An AI-based geographic information precise navigation method and system

Abstract

This invention discloses an AI-based geographic information precision navigation method and system, relating to the technical field of satellite sensing measurement. It establishes a sensor database; based on a hierarchical LSTM structure, it performs branching processing on the temporal and spatial feature sets of the sliding selection, constructing a dynamic time-delay observation correction model. Spatiotemporal features are fused into the dynamic time-delay observation correction model and physical constraints are embedded; the fusion algorithm maintains positioning accuracy and optimizes navigation and positioning in extreme weather. Employing a four-level redundant tight coupling of satellite navigation GNSS, ground IMU, vision, and ground sensors, the dynamic time-delay observation correction model is fused with satellite navigation GNSS to solve the long-term drift problem of ground IMU under extreme weather conditions. Deep integration with the physical model and tightly coupled navigation algorithm achieves sub-meter real-time positioning accuracy under extreme weather conditions. This realizes deep collaboration of multi-source data, dynamic environment adaptive modeling, and computational resource optimization, providing a systematic solution for high-precision positioning under extreme weather conditions.

Description

Technical Field

[0001] This invention relates to the technical field of satellite sensing and measurement, and in particular to an AI-based method and system for precise navigation using geographic information. Background Technology

[0002] With the widespread adoption of smartphones and mobile devices, and the development of artificial intelligence (AI) technology, AI-based geographic information-based precise navigation methods and systems have become a cutting-edge development direction in modern navigation technology. Traditional navigation systems mainly rely on the Global Positioning System (GPS) and inertial navigation systems (INS). While these systems provide basic positioning services, their accuracy and reliability are often significantly affected in certain environments, such as densely populated areas with tall buildings, underground environments, severe weather conditions, and areas where GPS signals are interfered with. AI technologies, especially deep learning, machine learning, image recognition, sensor fusion, and big data analytics, can significantly improve the accuracy, robustness, and real-time performance of geographic information navigation systems. AI-based navigation methods, by integrating multi-source data and combining intelligent algorithms, can effectively overcome the shortcomings of traditional navigation systems and provide more accurate, real-time, and reliable navigation services.

[0003] Currently, Chinese invention patent application number CN202211648354.8 discloses a precise navigation method and system. Based on user needs and actual conditions, it generates precise navigation path information to provide parking navigation services. It also generates detailed and precise 3D navigation information based on 3D data, based on the user's actual destination location, providing precise destination navigation and avoiding issues caused by poor signal. It adapts to the actual environment and user needs, thus achieving efficient and accurate navigation. However, existing technologies use single satellite navigation, static navigation models, and fixed-parameter ground correction, and cannot improve navigation accuracy in extreme weather. They also cannot perform deep collaboration of multi-source data, dynamic environment adaptive modeling, or computational resource optimization, and therefore cannot provide a systematic solution for high-precision positioning in extreme weather conditions. Summary of the Invention

[0004] The technical problem solved by this invention is that existing technologies use a single satellite navigation system, a static navigation model, and fixed parameter ground correction, which cannot improve navigation accuracy in extreme weather, cannot perform deep collaboration of multi-source data, dynamic environment adaptive modeling and computing resource optimization, and cannot provide a systematic solution for high-precision positioning under extreme weather conditions.

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: an AI-based method for precise geographic information navigation, comprising the following steps:

[0006] Step S1: Acquire data through sensors and establish a sensor database;

[0007] Step S2: Based on the sensor database, the time feature set and spatial feature set of the sliding selection are branched based on the hierarchical LSTM structure to construct a dynamic time delay observation correction model, and the spatiotemporal features are fused into the dynamic time delay observation correction model and physical constraints are embedded.

[0008] Step S3: Maintain positioning accuracy by applying a fusion algorithm to the dynamic time delay observation correction model and optimize navigation and positioning in extreme weather.

[0009] Preferably, the sensors include a satellite navigation GNSS, a fiber optic grating (FBG) strain sensor, a ground-based IMU, and a vision system. A low-cost dual-antenna satellite navigation GNSS receiver is used to acquire the signal strength and raw observation values ​​in the L1 or L2 band in real time. A fiber optic FBG strain sensor network is deployed to monitor micro-deformation of the ground surface. Continuous operating data from the ground-based IMU is collected. The continuous operating data includes data from accelerometers, gyroscopes, and temperature sensors, including drift characteristic data under normal conditions and extreme weather conditions. The extreme weather conditions include heavy rain and high temperatures. The ground-based IMU is integrated as an inertial measurement unit to track the dynamic attitude of the target object. A binocular vision system is used to acquire high-resolution terrain matching data.

[0010] Maintaining time synchronization and spatial alignment of data acquired through sensors includes:

[0011] By using a unified clock system and data synchronization mechanism, the data obtained from each sensor are accurately aligned in time. The Network Time Protocol (NTP) is used to synchronize the clocks of sensors that do not have the ability to directly receive GNSS satellite navigation signals, giving all sensors a unified trigger signal and unifying the data acquisition time of all sensors to the same point in time.

[0012] Each sensor is assigned a timestamp during data acquisition. The logic for adding the timestamp includes: recording a timestamp for each acquired data point, the timestamps originating from the same clock source, and recording the spatiotemporal reference, accuracy indicators, and real-time operating status of the sensor corresponding to each data point. The spatiotemporal reference includes a reference coordinate system and a timestamp. The accuracy indicators include positioning accuracy and sensor sensitivity. The real-time operating status includes the sensor's configuration information, operating status, health status, and performance indicators.

[0013] The spatial registration algorithm between sensors is used to spatially align the positioning data, visual data, ground IMU data, and FBG sensor data of satellite navigation GNSS. The spatial alignment includes placing the reference coordinates of the positioning data, visual data, ground IMU data, and FBG sensor data of satellite navigation GNSS on the same reference coordinate system.

[0014] Kalman filtering is used to optimize and fuse time-synchronized and spatially aligned data, and the optimized and fused dataset becomes a sensor database.

[0015] Preferably, step S2 includes the following sub-steps:

[0016] Step S21: Using the TOPSIS multi-criteria decision algorithm, select initial correction factors from the sensor database based on data timeliness, spatial resolution and reliability, use a sliding window to filter the initial correction factors to obtain the optimal correction factors, and divide the initial correction factors into a time feature set and a spatial feature set;

[0017] Step S22: Based on the sensor database, calculate the prior dry delay and wet delay using the Saastamoinen model;

[0018] Step S23: Use the ID convolutional layer to extract the local meteorological gradient and construct a dynamic time-delay observation correction model.

[0019] Preferably, step S21 includes the following sub-steps:

[0020] Step S211: Select an initial correction factor from the sensor database using the TOPSIS multi-criteria decision algorithm;

[0021] Step S212: Calculate the time delay and update time interval of the initial correction factor, measure the standard deviation of the data update time interval, and use the standard deviation as an evaluation index of timeliness;

[0022] Step S213: Extract the temporal and spatial feature sets based on the sliding window, and construct the decision matrix, including:

[0023] A time feature set is obtained based on a sliding window with 5-minute intervals and a total of 1 hour, and a spatial feature set is obtained based on a sliding window with 10-minute intervals and a total of 1 hour. The performance of each correction factor on each feature will be used as the row of the decision matrix, and the feature will be used as the column. The feature is a random feature value in the time feature set and the spatial feature set. The time feature set includes data timeliness, and the spatial feature set includes spatial resolution and reliability.

[0024] Step S214: Normalize the decision matrix, and obtain the positive ideal solution and the negative ideal solution based on the normalized decision matrix. The positive ideal solution is the maximum value of each column, and the negative ideal solution is the minimum value of each column.

[0025] Step S215: Calculate the Euclidean distance between each correction factor and the positive ideal solution and the negative ideal solution respectively. Calculate the relative proximity between each correction factor and the positive ideal solution based on the Euclidean distance. Select the correction factor with the largest relative proximity as the optimal correction factor. Discard the numerical nodes corresponding to the features of non-optimal correction factors from the time feature set and the spatial feature set to obtain the final time feature set and spatial feature set.

[0026] Preferably, step S23 includes the following sub-steps:

[0027] Step S231: Extract local meteorological gradients using ID convolutional layers. The local meteorological gradients include the spatial rate of temperature change. The temporal feature set and spatial feature set are then processed by branching based on a hierarchical LSTM structure.

[0028] Step S232: Construct a dynamic time delay observation correction model and perform adversarial training on the dynamic time delay observation correction model using gradient penalty.

[0029] Preferably, step S231 includes:

[0030] The time and spatial feature sets are processed at high frequency intervals through the first layer of the hierarchical LSTM structure to capture turbulent fluctuations. The captured turbulent fluctuations are then input into the second layer of the hierarchical LSTM structure for mid-frequency aggregation processing to extract weather system evolution feature values. The weather system evolution feature values ​​are then input into the third layer of the hierarchical LSTM structure for low-frequency learning, and seasonal pattern features are learned in daily periodic components.

[0031] Preferably, step S232 includes:

[0032] Using the wet delay output by Saastamoinen as a prior value, the residual term of the wet delay is adaptively learned by a neural network. The atmospheric thermodynamic equation is used as a physical loss function. MC-Dropout variational inference is added to the output layer to predict the mean and standard deviation of the wet delay.

[0033] Embedding the wet delay predicted by LSTM into the GNSS pseudorange equation yields the tightly coupled observation equation, the mathematical expression of which is:

[0034] ;

[0035] in, For GNSS pseudorange coordinates, For the distance to receive the signal, At the speed of light, t represents the delay correction error between the sensor and the original coordinate point in the current coordinate system. This delay correction error is the error between the initial and predicted coordinate parameters calculated using the least squares iterative method. For dry delay, For the wet delay predicted by LSTM, For noise;

[0036] The weights of the dynamic time delay observation correction model are set to those calculated using dynamic uncertainty, and their mathematical expression is:

[0037] ;

[0038] in, As weight, The standard deviation of the wet delay, The observation weights for GNSS satellite navigation.

[0039] Preferably, according to step S21, the sliding window is set to a 1-hour sliding window with a 10-minute interval. An initial correction factor is extracted from the continuous operating data and dynamic attitude obtained by the ground IMU in the sensor database to obtain the temperature drift correlation feature and the angular velocity integral error accumulation rate feature.

[0040] The Euler dynamics equations are added as regularization terms to the loss function of the dynamic time-delay observation correction model.

[0041] The difference between the predicted values ​​of the output temperature drift correlation feature and the cumulative rate of angular velocity integral error feature and the actual temperature drift correlation feature and the cumulative rate of angular velocity integral error feature calculated per unit time is used as the prior error to establish the ground IMU error state vector. The ground IMU error state vector is obtained by convolving the prior error with the state transition matrix. The state transition matrix is ​​derived from the kinematics of the ground IMU. The kinematics derivation formulas of the ground IMU include the position update formula, the velocity update formula, and the rotation matrix update formula.

[0042] Preferably, when data is acquired based on extreme weather, the reciprocal of the negative indicators in the sensor database is taken to normalize the data in the sensor database, and the L1 or L2 band signal strength and original observation values ​​of the satellite navigation GNSS output are adjusted in real time to construct an evaluation matrix. The index function of the evaluation matrix includes GNSS signal-to-noise ratio, visual feature matching rate and ground IMU confidence.

[0043] The index weights of the evaluation matrix are calculated based on the RSR method. The optimal correction factor is obtained through step S21. The optimal correction factor is weighted and added to the index weights to obtain the fusion weights of each sensor. The fusion weights are used as the extreme weather weight combination of the dynamic time-delay observation correction model. The dynamic time-delay observation correction model is iteratively trained based on the real-time updated sensor database to obtain the optimal extreme weather weight combination. The historical data output by the ground IMU is analyzed using a hierarchical LSTM model to identify abnormal data. The abnormal data is used as drift pattern data and input into the dynamic time-delay observation correction model. The dynamic equation is used as the discrete loss function for real-time data correction. Error compensation is achieved through a Kalman filter.

[0044] An AI-based geographic information precision navigation system includes a sensor module, a model building module, and an extreme optimization module.

[0045] The sensor module includes acquiring data through sensors and establishing a sensor database;

[0046] The model building module includes, based on the sensor database, performing branching processing on the sliding selection time feature set and spatial feature set based on the hierarchical LSTM structure, constructing a dynamic time delay observation correction model, integrating spatiotemporal features into the dynamic time delay observation correction model and embedding physical constraints.

[0047] The extreme optimization module includes maintaining positioning accuracy by using a fusion algorithm on the dynamic time-delay observation correction model, and optimizing navigation and positioning in extreme weather.

[0048] The beneficial effects of this invention are as follows: It employs a four-level redundant tight coupling of satellite navigation GNSS, ground IMU, vision, and ground sensors. A hierarchical LSTM structure dynamically corrects Saastamoinen (the system's latency), adapting to sudden weather changes. By embedding physical constraints through sliding window branching features, a dynamic time-delay observation correction model is constructed. This model includes more extreme weather samples in the training data, improving its generalization ability. It can also predict wet delay and other factors affecting positioning, such as ionospheric delay, thereby improving overall navigation accuracy. LSTM corrects the residuals of traditional models rather than completely replacing them, combining the advantages of both. By integrating the dynamic time-delay observation correction model with satellite navigation GNSS, the long-term drift problem of ground IMUs under extreme weather conditions is solved. Deep integration with the physical model and tightly coupled navigation algorithm achieves sub-meter real-time positioning accuracy under extreme weather conditions. This enables deep collaboration of multi-source data, dynamic environment adaptive modeling, and computational resource optimization, providing a systematic solution for high-precision positioning under extreme weather conditions. Attached Figure Description

[0049] Figure 1This is a basic flowchart illustrating an AI-based geographic information-based precise navigation method provided in one embodiment of the present invention. Detailed Implementation

[0050] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0051] Reference Figure 1 As an embodiment of the present invention, an AI-based geographic information-based precise navigation method is provided, comprising the following steps:

[0052] Step S1: Acquire data through sensors and establish a sensor database;

[0053] Step S2: Based on the sensor database, the time feature set and spatial feature set of the sliding selection are branched based on the hierarchical LSTM structure to construct a dynamic time delay observation correction model, and the spatiotemporal features are fused into the dynamic time delay observation correction model and physical constraints are embedded.

[0054] Step S3: Maintain positioning accuracy by using a fusion algorithm on the dynamic time delay observation correction model and optimize navigation and positioning in extreme weather.

[0055] The sensors include satellite navigation GNSS, fiber Bragg grating (FBG) strain sensors, ground-based IMUs, and a vision system. A low-cost dual-antenna satellite navigation GNSS receiver is used to acquire signal strength and raw observations in the L1 or L2 bands in real time. A fiber Bragg grating (FBG) strain sensor network is deployed to monitor micro-deformation of the ground surface. Continuous operational data from the ground-based IMU is collected, including data from accelerometers, gyroscopes, and temperature sensors. This data includes drift characteristics under normal conditions and extreme weather conditions, including heavy rain and high temperatures. The ground-based IMU is integrated as an inertial measurement unit to track the dynamic attitude of targets. A binocular vision system is used to acquire high-resolution terrain matching data.

[0056] Maintaining time synchronization and spatial alignment of data acquired through sensors includes:

[0057] By using a unified clock system and data synchronization mechanism, the data obtained from each sensor are accurately aligned in time, avoiding data errors caused by time differences in the acquisition of different sensors. The Network Time Protocol (NTP) is used to synchronize the clocks of sensors that do not have the ability to directly receive satellite navigation GNSS signals. Through network connection, the sensors can periodically obtain the current accurate time from the central time server to ensure that the clocks of each sensor are always synchronized, giving all sensors a unified trigger signal and unifying the data acquisition time of all sensors to the same point in time.

[0058] Each sensor is given a precise timestamp during data acquisition. The logic for adding the timestamp includes: recording a timestamp for each acquired data point, with the timestamps originating from the same clock source, and recording the spatiotemporal reference, accuracy indicators, and real-time operating status of the sensor corresponding to each data point. The spatiotemporal reference includes the reference coordinate system and the timestamp, the accuracy indicators include positioning accuracy and sensor sensitivity, and the real-time operating status includes the sensor's configuration information, operating status, health status, and built-in performance indicators.

[0059] The spatial registration algorithm between sensors is used to spatially align the positioning data, visual data, ground IMU data and FBG sensor data of satellite navigation GNSS. Spatial alignment includes placing the reference coordinates of the positioning data, visual data, ground IMU data and FBG sensor data of satellite navigation GNSS on the same reference coordinate system.

[0060] By using Kalman filtering to optimize and fuse time-synchronized and spatially aligned data, the optimized and fused dataset becomes a sensor database, improving the accuracy and reliability of the data. In the presence of noise or uncertainty, data from multiple sensors can be used to complement each other, thereby obtaining more accurate real-time data.

[0061] Step S2 includes the following sub-steps:

[0062] Step S21: Using the TOPSIS multi-criteria decision algorithm, select initial correction factors from the sensor database based on data timeliness, spatial resolution and reliability. Use a sliding window to filter the initial correction factors to obtain the optimal correction factors. Divide the initial correction factors into a time feature set and a spatial feature set.

[0063] Step S22: Based on the sensor database, calculate the prior dry delay and wet delay using the Saastamoinen model;

[0064] Step S23: Use the ID convolutional layer to extract the local meteorological gradient and construct a dynamic time-delay observation correction model.

[0065] By deeply integrating LSTM with physical models and tightly coupled navigation algorithms, sub-meter level real-time positioning accuracy can be achieved under extreme weather conditions, which is 3-5 times higher than traditional methods, providing reliable positioning support for scenarios such as autonomous driving and disaster emergency response.

[0066] Step S21 includes the following sub-steps:

[0067] Temporal and spatial feature sets are used to capture patterns in data changes over time.

[0068] Step S211: Select initial correction factors from the sensor database using the TOPSIS multi-criteria decision algorithm;

[0069] Step S212: Calculate the time delay and update interval of the initial correction factor, measure the standard deviation of the data update interval, and use the standard deviation as an evaluation index of timeliness;

[0070] Step S213: Extract the temporal and spatial feature sets based on the sliding window, and construct the decision matrix, including:

[0071] The time feature set is obtained based on a sliding window with a 5-minute interval and a total of 1 hour, and the spatial feature set is obtained based on a sliding window with a 10-minute interval and a total of 1 hour. The performance of each correction factor on each feature will be used as the row of the decision matrix, and the feature will be used as the column. The feature is a random feature value in the time feature set and the spatial feature set. The time feature set includes data timeliness, and the spatial feature set includes spatial resolution and reliability.

[0072] The decision matrix is ​​shown in the table below:

[0073] Correction factor Timeliness Spatial resolution reliability Correction factor 1 0.9 0.85 0.8 Correction factor 2 0.8 0.9 0.9 Correction factor 3 0.85 0.8 0.85 ... ... ... ...

[0074] Step S214: Normalize the decision matrix, and obtain the positive ideal solution and negative ideal solution based on the normalized decision matrix. The positive ideal solution is the maximum value of each column, and the negative ideal solution is the minimum value of each column.

[0075] Step S215: Calculate the Euclidean distance between each correction factor and the positive ideal solution and the negative ideal solution respectively. Calculate the relative proximity between each correction factor and the positive ideal solution based on the Euclidean distance. Select the correction factor with the largest relative proximity as the optimal correction factor. Discard the numerical nodes corresponding to the features of non-optimal correction factors from the time feature set and the spatial feature set to obtain the final time feature set and spatial feature set.

[0076] Step S23 includes the following sub-steps:

[0077] Step S231: Extract local meteorological gradients using ID convolutional layers. Local meteorological gradients include the spatial rate of temperature change. The temporal feature set and spatial feature set are branched based on a hierarchical LSTM structure.

[0078] Step S232: Construct a dynamic time-delay observation correction model, and use gradient penalty to perform adversarial training on the dynamic time-delay observation correction model to improve the model's generalization ability when data is missing.

[0079] Step S231 includes:

[0080] The temporal and spatial feature sets are processed at high frequency intervals through the first layer of the hierarchical LSTM structure to capture turbulent fluctuations. The captured turbulent fluctuations are then input into the second layer of the hierarchical LSTM structure for mid-frequency aggregation to extract weather system evolution feature values. Finally, the weather system evolution feature values ​​are input into the third layer of the hierarchical LSTM structure for low-frequency learning, and seasonal model features are learned in daily periodic components.

[0081] Step S232 includes:

[0082] Using the wet delay output by Saastamoinen as a prior value, the residual term of the wet delay is adaptively learned by a neural network. The atmospheric thermodynamic equation is used as a physical loss function. MC-Dropout variational inference is added to the output layer to predict the mean and standard deviation of the wet delay.

[0083] Embedding the wet delay predicted by LSTM into the GNSS pseudorange equation yields the tightly coupled observation equation, the mathematical expression of which is:

[0084] ;

[0085] in, For GNSS pseudorange coordinates, For the distance to receive the signal, At the speed of light, t represents the delay correction error between the sensor and the original coordinate point in the current coordinate system. This delay correction error is the error between the initial and predicted coordinate parameters calculated using the least squares iterative method. For dry delay, For the wet delay predicted by LSTM, For noise;

[0086] The weights of the dynamic time-delay observation correction model are set to those calculated using the dynamic uncertainty formula, and its mathematical expression is:

[0087] ;

[0088] in, As weight, The standard deviation of the wet delay, The observation weights for satellite navigation GNSS are manually set.

[0089] According to step S21, the sliding window is set to a 1-hour sliding window with a 10-minute interval. The initial correction factor is extracted from the continuous operating data and dynamic attitude obtained by the ground IMU in the sensor database to obtain the temperature drift correlation characteristic and the angular velocity integral error accumulation rate characteristic.

[0090] Euler's dynamic equations are added as regularization terms to the loss function of the dynamic time-delay observation correction model to ensure that the predicted values ​​conform to physical laws.

[0091] The difference between the predicted values ​​of the temperature drift correlation feature and the cumulative rate of angular velocity integral error feature and the actual temperature drift correlation feature and the cumulative rate of angular velocity integral error feature calculated per unit time is used as the prior error to establish the ground IMU error state vector. The ground IMU error state vector is obtained by convolving the prior error with the state transition matrix. The state transition matrix is ​​derived from the kinematics of the ground IMU. The kinematics derivation formulas of the ground IMU include the position update formula, the velocity update formula, and the rotation matrix update formula.

[0092] When acquiring data based on extreme weather, the reciprocal of the negative indicators in the sensor database is taken to normalize the data in the sensor database. The signal strength of the L1 or L2 band output of the satellite navigation GNSS and the original observation value are adjusted in real time to avoid the impact of signal error accumulation on positioning accuracy. An evaluation matrix is ​​constructed. The index function of the evaluation matrix includes GNSS signal-to-noise ratio, visual feature matching rate and ground IMU confidence.

[0093] The GNSS signal-to-noise ratio, visual feature matching rate, and ground IMU confidence level are calculated from data in the sensor database using formulas available in the prior art.

[0094] The index weights of the evaluation matrix are calculated based on the RSR method. The optimal correction factor is obtained through step S21. The optimal correction factor is weighted and added to the index weights to obtain the fusion weights of each sensor. The fusion weights are used as the extreme weather weight combination of the dynamic time-delay observation correction model. The dynamic time-delay observation correction model is iteratively trained based on the real-time updated sensor database to obtain the optimal extreme weather weight combination. The historical data output by the ground IMU is analyzed using a hierarchical LSTM model to identify abnormal data. The abnormal data is used as drift mode data. The drift mode data is input into the dynamic time-delay observation correction model. The dynamic equation is used as the discrete loss function for real-time data correction. Error compensation is achieved through a Kalman filter.

[0095] An AI-based geographic information precision navigation system includes a sensor module, a model building module, and an extreme optimization module.

[0096] The sensor module includes acquiring data through sensors and establishing a sensor database;

[0097] The model building module includes, based on the sensor database, performing branching processing on the time feature set and spatial feature set of sliding selection based on a hierarchical LSTM structure, constructing a dynamic time delay observation correction model, integrating spatiotemporal features into the dynamic time delay observation correction model and embedding physical constraints;

[0098] The extreme optimization module includes maintaining positioning accuracy by fusing dynamic time-delay observation correction models with algorithms, and optimizing navigation and positioning in extreme weather.

[0099] This invention employs a four-level redundant tight coupling of satellite navigation GNSS, ground IMU, vision, and ground sensors. It dynamically corrects Saastamoinen (the system's latency) through a hierarchical LSTM structure to adapt to sudden weather changes. By embedding physical constraints through sliding window branching of features, a dynamic time-delay observation correction model is constructed. The training data includes more extreme weather samples, improving the generalization ability of the dynamic time-delay observation correction model. It can also predict wet delay and other factors affecting positioning, such as ionospheric delay, thereby improving overall navigation accuracy. It uses LSTM to correct the residuals of traditional models, rather than completely replacing them, combining the advantages of both. By fusing the dynamic time-delay observation correction model with satellite navigation GNSS, it solves the long-term drift problem of ground IMUs under extreme weather conditions. Deep integration with the physical model and tightly coupled navigation algorithm achieves sub-meter real-time positioning accuracy under extreme weather conditions. It realizes deep collaboration of multi-source data, dynamic environment adaptive modeling, and computational resource optimization, providing a systematic solution for high-precision positioning under extreme weather conditions.

[0100] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0101] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An AI-based method for precise geographic information navigation, characterized in that, Includes the following steps: Step S1: Acquire data through sensors and establish a sensor database; Step S2: Construct a dynamic time-delay observation correction model; Step S2 includes the following sub-steps: Step S21: Using the TOPSIS multi-criteria decision algorithm, select initial correction factors from the sensor database based on data timeliness, spatial resolution and reliability, use a sliding window to filter the initial correction factors to obtain the optimal correction factors, and divide the initial correction factors into a time feature set and a spatial feature set; Step S22: Based on the sensor database, calculate the prior dry delay and wet delay using the Saastamoinen model; Step S23: Extract local meteorological gradients using ID convolutional layers and construct a dynamic time-delay observation correction model; Step S23 includes the following sub-steps: Step S231: Extract local meteorological gradients using ID convolutional layers. The local meteorological gradients include the spatial rate of temperature change. The temporal feature set and spatial feature set are branched based on a hierarchical LSTM structure. Step S232: Construct a dynamic time delay observation correction model and perform adversarial training on the dynamic time delay observation correction model using gradient penalty; Step S231 includes: The temporal and spatial feature sets are processed at high frequency intervals through the first layer of the hierarchical LSTM structure to capture turbulent fluctuations. The captured turbulent fluctuations are then input into the second layer of the hierarchical LSTM structure for mid-frequency aggregation processing to extract weather system evolution feature values. These weather system evolution feature values ​​are then input into the third layer of the hierarchical LSTM structure for low-frequency learning, and seasonal pattern features are learned in daily periodic components. Step S232 includes: Using the wet delay output by Saastamoinen as a prior value, the residual term of the wet delay is adaptively learned by a neural network. The atmospheric thermodynamic equation is used as a physical loss function. MC-Dropout variational inference is added to the output layer to predict the mean and standard deviation of the wet delay. Embedding the wet delay predicted by LSTM into the GNSS pseudorange equation yields the tightly coupled observation equation, the mathematical expression of which is: ； in, For GNSS pseudorange coordinates, For the distance to receive the signal, At the speed of light, t represents the delay correction error between the sensor and the original coordinate point in the current coordinate system. This delay correction error is the error between the initial and predicted coordinate parameters calculated using the least squares iterative method. For dry delay, For the wet delay predicted by LSTM, For noise; The weights of the dynamic time delay observation correction model are set to those calculated using dynamic uncertainty, and their mathematical expression is: ； in, As weight, The standard deviation of the wet delay, Weighting of GNSS observations for satellite navigation; Step S3: Maintain positioning accuracy by applying a fusion algorithm to the dynamic time delay observation correction model, and optimize navigation and positioning in extreme weather. According to step S21, the sliding window is set to a 1-hour sliding window with a 10-minute interval. The initial correction factor is extracted from the continuous operating data and dynamic attitude obtained by the ground IMU in the sensor database to obtain the temperature drift correlation characteristic and the angular velocity integral error accumulation rate characteristic. The Euler dynamics equations are added as regularization terms to the loss function of the dynamic time-delay observation correction model. The difference between the predicted values ​​of the output temperature drift correlation feature and the cumulative rate of angular velocity integral error feature and the actual temperature drift correlation feature and the cumulative rate of angular velocity integral error feature calculated per unit time is used as the prior error to establish the ground IMU error state vector. The ground IMU error state vector is obtained by convolving the prior error with the state transition matrix. The state transition matrix is ​​derived from the kinematics of the ground IMU. The kinematics derivation formulas of the ground IMU include the position update formula, the velocity update formula, and the rotation matrix update formula.

2. The AI-based geographic information-based precise navigation method as described in claim 1, characterized in that: The sensors include a satellite navigation GNSS, a fiber optic grating (FBG) strain sensor, a ground-based IMU, and a vision system. A low-cost dual-antenna satellite navigation GNSS receiver is used to acquire the signal strength and raw observation values ​​in the L1 or L2 band in real time. A fiber optic grating (FBG) strain sensor network is deployed to monitor micro-deformation of the ground surface. Continuous operational data from the ground-based IMU is collected, including data from accelerometers, gyroscopes, and temperature sensors, as well as drift characteristic data under normal conditions and extreme weather conditions, including heavy rain and high temperatures. The ground-based IMU is integrated as an inertial measurement unit to track the dynamic attitude of the target object. A binocular vision system is used to acquire high-resolution terrain matching data. Maintaining time synchronization and spatial alignment of data acquired through sensors includes: By using a unified clock system and data synchronization mechanism, the data obtained from each sensor are accurately aligned in time. The Network Time Protocol (NTP) is used to synchronize the clocks of sensors that do not have the ability to directly receive GNSS satellite navigation signals, giving all sensors a unified trigger signal and unifying the data acquisition time of all sensors to the same point in time. Each sensor is assigned a timestamp during data acquisition. The logic for adding the timestamp includes: recording a timestamp for each acquired data point, the timestamps originating from the same clock source, and recording the spatiotemporal reference, accuracy indicators, and real-time operating status of the sensor corresponding to each data point. The spatiotemporal reference includes a reference coordinate system and a timestamp. The accuracy indicators include positioning accuracy and sensor sensitivity. The real-time operating status includes the sensor's configuration information, operating status, health status, and performance indicators. The spatial registration algorithm between sensors is used to spatially align the positioning data, visual data, ground IMU data, and FBG sensor data of satellite navigation GNSS. The spatial alignment includes placing the reference coordinates of the positioning data, visual data, ground IMU data, and FBG sensor data of satellite navigation GNSS on the same reference coordinate system. Kalman filtering is used to optimize and fuse time-synchronized and spatially aligned data, and the optimized and fused dataset becomes a sensor database.

3. The AI-based geographic information-based precise navigation method as described in claim 2, characterized in that, Step S21 includes the following sub-steps: Step S211: Select an initial correction factor from the sensor database using the TOPSIS multi-criteria decision algorithm; Step S212: Calculate the time delay and update time interval of the initial correction factor, measure the standard deviation of the data update time interval, and use the standard deviation as an evaluation index of timeliness; Step S213: Extract the temporal and spatial feature sets based on the sliding window, and construct the decision matrix, including: A time feature set is obtained based on a sliding window with 5-minute intervals and a total of 1 hour, and a spatial feature set is obtained based on a sliding window with 10-minute intervals and a total of 1 hour. The performance of each correction factor on each feature will be used as the row of the decision matrix, and the feature will be used as the column. The feature is a random feature value in the time feature set and the spatial feature set. The time feature set includes data timeliness, and the spatial feature set includes spatial resolution and reliability. Step S214: Normalize the decision matrix, and obtain the positive ideal solution and the negative ideal solution based on the normalized decision matrix. The positive ideal solution is the maximum value of each column, and the negative ideal solution is the minimum value of each column. Step S215: Calculate the Euclidean distance between each correction factor and the positive ideal solution and the negative ideal solution respectively. Calculate the relative proximity between each correction factor and the positive ideal solution based on the Euclidean distance. Select the correction factor with the largest relative proximity as the optimal correction factor. Discard the numerical nodes corresponding to the features of non-optimal correction factors from the time feature set and the spatial feature set to obtain the final time feature set and spatial feature set.

4. The AI-based geographic information-based precise navigation method as described in claim 3, characterized in that, When acquiring data based on extreme weather, the inverse of the negative indicators in the sensor database is taken to normalize the data in the sensor database, and the L1 or L2 band signal strength of the satellite navigation GNSS output and the original observation value are adjusted in real time to construct an evaluation matrix. The index function of the evaluation matrix includes GNSS signal-to-noise ratio, visual feature matching rate and ground IMU confidence. The index weights of the evaluation matrix are calculated based on the RSR method. The optimal correction factor is obtained through step S21. The optimal correction factor is weighted and added to the index weights to obtain the fusion weights of each sensor. The fusion weights are used as the extreme weather weight combination of the dynamic time-delay observation correction model. The dynamic time-delay observation correction model is iteratively trained based on the real-time updated sensor database to obtain the optimal extreme weather weight combination. The historical data output by the ground IMU is analyzed using a hierarchical LSTM model to identify abnormal data. The abnormal data is used as drift pattern data and input into the dynamic time-delay observation correction model. The dynamic equation is used as the discrete loss function for real-time data correction. Error compensation is achieved through a Kalman filter.

5. An AI-based geographic information precision navigation system, characterized in that, It includes a sensor module, a model building module, and an extreme optimization module: The sensor module includes acquiring data through sensors and establishing a sensor database; The model building module includes constructing a dynamic time-delay observation correction model; Constructing a dynamic time-delay observation correction model includes the following sub-steps: Step S21: Using the TOPSIS multi-criteria decision algorithm, select initial correction factors from the sensor database based on data timeliness, spatial resolution and reliability, use a sliding window to filter the initial correction factors to obtain the optimal correction factors, and divide the initial correction factors into a time feature set and a spatial feature set; Step S22: Based on the sensor database, calculate the prior dry delay and wet delay using the Saastamoinen model; Step S23: Extract local meteorological gradients using ID convolutional layers and construct a dynamic time-delay observation correction model; Step S23 includes the following sub-steps: Step S231: Extract local meteorological gradients using ID convolutional layers. The local meteorological gradients include the spatial rate of temperature change. The temporal feature set and spatial feature set are branched based on a hierarchical LSTM structure. Step S232: Construct a dynamic time delay observation correction model and perform adversarial training on the dynamic time delay observation correction model using gradient penalty; Step S231 includes: The temporal and spatial feature sets are processed at high frequency intervals through the first layer of the hierarchical LSTM structure to capture turbulent fluctuations. The captured turbulent fluctuations are then input into the second layer of the hierarchical LSTM structure for mid-frequency aggregation processing to extract weather system evolution feature values. These weather system evolution feature values ​​are then input into the third layer of the hierarchical LSTM structure for low-frequency learning, and seasonal pattern features are learned in daily periodic components. Step S232 includes: Using the wet delay output by Saastamoinen as a prior value, the residual term of the wet delay is adaptively learned by a neural network. The atmospheric thermodynamic equation is used as a physical loss function. MC-Dropout variational inference is added to the output layer to predict the mean and standard deviation of the wet delay. Embedding the wet delay predicted by LSTM into the GNSS pseudorange equation yields the tightly coupled observation equation, the mathematical expression of which is: ； in, For GNSS pseudorange coordinates, For the distance to receive the signal, At the speed of light, t represents the delay correction error between the sensor and the original coordinate point in the current coordinate system. This delay correction error is the error between the initial and predicted coordinate parameters calculated using the least squares iterative method. For dry delay, For the wet delay predicted by LSTM, For noise; The weights of the dynamic time delay observation correction model are set to those calculated using dynamic uncertainty, and their mathematical expression is: ； in, As weight, The standard deviation of the wet delay, Weighting of GNSS observations for satellite navigation; The extreme optimization module includes maintaining positioning accuracy by using a fusion algorithm on the dynamic time-delay observation correction model, and optimizing navigation and positioning in extreme weather. According to step S21, the sliding window is set to a 1-hour sliding window with a 10-minute interval. The initial correction factor is extracted from the continuous operating data and dynamic attitude obtained by the ground IMU in the sensor database to obtain the temperature drift correlation characteristic and the angular velocity integral error accumulation rate characteristic. The Euler dynamics equations are added as regularization terms to the loss function of the dynamic time-delay observation correction model. The difference between the predicted values ​​of the output temperature drift correlation feature and the cumulative rate of angular velocity integral error feature and the actual temperature drift correlation feature and the cumulative rate of angular velocity integral error feature calculated per unit time is used as the prior error to establish the ground IMU error state vector. The ground IMU error state vector is obtained by convolving the prior error with the state transition matrix. The state transition matrix is ​​derived from the kinematics of the ground IMU. The kinematics derivation formulas of the ground IMU include the position update formula, the velocity update formula, and the rotation matrix update formula.

Images