A Strapdown Inertial Navigation System Error Compensation Method Based on Neural Network

Abstract

The invention discloses a strapdown inertial navigation system error compensation method based on a neural network, including: S1, constructing a deep neural network model including multiple neural network layers; S2, collecting actual operation data of the strapdown inertial navigation system under different conditions, the actual operation data including the measurement values ​​of the accelerometer and the gyroscope, and position information; S3, pre-processing the collected actual operation data of the strapdown inertial navigation system under different conditions; S4, dividing the pre-processed operation data into a certain proportion of training set, test set and verification set; S5, using the training set to train the deep neural network model constructed in step S1; S6, integrating the final deep neural network model into the strapdown inertial navigation system, collecting the output data of the strapdown inertial navigation system in real time, and inputting it into the deep neural network model, so as to obtain error estimation values ​​and adjust navigation parameters. The invention has the advantages of strong generalization ability and good adaptability.

Description

Technical Field

[0001] The invention relates to the field of satellite navigation technology, and more specifically, to a strapdown inertial navigation system error compensation method based on a neural network. Background Art

[0002] As an autonomous navigation system, the strapdown inertial navigation system plays an important role in modern navigation technology. It measures the acceleration of the carrier in the inertial reference system and, after a series of complex integral operations, accurately obtains the carrier's speed, position and attitude information, thereby achieving navigation and positioning. With its unique autonomy and independence, the inertial navigation system has been widely used in military, aviation, aerospace and other fields. The basic principle of the inertial navigation system is based on Newton's laws of motion. Inertial sensors installed on the carrier are used to sense the motion state of the carrier. These sensors mainly include gyroscopes and accelerometers. The gyroscope, with its unique rotation characteristics, can accurately measure the angular velocity of the carrier and provide key attitude information for the navigation system. The accelerometer can sense the linear acceleration of the carrier in real time, providing data support for the calculation of speed and position.

[0003] However, during long-term operation, the strapdown inertial navigation system will inevitably produce certain errors due to the interference of many internal and external factors, such as the performance limitations of the equipment itself, the influence of the external environment, and the defects of the data processing algorithm. These errors will continue to accumulate during the operation of the system, causing the positioning information it provides to gradually deviate from the true value, which may eventually make the navigation results unreliable or even completely ineffective. As the core component of the inertial navigation system, the performance of the inertial sensor directly determines the positioning accuracy of the entire system. In practical applications, inertial sensors will inevitably have certain errors due to the influence of various factors such as manufacturing process, material properties, and working environment. These errors include zero bias error, scale factor error, cross-coupling error, etc., which will gradually accumulate over time and have a serious impact on the positioning accuracy of the inertial navigation system. In response to this problem, how to take effective measures to suppress the accumulation of inertial sensor errors and thus improve the positioning accuracy of the inertial navigation system has become a research focus that has received much attention in the current field of navigation technology. Researchers are committed to exploring more advanced algorithms, optimizing data processing processes, improving hardware performance, and trying to introduce auxiliary navigation systems, in order to significantly reduce the negative impact of error accumulation and ensure that the inertial navigation system can still provide high-reliability positioning services in complex and changing environments. To this end, it is necessary to develop a strapdown inertial navigation system error compensation method based on neural networks. Summary of the invention

[0004] The purpose of the present invention is to provide a strapdown inertial navigation system error compensation method based on neural network to overcome the defects of the prior art.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0006] A strapdown inertial navigation system error compensation method based on neural network comprises the following steps:

[0007] S1. Build a deep neural network model including multiple neural network layers;

[0008] S2. collecting actual operation data of the strapdown inertial navigation system under different conditions, the actual operation data including measurement values ​​of the accelerometer and the gyroscope, and position information;

[0009] S3, preprocessing the collected actual operation data of the strapdown inertial navigation system under different conditions;

[0010] S4, dividing the preprocessed running data into a training set, a test set, and a validation set with a certain proportion;

[0011] S5, using the training set to train the deep neural network model constructed in step S1, by minimizing the difference between the error estimate value output by the deep neural network model and the true error value to optimize the parameters of the deep neural network model, and using the validation set to validate the trained deep neural network model to obtain the final deep neural network model;

[0012] S6. Integrate the final deep neural network model into the strapdown inertial navigation system, collect the output data of the strapdown inertial navigation system in real time, and input it into the deep neural network model to obtain an error estimate, and adjust the navigation parameters according to the error estimate.

[0013] Furthermore, the neural network model in step S1 adopts a long short-term memory network, and the neural network model adopts regularization technology.

[0014] Furthermore, the measurement value of the gyroscope in step S2 is the angular velocity information between the carrier coordinate system and the navigation coordinate system, and the acquired angular velocity information is corrected, and the attitude quaternion is updated based on the corrected data to solve the converted attitude matrix to obtain accurate attitude angle information; the accelerometer is used to measure the specific force information inside the carrier coordinate system, and the specific force information is converted to the navigation coordinate system using the attitude matrix, and then processed through integral operation to derive the speed and position information of the carrier.

[0015] Furthermore, the obtained speed information is combined with the position rate equation to solve to obtain the angular velocity of rotation relative to the ground in the navigation coordinate system.

[0016] Furthermore, the step S2 further includes:

[0017] Construct gyroscope error model and accelerometer error model;

[0018] The measurements of the accelerometer and gyroscope are corrected by the gyroscope error model and the accelerometer error model respectively.

[0019] Furthermore, the detailed formula of the gyroscope error model is: ;

[0020] In the formula, is the scale factor error, is the zero bias error, Non-rectangular coordinate system caused by installation angle error Coordinate system and rectangular coordinate system The coordinate transformation matrix between systems is obtained by parallelizing the above gyroscope error model:

[0021] ;

[0022] Where: ;

[0023] In the formula, is the theoretical angular velocity, is the angular velocity actually measured by the gyroscope, is the zero drift value, is the scale error matrix, is the scale factor error, For installation error, is the misalignment angle error;

[0024] Will By transposing the terms, the error model expression of the gyroscope can be obtained as follows:

[0025] ;

[0026] Similarly, the expression of the accelerometer error model is: ;

[0027] In the formula, ;

[0028] In the formula, is the theoretical specific force increment, is the specific force increment actually measured by the accelerometer, Measure the zero bias of the accelerometer, is the accelerometer scale error matrix, is the scale factor error, is the misalignment angle error, This is installation error.

[0029] Furthermore, the step S3 is specifically as follows:

[0030] Subtract the mean value of each feature from the value of that feature and divide it by its standard deviation. The mathematical formula is as follows:

[0031] ;

[0032] In the formula, and Represent the mean and standard deviation of the data set, respectively. It is the new data after standardization.

[0033] Furthermore, in step S4, the training set accounts for 60%, the validation set accounts for 20%, and the test set accounts for 20%.

[0034] Furthermore, the loss function used in step S5 is mean square error loss, which is expressed as: .

[0035] Furthermore, in step S5, an evaluation index is used to evaluate the deep neural network model, and the expression of the evaluation index is:

[0036] ;

[0037] In the formula, y i is the true value corresponding to the i-th sample, is the predicted value of the i-th sample obtained by the model, is the sample mean, n is the number of samples, TP, FP and FN represent true positive, false positive and false negative respectively;

[0038] Evaluation criteria: R 2 The closer it is to 1, the better the model fitting effect is. 2 The closer it is to 0, the worse the model fitting effect is; the closer ACC is to 1, the higher the prediction accuracy of the model is.

[0039] Compared with the prior art, the advantages of the present invention are: the present invention uses a neural network to estimate and compensate for the errors of the strapdown inertial navigation system in real time, effectively reducing the impact of the errors on the positioning accuracy of the navigation system. The neural network model can learn the complex relationship between the error and the motion state of the carrier, and predict the error in real time during actual operation, thereby achieving accurate compensation for the error. This method improves the stability and reliability of the navigation system, and is particularly suitable for application scenarios that require long-term continuous operation and high-precision positioning.

[0040] The present invention has the advantages of strong generalization ability and good adaptability. By collecting actual operation data under different conditions, a neural network model that adapts to various working environments can be trained. Compared with the traditional error compensation method based on mathematical models or physical models, the present invention adopts a neural network model for error estimation and compensation, without the need to establish a complex mathematical model, thereby reducing the difficulty of algorithm design and implementation. At the same time, the neural network model can handle nonlinear and non-stationary error characteristics and better adapt to complex and changeable navigation environments. This makes the present invention have strong versatility and scalability, and can be widely used in various navigation systems.

[0041] The present invention is flexible and scalable. In practical applications, with the continuous development of technology and the continuous improvement of system requirements, the neural network model can be improved and optimized as needed. For example, more network layers or neurons can be added to improve the complexity and learning ability of the model; different neural network architectures can be used, such as recurrent neural networks (RNNs), long short-term memory networks (LSTMs), etc., to better process time series data; advanced concepts such as transfer learning can also be introduced to use the trained model for fine-tuning to accelerate the training process of the new model and improve performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0043] Figure 1 The invention discloses a flow chart of a strapdown inertial navigation system error compensation method based on a neural network. DETAILED DESCRIPTION

[0044] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby making a clearer and more definite definition of the protection scope of the present invention.

[0045] See also Figure 1 As shown, this embodiment discloses a strapdown inertial navigation system error compensation method based on a neural network, comprising the following steps:

[0046] Step S1, constructing a deep neural network model including multiple neural network layers;

[0047] Step S2, collecting actual operation data of the strapdown inertial navigation system under different conditions, the actual operation data including measurement values ​​of the accelerometer and the gyroscope, and position information;

[0048] Step S3, preprocessing the collected actual operation data of the strapdown inertial navigation system under different conditions;

[0049] Step S4, dividing the preprocessed running data into a training set, a test set and a validation set with a certain proportion;

[0050] Step S5, using the training set to train the deep neural network model constructed in step S1, optimizing the parameters of the deep neural network model by minimizing the difference between the error estimate value output by the deep neural network model and the true error value, and using the validation set to validate the trained deep neural network model to obtain a final deep neural network model;

[0051] Step S6: Integrate the final deep neural network model into the strapdown inertial navigation system, collect the output data of the strapdown inertial navigation system in real time, and input it into the deep neural network model to obtain an error estimate, and adjust the navigation parameters according to the error estimate.

[0052] In this embodiment, the neural network model in step S1 adopts a long short-term memory network (LSTM). Through such a multi-layer structure, the model can learn the complex nonlinear relationship between the strapdown inertial navigation system error and the system input. This model should have sufficient complexity to accurately capture the subtle changes in the error, and it should also have good generalization ability, so that it can effectively adapt and process under various working conditions.

[0053] The long short-term memory network (LSTM) has a network structure with memory function. LSTM consists of three parts: forget gate, input gate and output gate:

[0054] Input Gate: The key role of the input gate is to determine which information should be updated to the cell state. Its input also contains the hidden state of the previous time step and the input data of the current time step. The input gate consists of two parts: one is the sigmoid activation function, which is used to determine which information needs to be updated; the other is the tanh activation function, which is used to generate candidate cell state values. Subsequently, the output of the sigmoid is multiplied element by element with the output of the tanh to obtain the cell state value that needs to be updated.

[0055] Output Gate: The main responsibility of the output gate is to decide which information should be passed from the cell state to the hidden state of the next time step. Its input also contains the hidden state of the previous time step and the input data of the current time step. First, the output gate value is calculated by the sigmoid activation function, then the cell state is activated by tanh, and finally the two are multiplied element by element to obtain the hidden state of the next time step.

[0056] Through the interaction between these three gating units, the LSTM network can effectively learn the long-term dependencies in the sequence and retain or forget relevant information as needed. This makes LSTM perform better when processing long sequence data. These three together constitute a complete gating mechanism, which is expressed as:

[0057] ;

[0058] ;

[0059] ;

[0060] in, W f 、W i 、W C and W o They represent the weight matrices of the neural network, and they each carry the weight information of the network in key links such as the forget gate, input gate, memory gate, and output gate. b f 、b i 、b C 、b o They correspond to the coefficients of the forget gate, input gate, memory gate and output gate respectively, which play an important regulatory role in the network. In addition, ht-1 and ht represent the output of the previous moment and the current moment respectively, and they record the state information of the network at different time steps. And xt represents the input at the current moment, which provides new data for the network to learn and process. In order to further improve the generalization ability of the model, this embodiment adopts a regularization technology - Dropout to prevent the occurrence of overfitting. Dropout technology effectively reduces the complexity of the model by randomly discarding the connections of some neurons during the training process, thereby improving its adaptability to new data.

[0061] In this embodiment, in step S2, the strapdown inertial navigation system is equipped with a gyroscope and an accelerometer, both of which are installed and arranged strictly in accordance with the carrier coordinate system. The gyroscope is mainly responsible for accurately measuring the angular velocity information between the carrier coordinate system (b system) and the navigation coordinate system (n system), and correcting the acquired angular velocity information, updating the attitude quaternion based on the corrected data to solve the converted attitude matrix to obtain accurate attitude angle information; the accelerometer is used to measure the specific force information inside the carrier coordinate system, and the specific force information is converted to the navigation coordinate system using the attitude matrix, and then processed by integral operation to derive the speed and position information of the carrier. In addition, the system will use the obtained velocity information, combined with the position rate equation for accurate solution, to obtain the angular velocity of rotation relative to the ground in the navigation coordinate system, thereby providing crucial data support for subsequent attitude updates.

[0062] Attitude parameters are the core theoretical basis of the strapdown inertial navigation system, among which: Euler angles and quaternions are two commonly used description methods. Here, this embodiment uses the Euler angle method as the main description method. According to the precise definition of the Euler angle method, when this embodiment transforms a reference coordinate system to a target coordinate system, the angle of rotation around a coordinate axis of the passive coordinate system is the Euler angle. Therefore, this method occupies an important position in attitude description with its intuitive and clear characteristics. The direction cosine matrix based on the Euler angle method is expressed as:

[0063] ;

[0064] In the formula, Represents an orthogonal matrix. The transformation from the platform coordinate system to the carrier coordinate system is generally through the Z axis → X axis → Y axis. The rotation angles are the heading angle, pitch angle, and roll angle, respectively. , and In discussing the influence of rotation angle on the direction cosine matrix, when the rotation angles are all large, the specific form of the direction cosine matrix will directly depend on the order of rotation. However, when the three rotation angles involved are all small, by applying the linear approximation method, this embodiment can derive a direction cosine matrix with the same characteristics. In this case, The calculation process can be further simplified into a more concise form, which is expressed as:

[0065] ;

[0066] In the formula, , Represents an antisymmetric matrix. The main values ​​of the three platform attitude angles can be accurately calculated and determined by the following formulas, and their expressions are:

[0067] .

[0068] The following is the derivation process of the Euler angle differential equation. Based on the rotation order of the Euler angle, the following formula can be obtained:

[0069] ;

[0070] Taking the inverse of it, we get the following formula: ;

[0071] Based on this, the present embodiment is able to derive the differential equation of the Euler angle. The Euler angle method can be effectively implemented by solving three differential equations. However, it should be noted that the transcendental function operations involved in this process will significantly increase the complexity and amount of calculation. When , the differential equation of the direction cosine matrix is ​​as follows: ;

[0072] In the solution process, this expression involves solving nine differential equations, however, by applying simplified mathematical operations, the solution of the direction cosine matrix can be directly obtained without the need for complex calculation steps.

[0073] Wherein, the step S2 further includes:

[0074] Construct gyroscope error model and accelerometer error model;

[0075] The measurements of the accelerometer and gyroscope are corrected by the gyroscope error model and the accelerometer error model respectively.

[0076] Among them, the detailed formula of the gyroscope error model is: ;

[0077] In the formula, The scale factor error, also known as the scale coefficient error, is an important form of gyroscope error. It mainly comes from the manufacturing accuracy of the dial or scale inside the sensor, resulting in a fixed proportional deviation between the measured value and the true value. This error can be reduced by accurately calibrating the sensor. Bias error is also called gyro drift. It means that the gyro outputs a non-zero angular velocity value without external input. This error may be caused by many factors, such as temperature change, mechanical vibration, electromagnetic interference, etc. In order to reduce the bias error, technical means such as temperature compensation and filtering can be used. Non-rectangular coordinate system caused by installation angle error Coordinate system and rectangular coordinate system The coordinate transformation matrix between the coordinate systems and the installation angle error are also important factors that cause gyroscope errors. Due to inaccuracies or errors in the installation process, there is a certain angle deviation between the sensitive axis of the gyroscope and the theoretical rectangular coordinate system. This deviation will cause errors in the measurement data during the conversion process between coordinate systems. The above gyroscope error model is parallelized to obtain: ;

[0078] In the formula, ;

[0079] Where: is the theoretical angular velocity, It is the angular velocity actually measured by the gyroscope. This parameter reflects the angular velocity value measured by the gyroscope under actual working conditions. It is an important basis for error analysis and correction. is the zero drift value, is the scale error matrix, is the scale factor error, For installation error, is the misalignment angle error.

[0080] Will By transposing the terms, the error model expression of the gyroscope can be obtained as follows: ;

[0081] Similarly, the expression of the accelerometer error model is: ;

[0082] In the formula, ;

[0083] In the formula, is the theoretical specific force increment, The specific force increment actually measured by the accelerometer. The theoretical specific force increment and the actual measured specific force increment are two crucial parameters, which together form the basis for the accuracy evaluation of the navigation system. Measure the zero bias of the accelerometer, is the accelerometer scale error matrix, is the scale factor error, is the misalignment angle error, This is installation error.

[0084] In step S3, in order to train the neural network model, the present embodiment needs to collect a large amount of actual operation data of the strapdown inertial navigation system under different conditions. These data should include the measurements of the accelerometer and the gyroscope, and the corresponding reference position information. By collecting data in different motion states and different working environments, the present embodiment can enable the model to better learn the variation law of the error, thereby improving its generalization ability. In the process of collecting data, the present embodiment also needs to pre-process the data. As a data pre-processing method, data standardization has the core purpose of eliminating the dimension and scale differences between different features. Through this processing method, the data can be unified in specifications, which helps the LSTM model to converge more efficiently and significantly improve its performance.

[0085] The data preprocessing is as follows:

[0086] Subtract the mean value of each feature from the value of that feature and divide it by its standard deviation. The mathematical formula is as follows: ;

[0087] In the formula, and Represent the mean and standard deviation of the data set, respectively. It is new data after standardization. The core purpose of data standardization is to adjust data of different units or magnitudes to a unified scale, that is, to ensure that the mean of the data is 0 and the standard deviation is 1, so as to facilitate subsequent comparison and analysis. This process aims to eliminate the deviation of model results that may be caused by differences in magnitude between data, thereby improving the accuracy and stability of the model.

[0088] In the present embodiment, in step S4, the division of the data set is crucial to the prediction performance of the model. When the proportion of the training set in the data set is too large, the scale of the test set will be relatively small, which will cause the evaluation index of the prediction results in different experiments to fluctuate greatly, affecting the accuracy of the prediction results. On the contrary, when the proportion of the training set is too small, the training of the neural network model will be insufficient, and the prediction results will have a large gap with the true value. Therefore, selecting an appropriate data set division ratio can not only effectively reduce the error in the training process, but also minimize the time and energy required for the experiment, reducing unnecessary workload. The present embodiment selects a common data set division, the first 60% of the data is used as the training set, the middle 20% of the data is the verification set, and the last 20% of the data is the test set, to test the generalization ability of the model.

[0089] In this embodiment, the loss function used in step S5 is mean square error loss, which is expressed as: .

[0090] Optimization algorithm plays a vital role in the field of machine learning. It mainly adjusts the parameters of the model by evaluating the value of the loss function to improve the prediction performance of the model. In actual operation, this embodiment selects an optimization algorithm called Adam. Specifically, the Adam algorithm can dynamically adjust the learning rate according to the historical gradient information of the model parameters, which not only makes the update of the model parameters more efficient, but also helps the algorithm to avoid convergence to the local minimum during the training process. In the process of model training using the Adam algorithm, this embodiment continuously performs iterative training, which means that this embodiment repeatedly updates and evaluates the model parameters. Each iteration adjusts the model parameters based on the feedback of the loss function so that it gradually approaches the optimal solution. The goal of this embodiment is that when the model reaches a convergence state, the accuracy of its prediction results can meet business needs. In this way, not only can the performance of the model be improved, but also the effectiveness and reliability of the model in practical applications can be ensured.

[0091] In this embodiment, the evaluation index is used to evaluate the deep neural network model in step S5, and the evaluation index includes goodness of fit (R2), accuracy (ACC), recall (R) and precision (P). Each index measures the performance of the model from a different perspective, thereby providing multiple directions for model optimization. The expression of the evaluation index is:

[0092] ;

[0093] In the formula, y i is the true value corresponding to the i-th sample, is the predicted value of the i-th sample obtained by the model, is the sample mean, n is the number of samples, TP, FP and FN represent true positive, false positive and false negative respectively;

[0094] Evaluation criteria: R 2 The closer it is to 1, the better the model fitting effect is. 2 The closer it is to 0, the worse the model fitting effect is; the closer ACC is to 1, the higher the prediction accuracy of the model is.

[0095] In the step S6 of the present embodiment, the integrated process can be realized by writing an interface program or utilizing an existing integrated framework. In the actual operation process, the strapdown inertial navigation system can obtain the angle information of the carrier in real time, and pass it to the neural network model as input data. The model will output corresponding error estimates according to the input data, and these error estimates can be used for real-time compensation of the output of the strapdown inertial navigation system. Through the integrated neural network model, real-time estimation and compensation of the strapdown inertial navigation system error can be realized. This compensation method can effectively improve the accuracy and stability of the navigation system, and reduce the accumulation and propagation of errors. Meanwhile, because the neural network model has powerful learning and adaptability, it can also adaptively adjust the error compensation strategy according to the change of the working environment, further improving the performance of the navigation system.

[0096] During the navigation process, the strapdown inertial navigation system will use the trained neural network model to predict the error in real time according to the current working environment and measurement data. The prediction results will be used as the basis for error compensation to compensate the output of the strapdown inertial navigation system in real time. In this way, the impact of errors on navigation positioning accuracy can be effectively reduced, and the stability and reliability of the system can be improved. In addition, with the continuous operation of the navigation system and the continuous accumulation of data, online learning technology is also used to update and optimize the neural network model in real time. Through online learning, the model can better adapt to the new working environment and error characteristics, further improving the performance of the navigation system.

[0097] The present invention uses a neural network to estimate and compensate for the errors of the strapdown inertial navigation system in real time, effectively reducing the impact of the errors on the positioning accuracy of the navigation system. The neural network model can learn the complex relationship between the error and the motion state of the carrier, and predict the error in real time during actual operation, thereby achieving accurate compensation for the error. This method improves the stability and reliability of the navigation system, and is particularly suitable for application scenarios that require long-term continuous operation and high-precision positioning.

[0098] The present invention has the advantages of strong generalization ability and good adaptability. By collecting actual operation data under different conditions, a neural network model that adapts to various working environments can be trained. Compared with the traditional error compensation method based on mathematical models or physical models, the present invention adopts a neural network model for error estimation and compensation, without the need to establish a complex mathematical model, thereby reducing the difficulty of algorithm design and implementation. At the same time, the neural network model can handle nonlinear and non-stationary error characteristics and better adapt to complex and changeable navigation environments. This makes the present invention have strong versatility and scalability, and can be widely used in various navigation systems.

[0099] The present invention is flexible and scalable. In practical applications, with the continuous development of technology and the continuous improvement of system requirements, the neural network model can be improved and optimized as needed. For example, more network layers or neurons can be added to improve the complexity and learning ability of the model; different neural network architectures can be used, such as recurrent neural networks (RNNs), long short-term memory networks (LSTMs), etc., to better process time series data; advanced concepts such as transfer learning can also be introduced to use the trained model for fine-tuning to accelerate the training process of the new model and improve performance.

[0100] Although the embodiments of the present invention are described in conjunction with the accompanying drawings, the patent owner may make various variations or modifications within the scope of the appended claims. As long as they do not exceed the protection scope described in the claims of the present invention, they should be within the protection scope of the present invention.

Claims

1. A strapdown inertial navigation system error compensation method based on neural network, characterized in that: The following steps are involved: S1. Build a deep neural network model including multiple neural network layers; S2. collecting actual operation data of the strapdown inertial navigation system under different conditions, the actual operation data including measurement values ​​of the accelerometer and the gyroscope, and position information; S3, preprocessing the collected actual operation data of the strapdown inertial navigation system under different conditions; S4, dividing the preprocessed running data into a training set, a test set, and a validation set with a certain proportion; S5, using the training set to train the deep neural network model constructed in step S1, by minimizing the difference between the error estimate value output by the deep neural network model and the true error value to optimize the parameters of the deep neural network model, and using the validation set to validate the trained deep neural network model to obtain the final deep neural network model; S6, integrating the final deep neural network model into the strapdown inertial navigation system, collecting the output data of the strapdown inertial navigation system in real time, and inputting the data into the deep neural network model to obtain an error estimate, and adjusting the navigation parameters according to the error estimate; The step S2 further comprises: Construct gyroscope error model and accelerometer error model; The measurement values ​​of the accelerometer and gyroscope are corrected by the gyroscope error model and the accelerometer error model respectively; The gyroscope error model The detailed formula is: In the formula, δk gpp is the scale factor error, is the zero bias error, p = x, y, z, b is the non-rectangular coordinate system caused by the installation angle error g The coordinate transformation matrix between the rectangular coordinate system b and the gyroscope error model is parallelized to obtain: In the formula, is the theoretical angular velocity, is the angular velocity actually measured by the gyroscope, ε b is the zero drift value, δK G is the scale error matrix, δk g is the scale coefficient error of the gyroscope error model, is the installation error of the gyroscope error model, μ g is the misalignment angle error of the gyroscope error model, By transposing the terms, the error model expression of the gyroscope can be obtained as follows: Similarly, the expression of the accelerometer error model is: In the formula, In the formula, is the theoretical specific force increment, is the specific force increment actually measured by the accelerometer, is the accelerometer measurement bias, δK A is the accelerometer calibration error matrix, δk A is the scale coefficient error of the accelerometer error model, μ a is the misalignment angle error of the accelerometer error model, is the installation error of the accelerometer error model.

2. The error compensation method for a strapdown inertial navigation system based on a neural network according to claim 1, characterized in that: The neural network model in step S1 adopts a long short-term memory network, and the neural network model adopts regularization technology.

3. The error compensation method for a strapdown inertial navigation system based on a neural network according to claim 1, characterized in that: The measurement value of the gyroscope in step S2 is the angular velocity information between the carrier coordinate system and the navigation coordinate system, and the acquired angular velocity information is corrected, and the attitude quaternion is updated based on the corrected data to solve the converted attitude matrix to obtain accurate attitude angle information; The accelerometer is used to measure the specific force information inside the carrier coordinate system, and the specific force information is converted to the navigation coordinate system using the attitude matrix, and then processed by integral operation to derive the speed and position information of the carrier.

4. The error compensation method for a strapdown inertial navigation system based on a neural network according to claim 3, characterized in that: The obtained velocity information is combined with the position rate equation to solve to obtain the angular velocity of rotation relative to the ground in the navigation coordinate system.

5. The error compensation method for strapdown inertial navigation system based on neural network according to claim 1, characterized in that: The step S3 is specifically as follows: Subtract the mean value of each feature from the value of that feature and divide it by its standard deviation. The mathematical formula is as follows: In the formula, μ and σ represent the mean and standard deviation of the data set respectively, x data It is the new data after standardization.

6. The error compensation method of strapdown inertial navigation system based on neural network according to claim 1, characterized in that: In step S4, the training set accounts for 60%, the validation set accounts for 20%, and the test set accounts for 20%.

7. The error compensation method of strapdown inertial navigation system based on neural network according to claim 1, characterized in that: The loss function used in step S5 is mean square error loss, which is expressed as:

8. The error compensation method of strapdown inertial navigation system based on neural network according to claim 1, characterized in that: In step S5, an evaluation index is used to evaluate the deep neural network model, and the expression of the evaluation index is: In the formula, y i is the true value corresponding to the i-th sample, is the predicted value of the i-th sample obtained by the model, is the sample mean, n is the number of samples, TP, FP and FN represent true positive, false positive and false negative respectively; Evaluation criteria: R 2 The closer it is to 1, the better the model fitting effect is. 2 The closer it is to 0, the worse the model fit is. The closer ACC is to 1, the higher the prediction accuracy of the model.

Images