An adaptive filtering method for precise airdrop integrated navigation data

Abstract

The present invention discloses a method for adaptive filtering of precision airdrop combined navigation data, which is applied to the field of airborne airdrop navigation technology, including: obtaining combined navigation data of two Beidou-inertial navigation combined navigation devices installed on the payload of a precision airdrop system in different airdrop environments; using the two-way combined navigation data as input, constructing a filtering model based on a bidirectional two-input LSTM deep neural network, and using the Noise2Noise algorithm to train the filtering model offline; using the two-way combined navigation data collected in real time, using a smaller learning rate to update the weight of the filtering model in real time online, inputting the two-way combined navigation data into the filtering model, and obtaining an adaptive filtering result. The present invention can take into account both efficiency and effect, better estimate system noise, and better adapt to the external airdrop environment, achieve more accurate and efficient adaptive filtering, and has certain generalization.

Description

Technical Field

[0001] The present invention relates to the field of airborne airdrop navigation technology, and in particular to a method for adaptively filtering precise airdrop combined navigation data. Background Art

[0002] The goal of precision airdrop is to provide fast, accurate and low-cost airdrop, and precision navigation is the first step to achieve precision airdrop. The mainstream in the existing technology is to measure navigation parameters by low-cost Beidou-INS combination, and achieve higher-precision navigation by fusing and complementing navigation information from different sensors. However, in the actual airdrop process, due to the influence of factors such as the instability of navigation components and the uncertainty of the external airdrop environment, it will bring difficulties to the accurate description of the statistical characteristics of system noise and measurement noise. Therefore, the denoising and fusion of the combined navigation signal of the precision airdrop system is an important problem facing this technology.

[0003] For the denoising and fusion of the combined navigation signal of the precision airdrop system, the adaptive Kalman filter and other methods widely used in the prior art have some problems, such as high algorithm complexity, poor real-time performance, or weak online noise estimation ability, which reduces the fusion effect. Conventional signal filtering methods based on machine learning also require the use of higher-precision data as target data for model training, which has certain limitations in implementation and use.

[0004] Therefore, how to provide a method for adaptive filtering of precise airdrop combined navigation data that can balance efficiency and effect, achieve higher precision and more efficient adaptive filtering, reduce the dependence on using higher precision data as target data for model training, and improve the generalization ability of the method is an urgent problem that technicians in this field need to solve. Summary of the invention

[0005] In view of this, the present invention proposes a method for adaptive filtering of precise airdrop combined navigation data. On the basis of installing two low-cost Beidou-INS combined navigation devices, a deep neural network is used as a filter, and the network weights are updated in the way of "offline learning + online update". Offline learning ensures that the filter can accurately estimate the system noise, and online migration ensures that the filter can adapt to the external airdrop environment.

[0006] In order to achieve the above object, the present invention adopts the following technical solution:

[0007] A precise airdrop combined navigation data adaptive filtering method, comprising:

[0008] Step (1): Obtaining the combined navigation data of two Beidou-INS combined navigation devices installed on the payload of the precision airdrop system in different airdrop environments;

[0009] Step (2): Using the two-way combined navigation data as input, a filtering model based on a bidirectional two-input LSTM deep neural network is constructed, and the filtering model is trained offline using the Noise2Noise algorithm;

[0010] Step (3): using the two-way combined navigation data collected in real time, using a smaller learning rate to update the weight of the filter model in real time online, inputting the two-way combined navigation data into the filter model, and obtaining an adaptive filtering result.

[0011] Optionally, in step (1), the distance between the two Beidou-INS integrated navigation devices should be less than or equal to a preset device spacing.

[0012] Optionally, in step (1), different airdrop environments include: different weather environments, different airdrop altitude conditions, and different flight modes.

[0013] Optional, different flight modes, including: stable flight, turning, bird landing.

[0014] Optionally, in step (1), the combined navigation data includes: the position, attitude, and acceleration of two Beidou-INS combined navigation devices.

[0015] Optionally, in step (2), the bidirectional two-input LSTM deep neural network includes:

[0016] Input layer: The input data dimension is B×T×2N; where B is the variable number of samples; T is the length of the input time series; N is the dimension of the single-channel combined navigation data at a certain moment; 2N is two channels;

[0017] Forward and reverse LSTM layers: the activation function of input information and current hidden state output is tanh; the activation function of forget gate, input gate, and output gate is sigmoid; the number of units is 16, the bidirectional LSTM output is concatenated on the feature dimension, and the output data dimension is B×T×32; the Dropout probability is set to 0.2;

[0018] Fully connected layer: two layers of fully connected layers are used, with 16 and 3 neurons respectively; ReLU is used as the activation function;

[0019] Output layer: The output data dimension is B×T×N, which is the combined navigation data after filtering and fusion.

[0020] Optionally, in step (2), the filter model is trained offline using the Noise2Noise algorithm, specifically:

[0021] The objective function of the filtering model is as follows:

[0022]

[0023] Among them, f θ is the filtering model; L is the training loss function; are sample pairs participating in model training; among them, is the two-way combined navigation data containing noise; y is the expected combined navigation data;

[0024] Using the Noise2Noise algorithm, the objective function is transformed. y are not independent of each other, we get:

[0025]

[0026] Add a zero-mean noise as disturbance to the expected combined navigation data y, and we get Transformed into the following optimization problem:

[0027]

[0028] in, Input data for training; Output data for training;

[0029] During the training process, the loss function selects L2 error, the optimizer selects the gradient descent method with momentum, and adopts a speed reduction mechanism to speed up the efficiency in the initial stage of training.

[0030] Optional, training input data The generation of is as follows:

[0031] The two-way combined navigation data collected offline in the same time period are randomly cropped as training input data, and the two-way cropped combined navigation data are randomly weighted averaged with a selection probability of 50%, and the generated two-way weighted data are also used as training input data.

[0032] Optionally, training output data The generation of is as follows:

[0033] A certain path of cropped combined navigation data is randomly selected as training output data, and the selection probability of each path is 50%.

[0034] Optionally, in step (3), the weights of the filter model are updated online in real time using the two-way combined navigation data collected in real time and a smaller learning rate, specifically:

[0035] Using the two-way combined navigation data collected in real time, the weight of the filtering model is updated every preset update time, and the data of the previous preset time length is selected as the migration training data set. The training method is the same as the method of offline training of the filtering model using the Noise2Noise algorithm, and the learning rate is reduced to 10%.

[0036] It can be seen from the above technical solutions that compared with the prior art, the present invention proposes a method for adaptive filtering of precise airdrop combined navigation data. Taking advantage of the two-way low-cost Beidou-INS combined navigation equipment, combined with the two-way two-input LSTM model and the "offline learning + online update" method, it is possible to take into account both efficiency and effect, better estimate system noise, and better adapt to the external airdrop environment, to achieve more accurate and efficient adaptive filtering. Taking advantage of the Noise2Noise training method that does not require clean target data, the disadvantage of the traditional machine learning adaptive filtering method that must obtain clean data in advance for training is simplified, thereby improving the versatility of this method. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] 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 embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0038] Figure 1 It is a schematic diagram of the method flow of the present invention.

[0039] Figure 2 It is a schematic diagram of the bidirectional two-input LSTM deep neural network structure of the present invention. DETAILED DESCRIPTION

[0040] 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.

[0041] Embodiment 1:

[0042] Embodiment 1 of the present invention discloses a method for adaptive filtering of precise airdrop combined navigation data, such as Figure 1 As shown, including:

[0043] Step (1): Obtain the combined navigation data of two Beidou-INS integrated navigation devices (same model) installed on the payload of the precision airdrop system under different airdrop environments.

[0044] The distance between the two Beidou-INS integrated navigation devices should be less than or equal to the preset device spacing.

[0045] Different airdrop environments include: different weather environments, different airdrop altitude conditions, different flight modes, etc.

[0046] Different flight modes, including: stable flight, turning, bird landing, etc.

[0047] Combined navigation data includes: position, attitude, acceleration, etc. of two Beidou-INS combined navigation devices.

[0048] Step (2): Using the two-way combined navigation data as input, a filtering model based on a bidirectional two-input LSTM deep neural network is constructed, and the Noise2Noise algorithm is used to perform offline training on the filtering model.

[0049] Bidirectional dual-input long short-term memory network (Bidirectional, BiLSTM) is a time recursive network suitable for time series processing. It accesses and combines navigation data by combining forward hidden layers and reverse hidden layers, fully mining the implicit state information in the time series, and thereby realizing the correction of the combined navigation data. Compared with traditional algorithms such as Kalman filtering and particle filtering, deep learning algorithms such as BiLSTM require fewer assumptions, have stronger representation capabilities, and have better state estimation efficiency and effects. Bidirectional two-input LSTM deep neural network, such as Figure 2 As shown, including:

[0050] Input layer: The input data dimension is B×T×2N; where B is the variable number of samples; T is the length of the input time series; N is the dimension of the single-channel combined navigation data at a certain moment; 2N is two channels;

[0051] Forward and reverse LSTM layers: the activation function of input information and current hidden state output is tanh; the activation function of forget gate, input gate, and output gate is sigmoid; the number of units is 16, the bidirectional LSTM output is concatenated on the feature dimension, and the output data dimension is B×T×32; at the same time, in order to improve the generalization ability of the model, the Dropout probability is set to 0.2;

[0052] Fully connected layer: two layers of fully connected layers are used, with 16 and 3 neurons respectively; ReLU is used as the activation function;

[0053] Output layer: The output data dimension is B×T×N, which is the combined navigation data after filtering and fusion.

[0054] The Noise2Noise algorithm is used to train the filter model offline, specifically:

[0055] The objective function of the filtering model is as follows:

[0056]

[0057] Among them, f θ is the filtering model; L is the training loss function; are sample pairs participating in model training; among them, is the two-way combined navigation data containing noise; y is the expected combined navigation data;

[0058] In the actual airdrop process, it is difficult to obtain high-precision integrated navigation data at low cost, that is, only the input data for training is available. Without the real label y, conventional training methods cannot be used for training. Therefore, the "Noise to Noise" (Noise2Noise) training strategy is adopted to restore and reconstruct the real data only through the noise input and output signal pairs.

[0059] Using the Noise2Noise algorithm (drawing on the above ideas), the objective function is transformed. y are not independent of each other, we get:

[0060]

[0061] Change The distribution of , the conditional expectation remains unchanged, and the final θ remains unchanged. Therefore, a zero-mean noise is added to the expected combined navigation data y as a disturbance, and we get It can be transformed into the following optimization problem:

[0062]

[0063] in, Input data for training; Output data for training;

[0064] During the training process, the loss function selects L2 error, the optimizer selects the gradient descent method with momentum, and adopts a suitable speed reduction mechanism to speed up the efficiency in the initial stage of training and ensure the accuracy in the convergence stage.

[0065] Training input data The generation of is as follows:

[0066] The two-way combined navigation data collected offline in the same time period are randomly cropped as training input data, and the two-way cropped combined navigation data are randomly weighted averaged with a selection probability of 50%, and the generated two-way weighted data are also used as training input data.

[0067] Training output data The generation of is as follows:

[0068] A certain path of cropped combined navigation data is randomly selected as training output data, and the selection probability of each path is 50%.

[0069] Step (3): using the two-way combined navigation data collected in real time, using a smaller learning rate to update the weight of the filter model in real time online, inputting the two-way combined navigation data into the filter model, and obtaining an adaptive filtering result.

[0070] Using the two-way combined navigation data collected in real time, a small learning rate is used to update the weight of the filter model in real time online, specifically:

[0071] Using the two-way combined navigation data collected in real time, the weight of the filter model is updated every preset update time (the default is 1 second, which can be determined according to the performance of the computing unit), and the data of the previous preset time (the default is 20 seconds, which can be determined according to the performance of the computing unit) is selected as the migration training data set. The training method is the same as the method of offline training the filter model using the Noise2Noise algorithm in step (2), and the learning rate is reduced to 10% of step (2).

[0072] The embodiment of the present invention discloses a method for adaptive filtering of precise airdrop combined navigation data. By taking advantage of two low-cost Beidou-INS combined navigation devices, combined with a bidirectional two-input LSTM model and an "offline learning + online update" method, it is possible to balance efficiency and effect to better estimate system noise, and better adapt to the external airdrop environment, thereby achieving more accurate and efficient adaptive filtering. By taking advantage of the fact that the Noise2Noise training method does not require clean target data, the disadvantage of the traditional machine learning adaptive filtering method that clean data must be obtained in advance for training is simplified, thereby improving the versatility of the method.

[0073] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0074] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A precise airdrop integrated navigation data adaptive filtering method, characterized in that: include: Step (1): Obtaining the combined navigation data of two Beidou-INS combined navigation devices installed on the payload of the precision airdrop system in different airdrop environments; Step (2): using the two channels of combined navigation data as input, constructing a filtering model based on a bidirectional two-input LSTM deep neural network, and using the Noise2Noise algorithm to perform offline training on the filtering model; Step (3): using the two channels of combined navigation data collected in real time, using a relatively small learning rate to perform real-time online updating on the weights of the filtering model, inputting the two channels of combined navigation data into the filtering model, and obtaining an adaptive filtering result; In step (2), the bidirectional two-input LSTM deep neural network includes: Input layer: The input data dimension is B×T×2N; where B is the variable number of samples; T is the length of the input time series; N is the dimension of the single-channel combined navigation data at a certain moment; 2N is two channels; Forward and reverse LSTM layers: the activation function of input information and current hidden state output is tanh; the activation function of forget gate, input gate, and output gate is sigmoid; the number of units is 16, the bidirectional LSTM output is concatenated on the feature dimension, and the output data dimension is B×T×32; the Dropout probability is set to 0.2; Fully connected layer: two layers of fully connected layers are used, with 16 and 3 neurons respectively; ReLU is used as the activation function; Output layer: The output data dimension is B×T×N, which is the combined navigation data after filtering and fusion; In step (2), the filter model is trained offline using the Noise2Noise algorithm, specifically: The objective function of the filtering model is as follows: Among them, f θ is the filtering model; L is the training loss function; are sample pairs participating in model training; among them, are the two-way combined navigation data containing noise; y is the desired combined navigation data; The objective function is transformed using the Noise2Noise algorithm. y are not independent of each other, we get: Add a zero-mean noise as disturbance to the expected combined navigation data y, and we get Transformed into the following optimization problem: in, Input data for training; Output data for training; During the training process, the loss function selects L2 error, the optimizer selects the gradient descent method with momentum, and adopts the speed reduction mechanism to speed up the initial training efficiency; The training input data The generation of is as follows: Randomly cutting two paths of the combined navigation data collected offline in the same time period as the training input data, and randomly performing a weighted average on the two paths of the cut combined navigation data with a selection probability of 50%, and the generated two paths of weighted data are also used as the training input data; The training output data The generation of is as follows: Randomly select a certain path of the cropped combined navigation data as training output data, and the selection probability of each path is 50%; In step (3), the weights of the filtering model are updated online in real time using the two channels of combined navigation data collected in real time and a relatively small learning rate, specifically: Using the two channels of combined navigation data collected in real time, the weight of the filtering model is updated every preset update time, and the data of the previous preset time period is selected as the migration training data set. The training method is the same as the method of offline training of the filtering model using the Noise2Noise algorithm, and the learning rate is reduced to 10%.

2. According to claim 1, a precise airdrop integrated navigation data adaptive filtering method is characterized in that: In step (1), the distance between the two Beidou-INS integrated navigation devices should be less than or equal to the preset device spacing.

3. The method for adaptive filtering of precise airdrop integrated navigation data according to claim 1, characterized in that: In step (1), the different airdrop environments include: different meteorological environments, different airdrop altitude conditions, and different flight modes.

4. The method for adaptive filtering of precise airdrop integrated navigation data according to claim 3, characterized in that: The different flight modes include: stable flight, turning, and bird landing.

5. The method for adaptive filtering of precise airdrop integrated navigation data according to claim 1, characterized in that: In step (1), the combined navigation data includes: the position, attitude, and acceleration of the two Beidou-INS combined navigation devices.

Images