposed AT-GRU algorithm-based predictive compensation network localization method demonstrates significant advantages in localization accuracy, environmental adaptability, and system robustness compared to existing tightly coupled localization techniques based on LSSVM (Least Squares Support Vector Machine) and ESKF (Extended Kalman Filter). While traditional LSSVM schemes achieve 5G and IMU data fusion, they rely solely on ESKF inertial estimation when 5G signals are missing, leading to error accumulation and a lack of active compensation mechanisms, resulting in a sharp decline in localization performance. In contrast, this invention innovatively introduces the AT-GRU deep learning model, dynamically capturing motion features through an attention mechanism to provide high-precision predictive compensation when 5G signals are interrupted. This effectively solves the information interruption problem, overcomes the aforementioned shortcomings, and fills the gap in the field of compensation using inertial navigation information in the absence of 5G signals. Furthermore, the superiority of the algorithm is verified by selecting equipment with higher sampling frequencies as the complexity of the test scenarios increases.

[0090] This invention first theoretically derives the input-output relationship of the model, proving the feasibility of the algorithm. The overall logic of the algorithm is then detailed. Finally, experiments demonstrate that the proposed algorithm can still promptly correct inertial navigation (INS) prediction results even in the absence of 5G signal. Compared with other prediction networks, the algorithm proposed in this invention exhibits better positioning accuracy. In complex positioning environments, existing 5G signals used for positioning may be lost due to obstructions or multipath effects. Without timely compensation for this signal, high-precision positioning requirements cannot be met. This invention utilizes features from INS data to track and predict 5G signals in real time, ensuring the continuity and robustness of the positioning process and improving positioning accuracy in complex scenarios.

[0091] This invention derives the model's input and output based on the positioning principles of 5G and IMU positioning, enriching the content of this research direction. A self-attention mechanism is added to the GRU network to improve its predictive ability. Comparative experiments with other prediction algorithms and unpredictable algorithms verify that the proposed algorithm has better positioning accuracy.

[0092] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An inertial navigation compensation method for 5G signal loss, characterized in that: Includes the following steps: The inertial measurement unit (IMU) collects raw inertial data, including data from the three-axis accelerometer and the three-axis gyroscope. The inertial navigation unit (INS) uses the data provided by the IMU to calculate the position, velocity and attitude of the object through integration. The 5G data and the data calculated by the IMU are fused and optimized by Kalman filtering. The optimized data and the 5G position increment information are used as the input and output of the AT-GRU algorithm to achieve model training and prediction. When the 5G positioning is good, it is in training mode. At this time, inertial navigation data and 5G positioning results are collected. The inertial navigation acceleration, angular velocity, and the calculated velocity and attitude angle are used as the model input, and the position increment of 5G in the n-frame at the corresponding time is used as the model output to train the network. When 5G positioning results are missing due to environmental reasons, a trained model is used for prediction. The input is inertial navigation acceleration, angular velocity, and calculated velocity. The output is the predicted position increment at time k in the n-frame, which is then added to the position at time k-1 to obtain the predicted position at time k and participate in system fusion.

2. The inertial navigation compensation method for 5G signal loss as described in claim 1, characterized in that: The self-attention mechanism in the AT-GRU algorithm allows the model to dynamically pay attention to all other elements in the sequence while processing each element, and to assign different weights based on the correlation between them.

3. The inertial navigation compensation method for 5G signal loss as described in claim 1, characterized in that: Inertial navigation compensation uses O INS -ΔP 5G The model takes data from inertial navigation as input and outputs the 5G position increment in the station-centric coordinate system. The input and output of the prediction model are determined based on the principles of inertial navigation positioning and 5G positioning. Among them, λ, h represents latitude, longitude, and altitude, v N ,v E ,v D R represents the northward velocity, eastward velocity, and ground velocity. M,k-1 R represents the radius of the meridian calculated using the position at time k-1. M,k-1 / 2 This represents the meridian radius calculated using the latitude at the midpoint of the time interval. Position updates are primarily related to velocity. According to the mechanized velocity update algorithm, the velocity at time k is updated using the following expression: i represents the inertial coordinate system, e represents the geocentric coordinate system, n represents the navigation coordinate system, b represents the body coordinate system, and the value in the upper right corner of the variable represents the projection in that coordinate system. ω ie It is the Earth's rotational angular velocity vector, ω en It is the angular velocity vector of the n-frame relative to the e-frame, and C is the transformation matrix. It is a gravity / Gorbachev integral term. It is a specific force measured by the accelerometer. It is the angular velocity measured by the gyroscope; According to the above formula, the position increment of 5G in the n-frame is only related to the inertial navigation acceleration, angular velocity, calculated velocity, and calculated attitude angle. Therefore, the input of the prediction model is the inertial navigation acceleration, angular velocity, calculated velocity, and calculated attitude angle, and the output is the position increment of 5G in the n-frame.

4. The inertial navigation compensation method for 5G signal loss as described in claim 1, characterized in that: The GRU network structure and operation process of the AT-GRU algorithm are as follows: First, the reset gate r1 and update gate Z1 are calculated using the current input x1 and the previous hidden state h1, respectively determining whether to ignore past information and adjust the state update level. Then, candidate hidden states are generated using the historical information filtered by the reset gate and the current input. Finally, update the goalkeeper's previous state h1 and the candidate state. Weighted fusion is performed to output a new hidden state h2.

5. The inertial navigation compensation method for 5G signal loss as described in claim 4, characterized in that: The specific calculation formula for the GRU structure is as follows: Z t =σ(W z ·[h t-1 ,x t ]+b z ); r t =σ(W r ·[h t-1 ,x t ]+b r ); Where W is the weight matrix, b is the bias term, σ is the sigmoid function with an output value between 0 and 1, θ is the convolution, and Z is the update gate. t Control the hidden state h at the current moment t In the middle, the hidden state h in the previous time step t-1 With the current candidate hidden state The weight ratio, when Z t As Z approaches 1, the model relies more on historical information; when Z... t When the value is close to 0, the model focuses more on the current input and resets the gate r. t This determines the degree of influence of the previous hidden state on the current candidate hidden state. If r t When r approaches 0, the model resets historical information and generates a new state based solely on the current input; if r... t If the value is close to 1, the historical context is preserved.

Images