Inertial navigation system calibration method and device based on grating sensor

Abstract

The invention discloses an inertial navigation system calibration method and device based on a grating sensor. The method comprises the following steps: dynamically adjusting sampling frequencies and data acquisition strategies of the grating sensor and an IMU (Inertial Measurement Unit); the error correction weight of the sensor is dynamically adjusted according to the motion state and the environment change, the error compensation effect is optimized, and an error correction strategy is provided for calibration; correcting a partial differential equation and a self-adaptive correction coefficient of the error model in real time by utilizing a self-adaptive partial differential equation algorithm, and dynamically adjusting an error compensation strategy; a virtual feedback correction ring mechanism and a deep reinforcement learning strategy are introduced, a calibration behavior is optimized by using a virtual environment model and a real-time feedback signal, the weight of the sensor is dynamically adjusted, and self-adaptive calibration is realized; an incremental self-learning calibration mechanism is adopted, self-adaptive incremental optimization is carried out based on error accumulation monitoring and historical data in operation, and calibration parameters are dynamically adjusted. According to the invention, the limitation of static calibration is broken through, and high-precision error compensation can be realized in a complex dynamic environment.

Description

Technical Field

[0001] The present invention relates to the field of navigation technology, and in particular to a grating sensor-based inertial navigation system calibration method and device. Background Art

[0002] As an autonomous positioning and navigation technology, the Inertial Navigation System (INS) has a wide range of applications in aviation, aerospace, unmanned driving, industrial measurement and other fields. INS mainly relies on the accelerometer and gyroscope in the inertial measurement unit (IMU) to calculate attitude, velocity and position information. However, due to the inherent errors of IMU, such as zero bias drift, scale factor error, installation error, etc., these errors will continue to accumulate during long-term operation, resulting in a decrease in navigation accuracy. Therefore, high-precision calibration and error compensation of INS is the key to improving system stability and accuracy.

[0003] As a high-precision displacement measurement device, grating sensors have been introduced into the error correction process of inertial navigation systems in recent years due to their strong anti-interference ability, high precision, and good long-term stability. Grating sensors can provide high-resolution displacement information and provide a reference benchmark for IMU error calibration. By fusing grating sensors with inertial navigation data, the cumulative error of IMU can be effectively reduced and the accuracy of the system can be improved. However, how to achieve high-precision fusion of the two in a complex environment remains a challenging problem. At present, the traditional inertial navigation system calibration method mainly adopts static or quasi-static calibration technology. For example, common methods include static zero bias compensation, earth rotation compensation, standard displacement stage calibration, etc. These methods are usually based on laboratory environment and are performed under ideal static conditions. However, in actual application scenarios, INS is often in a dynamic environment, such as severe vibration during the flight of drones and complex road conditions during vehicle navigation. In this case, the effectiveness of the static calibration method is greatly reduced and cannot fully reflect the impact of the dynamic environment on sensor errors. Therefore, how to realize the joint calibration of grating sensors and INS under dynamic conditions has become the frontier direction of current research. Summary of the invention

[0004] In order to solve the above problems, the present invention provides a method and device for calibrating an inertial navigation system based on a grating sensor.

[0005] To achieve the above purpose, the technical solution adopted by the present invention is as follows: On the one hand, the present invention discloses a method for calibrating an inertial navigation system based on a grating sensor, comprising: Step 1: Dynamically adjust the sampling frequency and data acquisition strategy of the grating sensor and IMU through the adaptive adjustment mechanism of the perceived environment, the real-time sensor health assessment model and the real-time motion state classifier; Step 2: Based on the real-time data interaction between the grating sensor and the IMU, an interactive dynamic error allocation algorithm is designed to dynamically adjust the sensor error correction weight according to the motion state and environmental changes, optimize the error compensation effect, and provide an error correction strategy for calibration; Step 3: Using the adaptive partial differential equation algorithm, the error compensation strategy is dynamically adjusted by correcting the partial differential equation of the error model and the adaptive correction coefficient in real time; Step 4: Introduce a virtual feedback correction loop mechanism and deep reinforcement learning strategy, use the virtual environment model and real-time feedback signals to optimize the calibration behavior, dynamically adjust the sensor weights, and achieve adaptive calibration; Step 5: Adopt an incremental self-learning calibration mechanism to perform adaptive incremental optimization and dynamically adjust calibration parameters based on error accumulation monitoring and historical data during operation.

[0006] Further: Step 1 comprises: The environmental perception module monitors environmental changes in real time and dynamically adjusts the sampling frequency of the grating sensor and IMU according to the intensity of environmental changes; Quantify the health of the sensor based on the sensor's offset, drift, and noise level indicators; if the health is below the threshold, mark the low-quality data through the health alert mechanism, and adjust the sampling frequency or trigger recalibration; Use machine learning methods to identify the current motion state in real time, and adjust the sampling frequency and calibration strategy according to different motion states.

[0007] Further: Step 2 comprises: Determine the current motion state in real time and decide how to adjust the sensor's error correction weight; According to the real-time motion state, the error correction ratio of the grating sensor and IMU is dynamically adjusted. In the rotational motion stage, the correction is based on the IMU data, and in the linear motion stage, the correction is based on the grating sensor data. Based on real-time sensor data and historical error data, the dynamic error allocation model algorithm monitors the errors of the grating sensor and IMU, and optimizes them using an adaptive weight adjustment model; when a large sensor error is detected, the correction weight of the sensor is increased and the correction weight of another sensor is reduced accordingly.

[0008] Further: Step 3 includes: Partial differential equations are introduced to describe the errors of grating sensors and IMU in time and space to dynamically reflect the changes of sensor errors over time while considering the interactions between sensors. Introduce adaptive correction factors to adjust the error correction strength in real time. When the acceleration or angular velocity changes suddenly, the correction weight of the IMU is adjusted first. When the environmental changes are small, the center of gravity is corrected to the grating sensor. The finite difference method is used to discretize and solve the partial differential equation to achieve dynamic adjustment of the error correction coefficient. By setting the time step, the error correction model is updated in each time step to track and correct the sensor error in real time. At the same time, a dynamic error correction rate adjustment mechanism is introduced to adjust the correction speed according to the overall error feedback. When the error reaches the set threshold, the correction rate is accelerated. When the error is small, the correction rate is moderately slowed down to avoid error oscillation caused by excessive correction.

[0009] Further: Step 4 includes: Construct a virtual environment model to simulate the impact of various factors in the real environment on the sensor output; generate an error feedback signal by comparing the actual output of the sensor with the expected output of the virtual model; the feedback signal is passed to the artificial intelligence algorithm to simulate and optimize the next calibration behavior; The deep reinforcement learning algorithm is used to adaptively adjust the calibration strategy according to the real-time error and environmental feedback. The state space includes the current output of the sensor, error feedback, sensor health, acceleration and angular velocity. The action space includes the intensity, direction and weight distribution of the sensor error correction. The reward function is defined as the reduction in error, and the calibration process is optimized by maximizing the long-term reward. Deep reinforcement learning determines the correction action through the policy network, and the value network evaluates the expected effect, thereby achieving dynamic optimization.

[0010] Further: Step 5 comprises: Through the incremental calculation method based on the time window, the changes of system errors are monitored periodically; a dynamic time window is set to accumulate and statistically analyze the errors in each calibration cycle. If the accumulated error reaches the preset threshold, the self-learning update mechanism is triggered to start the incremental adjustment of the calibration parameters; Based on historical data and real-time data, an incremental optimization process is started. By analyzing the changing trend of historical errors, a trend prediction algorithm is used to predict future error changes and dynamically adjust the calibration parameters.

[0011] On the other hand, the present invention discloses an inertial navigation system calibration device based on a grating sensor, comprising: Environmental perception and raw data acquisition module: Dynamically adjust the sampling frequency and data acquisition strategy of grating sensors and IMU through adaptive adjustment mechanism of perceived environment, real-time sensor health assessment model and real-time motion state classifier; Interactive dynamic error allocation module: Based on the real-time data interaction between the grating sensor and the IMU, an interactive dynamic error allocation algorithm is designed to dynamically adjust the sensor error correction weight according to the motion state and environmental changes, optimize the error compensation effect, and provide an error correction strategy for calibration; Dynamic error correction module: using adaptive partial differential equation algorithm, dynamically adjusts the error compensation strategy by correcting the partial differential equation of the error model and adaptive correction coefficient in real time; Adaptive calibration module: introduces a virtual feedback correction loop mechanism and deep reinforcement learning strategy, uses virtual environment models and real-time feedback signals to optimize calibration behavior, dynamically adjusts sensor weights, and achieves adaptive calibration; Calibration parameter optimization module: It adopts an incremental self-learning calibration mechanism, performs adaptive incremental optimization based on error accumulation monitoring and historical data during operation, and dynamically adjusts calibration parameters.

[0012] Further: the environment perception and raw data acquisition module includes: Adaptive adjustment of perceived environment: The environmental perception module monitors environmental changes in real time and dynamically adjusts the sampling frequency of the grating sensor and IMU according to the intensity of environmental changes; Real-time sensor health assessment module: quantifies sensor health based on sensor offset, drift, and noise level indicators; if the health is below the threshold, the health alarm mechanism marks low-quality data and adjusts the sampling frequency or triggers recalibration; Real-time motion state classification module: uses machine learning methods to identify the current motion state in real time, and adjusts the sampling frequency and calibration strategy according to different motion states; The interactive dynamic error allocation module includes: Motion state recognition and sensor interaction analysis module: determines the current motion state in real time and decides how to adjust the sensor's error correction weight; Building blocks of the error allocation model: dynamically adjust the error correction ratio of the grating sensor and IMU according to the real-time motion state. In the rotational motion stage, the error correction is based on the IMU data, and in the linear motion stage, the error correction is based on the grating sensor data. Real-time adjustment mechanism module of dynamic error correction weight: Based on real-time sensor data and historical error data, the dynamic error allocation model algorithm monitors the errors of the grating sensor and IMU, and optimizes them using the adaptive weight adjustment model; when a large sensor error is detected, the correction weight of the sensor is increased, and the correction weight of another sensor is reduced accordingly.

[0013] Further: the dynamic error correction module includes: Partial differential equation building blocks of the error model: Partial differential equations are introduced to describe the errors of the grating sensor and IMU in time and space to dynamically reflect the changes in sensor errors over time while considering the interactions between sensors; Adaptive adjustment module of error correction coefficient: introduce adaptive correction factor to adjust the error correction strength in real time. When the acceleration or angular velocity changes suddenly, the correction weight of IMU is adjusted first. When the environmental change is small, the center of gravity is corrected to the grating sensor. Real-time solution and system update module of partial differential equations: The finite difference method is used to discretize and solve the partial differential equations to achieve dynamic adjustment of the error correction coefficient; by setting the time step, the error correction model is updated in each time step to track and correct the sensor error in real time; at the same time, a dynamic error correction rate adjustment mechanism is introduced to adjust the correction speed according to the overall error feedback; when the error reaches the set threshold, the correction rate is accelerated; when the error is small, the correction rate is moderately slowed down to avoid error oscillation caused by excessive correction.

[0014] Further: the adaptive calibration module includes: Virtual feedback correction loop mechanism module: builds a virtual environment model to simulate the impact of various factors in the real environment on the sensor output; generates an error feedback signal by comparing the actual output of the sensor with the expected output of the virtual model; the feedback signal is passed to the artificial intelligence algorithm to simulate and optimize the next calibration behavior; Deep reinforcement learning strategy module: uses deep reinforcement learning algorithm to adaptively adjust the calibration strategy according to real-time error and environmental feedback; the state space includes the current output of the sensor, error feedback, sensor health, acceleration and angular velocity; the action space includes the intensity, direction and weight distribution of sensor error correction; the reward function is defined as the amount of error reduction, and the calibration process is optimized by maximizing long-term rewards; deep reinforcement learning determines the correction action through the policy network, and the value network evaluates the expected effect, thereby achieving dynamic optimization; The calibration parameter optimization module includes: Error accumulation monitoring and self-learning triggering module: Through the incremental calculation method based on the time window, the system error changes are periodically monitored; a dynamic time window is set to accumulate and statistically analyze the errors in each calibration cycle. If the accumulated error reaches the preset threshold, the self-learning update mechanism is triggered to start the incremental adjustment of the calibration parameters; Incremental optimization strategy module: Based on historical data and real-time data, the incremental optimization process is started. By analyzing the changing trend of historical errors, the trend prediction algorithm is used to predict future error changes and dynamically adjust the calibration parameters.

[0015] Compared with the prior art, the present invention has the following technical advances: Traditional INS calibration methods are usually based on static or quasi-static environments, and are difficult to cope with error changes in dynamic environments. This method uses an adaptive adjustment mechanism for perceiving the environment, which can adjust the data acquisition frequency in real time, and combines a real-time motion state classifier to analyze the current motion state to achieve adaptive adaptation to the dynamic environment. Traditional methods often use fixed weights to correct errors in the data of IMU and grating sensors, which is difficult to adapt to the complex changes in the dynamic environment. This method introduces an interactive dynamic error allocation algorithm, which can dynamically allocate the error correction ratio in different motion stages according to the real-time data interaction relationship between the grating sensor and the IMU. For example, during the rotation process, the grating sensor data is greatly affected by the moment of inertia, and the IMU error is easily amplified in linear motion. The algorithm can intelligently adjust the correction weight to achieve optimal error compensation between multiple sensors.

[0016] Traditional error compensation methods mainly rely on discrete methods such as global optimization or Kalman filtering, which are difficult to cope with continuous error changes in complex dynamic environments. This method proposes an adaptive partial differential equation algorithm, which can dynamically adjust the error correction model in the continuous time domain by solving nonlinear partial differential equations. The adaptive partial differential equation algorithm combines the real-time motion state of the system to automatically adjust the correction coefficient, making the error correction more accurate, especially suitable for high-speed changing environments, such as drones, rapid maneuvers or vehicles, etc. Traditional calibration methods are often performed offline and are difficult to adapt to real-time error changes. This method proposes a virtual feedback correction loop, which builds a virtual environment model, compares sensor errors with ideal outputs, and uses deep reinforcement learning strategies for error trend prediction and dynamic optimization. The virtual feedback correction loop can continuously simulate future error changes and optimize the compensation strategy in advance, thereby achieving adaptive real-time calibration and improving the accuracy and stability of INS in long-term operation.

[0017] In summary, the present invention extends from static calibration to dynamic calibration, from fixed error correction to interactive error allocation, and from discrete error compensation to dynamic correction of partial differential equations. Through a virtual feedback correction loop and a self-learning mechanism, continuous optimization of the system is achieved, breaking through the limitations of static calibration and realizing high-precision error compensation in complex dynamic environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0019] In the attached picture: Figure 1 This is a flow chart of Embodiment 1 of the present invention; Figure 2 This is a system structure diagram of Example 2 of the present invention. DETAILED DESCRIPTION

[0020] The following specific embodiments may be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments. The embodiments of the present invention will be described below in conjunction with the accompanying drawings.

[0021] Example 1 like Figure 1 As shown, the present invention discloses a calibration method of an inertial navigation system based on a grating sensor, comprising: step 1: dynamically adjusting the sampling frequency and data acquisition strategy of the grating sensor and the IMU through a perception environment adaptive adjustment mechanism, a real-time sensor health evaluation model and a real-time motion state classifier; Step 2: Based on the real-time data interaction between the grating sensor and the IMU, an interactive dynamic error allocation algorithm is designed to dynamically adjust the sensor error correction weight according to the motion state and environmental changes, optimize the error compensation effect, and provide an error correction strategy for calibration; Step 3: Using the adaptive partial differential equation algorithm, the error compensation strategy is dynamically adjusted by correcting the partial differential equation of the error model and the adaptive correction coefficient in real time; Step 4: Introduce a virtual feedback correction loop mechanism and deep reinforcement learning strategy, use the virtual environment model and real-time feedback signals to optimize the calibration behavior, dynamically adjust the sensor weights, and achieve adaptive calibration; Step 5: Adopt an incremental self-learning calibration mechanism to perform adaptive incremental optimization and dynamically adjust calibration parameters based on error accumulation monitoring and historical data during operation.

[0022] Specifically, the goal of step 1 is to obtain raw sensor data in a dynamic environment to provide necessary input data for the subsequent calibration process, especially to effectively collect data from grating sensors and inertial measurement units (IMUs) in a dynamic environment, and to adjust the data acquisition strategy in real time through intelligent mechanisms, so as to provide the most accurate and reliable raw data for the subsequent error modeling and calibration process.

[0023] 1.1 Adaptive adjustment mechanism of perceived environment: The core of the adaptive adjustment mechanism of the perception environment is to dynamically adjust the sampling frequency of the grating sensor and IMU according to the instantaneous changes in the environment (such as vibration, acceleration, temperature change, etc.). In this way, the system can adaptively adjust the working parameters of the sensor under different environmental conditions to reduce redundant data when the environment is stable, and at the same time increase the sampling frequency when the environment changes drastically to ensure the accuracy of the data.

[0024] Specific implementation: 1.1.1 Environmental monitoring and sampling adjustment strategy: When environmental conditions change (such as external vibration, acceleration, temperature change, etc.), the environmental perception adaptive adjustment mechanism detects changes in the current environment through the sensor's own environmental perception module (such as temperature and humidity sensor, accelerometer, etc.).

[0025] When the environmental vibration intensifies or the acceleration changes greatly, the system will increase the sampling frequency to capture the rapid dynamic characteristics of the environmental changes. For example, in a strong vibration environment, the system may increase the sampling frequency of the grating sensor to 10kHz and the sampling frequency of the IMU to 1kHz; while in stable and smooth motion, the sampling frequency is reduced to 1kHz and 100Hz.

[0026] 1.2.2 Adaptive frequency adjustment: According to changes in the environment and sensor output, the perception environment adaptive adjustment mechanism uses dynamic threshold judgment to automatically adjust the sensor sampling frequency and data update frequency. The frequency adjustment strategy takes into account information such as current acceleration, vibration intensity, and noise output by the sensor. For example, when the standard deviation of acceleration is greater than a certain threshold, the perception environment adaptive adjustment mechanism will automatically trigger the high-frequency sampling mode.

[0027] Set the environment parameters to , which includes factors such as acceleration, vibration intensity, temperature, etc., and the sensor sampling frequency The change formula is: in, is the base sampling frequency, is the sensitivity coefficient, is a measure of environmental change. Automatic adjustment.

[0028] 1.1.3 Data redundancy control: When the environment is relatively stable, the system will enter the "low-frequency sampling mode" to reduce redundant data and extend the service life of the sensor. In low-frequency sampling mode, the data storage and transmission load is low, which helps to reduce the storage space occupied.

[0029] For grating sensors, the sampling frequency in steady state may be reduced to once per second, while the sampling frequency of IMU is reduced to about 10Hz.

[0030] 1.2 Real-time sensor health assessment model: To ensure the credibility and accuracy of the data, the real-time sensor health assessment model will evaluate the operating status of the grating sensor and IMU at every moment of data collection to determine whether the sensor is working properly and effectively. This model will dynamically adjust the data collection strategy based on the data features detected in real time to further improve the robustness of the system.

[0031] 1.2.1 Health Assessment Framework: Sensor health assessment models rely on a set of metrics to quantify the health of a sensor, including sensor offset, drift, noise level, and sensor deviation from ideal values.

[0032] Assume the health of the grating sensor and the health of the IMU In time The moments are defined by the following formulas: in, is the offset of the grating sensor, is the drift of the IMU, and is a constant determined according to the sensor performance.

[0033] If the health assessment value is lower than a certain threshold, the system will mark the current data as low-quality data through the "health alert mechanism" and adjust the sampling frequency in a timely manner, or trigger a recalibration process.

[0034] 1.2.2 Sensor health status monitoring: The results of sensor health assessment are used to guide changes in sensor sampling frequency. When the sensor health is high, the system can choose a lower sampling frequency to reduce the waste of computing resources. When the sensor health is low, the system will increase the sampling frequency or use a more accurate calibration method.

[0035] 1.3 Real-time motion state classifier: In order to effectively identify the current motion state (such as uniform linear motion, acceleration, rotation, etc.), this embodiment designs a "real-time motion state classifier" based on machine learning. The classifier identifies the current motion state in real time by analyzing the sensor data, so as to adaptively adjust the sampling frequency and calibration process under different motion states.

[0036] 1.3.1 State Classification Model: A classifier is trained using machine learning methods such as deep neural network (DNN) or support vector machine (SVM) to predict the current motion state in real time based on the sensor's acceleration, angular velocity, grating sensor displacement data and other features. The model used in this embodiment inputs the real-time data of the grating sensor and IMU, and after preprocessing, inputs these features into the trained classifier to obtain the motion state prediction.

[0037] Assume that the input features are (including acceleration, angular velocity, etc.), the output of the classifier is the state category , which can be achieved through deep learning networks To obtain: in Represents the neural network output nodes, representing different motion states (such as uniform linear, acceleration, rotation, etc.).

[0038] 2. State adjustment logic: After the classifier identifies the current state at each moment, it will output the motion state label, which will trigger the corresponding adaptive adjustment strategy. For example, during rotation, the system will pay more attention to the stability of the IMU; during uniform motion, the accuracy of the grating sensor will be given priority.

[0039] The system will adjust the data collection strategy in time according to the changes in the motion state, such as increasing the sampling frequency under conditions of acceleration or large vibration, and reducing the frequency under uniform speed conditions.

[0040] In step 1, by introducing the perception environment adaptive adjustment mechanism, the real-time sensor health assessment model and the real-time motion state classifier, the system achieves sensitive response to environmental changes and motion states. The perception environment adaptive adjustment mechanism ensures the data collection efficiency of the sensor in different dynamic environments. The health assessment model judges the reliability of the sensor in real time and adjusts the data collection strategy in time, while the real-time motion state classifier ensures that the system can make the best data collection decision under different motion states.

[0041] The specific goal of step 2 is to design a new dynamic error allocation model through real-time interactive analysis of grating sensor and inertial navigation system (INS) data, so that the system can dynamically adjust the error correction ratio of each sensor according to the motion state, environmental changes and interaction characteristics between sensors, optimize the error compensation effect, and improve the overall accuracy of the system.

[0042] 2.1 Interactive dynamic error allocation algorithm: Traditional error correction methods usually rely on linear superposition models, which simply add up the errors of each sensor and then make corrections, but this method fails to take into account the dynamic interaction between sensors and changes in the motion environment. In order to address this limitation, this embodiment proposes an "interactive dynamic error allocation algorithm", the core idea of ​​which is to dynamically adjust the weight of each sensor error correction based on the real-time data interaction relationship between the grating sensor and the IMU, the environmental state, and the different motion stages. Under different motion states, the errors of the grating sensor and the IMU have different effects on the final navigation accuracy. Therefore, the error correction weight needs to be adaptively adjusted according to the motion state at each moment.

[0043] 2.1.1 Motion state recognition and sensor interaction analysis: Based on the real-time motion state classifier introduced in step 1, the dynamic error allocation model algorithm can determine the motion state of the system in real time (such as uniform linear motion, acceleration, rotation, etc.). Under different motion states, the errors of the grating sensor and IMU are different. For example, in rotational motion, the error of the grating sensor may be amplified due to the moment of inertia effect, while in linear uniform motion, the acceleration error of the IMU usually becomes dominant. The dynamic error allocation model uses the output information of the motion state classifier to determine the current motion mode, thereby deciding how to adjust the error correction weight of the sensor.

[0044] 2.1.2 Construction of error allocation model: In each motion state, the dynamic error allocation model algorithm will dynamically adjust the error correction ratio of the grating sensor and IMU. For example, in the rotational motion stage, the system may rely on IMU data for correction first, while in the linear motion stage, it will rely more on grating sensor data for correction. The dynamic error allocation model allocates the correction weight of each sensor based on real-time acceleration, rotational angular velocity, displacement and other sensor data through the following mathematical model.

[0045] 2.1.3 Error correction weight allocation formula: Assume that the errors of the grating sensor and IMU are and , the total error correction of the system The dynamic distribution of weights depends not only on the real-time data of the sensor, but also on the state of motion. Related to environmental feedback information (such as vibration, temperature, etc.), the specific formula is as follows: in: is the error correction weight of the grating sensor, is the error correction weight of the IMU, , ensuring that the sum of the correction weights of the two sensors is 1.

[0046] Dynamic weight update rules: Weight and According to the current motion status , acceleration , angular velocity and the health of the sensor and Dynamic update, the specific update rules are as follows: in, and It is a function calculated based on the sensor's motion state, acceleration, rotation information and other factors, and is defined as follows: These two functions measure the relative importance of the grating sensor and IMU in the current environment and motion state, and the health value and The values ​​of these functions are dynamically affected, thereby adjusting the weight of each sensor. For example, in acceleration motion, the IMU may be more important; while in rotational motion, the influence of the grating sensor may be greater.

[0047] 2.1.4 Error collaborative correction: Once the weight and Once the errors are calculated, the system can perform weighted correction on the errors of the grating sensor and IMU to obtain the optimal error correction value. By continuously iteratively updating the weights and correction values, the system can dynamically adjust the error allocation strategy under different motion stages and environmental conditions, and ultimately achieve coordinated error correction of multiple sensors.

[0048] 2.2 Real-time adjustment mechanism of dynamic error correction weight: In order to better adapt to different motion states and environmental changes, the dynamic error allocation model algorithm not only relies on the initial weight allocation rules, but also adjusts itself according to real-time sensor data and system errors. As the system runs over time, the error relationship between sensors may change, so an adaptive weight adjustment mechanism is needed to optimize the error correction process.

[0049] 2.2.1 Dynamic weight adaptive adjustment: Based on real-time sensor data, the dynamic error allocation model algorithm continuously monitors the errors of the grating sensor and IMU, and uses historical data to establish an adaptive weight adjustment model. When a large sensor error is detected, the system automatically increases the correction weight of the sensor and reduces the correction weight of another sensor accordingly. This adaptive mechanism updates the weight in real time through the following formula: in, is the adjustment coefficient, and It is the error change between the grating sensor and the IMU, which indicates the error correction amplitude of the current sensor relative to the previous moment. In this way, the system can adjust the weight distribution in real time and optimize the error correction.

[0050] By designing an interactive dynamic error allocation algorithm, this embodiment implements a mechanism for dynamically adjusting the error correction weights between the grating sensor and the IMU, so that the system can intelligently optimize the error correction strategy under different motion states and environmental conditions to achieve the best error collaborative correction effect. The algorithm not only takes into account the interaction between sensors, but also makes adaptive adjustments based on the real-time sensor health status and motion status, providing a reliable foundation for subsequent precise calibration.

[0051] Specifically, the goal of step 3 is to ensure that the system can maintain high accuracy in different motion environments and overcome the limitations of traditional global optimization or filtering methods. An adaptive partial differential equation algorithm is proposed, which can dynamically adjust the system's error correction strategy according to the real-time changing environment and sensor errors, ensuring the real-time adaptability and accuracy of the system in high dynamic environments.

[0052] 3.1 Adaptive partial differential equation algorithm: Traditional error correction methods, such as global optimization or filtering methods, usually assume that environmental changes are static or gradual, and are therefore not suitable for highly dynamic application scenarios, especially under high-speed acceleration, rapid rotation, or sudden environmental changes. To this end, this embodiment proposes an error correction model based on partial differential equations, an adaptive partial differential equation algorithm, which has a high degree of dynamic adaptability and can dynamically adjust the error compensation model through real-time correction rules to cope with rapidly changing motion states and environmental factors.

[0053] 3.1.1 Construction of partial differential equation of error model: Based on the data of the grating sensor and IMU, this embodiment assumes that the system error changes dynamically in time and space, so a partial differential equation is introduced to describe the time variation of the error and the spatial interaction between sensors. The state variable of the sensor error is set to and , the dynamics of these two errors are described by the following partial differential equations: in, and Represent the errors of grating sensor and IMU respectively, and are acceleration and angular velocity, Indicates the influencing factors of the external environment (such as temperature, vibration, etc.).

[0054] function and It is a nonlinear function based on factors such as sensor error, motion state and environmental change, which describes the dynamic process of sensor error changing over time. This embodiment assumes that these functions have different dynamic behaviors under specific motion states, so they will be adaptively adjusted under different environmental conditions.

[0055] 3.1.2 Adaptive adjustment of error correction coefficient: In a dynamic environment, the adjustment of the error correction coefficient must respond to changes in the environment in real time, especially in fast-changing motion states such as acceleration, deceleration, and rotation. The adaptive partial differential equation algorithm adjusts the correction coefficient through adaptive correction rules to ensure that the error compensation of each sensor remains optimal under instantaneous environmental changes. For example, when the acceleration changes suddenly, the system should give priority to adjusting the correction weight of the IMU; and when the error of the grating sensor increases, its correction weight should be adjusted.

[0056] Specifically, this embodiment introduces an adaptive correction factor , which depends on factors such as acceleration, angular velocity, and sensor health, and adjusts the error correction strength of each sensor in real time. This factor The calculation is done by the following formula: in, and denote the magnitude of acceleration and angular velocity, respectively. is a constant, representing the correction benchmark when the environment is static. This factor Used to adjust the strength of error correction, for example: if If it becomes larger, it means that the motion state has undergone a large acceleration or rotation change, and the system should increase the proportion of IMU correction; if It becomes smaller, which means that the environmental change is small, and the system should shift the correction focus to the grating sensor.

[0057] At the same time, the correction factor The health of the sensor is also taken into account ( and ), when a sensor fails or the data quality deteriorates, the correction weight of the sensor is automatically reduced.

[0058] 3.1.3 Real-time solution and system update of partial differential equations: In order to correct errors in real time in a dynamic environment, the adaptive partial differential equation algorithm needs to quickly solve the partial differential equation to update the sensor error correction model in real time. This embodiment uses the finite difference method in the numerical solution to discretize and solve the above partial differential equation to achieve dynamic adjustment of the error correction coefficient.

[0059] Set the time step , this embodiment approximately solves the partial differential equation through the following discretization form: By continuously iterating the above discretized equations, the adaptive partial differential equation algorithm can update the error correction model in each time step, thereby tracking and correcting the error between the grating sensor and the IMU in real time.

[0060] 3.2 Adaptive error correction rate: In order to further improve the system's correction capability in high dynamic environments, the adaptive partial differential equation algorithm also introduces a dynamic error correction rate adjustment mechanism, which adjusts the correction speed within each time step based on the system's overall error feedback to cope with large fluctuations in system errors.

[0061] When the system error reaches the set threshold, the correction rate will automatically increase to quickly correct the system state; conversely, when the error is small, the correction rate will be moderately slowed down to avoid error oscillation caused by over-correction. It can be adjusted by the following formula: in, is the adjustment factor, is the total error at the current moment. Through this mechanism, the adaptive partial differential equation algorithm can adjust the intensity and speed of correction according to real-time error feedback, ensuring that the system always maintains high accuracy in different environments.

[0062] By introducing an adaptive partial differential equation algorithm, this embodiment constructs an error correction model with high dynamic adaptability. The model can dynamically correct errors through nonlinear partial differential equations according to real-time motion states, sensor data, and environmental changes, and adjust the error correction strategy in real time by combining adaptive correction factors and correction rate mechanisms, thereby effectively responding to challenges in various dynamic environments and ensuring that the system always maintains high accuracy and stability.

[0063] Specifically, the goal of step 4 is to perform real-time calibration through a virtual model and optimize the accuracy of sensor data through step-by-step corrections in combination with real-time environmental feedback. This process can not only simulate the behavior of sensors in a changing environment, but also intelligently adjust the calibration behavior of the system through deep reinforcement learning strategies to achieve efficient and dynamic adaptive calibration.

[0064] 4.1 Introducing a virtual feedback correction loop mechanism: The virtual feedback correction loop mechanism creates a virtual environment model, which enables the system to generate feedback signals in real time according to the deviation between the actual output of the sensor and the expected output at each moment. This feedback signal is not directly fed back to the control system, but is simulated and adjusted through a set of artificial intelligence algorithms to perform a more accurate calibration process. Specifically, the workflow of the virtual feedback correction loop mechanism can be divided into the following steps: 4.1.1 Virtual environment modeling: This embodiment first constructs a virtual environment model based on the working characteristics of the sensor, environmental changes and dynamic error model The model can simulate the impact of various factors in the real environment (such as vibration, acceleration, temperature change, etc.) on the sensor output and provide ideal output for subsequent error correction. Error of the virtual environment model The actual output of the sensor With the expected output Definition of the difference between: The error It is the core information in the virtual feedback correction loop, reflecting the deviation between the sensor and the ideal environment model.

[0065] 4.1.2 Virtual feedback signal generation and simulation: Based on error , this embodiment uses a feedback mechanism to generate a virtual feedback signal This signal is used to guide the next correction strategy. Different from the traditional error feedback method, this feedback signal is not directly used in the control system, but is transmitted to an artificial intelligence system to simulate how to optimize the next calibration behavior.

[0066] 4.2 Deep reinforcement learning strategy: Deep reinforcement learning is the core technology for realizing the virtual feedback correction loop mechanism. It can adaptively adjust the calibration strategy according to real-time errors and environmental feedback. Deep reinforcement learning simulates and predicts the future state of the environment and gradually optimizes the calibration strategy based on feedback signals.

[0067] 4.2.1 State space definition: In deep reinforcement learning, the state space of the system Including multi-dimensional factors such as the current output of the sensor, environmental status, sensor health, error feedback information, etc., this embodiment defines the state space as: in, is the current output of the sensor, is the error with the virtual environment model, and Represent the health of the grating sensor and IMU respectively, and It is the acceleration and angular velocity information, etc. This multi-dimensional state space can reflect the complex interaction between the sensor and the environment.

[0068] 4.1.2 Action Space Definition: Action Space It includes the decision space of the correction strategy, that is, the intensity, correction direction, correction coefficient, etc. of the sensor error correction. The action space can be expressed as: in, and They are the corrections for the grating sensor and IMU errors respectively. and It is the weight distribution of grating sensor and IMU.

[0069] 4.1.3 Reward function and optimization objective: The goal of deep reinforcement learning is to minimize the error in the virtual environment and optimize the calibration accuracy by adjusting the system's actions. Take Action , and reward according to the correction results , the reward function Defined as the reduction in systematic error: The goal is to maximize long-term rewards To gradually optimize the calibration process, this process is achieved by the joint action of the policy network and the value network of deep reinforcement learning. The policy network determines the corrective action to be taken, while the value network evaluates the expected effect of taking the action in the current state.

[0070] 4.1.4 Training and adjustment: The system uses an environmental simulator and historical data to train a deep reinforcement learning model. In each training step, a virtual feedback signal is used as input to drive the policy network to learn, and the reward function guides the optimization of the learning process. As the training progresses, the model can automatically adapt to dynamic changes in the environment and accurately adjust the calibration parameters of the sensor.

[0071] 4.3 Adaptive adjustment of sensor weights: During the virtual feedback correction loop, the system can adjust the weight distribution of different sensors in real time according to the environment and sensor performance, ensuring that the overall calibration accuracy of the system is maximized under specific motion and environmental conditions. For example, in a high-speed acceleration environment, the IMU may show higher accuracy than the grating sensor, while in a low-speed stable motion, the grating sensor may perform better. Therefore, through the virtual feedback correction loop, the system automatically adjusts the weight of the sensor according to the current state and feedback signal and , so that each sensor can play its best role in a dynamic environment.

[0072] The weight adjustment formula is as follows: in, and It is a dynamic weight adjustment factor related to the performance of each sensor, which is adjusted in real time based on sensor health and virtual error feedback.

[0073] By introducing a virtual feedback correction loop mechanism and a deep reinforcement learning strategy (deep reinforcement learning), this embodiment constructs a highly adaptive calibration process. This process enables the system to maintain accurate calibration accuracy in various dynamic environments by simulating and optimizing calibration behaviors in real time. Through modeling of virtual environments, generation of error feedback, and optimization of deep learning, the virtual feedback correction loop mechanism can not only predict and correct sensor errors, but also adjust the weight distribution of sensors based on real-time feedback to ensure the accuracy of the entire system in complex environments.

[0074] Specifically, the goal of step 5 is to continuously optimize the calibration results using an incremental self-learning mechanism during the long-term operation of the system, ensuring that the errors of the system in actual use are continuously corrected, maintaining high accuracy while adapting to changes in the environment and sensor performance.

[0075] 5.1 Incremental self-learning calibration mechanism: The core goal of the incremental self-learning calibration mechanism is to continuously monitor the accumulation of system errors and perform adaptive incremental optimization based on historical data and real-time feedback. This process is not limited to traditional error compensation model updates, but combines the feedback from the system during actual use to dynamically adjust the calibration strategy and algorithm parameters to ensure the long-term stability and high precision of the system.

[0076] 5.1.1 Error accumulation monitoring and self-learning triggering: During the system operation, the incremental self-learning calibration mechanism will periodically detect the changes in system errors and analyze the error trends. For error monitoring, this embodiment adopts an incremental calculation method based on a time window and sets a dynamic time window. , accumulate the errors in each calibration cycle and perform statistical analysis: in, Indicates that the system is at time The total error, is the cumulative error within the time window. If the cumulative error reaches the preset threshold , the self-learning update mechanism is triggered and the incremental adjustment of the calibration parameters is started.

[0077] 5.1.2 Incremental optimization strategy: Once the error accumulation reaches the trigger condition, the incremental self-learning calibration mechanism will start the incremental optimization process based on historical data (including calibration error, environmental feedback, sensor health, etc.) and real-time data. Specifically, the incremental self-learning calibration mechanism algorithm adopts an incremental learning method to gradually adjust the parameters of the calibration algorithm. This embodiment is optimized in the following ways: Error trend analysis: The system will analyze the changing trends of historical errors and use trend prediction algorithms to predict future error changes. This analysis can be achieved through regression models (such as historical data regression of acceleration error, angular velocity error, grating sensor error, etc.).

[0078] Dynamic adjustment of calibration parameters: According to the error trend and current error accumulation, the incremental self-learning calibration mechanism adjusts the calibration parameters in the error compensation model. Incremental adjustments are made. The adjustment is made through a gradient descent algorithm or an adaptive method based on Bayesian inference. , the incremental self-learning calibration mechanism calculates the increment of the parameters : in, is the learning rate, It represents the sensitivity to the current error, that is, the gradient of the calibration parameter. Through this incremental adjustment, the calibration process can gradually optimize the calibration results without interfering with the system operation.

[0079] 5.1.3 Incremental error correction model: In the incremental optimization process, the incremental self-learning calibration mechanism does not simply compensate for errors, but deeply integrates information such as interaction with the dynamic environment and changes in sensor performance to adjust multiple parameters in the calibration model. Specifically, the incremental self-learning calibration mechanism combines the grating sensor and IMU error models in the previous steps and dynamically corrects each error model during the incremental update. Assuming the system error is the weighted sum of multiple sensor errors, then the incremental optimization process can be expressed by the following formula: in, For each sensor The weight of For sensor The incremental self-learning calibration mechanism updates the weight of each sensor incrementally. , thereby dynamically adjusting the contribution ratio of different sensors during the calibration process.

[0080] 5.2 Self-learning update and algorithm parameter adjustment: The incremental self-learning calibration mechanism system not only makes incremental corrections to the error compensation model, but also gradually adjusts the core parameters of the algorithm through self-learning. Traditional calibration methods usually rely on static model parameters, while the incremental self-learning calibration mechanism can adaptively adjust the algorithm parameters based on the actual operation history of the system and long-term data accumulation through the self-learning mechanism, ensuring the stability and accuracy of the system during long-term use.

[0081] 5.2.1 Self-learning update based on feedback: At the end of each calibration cycle, the incremental self-learning calibration mechanism optimizes the algorithm parameters during the calibration process based on the current error and compensation results. This process is done by updating the calibration algorithm based on historical data feedback (e.g., dynamically adjusting the calibration learning rate). , allowing the system to continuously adjust itself based on long-term data.

[0082] in, is the initial learning rate, is the adjustment factor. As the error accumulates, the learning rate It will gradually decrease, reflecting that the system has entered a more stable learning state.

[0083] 5.2.2 Adaptive algorithm parameter update: In addition to the calibration parameters in the error compensation model, the incremental self-learning calibration mechanism will also adaptively adjust other core parameters of the algorithm, such as sensor weight distribution, error correction speed, etc., based on the historical data of the system operation. The update of these parameters depends not only on error feedback, but also on the dynamic changes of the environment and the specific conditions of the system operation. For example, a grating sensor may perform better in low-speed stable motion, while an IMU is more accurate in high-speed motion. The incremental self-learning calibration mechanism will dynamically adjust the weights of each sensor based on this real-time information to ensure the calibration accuracy of the system in different scenarios, by dynamically adjusting the weight function To achieve: in, It is a dynamic adjustment factor related to the performance of each sensor, updated in real time based on the historical trend of errors and the current operating environment.

[0084] By introducing an incremental self-learning calibration mechanism, this embodiment designs a calibration optimization mechanism that can adapt to environmental changes and sensor performance evolution during long-term operation. The incremental self-learning calibration mechanism system ensures that the system can continuously correct and optimize the calibration error during long-term operation and always maintain high precision through steps such as periodic error accumulation monitoring, incremental optimization strategy, and real-time feedback self-learning.

[0085] Example 2 like Figure 2 As shown, the present invention discloses an inertial navigation system calibration device based on a grating sensor, comprising: Environmental perception and raw data acquisition module: Dynamically adjust the sampling frequency and data acquisition strategy of grating sensors and IMU through adaptive adjustment mechanism of perceived environment, real-time sensor health assessment model and real-time motion state classifier; Interactive dynamic error allocation module: Based on the real-time data interaction between the grating sensor and the IMU, an interactive dynamic error allocation algorithm is designed to dynamically adjust the sensor error correction weight according to the motion state and environmental changes, optimize the error compensation effect, and provide an error correction strategy for calibration; Dynamic error correction module: using adaptive partial differential equation algorithm, dynamically adjusts the error compensation strategy by correcting the partial differential equation of the error model and adaptive correction coefficient in real time; Adaptive calibration module: introduces a virtual feedback correction loop mechanism and deep reinforcement learning strategy, uses virtual environment models and real-time feedback signals to optimize calibration behavior, dynamically adjusts sensor weights, and achieves adaptive calibration; Calibration parameter optimization module: It adopts an incremental self-learning calibration mechanism, performs adaptive incremental optimization based on error accumulation monitoring and historical data during operation, and dynamically adjusts calibration parameters.

[0086] Specifically, the environmental perception and raw data acquisition module is divided into three sub-modules: the perception environment adaptive adjustment module, the real-time sensor health assessment module, and the real-time motion state classification module. First, the perception environment adaptive adjustment module monitors environmental changes (such as vibration, acceleration, temperature, etc.) in real time through the sensor's own environmental perception module (such as accelerometer, temperature and humidity sensor, etc.). According to the intensity of environmental changes, the sampling frequency of the grating sensor and IMU is dynamically adjusted. For example, when there is strong vibration or large acceleration changes, the sampling frequency of the grating sensor is increased to 10kHz, and the sampling frequency of the IMU is increased to 1kHz; when the environment is stable, the sampling frequency is reduced to reduce redundant data and extend the life of the sensor. The change in sampling frequency is achieved through dynamic threshold judgment, taking into account factors such as acceleration, vibration intensity, and sensor output noise.

[0087] Secondly, the real-time sensor health assessment module quantifies the health of the sensor based on indicators such as sensor offset, drift, noise level, etc. If the health is lower than the threshold, the system will mark low-quality data through the "health alarm mechanism" and adjust the sampling frequency or trigger recalibration to ensure the credibility and accuracy of the data.

[0088] Finally, the real-time sensor health assessment module includes a real-time motion state classifier, which uses machine learning methods such as deep neural networks (DNN) or support vector machines (SVM) to identify the current motion state (such as uniform linear motion, acceleration, rotation, etc.) in real time based on the sensor's acceleration, angular velocity, and displacement data. According to different motion states, the system adjusts the sampling frequency and calibration strategy, for example, giving priority to the stability of the IMU in rotational motion, and giving priority to the accuracy of the grating sensor in uniform motion.

[0089] Specifically, the interactive dynamic error allocation module is divided into three submodules: the motion state recognition and sensor interaction analysis module, the error allocation model construction module, and the real-time adjustment mechanism module of the dynamic error correction weight. First, in the motion state recognition and sensor interaction analysis module, based on the real-time motion state classifier introduced in the environment perception and raw data acquisition module, the system can judge the current motion state (such as uniform linear motion, acceleration, rotation, etc.) in real time. Under different motion states, the errors of the grating sensor and IMU behave differently. For example, in rotational motion, the error of the grating sensor may be amplified due to the moment of inertia effect, while in linear uniform motion, the acceleration error of the IMU usually becomes dominant. Through the output information of the motion state classifier, the system decides how to adjust the error correction weight of the sensor.

[0090] Secondly, in the construction module of the error allocation model, the system dynamically adjusts the error correction ratio of the grating sensor and IMU based on real-time sensor data such as acceleration, rotational angular velocity, displacement, etc. The dynamic allocation of weights depends not only on the real-time data of the sensor, but also on the motion state and environmental feedback information (such as vibration, temperature, etc.). For example, in the rotational motion stage, the system may rely on IMU data for correction first, while in the linear motion stage, it will rely more on grating sensor data for correction.

[0091] Finally, in the real-time adjustment mechanism module of dynamic error correction weights, the system optimizes the error correction process through an adaptive weight adjustment mechanism. Based on real-time sensor data and historical data of system errors, the dynamic error allocation model algorithm continuously monitors the errors of the grating sensor and IMU, and optimizes them using an adaptive weight adjustment model. When a sensor error is detected to be large, the system automatically increases the correction weight of the sensor and reduces the correction weight of another sensor accordingly.

[0092] Specifically, the dynamic error correction module is divided into three submodules: the partial differential equation construction module of the error model, the adaptive adjustment module of the error correction coefficient, and the real-time solution and system update module of the partial differential equation. First, in the partial differential equation construction module of the error model, the system assumes that the errors of the grating sensor and IMU change dynamically in time and space, and introduces partial differential equations to describe the evolution of these errors. The partial differential equations take into account the influence of acceleration, angular velocity, and external environmental factors (such as temperature, vibration, etc.) on the error.

[0093] Secondly, in the adaptive adjustment module of the error correction coefficient, the system introduces an adaptive correction factor to adjust the error correction strength in real time according to factors such as acceleration, angular velocity and sensor health. For example, when the acceleration or angular velocity changes suddenly, the system will give priority to adjusting the correction weight of the IMU; when the environmental changes are small, the correction center of gravity will be turned to the grating sensor.

[0094] Finally, in the real-time solution and system update module of partial differential equations, the system uses the finite difference method to discretize and solve the partial differential equations to achieve dynamic adjustment of the error correction coefficient. By setting the time step, the system updates the error correction model within each time step to track and correct the sensor error in real time.

[0095] Specifically, the adaptive calibration module is divided into two submodules: the virtual feedback correction loop mechanism module and the deep reinforcement learning strategy module. First, in the virtual feedback correction loop mechanism module, the system builds a virtual environment model to simulate the impact of various factors in the real environment (such as vibration, acceleration, temperature change, etc.) on the sensor output. By comparing the actual output of the sensor with the expected output of the virtual model, an error feedback signal is generated. This feedback signal is not directly used for the control system, but is passed to the artificial intelligence algorithm to simulate and optimize the next calibration behavior.

[0096] Secondly, in the deep reinforcement learning strategy module, the system uses the deep reinforcement learning algorithm to adaptively adjust the calibration strategy based on real-time errors and environmental feedback. The state space includes multi-dimensional information such as the current output of the sensor, error feedback, sensor health, acceleration and angular velocity. The action space includes the intensity, direction and weight distribution of the sensor error correction. The reward function is defined as the reduction in system error, and the calibration process is optimized by maximizing long-term rewards. Deep reinforcement learning determines the correction action through the policy network, and the value network evaluates the expected effect, thereby achieving dynamic optimization.

[0097] Specifically, the calibration parameter optimization module includes two submodules: the error accumulation monitoring and self-learning triggering module and the incremental optimization strategy module. First, in the error accumulation monitoring and self-learning triggering module, the system periodically monitors the changes in system errors through an incremental calculation method based on a time window. A dynamic time window is set to accumulate and statistically analyze the errors within each calibration cycle. If the accumulated error reaches the preset threshold, the self-learning update mechanism is triggered to start the incremental adjustment of the calibration parameters.

[0098] Secondly, in the incremental optimization strategy module, the system starts the incremental optimization process based on historical data (including calibration error, environmental feedback, sensor health, etc.) and real-time data. By analyzing the trend of historical errors, trend prediction algorithms (such as regression models) are used to predict future error changes and dynamically adjust calibration parameters.

[0099] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein with equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the scope of protection of the claims of the present invention.

Claims

1. A calibration method for an inertial navigation system based on a grating sensor, characterized in that: include: Step 1: Dynamically adjust the sampling frequency and data acquisition strategy of the grating sensor and IMU through the adaptive adjustment mechanism of the perceived environment, the real-time sensor health assessment model and the real-time motion state classifier; Step 2: Based on the real-time data interaction between the grating sensor and the IMU, an interactive dynamic error allocation algorithm is designed to dynamically adjust the sensor error correction weight according to the motion state and environmental changes, optimize the error compensation effect, and provide an error correction strategy for calibration; Step 3: Using the adaptive partial differential equation algorithm, the error compensation strategy is dynamically adjusted by correcting the partial differential equation of the error model and the adaptive correction coefficient in real time; Step 4: Introduce a virtual feedback correction loop mechanism and deep reinforcement learning strategy, use the virtual environment model and real-time feedback signals to optimize the calibration behavior, dynamically adjust the sensor weights, and achieve adaptive calibration; Step 5: Adopt an incremental self-learning calibration mechanism to perform adaptive incremental optimization and dynamically adjust calibration parameters based on error accumulation monitoring and historical data during operation.

2. The method for calibrating an inertial navigation system based on a grating sensor according to claim 1, characterized in that: The step 1 comprises: The environmental perception module monitors environmental changes in real time and dynamically adjusts the sampling frequency of the grating sensor and IMU according to the intensity of environmental changes; Quantify the health of the sensor based on the sensor's offset, drift, and noise level indicators; if the health is below the threshold, mark the low-quality data through the health alert mechanism, and adjust the sampling frequency or trigger recalibration; Use machine learning methods to identify the current motion state in real time, and adjust the sampling frequency and calibration strategy according to different motion states.

3. The method for calibrating an inertial navigation system based on a grating sensor according to claim 2, characterized in that: The step 2 comprises: Determine the current motion state in real time and decide how to adjust the sensor's error correction weight; According to the real-time motion state, the error correction ratio of the grating sensor and IMU is dynamically adjusted. In the rotational motion stage, the correction is based on the IMU data, and in the linear motion stage, the correction is based on the grating sensor data. Based on real-time sensor data and historical error data, the dynamic error allocation model algorithm monitors the errors of the grating sensor and IMU, and optimizes them using an adaptive weight adjustment model; when a large sensor error is detected, the correction weight of the sensor is increased and the correction weight of another sensor is reduced accordingly.

4. The method for calibrating an inertial navigation system based on a grating sensor according to claim 3, characterized in that: The step 3 comprises: Partial differential equations are introduced to describe the errors of grating sensors and IMU in time and space to dynamically reflect the changes of sensor errors over time while considering the interactions between sensors. Introduce adaptive correction factors to adjust the error correction strength in real time. When the acceleration or angular velocity changes suddenly, the correction weight of the IMU is adjusted first. When the environmental changes are small, the center of gravity is corrected to the grating sensor. The finite difference method is used to discretize and solve the partial differential equation to achieve dynamic adjustment of the error correction coefficient. By setting the time step, the error correction model is updated in each time step to track and correct the sensor error in real time. At the same time, a dynamic error correction rate adjustment mechanism is introduced to adjust the correction speed according to the overall error feedback. When the error reaches the set threshold, the correction rate is accelerated. When the error is small, the correction rate is moderately slowed down to avoid error oscillation caused by excessive correction.

5. The method for calibrating an inertial navigation system based on a grating sensor according to claim 4, characterized in that: The step 4 comprises: Construct a virtual environment model to simulate the impact of various factors in the real environment on the sensor output; generate an error feedback signal by comparing the actual output of the sensor with the expected output of the virtual model; the feedback signal is passed to the artificial intelligence algorithm to simulate and optimize the next calibration behavior; The deep reinforcement learning algorithm is used to adaptively adjust the calibration strategy according to the real-time error and environmental feedback. The state space includes the current output of the sensor, error feedback, sensor health, acceleration and angular velocity. The action space includes the intensity, direction and weight distribution of the sensor error correction. The reward function is defined as the reduction in error, and the calibration process is optimized by maximizing the long-term reward. Deep reinforcement learning determines the correction action through the policy network, and the value network evaluates the expected effect, thereby achieving dynamic optimization.

6. The method for calibrating an inertial navigation system based on a grating sensor according to claim 5, characterized in that: The step 5 comprises: Through the incremental calculation method based on the time window, the changes of system errors are monitored periodically; a dynamic time window is set to accumulate and statistically analyze the errors in each calibration cycle. If the accumulated error reaches the preset threshold, the self-learning update mechanism is triggered to start the incremental adjustment of the calibration parameters; Based on historical data and real-time data, an incremental optimization process is started. By analyzing the changing trend of historical errors, a trend prediction algorithm is used to predict future error changes and dynamically adjust the calibration parameters.

7. An inertial navigation system calibration device based on a grating sensor, characterized in that: include: Environmental perception and raw data acquisition module: Dynamically adjust the sampling frequency and data acquisition strategy of grating sensors and IMU through adaptive adjustment mechanism of perceived environment, real-time sensor health assessment model and real-time motion state classifier; Interactive dynamic error allocation module: Based on the real-time data interaction between the grating sensor and the IMU, an interactive dynamic error allocation algorithm is designed to dynamically adjust the sensor error correction weight according to the motion state and environmental changes, optimize the error compensation effect, and provide an error correction strategy for calibration; Dynamic error correction module: using adaptive partial differential equation algorithm, dynamically adjusts the error compensation strategy by correcting the partial differential equation of the error model and adaptive correction coefficient in real time; Adaptive calibration module: introduces a virtual feedback correction loop mechanism and deep reinforcement learning strategy, uses virtual environment models and real-time feedback signals to optimize calibration behavior, dynamically adjusts sensor weights, and achieves adaptive calibration; Calibration parameter optimization module: It adopts an incremental self-learning calibration mechanism, performs adaptive incremental optimization based on error accumulation monitoring and historical data during operation, and dynamically adjusts calibration parameters.

8. The grating sensor-based inertial navigation system calibration device according to claim 7, characterized in that: The environment perception and raw data acquisition module includes: Perception environment adaptive adjustment module: The environment perception module monitors environmental changes in real time and dynamically adjusts the sampling frequency of the grating sensor and IMU according to the intensity of environmental changes; Real-time sensor health assessment module: quantifies sensor health based on sensor offset, drift, and noise level indicators; if the health is below the threshold, the health alarm mechanism marks low-quality data and adjusts the sampling frequency or triggers recalibration; Real-time motion state classification module: uses machine learning methods to identify the current motion state in real time, and adjusts the sampling frequency and calibration strategy according to different motion states; The interactive dynamic error allocation module includes: Motion state recognition and sensor interaction analysis module: determines the current motion state in real time and decides how to adjust the sensor's error correction weight; Building blocks of the error allocation model: dynamically adjust the error correction ratio of the grating sensor and IMU according to the real-time motion state. In the rotational motion stage, the error correction is based on the IMU data, and in the linear motion stage, the error correction is based on the grating sensor data. Real-time adjustment mechanism module of dynamic error correction weight: Based on real-time sensor data and historical error data, the dynamic error allocation model algorithm monitors the errors of the grating sensor and IMU, and optimizes them using the adaptive weight adjustment model; when a large sensor error is detected, the correction weight of the sensor is increased, and the correction weight of another sensor is reduced accordingly.

9. The grating sensor-based inertial navigation system calibration device according to claim 8, characterized in that: The dynamic error correction module comprises: Partial differential equation building blocks of the error model: Partial differential equations are introduced to describe the errors of the grating sensor and IMU in time and space to dynamically reflect the changes in sensor errors over time while considering the interactions between sensors; Adaptive adjustment module of error correction coefficient: introduce adaptive correction factor to adjust the error correction strength in real time. When the acceleration or angular velocity changes suddenly, the correction weight of IMU is adjusted first. When the environmental change is small, the center of gravity is corrected to the grating sensor. Real-time solution and system update module of partial differential equations: The finite difference method is used to discretize and solve the partial differential equations to achieve dynamic adjustment of the error correction coefficient; by setting the time step, the error correction model is updated in each time step to track and correct the sensor error in real time; at the same time, a dynamic error correction rate adjustment mechanism is introduced to adjust the correction speed according to the overall error feedback; when the error reaches the set threshold, the correction rate is accelerated; when the error is small, the correction rate is moderately slowed down to avoid error oscillation caused by excessive correction.

10. The grating sensor-based inertial navigation system calibration device according to claim 9, characterized in that: The adaptive calibration module comprises: Virtual feedback correction loop mechanism module: builds a virtual environment model to simulate the impact of various factors in the real environment on the sensor output; generates an error feedback signal by comparing the actual output of the sensor with the expected output of the virtual model; the feedback signal is passed to the artificial intelligence algorithm to simulate and optimize the next calibration behavior; Deep reinforcement learning strategy module: uses deep reinforcement learning algorithm to adaptively adjust the calibration strategy according to real-time error and environmental feedback; the state space includes the current output of the sensor, error feedback, sensor health, acceleration and angular velocity; the action space includes the intensity, direction and weight distribution of sensor error correction; the reward function is defined as the amount of error reduction, and the calibration process is optimized by maximizing long-term rewards; deep reinforcement learning determines the correction action through the policy network, and the value network evaluates the expected effect, thereby achieving dynamic optimization; The calibration parameter optimization module includes: Error accumulation monitoring and self-learning triggering module: Through the incremental calculation method based on the time window, the system error changes are periodically monitored; a dynamic time window is set to accumulate and statistically analyze the errors in each calibration cycle. If the accumulated error reaches the preset threshold, the self-learning update mechanism is triggered to start the incremental adjustment of the calibration parameters; Incremental optimization strategy module: Based on historical data and real-time data, the incremental optimization process is started. By analyzing the changing trend of historical errors, the trend prediction algorithm is used to predict future error changes and dynamically adjust the calibration parameters.

Images