INS / SAR (inertial navigation system / synthetic aperture radar) integrated navigation method based on liquid time constant network

Abstract

The invention discloses an INS / SAR (inertial navigation system / synthetic aperture radar) integrated navigation method based on a liquid time constant network, which adopts a loose coupling integrated navigation strategy of INS and SAR and comprises a training mode and a prediction mode: the training mode: training an LTC auxiliary positioning prediction network by using obtained data in an offline stage, and loading a trained network model to a machine; then, the network model is continuously trained online by using the data acquired in real time, and the input and output relationship of the LTC assisted positioning prediction network is better fitted; meanwhile, the INS measurement information and the SAR measurement information are fused through a Kalman filter; the method has the beneficial effects that the neural network model based on LTC neurons and closed-form solution approximate solution is established, the solution speed is increased, and the problems of different and lagged SAR measurement output time intervals are solved by utilizing the capability of processing non-equal interval sampling time sequences and predicting the time sequences of the LTC neural network, so that the SAR measurement accuracy is improved. The INS / SAR integrated navigation positioning method is realized.

Description

Technical Field

[0001] The invention belongs to the technical field of aircraft navigation, and in particular relates to an INS / SAR combined navigation method based on a liquid time constant network. Background Art

[0002] The Inertial Navigation System (INS) can navigate autonomously without relying on external environmental information, but its errors will accumulate over time. The scene matching navigation system using Synthetic Aperture Radar (SAR) can provide the precise position of the aircraft, correct the INS navigation error, and improve navigation accuracy. However, the INS / SAR combined navigation system outputs information at different frequencies during navigation and positioning. Due to the time delay in SAR imaging and scene matching processing, the SAR system has two problems: non-uniformly spaced output and delayed measurement information. Therefore, if these two problems cannot be effectively solved, the SAR system will not be able to correct the INS navigation error in a timely manner, thereby affecting the accuracy of the combined navigation system.

[0003] At present, in order to solve the problem of measurement asynchrony, whether to update the measurement is selected according to whether there is SAR measurement information output; the SAR output measurement information is fitted with a parabola to obtain a compensation algorithm for the SAR measurement information lag. This type of method has improved the navigation accuracy to a certain extent. However, when there is no SAR measurement information, it is still impossible to correct the INS navigation error. In addition, high-precision fitting requires a high-order parabola, which will increase the amount of calculation. Summary of the invention

[0004] The object of the present invention is to provide an INS / SAR integrated navigation method based on a liquid time constant network to solve the problem of different time intervals and lags in SAR measurement output.

[0005] To achieve the above object, the present invention provides the following technical solution: an INS / SAR integrated navigation method based on a liquid time constant network, adopting a loosely coupled integrated navigation strategy of INS and SAR, including two modes: training and prediction:

[0006] Training mode: In the offline stage, the LTC-assisted positioning prediction network is trained using the acquired data, and the trained network model is loaded onto the machine; then, the network model is trained online using the real-time acquired data to better fit the input-output relationship of the LTC-assisted positioning prediction network; at the same time, the Kalman filter is used to fuse the INS and SAR measurement information;

[0007] Prediction mode: When SAR is observing and the SAR measurement information has not been output due to processing delay, the LTC-assisted positioning prediction network is used to predict the SAR measurement information, and then it is fused with the INS measurement information through the Kalman filter.

[0008] As a preferred technical solution of the present invention, it also includes a learning process of LTC-assisted positioning prediction network parameters in a training mode, and the corresponding content is as follows:

[0009] S1, standardize and preprocess the original data to obtain dimensionless training data, and send it to the network for training;

[0010] S2, feed the standardized preprocessed training data into the convolution module to extract the sparse feature representation between INS and SAR sensors and learn the implicit dependency between sensors;

[0011] S3, the output of the convolution module is sent to the recurrent module composed of LTC neurons to extract the time series features of INS and SAR sensor data and learn the time dynamic relationship of sensor data;

[0012] S4. Train the LTC-assisted positioning prediction network to obtain the optimal parameters of the network and make predictions for the test samples.

[0013] As a preferred technical solution of the present invention, the specific implementation steps of S1 are as follows:

[0014] S101. Assume that the original data input to the network can be expressed as S = [S1, S2, ..., S K ]∈R N×K ,in, represents the kth feature sequence, k = 1, 2, ..., K, K is the total number of features of the original data; since k represents the three-axis specific force of the IMU accelerometer and the three-axis angular velocity of the gyroscope, as well as the longitude and latitude information obtained by SAR, there are 8 features in total, so K = 8; represents the measured value of the kth feature at the nth moment, n = 1, 2, ..., N, N is the length of the sequence; R represents a set of real numbers;

[0015] S102. Standardize and preprocess the original data to obtain dimensionless data k=1,2,…,K, which can be expressed as:

[0016]

[0017] Among them, μ(*) and σ(*) represent the k Find the mean and standard deviation;

[0018] S103, data after standardization preprocessing It can be expressed as a row vector: in, Represents the eigenvector at the nth moment, n = 1, 2, ..., N; Represents the measurement value of the kth feature at the nth moment, k = 1, 2, …, K.

[0019] The specific implementation steps of S2 are as follows:

[0020] The convolution module consists of 4 one-dimensional convolutional layers and 2 batch normalization layers; the feature vector at the nth moment First, after two layers of convolution processing and one layer of batch normalization processing, and then after two layers of convolution processing and one layer of batch normalization processing, the sparse features of the sensor at the nth moment are obtained. in, represents the measured value of the cth feature at the nth moment, c = 1, 2, ..., C, C is the number of features output by the convolution module, and the process can be expressed as

[0021]

[0022] Among them, Conv i (*), (i=1,2,3,4) respectively represent the i-th layer convolution operation, BN j (*), (j=1,2) represent the batch normalization operation of the jth layer respectively.

[0023] As a preferred technical solution of the present invention, the specific parameters of each layer of the convolution module are as follows:

[0024] Convolution layer 1: The number of input channels is 8, the number of output channels is 64, and the convolution kernel size is 1;

[0025] The second convolution layer has 64 input channels, 128 output channels, and a convolution kernel size of 1.

[0026] Layer 1 batch normalization: the number of input features is 128;

[0027] The third convolution layer has 128 input channels, 256 output channels, and a kernel size of 1.

[0028] The fourth convolution layer has 256 input channels, 512 output channels, and a kernel size of 1.

[0029] Layer 2 batch normalization: the number of input features is 512,

[0030] As a preferred technical solution of the present invention, the circulation module of S3 is specifically:

[0031] The hidden state of the recurrent module based on LTC neurons can be solved by the following initial value problem:

[0032]

[0033] Among them, when the time step is t, h (D×1) (t) defines the hidden state of the LTC layer with D units, I (M×1) (t) is the exogenous input of the system, which has M features, is a time-invariant parameter vector, a (D×1) is the bias vector, f is the neural network parameterized by θ, and ⊙ is the Hadamard product;

[0034] In order to achieve efficient and fast solution, the closed-form analytical solution is used to approximate the solution of the above initial value problem. In each time step t, there are D hidden units with hidden states h (D×1) (t) can be expressed as:

[0035] h(t)=b⊙exp(-[w τ +f(h,I,θ)]t)⊙f(-h,-I,θ)+a,

[0036] Among them, b (D×1) is a parameter vector;

[0037] The bias parameter b is regarded as part of the trainable parameters of the neural network f(-h,-I,θ), and the inverse sine-cosine nonlinear function σ(*) is used to replace the exponential decay term. Therefore, the neural network f can be expressed in the form of an exponential decay factor; another neural network h is used to replace a to enhance the flexibility of the model; in order to obtain a more general network structure, the nonlinear f(-h,-I,θ) is made to have both shared (backbone) and independent (g) network parts; h(*) is multiplied by (1-σ(*)), and the time-decayed sigma term can play a gating role; thus, the time-decayed sigma function represents a gating mechanism;

[0038] After the above changes, the LTC-based convolution module can be expressed as:

[0039] h(t)=σ(-[f(h,I,θ f )]t)⊙g(h,I,θ g )+[1-σ(-[f(h,I,θ g )]t)]⊙h(h,I,θ g )

[0040] Among them, θ f ,θ g and θ h are the network parameters of neural networks f, g and h respectively, and t represents the time vector;

[0041] The neural networks f, g, and h share the first few layers in the form of a backbone network during the learning process and branch into these three functions. Therefore, the backbone network allows the model to learn shared representations, which speeds up and stabilizes the learning process.

[0042] As a preferred technical solution of the present invention, the S4 obtains the optimal parameters of the network, specifically:

[0043] The entire network is jointly trained in an end-to-end manner, that is, the convolutional module and the recurrent module are optimized simultaneously; the cross-entropy-based loss function and the Adaptive Moment Estimation with Weight Decay Fix (AdamW) optimizer are selected to train the LTC-assisted positioning prediction network and obtain the optimal parameters of the network.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] A neural network model based on LTC neurons and closed-form approximate solution was established to speed up the solution. The ability of LTC neural network to process non-uniformly spaced sampling time series and predict time series was utilized to solve the problem of different time intervals and lags in SAR measurement output, and an INS / SAR combined navigation and positioning method was realized. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 A training mode block diagram of the proposed method of the present invention;

[0047] Figure 2 A prediction model block diagram of the proposed method of the present invention;

[0048] Figure 3 This is a diagram of the LTC-assisted positioning prediction network architecture of the present invention. DETAILED DESCRIPTION

[0049] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0050] Example 1

[0051] See also Figure 1-3, which is the first embodiment of the present invention, provides an INS / SAR integrated navigation method based on a liquid time constant network, adopts a loosely coupled integrated navigation strategy of INS and SAR, and includes two modes: training and prediction:

[0052] Training mode: In the offline stage, the LTC-assisted positioning prediction network is trained using the acquired data, and the trained network model is loaded onto the machine; then, the network model is trained online using the real-time acquired data to better fit the input-output relationship of the LTC-assisted positioning prediction network; at the same time, the Kalman filter is used to fuse the INS and SAR measurement information;

[0053] Prediction mode: When SAR is observing and the SAR measurement information has not been output due to processing delay, the LTC-assisted positioning prediction network is used to predict the SAR measurement information, and then it is fused with the INS measurement information through the Kalman filter.

[0054] It is necessary to add that:

[0055] LTC-Assisted Position Prediction Network: LTC-Assisted Position Prediction Network, LTC-APPN;

[0056] Kalman filter: Kalman Filter, KF.

[0057] In this embodiment, preferably, it also includes a learning process of LTC-assisted positioning prediction network parameters in a training mode, and the corresponding content is as follows:

[0058] S1, standardize and preprocess the original data to obtain dimensionless training data, and send it to the network for training;

[0059] S2, feed the standardized preprocessed training data into the convolution module to extract the sparse feature representation between INS and SAR sensors and learn the implicit dependency between sensors;

[0060] S3, the output of the convolution module is sent to the recurrent module composed of LTC neurons to extract the time series features of INS and SAR sensor data and learn the time dynamic relationship of sensor data;

[0061] S4. Train the LTC-assisted positioning prediction network to obtain the optimal parameters of the network and make predictions for the test samples.

[0062] In this embodiment, preferably, the specific implementation steps of S1 are as follows:

[0063] S101. Assume that the original data input to the network can be expressed as S = [S1, S2, ..., SK ]∈R N×K ,in, represents the kth feature sequence, k = 1, 2, ..., K, K is the total number of features of the original data; since k represents the three-axis specific force of the IMU accelerometer and the three-axis angular velocity of the gyroscope, as well as the longitude and latitude information obtained by SAR, there are 8 features in total, so K = 8; represents the measured value of the kth feature at the nth moment, n = 1, 2, ..., N, N is the length of the sequence; R represents a set of real numbers;

[0064] S102. Standardize and preprocess the original data to obtain dimensionless data k=1,2,…,K, which can be expressed as:

[0065]

[0066] Among them, μ(*) and σ(*) represent the k Find the mean and standard deviation;

[0067] S103, data after standardization preprocessing It can be expressed as a row vector: in, Represents the eigenvector at the nth moment, n = 1, 2, ..., N; Represents the measurement value of the kth feature at the nth moment, k = 1, 2, …, K.

[0068] In this embodiment, preferably, the specific implementation steps of S2 are as follows:

[0069] The convolution module consists of 4 one-dimensional convolutional layers and 2 batch normalization layers; the feature vector at the nth moment First, after two layers of convolution processing and one layer of batch normalization processing, and then after two layers of convolution processing and one layer of batch normalization processing, the sparse features of the sensor at the nth moment are obtained. in, represents the measured value of the cth feature at the nth moment, c = 1, 2, ..., C, C is the number of features output by the convolution module, and the process can be expressed as

[0070]

[0071] Among them, Conv i (*), (i=1,2,3,4) respectively represent the i-th layer convolution operation, BN j (*), (j=1,2) represent the batch normalization operation of the jth layer respectively.

[0072] Example 2

[0073] See also Figure 1-3 , which is the second embodiment of the present invention, is based on the previous embodiment, except that:

[0074] The specific parameters of each layer of the convolution module in S2 are as follows:

[0075] Convolution layer 1: The number of input channels is 8, the number of output channels is 64, and the convolution kernel size is 1;

[0076] The second convolution layer has 64 input channels, 128 output channels, and a convolution kernel size of 1.

[0077] Layer 1 batch normalization: the number of input features is 128;

[0078] The third convolution layer has 128 input channels, 256 output channels, and a kernel size of 1.

[0079] The fourth convolution layer has 256 input channels, 512 output channels, and a kernel size of 1.

[0080] Layer 2 batch normalization: the number of input features is 512.

[0081] Example 3

[0082] See also Figure 1-3 , which is the third embodiment of the present invention, is based on the previous embodiment, except that:

[0083] The loop module of S3 is as follows:

[0084] The hidden state of the recurrent module based on LTC neurons can be solved by the following initial value problem:

[0085]

[0086] Among them, when the time step is t, h (D×1) (t) defines the hidden state of the LTC layer with D units, I (M×1) (t) is the exogenous input of the system, which has M features, is a time-invariant parameter vector, a (D×1) is the bias vector, f is the neural network parameterized by θ, and ⊙ is the Hadamard product;

[0087] In order to achieve efficient and fast solution, the closed-form analytical solution is used to approximate the solution of the above initial value problem. In each time step t, there are D hidden units with hidden states h (D×1) (t) can be expressed as:

[0088] h(t)=b⊙exp(-[w τ+f(h,I,θ)]t)⊙f(-h,-I,θ)+a,

[0089] Among them, b (D×1) is a parameter vector;

[0090] The bias parameter b is regarded as part of the trainable parameters of the neural network f(-h,-I,θ), and the inverse sine-cosine nonlinear function σ(*) is used to replace the exponential decay term. Therefore, the neural network f can be expressed in the form of an exponential decay factor; another neural network h is used to replace a to enhance the flexibility of the model; in order to obtain a more general network structure, the nonlinear f(-h,-I,θ) is made to have both shared (backbone) and independent (g) network parts; h(*) is multiplied by (1-σ(*)), and the time-decayed sigma term can play a gating role; thus, the time-decayed sigma function represents a gating mechanism;

[0091] After the above changes, the LTC-based convolution module can be expressed as:

[0092] h(t)=σ(-[f(h,I,θ f )]t)⊙g(h,I,θ g )+[1-σ(-[f(h,I,θ g )]t)]⊙h(h,I,θ g )

[0093] Among them, θ f ,θ g and θ h are the network parameters of neural networks f, g and h respectively, and t represents the time vector;

[0094] The neural networks f, g, and h share the first few layers in the form of a backbone network during the learning process and branch into these three functions.

[0095] Example 4

[0096] See also Figure 1-3 , which is the fourth embodiment of the present invention, is based on the previous embodiment, except that:

[0097] S4 obtains the optimal parameters of the network, which are:

[0098] The entire network is jointly trained in an end-to-end manner, that is, the convolutional module and the recurrent module are optimized simultaneously; a cross-entropy-based loss function and an adaptive momentum estimation optimizer with weight decay are selected to train the LTC-assisted positioning prediction network and obtain the optimal parameters of the network.

[0099] The specific steps of S4 predicting the test sample are:

[0100] S401, test sample data S val After standardization preprocessing,

[0101] S402: Standardize the pre-processed data Send the trained LTC auxiliary positioning prediction network for prediction to obtain the network predicted longitude and latitude values

[0102] S403: Obtain SAR output latitude and longitude prediction values ​​through standardized inverse transformation processing It can be expressed as:

[0103]

[0104] Among them, S lat ,S lon They represent the original data of SAR output latitude and longitude values, μ(*) and σ(*) represent the lat ,S lon Find the mean and standard deviation.

[0105] Although the embodiments of the present invention have been shown and described, as detailed above, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An INS / SAR integrated navigation method based on a liquid time constant network, characterized in that: The loosely coupled integrated navigation strategy of INS and SAR is adopted, including two modes: training and prediction: Training mode: In the offline stage, the LTC-assisted positioning prediction network is trained using the acquired data, and the trained network model is loaded onto the machine; then, the network model is trained online using the real-time acquired data to better fit the input-output relationship of the LTC-assisted positioning prediction network; at the same time, the Kalman filter is used to fuse the INS and SAR measurement information; Prediction mode: When SAR is observing and the SAR measurement information has not been output due to processing delay, the LTC-assisted positioning prediction network is used to predict the SAR measurement information, and then it is fused with the INS measurement information through the Kalman filter.

2. The INS / SAR integrated navigation method based on a liquid time constant network according to claim 1, characterized in that: It also includes the learning process of LTC-assisted positioning prediction network parameters in training mode, and the corresponding contents are as follows: S1, standardize and preprocess the original data to obtain dimensionless training data, and send it to the network for training; S2, feed the standardized preprocessed training data into the convolution module to extract the sparse feature representation between INS and SAR sensors and learn the implicit dependency between sensors; S3, the output of the convolution module is sent to the recurrent module composed of LTC neurons to extract the time series features of INS and SAR sensor data and learn the time dynamic relationship of sensor data; S4. Train the LTC-assisted positioning prediction network to obtain the optimal parameters of the network and make predictions for the test samples.

3. The INS / SAR integrated navigation method based on liquid time constant network according to claim 2, characterized in that: The specific implementation steps of S1 are as follows: S101. Assume that the original data input to the network can be expressed as S = [S1, S2, ..., S K ]∈R N×K ,in, represents the kth feature sequence, k = 1, 2, ..., K, K is the total number of features of the original data; represents the measured value of the kth feature at the nth moment, n = 1, 2, ..., N, N is the length of the sequence; R represents a set of real numbers; S102. Standardize and preprocess the original data to obtain dimensionless data k=1,2,…,K, which can be expressed as: Among them, μ(*) and σ(*) represent the k Find the mean and standard deviation; S103, data after standardization preprocessing It can be expressed as a row vector: in, Represents the eigenvector at the nth moment, n = 1, 2, ..., N; Represents the measurement value of the kth feature at the nth moment, k = 1, 2, …, K.

4. The INS / SAR integrated navigation method based on liquid time constant network according to claim 2, characterized in that: The specific implementation steps of S2 are as follows: The convolution module consists of 4 one-dimensional convolutional layers and 2 batch normalization layers; the feature vector at the nth moment First, after two layers of convolution processing and one layer of batch normalization processing, and then after two layers of convolution processing and one layer of batch normalization processing, the sparse features of the sensor at the nth moment are obtained. in, represents the measured value of the cth feature at the nth moment, c = 1, 2, ..., C, C is the number of features output by the convolution module, and the process can be expressed as Among them, Conv i (*), (i=1,2,3,4) respectively represent the i-th layer convolution operation, BN j (*), (j=1,2) represent the batch normalization operation of the jth layer respectively.

5. The INS / SAR integrated navigation method based on liquid time constant network according to claim 4, characterized in that: The specific parameters of each layer of the convolution module are as follows: Convolution layer 1: The number of input channels is 8, the number of output channels is 64, and the convolution kernel size is 1; The second convolution layer has 64 input channels, 128 output channels, and a convolution kernel size of 1. Layer 1 batch normalization: the number of input features is 128; The third convolution layer has 128 input channels, 256 output channels, and a kernel size of 1. The fourth convolution layer has 256 input channels, 512 output channels, and a kernel size of 1. Layer 2 batch normalization: the number of input features is 512.

6. The INS / SAR integrated navigation method based on liquid time constant network according to claim 2, characterized in that: The S4 obtains the optimal parameters of the network, specifically: The entire network is jointly trained in an end-to-end manner; a cross-entropy-based loss function and an adaptive momentum estimation optimizer with weight decay are selected to train the LTC-assisted positioning prediction network and obtain the optimal parameters of the network.

7. The INS / SAR integrated navigation method based on liquid time constant network according to claim 2, characterized in that: The specific steps of S4 for predicting the test sample are: S401, test sample data S val After standardization preprocessing, S402: Standardize the pre-processed data Send the trained LTC auxiliary positioning prediction network for prediction to obtain the network predicted longitude and latitude values S403: Obtain SAR output latitude and longitude prediction values ​​through standardized inverse transformation processing It can be expressed as: Among them, S lat ,S lon They represent the original data of SAR output latitude and longitude values, μ(*) and σ(*) represent the lat ,S lon Find the mean and standard deviation.

Images