Deep learning assisted extended Kalman dual-band infrared radiation attitude estimation method

Abstract

The invention discloses a dual-band infrared radiation attitude estimation method for deep learning assisted extended Kalman. The method comprises the following steps: analyzing the influence of geometric errors and random errors in dual-band infrared radiation measurement signals on attitude estimation reliability; the method comprises the following steps: constructing a CBGTE model to eliminate geometric errors and random errors in dual-band infrared radiation signals, firstly, collecting a dual-band infrared radiation signal data set by using a semi-physical simulation platform, cleaning the dual-band infrared radiation signal data set, inputting the dual-band infrared radiation signal data set into a BGT neural network for feature learning, and storing an optimal weight; and inputting data obtained by the dual-band infrared radiation sensor into the BGT neural network with the optimal weight to obtain infrared radiation data after compensation of random errors, and inputting the data after filtering of the random errors into an extended Kalman algorithm to filter geometric errors to obtain reliable attitude angle information. The invention provides a high-precision and robust attitude information acquisition and processing method for a spinning unmanned aerial vehicle taking infrared radiation as an attitude measurement method.

Description

Technical Field

[0001] The present invention relates to the technical field of deep learning feature extraction, and in particular to a deep learning assisted extended Kalman dual-band infrared radiation attitude estimation method. Background Art

[0002] With the advancement of infrared sensor technology, it has been optimized in terms of size, speed and energy efficiency. Therefore, infrared detection technology has been applied to various cutting-edge fields, including infrared thermal imaging technology, infrared communication technology and dual-wave infrared radiation attitude measurement technology. The principle of dual-band infrared radiation attitude measurement is to use dual-band infrared radiation sensors to extract navigation information from different radiation fields of the earth. Compared with mainstream attitude measurement technologies such as inertial system technology and global positioning system technology, dual-band infrared radiation attitude measurement technology has significant advantages, such as anti-electromagnetic interference, passive operation and low cost. Therefore, dual-band infrared radiation attitude measurement systems are widely used in spinning drones. However, spinning drones are complex nonlinear systems with multiple variables and strong coupling, and attitude information is easily interfered, thus affecting its reliability.

[0003] Although dual-band infrared radiation attitude measurement technology has many advantages, with further in-depth application, we find that it still has some disadvantages. Due to the nature of dual-band infrared radiation sensors, geometric errors and random errors are introduced into the data. In practical applications, all sensors are affected by geometric errors, including sensor measurement noise, assembly position, motor interference, etc. Scholars usually use various filtering algorithms to solve the geometric error problem, such as Bayesian filtering, Kalman filtering, unscented Kalman filtering, extended Kalman filtering, and particle filtering.

[0004] Since the flying background of the spinning UAV is the atmosphere, random errors are unique to dual-band infrared radiation sensors. The attitude information will inevitably be affected by various atmospheric radiations, such as solar interference and cloud interference. These interferences introduce random errors into the collected data, significantly affecting the attitude estimation and making the attitude information unreliable. Therefore, the current research focus of dual-band infrared radiation attitude measurement technology is to further reduce the attitude estimation error caused by random errors. Summary of the invention

[0005] The problem to be solved by the present invention is to provide a deep learning assisted extended Kalman dual-band infrared radiation attitude estimation method for reducing attitude estimation errors caused by random errors.

[0006] The present invention adopts the following technical solution: a deep learning assisted extended Kalman dual-band infrared radiation attitude estimation method, comprising the following steps:

[0007] S1. Analyze the causes of geometric errors and random errors in the dual-band infrared radiation measurement signals of the spinning UAV and their impact on the reliability of attitude estimation;

[0008] S2, collect dual-band infrared radiation data of the spinning UAV through a semi-physical simulation platform, and pre-process the collected data set;

[0009] S3, constructing a CBGTE model, including a BGT neural network and an extended Kalman filter, inputting the collected data set into the CBGTE model, training and learning the CBGTE model, and obtaining the optimal weight of the CBGTE model;

[0010] S4. The optimized weights of the CBGTE model are placed in the test equipment to verify the feasibility of the algorithm and estimate the dual-band infrared radiation attitude of the spinning UAV.

[0011] Preferably, in step S1, the geometric error is caused by sensor measurement noise, assembly position and motor interference factors; from the perspective of time series, the geometric error value is random and independent of time, and the amplitude is in the range of [-0.02, 0.02] V during the measurement process of the semi-physical platform;

[0012] Random errors are caused by solar interference and cloud interference. From the perspective of time series, random errors are limited by time intervals, and the impact range is within half a time period, with no limit on the amplitude.

[0013] Preferably, in step S2, the sampling rate and sampling time of the semi-physical simulation platform are set, dual-band infrared radiation data are collected under different atmospheric radiation backgrounds, and the collected data sets are preprocessed as follows:

[0014] S201, classifying and arranging the data set according to the interference type, and dividing the data set into: no interference, solar interference, and cloud interference;

[0015] S202, due to the inconsistency of the amplification coefficient of the conditioning circuit, the amplitudes of the long-wave infrared radiation data and the medium-wave infrared radiation data are inconsistent, so all the collected infrared radiation data are normalized to unify the dual-band infrared radiation data into the (-1,1) interval;

[0016] S203: Clean the normalized data set, delete unusable data, and ensure the reliability of the data set.

[0017] Preferably, in step S3, in the CBGTE model, the BGT neural network is used to compensate for random errors in the dual-band infrared attitude measurement process, and the extended Kalman filter is used to eliminate geometric errors in the dual-band infrared attitude measurement process;

[0018] The BGT neural network combines the Bi-GRU and Transformer Encoder networks to compensate for random errors in dual-band infrared radiation information. The BGT neural network consists of an encoder and a decoder. The encoder contains a Bi-GRU module and a Transformer Encoder module, and the decoder consists of a Bi-GRU and a fully connected layer.

[0019] The extended Kalman filter, which derives the nonlinear prediction equation from the observation equation and runs it with the tangent approximation, and the first-order Taylor expansion around the average state is used to eliminate the geometric error of the infrared radiation information after the BGT neural network compensation.

[0020] Preferably, in step S3, data enhancement is performed before training CBGTE, infrared data is sampled, and every 10s of data is used as a training sample. The loss function is the mean square error function, the optimizer is set to Adamw, the learning rate is 0.0005, the epoch is set to 200, the weight decay is set to 0.01, and the posture data curve is obtained based on the deep learning assisted extended Kalman algorithm.

[0021] The technical solution of the present invention also provides: an electronic device, comprising:

[0022] one or more processors;

[0023] a storage device having one or more programs stored thereon;

[0024] When the one or more programs are executed by the one or more processors, the one or more processors implement any of the above-mentioned deep learning assisted extended Kalman dual-band infrared radiation attitude estimation methods.

[0025] The technical solution of the present invention also provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps in any of the above-mentioned deep learning-assisted extended Kalman dual-band infrared radiation attitude estimation methods are implemented.

[0026] Compared with the prior art, the present invention adopts the above technical solution and has the following technical effects:

[0027] 1. Compared with the existing methods, the dual-band infrared radiation attitude estimation method of the present invention has superior performance through semi-physical experiment verification, and the calculated roll angle error is ±0.4° and the pitch angle error is ±0.2°.

[0028] 2. The dual-band infrared radiation attitude estimation method of the present invention uses an extended Kalman filter to eliminate geometric errors in the dual-band infrared attitude measurement process, and uses a BGT neural network to compensate for random errors in the dual-band infrared attitude measurement process.

[0029] 3. The present invention uses a BGT architecture. Different from the traditional Bi-GRU neural network, the dual-band infrared radiation posture estimation method of the present invention introduces a multi-head attention mechanism to capture the key information features in the infrared radiation data, thereby realizing a high-precision BGT neural network framework structure. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 This is a flow chart of the deep learning-assisted extended Kalman dual-band infrared radiation attitude estimation method of the present invention;

[0031] Figure 2 This is a schematic diagram of semi-physical experimental data of the present invention;

[0032] Figure 3 It is a structural diagram of the CBGTE model of the present invention;

[0033] Figure 4 This is a schematic diagram of the error value after solving the CBGTE model in the present invention. DETAILED DESCRIPTION

[0034] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the application is further elaborated in detail below in conjunction with the accompanying drawings. The described embodiments are only a part of the embodiments involved in the present invention. All non-innovative embodiments of other researchers in the field on this embodiment belong to the protection scope of the present invention. At the same time, for the step numbering in the embodiment of the present invention, it is only set for the convenience of explanation, and the order between the steps is not limited in any way. The execution order of each step in the embodiment can be adaptively adjusted according to the understanding of those skilled in the art.

[0035] In one embodiment of the present invention, a deep learning-assisted extended Kalman dual-band infrared radiation attitude estimation algorithm, such as Figure 1 As shown, the following steps are included:

[0036] S1. Analyze the influence of geometric error and random error in dual-band infrared radiation measurement signal on attitude estimation reliability, including: a) analyzing the causes of geometric error and its influence on attitude reliability, b) analyzing the causes of random error and its influence on attitude reliability.

[0037] Among them, a) geometric error is often caused by factors such as sensor measurement noise, assembly position and motor interference. Geometric error has the characteristics of numerical randomness, amplitude limitation and non-time sequence. The impact on the dual-wave infrared radiation attitude information is to superimpose a random and amplitude-limited noise on the true attitude value.

[0038] b) Random error is unique to the dual-band infrared radiation attitude measurement system. The core principle of this system is to use infrared radiation sensors to extract navigation information from infrared radiation fields in different areas. However, if the infrared radiation in the space changes, the output data of the infrared radiation sensor will also change accordingly.

[0039] Experiments show that the changes in radiation interference in space are mainly caused by solar interference and cloud interference. Medium-wave sensors are more sensitive to solar interference, and long-wave sensors are more sensitive to clouds. These interferences will cause random errors in infrared radiation data. Random errors cause the attitude angle value to deviate too much during the infrared radiation attitude solution process, which greatly affects the infrared radiation data within the corresponding range and makes the attitude angle information unreliable.

[0040] Preferably, in step S1 of this embodiment, from the perspective of time series, the geometric error is often random and has nothing to do with time, and the amplitude is in the range of [-0.02, 0.02] V during the semi-physical platform measurement process. From the perspective of time series, random errors are often limited by time intervals, and the impact range is within about half a time period, and there is no obvious limit on the amplitude.

[0041] S2. Collect dual-band infrared radiation data through a semi-physical simulation platform, and then process the data in the following steps: a) classify and organize the data set according to the interference type, and b) normalize and clean the data set through a preprocessing module.

[0042] Among them, a) dual-band infrared radiation data under different radiation backgrounds are collected through a semi-physical simulation platform, and the data set is divided into: no interference, solar interference and cloud interference according to the interference type.

[0043] b) Due to factors such as inconsistent amplification factors of the conditioning circuit, the amplitudes of the long-wave infrared radiation data and the medium-wave infrared radiation data are inconsistent. Therefore, all infrared radiation data need to be normalized and the normalized data set needs to be cleaned to ensure the reliability of the data set.

[0044] Preferably, in step S2 of this embodiment, Figure 2 As shown in the figure, the sampling rate of the semi-physical simulation platform is set to 10kHz and the sampling time is set to 10s to collect infrared radiation data. The data sets are collected under different atmospheric radiation backgrounds within one year and the data preprocessing is performed.

[0045] First, normalization is performed to unify the dual-band infrared radiation data into the (-1,1) interval. The formula is as follows:

[0046]

[0047] Where, X i After filtering the input data, Xmin and X max are the minimum and maximum values ​​in the input matrix respectively, and X is the output data after normalization.

[0048] Next, the data set is cleaned and unusable data is deleted.

[0049] S3, build a CBGTE model, input the collected dual-band infrared radiation data set into the CBGTE model, train and learn the CBGTE algorithm, and save the optimal weight of the model. Specifically, the CBGTECBGTE model includes two modules: a) BGT neural network model, b) extended Kalman filter.

[0050] In this embodiment, a) the BGT neural network is a fusion variant network of Bi-GRU and Transformer Encoder, which mainly includes two parts: encoder and decoder.

[0051] The BGT neural network absorbs the advantages of these two architectures: Bi-GRU effectively captures contextual information in the sequence by processing the sequence bidirectionally, thereby taking into account past and future states and enhancing the understanding of time dependencies. Transformer adopts a self-attention mechanism, which allows parallel processing of the entire sequence and is able to capture global contextual information, thereby increasing the weight of key information.

[0052] The main function of the BGT neural network is to compensate for the random errors in the dual-band infrared radiation information. The encoder mainly includes Bi-GRU and Transformer Encoder, and the decoder consists of Bi-GRU and fully connected layers.

[0053] First, Bi-GRU is used to extract time series information from two directions, and the bidirectional extracted feature information is superimposed according to the weights, which can maximize the data utilization.

[0054] Next, the extracted feature information is passed through the Transformer Encoder module, which can effectively extract the features of key information.

[0055] Then, the feature information encoded by the encoder is decoded by the decoder, first decoded by the Bi-GRU layer, and then the decoded features are feature mapped by the fully connected layer.

[0056] Finally, BGT can output dual-band infrared radiation information after compensating for random errors.

[0057] Preferably, in step S3 of this embodiment, Figure 3 As shown, the BGT neural network consists of an encoder and a decoder, and the specific processing is as follows:

[0058] The preprocessed infrared radiation information is input into the encoder. First, bidirectional information extraction is performed through Bi-GRU, where Bi-GRU is improved on the basis of GRU network. The principle of GRU is shown in the following formula:

[0059]

[0060] Where σ represents the sigmoid activation function, W r , W z and W represent the weights of the forget gate, update gate, and hidden layer, respectively. k represents the input of the model at time k, h k-1 represents the hidden layer state at time k-1, represents the sum of the past and current hidden layer states at time k, h k Represents the hidden layer output at time k.

[0061] Bi-GRU adds bidirectional feature extraction based on GRU, as shown in the following formula:

[0062]

[0063] In the formula, GRU(·) represents the GRU operation, and is the forward hidden sequence and the backward hidden sequence, W a and W b represents the weights of the forward and backward hidden sequences.

[0064] Next, the feature information extracted by Bi-GRU is passed through the Transformer Encoder module to extract key information features, where the Transformer Encoder mainly consists of two key parts: multi-layer perceptron (MLP) and multi-head self-attention mechanism.

[0065] The multi-layer perceptron consists of two linear fully connected layers, an activation function, and two Dropout layers. Unlike the traditional Transformer Encoder, changing the activation function from ReLu to SiLu will significantly improve the accuracy. The MLP calculation is shown in the following formula:

[0066]

[0067] Where w1 and b1 represent the weight and bias of the first fully connected layer, and w2 and b2 represent the weight and bias of the second fully connected layer.

[0068] The multi-head self-attention mechanism is an improvement of the self-attention mechanism. The self-attention mechanism enhances the internal feature correlation by minimizing external interference. The input vector is converted into a query matrix Q, a key matrix K, and a value matrix V. The input matrix is ​​weighted and passed through the softmax(·) function, and finally multiplied by the value matrix. The detailed formula is as follows:

[0069]

[0070] Where, d k Represents the dimension of the key matrix, T represents transpose;

[0071] The multi-head self-attention mechanism maps the self-attention mechanism into a multi-dimensional space, and concatenates the features extracted from each space through the Concat(·) function. The formula is as follows:

[0072]

[0073] Where W i Q , W i K and W i V They are the corresponding spatial mapping matrices of the Q, K and V matrices, head i is the head matrix of the i-th dimension space, the Attention(·) function is the self-attention mechanism, and the MHAT(·) function is the multi-head attention mechanism.

[0074] Therefore, the overall Transformer Encoder module can be expressed by the following formula:

[0075] X o =MLP(LN(MHAT(LN(X i ))+X i ))+MHAT(LN(X i ))+X i (7)

[0076] In the formula, X i and X o is the input and output of the Transformer Encoder layer, and the LN(·) function is the LayerNormalization function;

[0077] The overall BGT neural network process is shown in the following formula:

[0078]

[0079] Where X is the input of the neural network, Bi-GRU(·) function is the Bi-GRU operation, TransformerEncoder(·) function is the Transformer Encoder operation, FC(·) function is the fully connected operation, h0 is the bidirectional information feature passed through the Bi-GRU for the first time, h1 is the key feature extracted through the Transformer Encoder module, h2 is the decoded information feature through the Bi-GRU, and Y is the infrared information feature after compensating RE.

[0080] Furthermore, in this embodiment, b) geometric error poses a challenge to the dual-band infrared radiation attitude system in practical application. In the attitude measurement process, various factors will cause geometric errors, including sensor measurement noise, assembly position changes and motor interference, which reduce the reliability of attitude data.

[0081] This embodiment adopts an extended Kalman filter to meet this challenge during the attitude measurement process, and uses the extended Kalman filter to eliminate geometric errors of the infrared radiation information compensated by the BGT neural network.

[0082] The Extended Kalman Filter is a special Kalman filter that handles nonlinear systems by linearizing them locally. It operates by deriving the nonlinear prediction equation from the observation equation and using the tangent approximation. The core principle is the first-order Taylor expansion around the average state.

[0083] The state matrix of the extended Kalman filter is set to X = [γ ω θ] T , γ is the roll angle, ω is the rotation speed, and θ is the pitch angle.

[0084] The state transfer matrix is Among them, Δt is the time step.

[0085] The prediction noise covariance matrix can be obtained by the following formula:

[0086] P(k|k-1)=FP(k-1)F T +Q (9)

[0087] Where P(k|k-1) is the prediction noise covariance matrix, Q is the system noise covariance matrix, and P(k-1) is the covariance matrix at time k-1;

[0088] The observation matrix is ​​expressed as Y = [L1 L2 M1 M2] T According to the dual-band infrared attitude measurement model, the linearized output matrix can be obtained as

[0089] The update step involves predicting the output matrix H and the noise covariance matrix P(k|k-1) to calculate the Kalman gain K, as shown below:

[0090] K(k)=P(k|k-1)H(k) T [H(k)P(k|k-1)H(k) T +R] -1 (10)

[0091] Where R represents the observation noise variance matrix.

[0092] The gain matrix K adjusts the estimated posture parameters and updates the state matrix X. The formula is as follows:

[0093] X(k)=X(k-1)+K(k)[Y(k)-H(k)] (11)

[0094] Where Y(k) represents the actual measurement data of the sensor at time k, and H(k) represents the estimated data of the sensor at time k.

[0095] Finally, the noise covariance matrix P(k) of the filter is updated by the following formula:

[0096] P(k)=[IK(k)H(k)]P(k|k-1) (12)

[0097] It is particularly important to note that in step S3, data enhancement is required before training the CBGTE model, and infrared data is sampled, with every 10s of data as a training sample. The loss function is the mean square error function, the optimizer is set to Adamw, the learning rate is 0.0005, the epoch is set to 200, and the weight decay is set to 0.01. Finally, the posture data curve is obtained based on the deep learning-assisted extended Kalman algorithm.

[0098] Specifically, save the BGT best model weight: save the best weight of the BGT model, set the best loss (loss_best) to 100 before training. After each round of training, input the validation set into the BGT neural network to obtain the loss value on the validation set. If the current loss is less than loss_best, assign the loss to loss_best, save the model weight of this round of training as the best weight, and continue the next round of training; if the loss is greater than loss_best, there is no need to save, and proceed directly to the next round of training. Save the best model weight obtained in the end and verify it on the device.

[0099] S4. Put the optimal weights of the model into the FPGA (ZYNQ7020) device to verify the feasibility of the algorithm and estimate the dual-band infrared radiation attitude of the spinning UAV.

[0100] Specifically, based on the optimized weights of the CBGTE model obtained through optimization, the model weights are transplanted, and the dual-band infrared radiation attitude measurement data collected by the device is input into the FPGA (ZYNQ7020) device for model prediction output.

[0101] In summary, the present invention designs a deep learning assisted extended Kalman dual-band infrared radiation attitude estimation algorithm, and verifies its superior performance through semi-physical experiments. The roll angle error in the stable state is calculated to be ±0.2°, and the pitch angle error is ±0.5°. Figure 4 shown.

[0102] In an embodiment of the present invention, an electronic device is also provided, including: one or more processors; a storage device on which one or more programs are stored; when the one or more programs are executed by the one or more processors, the one or more processors implement the deep learning assisted extended Kalman dual-band infrared radiation attitude estimation method described in any of the above embodiments.

[0103] In an embodiment of the present invention, a computer-readable storage medium is also provided, on which a computer program is stored. When the program is executed by a processor, the steps in any one of the deep learning-assisted extended Kalman dual-band infrared radiation attitude estimation methods in the above embodiments are implemented.

[0104] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A deep learning assisted extended Kalman dual-band infrared radiation attitude estimation method, characterized in that: The steps include: S1. Analyze the causes of geometric errors and random errors in the dual-band infrared radiation measurement signals of the spinning UAV and their impact on the reliability of attitude estimation; S2, collect dual-band infrared radiation data of the spinning UAV through a semi-physical simulation platform, and pre-process the collected data set; S3, constructing a CBGTE model, including a BGT neural network and an extended Kalman filter, inputting the collected data set into the CBGTE model, training and learning the CBGTE model, and obtaining the optimal weight of the CBGTE model; S4. The optimized weights of the CBGTE model are placed in the test equipment to verify the feasibility of the algorithm and estimate the dual-band infrared radiation attitude of the spinning UAV.

2. The deep learning-assisted extended Kalman dual-band infrared radiation attitude estimation method according to claim 1 is characterized in that: In step S1, the geometric error is caused by sensor measurement noise, assembly position and motor interference factors; from the perspective of time series, the geometric error value is random and has nothing to do with time, and the amplitude is in the range of [-0.02, 0.02] V during the measurement process of the semi-physical platform; The random error is caused by solar interference and cloud interference. From the perspective of time series, the random error is limited by a time interval, and the impact range is within half a time period, with no limit on the amplitude.

3. The deep learning-assisted extended Kalman dual-band infrared radiation attitude estimation method according to claim 1 is characterized in that: In step S2, the sampling rate and sampling time of the semi-physical simulation platform are set, dual-band infrared radiation data are collected under different atmospheric radiation backgrounds, and the collected data sets are preprocessed as follows: S201, classifying and arranging the data set according to the interference type, and dividing the data set into: no interference, solar interference, and cloud interference; S202, due to the inconsistency of the amplification coefficient of the conditioning circuit, the amplitudes of the long-wave infrared radiation data and the medium-wave infrared radiation data are inconsistent, and all the collected infrared radiation data are normalized to unify the dual-band infrared radiation data into the interval [-1,1]; S203: Clean the normalized data set, delete unusable data, and ensure the reliability of the data set.

4. The deep learning-assisted extended Kalman dual-band infrared radiation attitude estimation method according to claim 3 is characterized in that: In step S202, all infrared radiation data are normalized, and the formula is as follows: In the formula, X i After filtering the input data, X min and X max are the minimum and maximum values ​​in the input matrix respectively, and X is the output data after normalization.

5. The deep learning-assisted extended Kalman dual-band infrared radiation attitude estimation method according to claim 3 is characterized in that: In step S3, in the CBGTE model, the BGT neural network is used to compensate for random errors in the dual-band infrared attitude measurement process, and the extended Kalman filter is used to eliminate geometric errors in the dual-band infrared attitude measurement process; The BGT neural network integrates the Bi-GRU and Transformer Encoder networks to compensate for random errors in dual-band infrared radiation information; the BGT neural network is composed of an encoder and a decoder, the encoder includes a Bi-GRU module and a Transformer Encoder module, and the decoder is composed of a Bi-GRU and a fully connected layer; The extended Kalman filter derives a nonlinear prediction equation from the observation equation and operates with a tangent approximation, and a first-order Taylor expansion around the average state is used to eliminate geometric errors of infrared radiation information after compensation by the BGT neural network.

6. The deep learning-assisted extended Kalman dual-band infrared radiation attitude estimation method according to claim 5 is characterized in that: In step S3, the collected data set is input into the CBGTE model, and the CBGTE model is trained and learned to obtain the optimal weight of the model, as follows: S301, inputting the preprocessed infrared radiation information into the encoder, performing bidirectional information extraction through the Bi-GRU module, and superimposing the bidirectional extracted feature information according to the weight to maximize data utilization; The Bi-GRU module is based on the improvement of the GRU network, and the formula is as follows: Where σ represents the sigmoid activation function, W r , W z and W represent the weights of the forget gate, update gate, and hidden layer, respectively. k represents the input of the model at time k, h k-1 represents the hidden layer state at time k-1, represents the sum of the past and current hidden layer states at time k, h k represents the hidden layer output at time k; Bidirectional feature extraction is performed through the Bi-GRU module, the formula is as follows: In the formula, GRU(·) represents the GRU operation, and is the forward hidden sequence and the backward hidden sequence, W a and W b represents the weights of the forward and backward hidden sequences; S302, inputting the bidirectional feature information extracted by the Bi-GRU module into the Transformer Encoder module to extract the features of key information; The Transformer Encoder module includes: a multi-layer perceptron and a multi-head self-attention mechanism; The multi-layer perceptron consists of two linear fully connected layers, an activation function and two Dropout layers. The activation function is the SiLu function. The MLP calculation is as follows: Where w1 and b1 represent the weight and bias of the first fully connected layer, w2 and b2 represent the weight and bias of the second fully connected layer; The multi-head self-attention mechanism is based on the improvement of the self-attention mechanism. The input vector is converted into a query matrix Q, a key matrix K and a value matrix V. The input matrix is ​​weighted and passed through the softmax(·) function, and then multiplied by the value matrix. The formula is as follows: Where, d k represents the dimension of the key matrix, T represents transpose; The self-attention mechanism is mapped into a multi-dimensional space through a multi-head self-attention mechanism, and the features extracted from each space are concatenated through the Concat(·) function. The formula is as follows: In the formula, and They are the corresponding spatial mapping matrices of the Q, K and V matrices, head i is the head matrix of the i-th dimension space, Attention(·) function is the self-attention mechanism, and MHAT(·) function is the multi-head attention mechanism; The overall Transformer Encoder module is obtained, which is expressed by the following formula: X o =MLP(LN(MHAT(LN(X i ))+X i ))+MHAT(LN(X i ))+X i ; In the formula, X i and X o is the input and output of the Transformer Encoder layer, and the LN(·) function is the LayerNormalization function; S303, the feature information encoded by the encoder is decoded by the decoder, firstly the decoding is performed through the Bi-GRU network to obtain the decoding features, and then the decoding features are feature mapped through the fully connected layer, and the dual-band infrared radiation information after the overall BGT neural network compensates for the random error is output, which is expressed as follows: Where X is the input of the neural network, Bi-GRU(·) function is the Bi-GRU operation, TransformerEncoder(·) function is the Transformer Encoder operation, FC(·) function is the fully connected operation, h0 is the bidirectional information feature passed through the Bi-GRU for the first time, h1 is the key feature extracted through the Transformer Encoder module, h2 is the decoded information feature through the Bi-GRU, and Y is the infrared information feature after compensating RE.

7. The deep learning assisted extended Kalman dual-band infrared radiation attitude estimation method according to claim 6 is characterized in that: In step S3, the state matrix of the extended Kalman filter is: X=[Goth] T ; The state transfer matrix is: Among them, γ is the roll angle, ω is the rotation speed, θ is the pitch angle; Δt is the time step The prediction noise covariance matrix is ​​obtained by the following formula: P(k|k-1)=FP(k-1)F T +Q; Where P(k|k-1) is the prediction noise covariance matrix, Q is the system noise covariance matrix, and P(k-1) is the covariance matrix at time k-1; The observation matrix is ​​expressed as: And=[L1 L2 M1 M2] T ; Wherein, L1, L2, M1, M2 represent the output data of long wave 1, long wave 2, medium wave 1 and medium wave 2 respectively; According to the dual-band infrared attitude measurement model, the linearized output matrix is ​​obtained as follows: The update step involves predicting the output matrix H and the noise covariance matrix P(k|k-1) to calculate the Kalman gain K, which is: K(k)=P(k|k-1)H(k) T [H(k)P(k|k-1)H(k) T +R] -1 ; Where R represents the observation noise variance matrix; The estimated posture parameters are adjusted by the gain matrix K, and the state matrix X is updated. The formula is as follows: X(k)=X(k-1)+K(k)[Y(k)-H(k)]; In the formula, Y(k) represents the actual measurement data of the sensor at time k, and H(k) represents the estimated data of the sensor at time k; The noise covariance matrix P(k) of the extended Kalman filter is updated by the following formula: P(k)=[IK(k)H(k)]P(k|k-1).

8. The deep learning-assisted extended Kalman dual-band infrared radiation attitude estimation method according to claim 1, characterized in that: In step S3, data enhancement is performed before training the CBGTE model, and the infrared data is sampled, with every 10 seconds of data being used as a training sample; the loss function is set to the mean square error function, the optimizer is Adamw, the learning rate is 0.0005, the epoch is 200, and the weight decay is set to 0.

01. The posture data curve is obtained based on the deep learning-assisted extended Kalman algorithm.

9. An electronic device, characterized in that: include: one or more processors; a storage device having one or more programs stored thereon; When the one or more programs are executed by the one or more processors, the one or more processors implement the deep learning assisted extended Kalman dual-band infrared radiation attitude estimation method as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and when the program is executed by a processor, the steps in the deep learning assisted extended Kalman dual-band infrared radiation attitude estimation method described in any one of claims 1 to 8 are implemented.

Images