Robust parallel polarization orientation error real-time processing method in severe environment

Abstract

The invention discloses a robust parallel polarization orientation error real-time processing method in a severe environment. The method comprises the steps of course data modeling, real-time monitoring and classification, real-time error compensation and denoising and component recombination. Carrying out modeling by adopting multi-scale dynamic kernel principal component analysis (MDKPCA); introducing a kernel density estimation (KDE) strategy to monitor and distinguish original dynamic course data states in real time; a self-adaptive time-frequency peak filtering method (LS-TFPF) with long and short window advantages is adopted to remove noise in the course state of the medium-high frequency band mixed data of the classification result; and meanwhile, a parallel layered transverse anti-Hebbian extra time gated long-short term memory neural network (HT-LSTM) performs error compensation on low-frequency course data of a classification result in real time. The problem that the course error cannot be corrected in real time in a severe environment by the existing polarized light orientation method is solved, and the orientation precision of the polarized light compass is effectively improved.

Description

Technical Field

[0001] The invention relates to the technical field of polarized light navigation, in particular to a robust parallel polarization orientation error real-time processing method under harsh environment. Background Art

[0002] Polarized light navigation system is a fully autonomous navigation system that relies on polarized light in the sky to solve the carrier's heading information. The system does not require any external information and does not radiate energy outward. Therefore, it is widely used in military fields such as aerospace, aviation, and navigation. The polarized light compass is a key component of the polarized light navigation system. The measurement accuracy of the polarized light compass directly determines the performance of the entire polarized light navigation system. Since the orientation accuracy of the polarized light compass is not only easily affected by external harsh environmental factors such as cloudy days, haze, dust and other external environmental factors and the noise generated by the compass itself, but also seriously affected by the sampling time and attitude angle changes during its actual working process, the above factors limit the application of atmospheric polarized light navigation orientation.

[0003] The existing methods that can reduce or eliminate the impact of noise on the orientation accuracy of polarized light compasses and solve the compass error caused by changes in the compass attitude angle are unable to process and monitor the output signal of the polarized light compass in real time, and do not fully solve the impact of the non-uniform sampling time of the compass on its orientation performance, which ultimately limits the effective improvement of the orientation accuracy of the polarized light compass. How to process the output signal of the polarized light compass in real time to ensure its orientation accuracy is a problem that needs to be solved at present. Summary of the invention

[0004] Purpose of the invention: In order to solve the problem that the polarized light navigation system in the prior art cannot correct the heading error in real time under harsh environments and the polarized light compass orientation error is large, the present invention provides a robust parallel polarization orientation error real-time processing method under harsh environments.

[0005] Technical solution: A robust parallel polarization orientation error real-time processing method in harsh environment, comprising the following steps:

[0006] Step 1: Obtain the original heading angle signal output by the polarized compass, model the original heading angle signal based on multi-scale dynamic kernel principal component analysis, and use the process parameters of various frequency heading data as training data;

[0007] Step 2: Calculate T using kernel density estimation strategy 2 and SPE statistics to determine modeling control limits and SPE q，lim ;

[0008] Step 3: Obtain dynamic heading angle signal and determine sample data Extract the nonlinear principal components of the sample data and calculate and SPE q，new Statistical value, based on the calculation results and the control limits of step 2 and SPE q，lim , monitor and classify the heading data status of the current heading angle signal in real time to obtain classification results, which include: low-frequency band heading data, medium-frequency band heading data, and high-frequency band heading data;

[0009] Step 4: Use a hierarchical transverse anti-Hebbian extra time gated long short-term memory neural network to perform error compensation on the low-frequency heading data of the classification results in real time to obtain the compensated heading error data component; in parallel, use an adaptive time-frequency peak filtering method with long and short window advantages to remove noise from the mixed data of the mid-frequency heading data and the high-frequency heading data of the classification results to obtain the denoised heading error data component;

[0010] Step 5: Recombining the compensated heading error data component and the denoised heading error data component to obtain the polarized light compass heading error data after parallel processing.

[0011] Furthermore, in step one, based on multi-scale dynamic kernel principal component analysis, according to the data state differences presented by the original heading angle signal, the time-delay data expansion matrix is ​​used to extract dynamic features from the original heading angle signal as training data.

[0012] Furthermore, in step 1, the modeling process includes extracting nonlinear principal components t of various modal data. q，k Specifically, the radial basis kernel function k(x, y) = <Φ(x), Φ(y)> in the form of dot product is substituted into formula (1) to extract the nonlinear principal components t of various modal data. q，k :

[0013]

[0014] Among them, Φ(·) represents the nonlinear mapping function that maps the input vector x(l) in the input space to the feature space F, is to make the input vector principal component The coefficient of validity; q represents the frequency corresponding to various modal states; Represents k∈R 1×N Perform mean centering processing; i represents the i-th data among the N data of heading data 1 to N; x q (l) represents the th standardized data of the mode corresponding to frequency q; x q，i (l) represents the frequency i-th x in the heading data 1 to N q (l) Data.

[0015] Furthermore, in step 2, T 2The calculation method of SPE statistics is:

[0016]

[0017] Among them, t represents the retention, principal component score vector, S is the retention, the principal component corresponding to the eigenvalue λ i A diagonal matrix composed of (i=1, 2, ..., p), They represent the sum of squares of all principal component score vectors, the retained sum of squares of principal component score vectors, and the sum of squares of the principal component score vectors, respectively.

[0018] Furthermore, in step 3, the sample data The method for determining is:

[0019] The sliding window technology is used to combine the current sample data with the previous sample data to form a new sample data with a window width of l+1. x q，new Indicates the current sampling data.

[0020] Furthermore, in step 4, the hierarchical architecture of the hierarchical lateral anti-Hebbian extra time gated long short-term memory neural network includes a direction encoding layer, a path integration layer, and a lateral anti-Hebbian extra time gated LSTM neural layer, wherein the extra time gated LSTM neural layer is achieved by adding a time gate as a nonlinear scaling factor to the classic LSTM network.

[0021] Furthermore, in step 4, the adaptive time-frequency peak filtering method with the advantages of long and short windows specifically includes:

[0022] (d1) defines the difference threshold DT,

[0023]

[0024] Where q represents L_S i (t) length, L_S i (t) represents the noise in the heading state of the mixed data in the middle and high frequency bands of the classification results;

[0025] (d2) Determine the standard threshold value ST according to the difference threshold value DT. The determination method is: if the lowest frequency heading data L_S i (t) and denoised heading data dL_S i (t) is greater than the difference threshold DT, and the adjustment is used for L_S i (t) The denoising threshold R1 is set until DV≤DT, and R1 is used as the standard threshold ST;

[0026] (d3) Calculate the adaptive threshold AT based on the standard threshold ST i , the calculation method is:

[0027]

[0028] Among them, CC j is the ,th principal element in the feature space F calculated in step 1; CC s is the sum of all principal elements in the feature space F;

[0029] (d4) Set the threshold value of the adaptive threshold. If AT i If the value is greater than the threshold, the long window time-frequency peak filtering method is used to remove the noise in the heading state of the mixed data in the middle and high frequency bands of the classification results; if AT i If the value is less than the threshold, the short window time-frequency peak filtering method is used to remove the noise in the heading state of the mixed data in the middle and high frequency bands of the classification results.

[0030] Compared with the prior art, the robust parallel polarization orientation error real-time processing method in a harsh environment provided by the present invention has the following beneficial effects:

[0031] (1) By modeling the original heading angle signal output by the polarized light compass, the current sampling data is combined with the previous sampling data as new sample data. The past state is also considered when judging the heading data state, which is more in line with the dynamic characteristics of the polarized light compass, making the judgment result more accurate, and being able to monitor and identify the original heading data state and category in real time;

[0032] (2) The hierarchical lateral inverse Hebbian extra time gated long short-term memory neural network is used to solve the problem of the influence of complex attitude changes of the carrier and non-uniform sampling of the compass on the orientation accuracy, and it can compensate the low-frequency heading data error of the above monitoring and classification results in real time;

[0033] (3) While compensating for the low-frequency heading data error, the adaptive time-frequency peak filtering method with the advantages of long and short windows is used in parallel to remove the noise in the mixed data heading state of the mid- and high-frequency bands of the monitoring classification results, which effectively improves the heading measurement accuracy of the polarized compass, especially in harsh environments such as complex weather or occlusion. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 A flowchart of a robust parallel polarization orientation error real-time processing method in harsh environments;

[0035] Figure 2 The result diagram of MKPCA modeling and classification of low-frequency data output by polarized compass raw heading angle data;

[0036] Figure 3 Output of the original heading angle data of the polarized light compass MKPCA modeling classification mid-frequency data result diagram;

[0037] Figure 4 The result diagram of MKPCA modeling and classification of high-frequency data output by the polarized compass;

[0038] Figure 5 Comparison of the effects of compensating low-frequency heading angle errors for various neural network models;

[0039] Figure 6 This is a comparison diagram of the error of heading data in the mid- and high-frequency bands processed by various denoising methods;

[0040] Figure 7 This is a comparison diagram of the real-time processing method for orientation error of the present invention and other heading error processing methods. DETAILED DESCRIPTION

[0041] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments.

[0042] A robust parallel polarization orientation error real-time processing method in harsh environments, such as Figure 1 As shown, the following steps are included:

[0043] Step 1: Obtain the original heading angle signal output by the polarized light compass. According to the data state differences presented by the original heading angle signal, the original heading angle signal is modeled based on multi-scale dynamic kernel principal component analysis. The dynamic features are extracted from the original heading angle signal using the time-delay data expansion matrix. The process parameters of the heading data of various frequencies are used as training data X. q (n q ×m), n q represents the number of samples, q represents the frequency, and m represents the number of variables. The implementation steps are as follows:

[0044] (a1) For training data X q (n q ×m) for standardization, and construct the augmented matrix X according to equation (1) q (l), and use the radial basis kernel function k(x, y) = exp(-||xy|| 2 / c) Calculate the kernel matrix K q ∈R N×N (N=n q -l), for the kernel matrix K q ∈R N×N Perform mean centering.

[0045]

[0046] (a2) Substitute the radial basis kernel function k(x, y) in dot product form into equation (2) to extract the nonlinear principal components of various modal data.

[0047]

[0048] Among them, Φ(·) represents the nonlinear mapping function that maps the input vector x(l) in the input space to the feature space F, is to make the input vector principal component The coefficient of validity; q represents the frequency corresponding to various modal states; Represents k∈R 1×N Perform mean centering processing; i represents the i-th data among the N data of heading data 1 to N; x q (l) represents the th standardized data of the mode corresponding to frequency q; x q,i (l) represents the frequency i-th x in the heading data 1 to N q (l) Data.

[0049] Step 2: Calculate T using kernel density estimation strategy 2 and SPE statistics to determine modeling control limits and SPE q,lim , T 2 The calculation methods of SPE statistics include:

[0050]

[0051] Among them, t represents the score vector of p principal components, and S represents the eigenvalue λ corresponding to the p principal components. i A diagonal matrix composed of (i=1, 2, ..., p), They represent the sum of squares of all principal component score vectors and the sum of squares of the retained p principal component score vectors respectively.

[0052] The control limits and SPE q,lim It can be set according to actual needs. For example, in this embodiment, and SPE q,lim The confidence level was chosen to be α = 0.99.

[0053] like Figure 2 to Figure 4 Shown are the results of modeling and classifying the low, medium and high frequency band data of the original heading angle data.

[0054] Step 3: Obtain dynamic heading angle signal and determine sample data Extract the nonlinear principal components of the sample data and calculate and SPE q，new Statistical value, based on the calculation results and the control limits of step 2 and SPE q,lim, monitor and classify the heading data status of the current heading angle signal in real time, obtain the classification result, and divide the heading data into low-frequency heading data, medium-frequency heading data, and high-frequency heading data. Specifically including:

[0055] (b1) As a dynamic navigation system, the current heading data state output by the polarized light compass is not completely independent between states, but depends largely on the past state. The process parameters at a single moment cannot accurately characterize the nature of data state changes to a certain extent. Therefore, when determining the sampling data, the sliding window technology is used to combine the current sampling data with the previous sampling data to form a new sample data with a window width of l+1. x q，new Indicates the current sampling data;

[0056] (b2) Using K in formula (1) p and 1 N Calculate the kernel vector k t ∈R 1×N And perform mean centering.

[0057]

[0058] (b3) Using equation (2) and setting x q (l) = x q，new Extract nonlinear principal components of sample data q，new .

[0059] (b4) Calculation and SPE q,new And distinguish the heading data modes. If the monitoring statistics and SPE q,new Both are below the modeling control limits and SPE q,new , then it is considered that the heading data state conforms to the modeling state and has consistent frequency characteristics. Let S q = 1. Otherwise, the current heading data state is considered not to be in conformity with the modeling state, and S q =0.

[0060] (b5) Comprehensive judgment. If The current data state is considered to be the corresponding qth heading data state; otherwise, it will be regarded as a judgment failure, and the heading data state of the sample data is classified into low-frequency band heading data, medium-frequency band heading data, and high-frequency band heading data.

[0061] Step 4: Use a layered transverse anti-Hebbian extra time gated long short-term memory neural network to perform error compensation on the low-frequency heading data of the classification results to obtain the compensated heading error data component. The layered architecture of the layered transverse anti-Hebbian extra time gated long short-term memory neural network includes a direction encoding layer, a path integration layer, and a transverse anti-Hebbian extra time gated LSTM neural layer. The extra time gated LSTM neural layer adds a time gate as a nonlinear scaling factor to the classic LSTM network to solve the problem of gradient vanishing and explosion, and can be used for non-uniformly sampled data. Specifically include:

[0062] (c1) The direction encoding layer takes the network input vectors: pitch angle δ, roll angle θ, angle between the solar meridian and the carrier body axis Adjust to the preferred direction so that the span reaches 360°, and the activities of the three types of neurons in this layer are calculated as follows:

[0063]

[0064] in, and Represent the i-th δ, θ and activities. and Represent the i-th δ, θ and preferred direction.

[0065] (c2) Each of the above neurons in the integration layer is equivalent to a phase oscillator. Due to the one-to-one connection with the upstream direction encoding layer, the previous direction encoding and speed information are integrated into the phase of the oscillator.

[0066]

[0067] Among them, ω j =2πf j and f j Denote the angular frequency of each of the above oscillators and the corresponding fundamental frequency of the oscillator. j and β j is the corresponding navigation speed and modulation factor. The oscillatory activity of the phase variable is used as the input of the next lateral anti-Hebbian neural network by integrating the sine function in the layer.

[0068] (c3) The lateral anti-Hebbian neural network response layer connects the integral layer with equation (7) and passes it to the downstream TGLSTM neural network layer, which ensures that the downstream neural network layer has integral information.

[0069]

[0070] The activity L of the i-th neuron in the downstream neural network layer at time t i (t) is as shown in formula (8).

[0071]

[0072] Among them, ω ij Represents the input L P The incoming weight of the jth component of to the i-th neuron of the neural network layer. ik represents the lateral weight from the kth neuron to the ith neuron. m and n represent the input dimension and the total number of neurons in the neural network, respectively.

[0073] (c4) Adjusting the neuroplasticity anti-Hebbian network. Introducing neuroplasticity makes the connection weights of the neural network more flexible to improve its changes and environmental adaptability, as shown below:

[0074]

[0075] Among them, Δω ij and Δω ik Plasticity of afferent and lateral weight was recorded using Hebbian and anti-Hebbian rules, respectively. A and η L Denote the incoming weights Δω ij and the lateral weight Δω ik The learning rate.

[0076] (c5) Training the lateral anti-Hebbian neural network. Continuously monitoring the incoming weights Δω ij and the lateral weight Δω ik Variations Train the Anti-Hebbian Network until the following criteria are met to stop training to ensure that all weights converge.

[0077] ∑|Δω ij |+∑|Δω ik |<T (10)

[0078] The symbol || is an absolute value operation, and T represents the tolerance value determined empirically.

[0079] (c6) Additional time-gated long short-term memory neural network (LSTM) layer, the input vector consists of an arbitrary time interval Δt and the output of the lateral inverse Hebbian neural network, and the output vector is the heading error.

[0080] like Figure 5 The figure shows the comparative effects of various neural network models in compensating low-frequency heading angle errors.

[0081] The adaptive time-frequency peak filtering method with the advantages of long and short windows is used in parallel to denoise the mixed data of the mid-frequency heading data and the high-frequency heading data of the classification results to obtain the denoised heading error data component. Specifically including:

[0082] A hierarchical transverse anti-Hebbian extra time gated long short-term memory neural network is used to compensate the low-frequency heading data error of the classification results in real time; and an adaptive time-frequency peak filtering method with the advantages of long and short windows is used in parallel to remove the noise in the mixed data heading state of the middle and high frequency bands of the classification results. The filtering method specifically includes:

[0083] (d1) define the difference threshold DT,

[0084]

[0085] Where q represents L_S i (t) length, L_S i (t) represents the noise in the heading state of the mixed data in the middle and high frequency bands of the classification results;

[0086] (d2) Determine the standard threshold value ST according to the difference threshold value DT. The determination method is: if the lowest frequency heading data L_S i (t) and denoised heading data dL_S i (t) is greater than the difference threshold DT, and the adjustment is used for L_S i (t) The denoising threshold R1 is set until DV≤≤DT, and R1 is used as the standard threshold ST;

[0087] (d3) Calculate the adaptive threshold AT based on the standard threshold ST i , the calculation method is:

[0088]

[0089] Among them, CC j is the jth principal element in the feature space F calculated in step 1; CC s is the sum of all principal elements in the feature space F;

[0090] (d4) Set the threshold value of the adaptive threshold. If AT i If the value is greater than the threshold, the long window time-frequency peak filtering method is used to remove the noise in the heading state of the mixed data in the middle and high frequency bands of the classification results; if AT i If the value is less than the threshold, the short window time-frequency peak filtering method is used to remove the noise in the heading state of the mixed data in the middle and high frequency bands of the classification results.

[0091] like Figure 6 The figure shows the comparative effect of various denoising methods on the heading data errors in the mid- and high-frequency bands.

[0092] Step 5: Recombine the compensated heading error data component and the denoised heading error data component to obtain the polarized compass heading error data after parallel processing. Figure 7 This is a comparison diagram of the heading angle error between the real-time processing method of parallel polarization orientation error, the single compensation method of low-frequency band heading angle error, the single denoising method of medium- and high-frequency band heading signal, and the existing representative error processing methods. It can be seen from the figure that the real-time processing method of parallel polarization orientation error has the best effect in harsh environments.

Claims

1. A robust parallel polarization orientation error real-time processing method in harsh environment, characterized in that: The following steps are involved: Step 1: Obtain the original heading angle signal output by the polarized compass, model the original heading angle signal based on multi-scale dynamic kernel principal component analysis, and use the process parameters of various frequency heading data as training data; Step 2: Calculate T using kernel density estimation strategy 2 and SPE statistics to determine modeling control limits and SPE q，lim ; Step 3: Obtain dynamic heading angle signal and determine sample data Extract the nonlinear principal components of the sample data and calculate and SPE q，new Statistical value, based on the calculation results and the control limits of step 2 and SPE q，lim , monitor and classify the heading data status of the current heading angle signal in real time to obtain classification results, which include: low-frequency band heading data, medium-frequency band heading data, and high-frequency band heading data; Step 4: Use a hierarchical transverse anti-Hebbian extra time gated long short-term memory neural network to perform error compensation on the low-frequency heading data of the classification results in real time to obtain the compensated heading error data component; in parallel, use an adaptive time-frequency peak filtering method with long and short window advantages to remove noise from the mixed data of the mid-frequency heading data and the high-frequency heading data of the classification results to obtain the denoised heading error data component; Step 5: Recombining the compensated heading error data component and the denoised heading error data component to obtain the polarized light compass heading error data after parallel processing.

2. The robust parallel polarization orientation error real-time processing method under harsh environment according to claim 1, characterized in that: In step 1, based on multi-scale dynamic kernel principal component analysis, according to the data state differences presented by the original heading angle signal, the time-delay data expansion matrix is ​​used to extract dynamic features from the original heading angle signal as training data.

3. The method for real-time processing of robust parallel polarization orientation errors in harsh environments according to claim 2, characterized in that: In step 1, the modeling process involves extracting the nonlinear principal components t of various modal data. q，k Specifically, the radial basis kernel function k(x, y) = <Φ(x), Φ(y)> in the form of dot product is substituted into formula (1) to extract the nonlinear principal components t of various modal data. q，k : Where Φ(·) represents the nonlinear mapping function that maps the input vector x(l) in the input space to the feature space F; is to make the input vector principal component The coefficient of validity; q represents the frequency corresponding to various modal states; Represents k∈R 1×N Perform mean centering processing; i represents the i-th data among the N data of heading data 1 to N; x q (l) represents the lth standardized data of the mode corresponding to frequency q; x q,i (l) represents the frequency i-th x in the heading data 1 to N q (l) Data.

4. The robust parallel polarization orientation error real-time processing method under harsh environment according to any one of claims 1 to 3, characterized in that: In step 2, T 2 The calculation method of SPE statistics is: Among them, t represents the score vector of p principal components, and S represents the eigenvalue λ corresponding to the p principal components. i A diagonal matrix composed of (i=1, 2, ..., p), They represent the sum of squares of all principal component score vectors and the sum of squares of p retained principal component score vectors respectively.

5. The robust parallel polarization orientation error real-time processing method in harsh environment according to any one of claims 1 to 3, characterized in that: In step 3, the sample data The method for determining is: The sliding window technology is used to combine the current sampling data with the previous l sampling data to form a new sample data with a window width of l+1. x q，new Indicates the current sampling data.

6. The robust parallel polarization orientation error real-time processing method in harsh environment according to any one of claims 1 to 3, characterized in that: In step 4, the hierarchical architecture of the hierarchical lateral anti-Hebbian extra time gated long short-term memory neural network includes a direction encoding layer, a path integration layer, and a lateral anti-Hebbian extra time gated LSTM neural layer, wherein the extra time gated LSTM neural layer is achieved by adding a time gate as a nonlinear scaling factor to the classic LSTM network.

7. The robust parallel polarization orientation error real-time processing method under harsh environment according to any one of claims 1 to 3, characterized in that: In step 4, the adaptive time-frequency peak filtering method with the advantages of long and short windows specifically includes: (d1) defines the difference threshold DT, Where q represents L_S i (t) length, L_S i (t) represents the noise in the heading state of the mixed data in the middle and high frequency bands of the classification results; (d2) Determine the standard threshold value ST according to the difference threshold value DT. The determination method is: if the lowest frequency heading data L_S i (t) and denoised heading data dL_S i (t) is greater than the difference threshold DT, and the adjustment is used for L_S i (t) The denoising threshold R1 is set until DV≤DT, and R1 is used as the standard threshold ST; (d3) Calculate the adaptive threshold AT based on the standard threshold ST i , the calculation method is: Among them, CC j is the jth principal element in the feature space F calculated in step 1; CC s is the sum of all principal elements in the feature space F; (d4) Set the threshold value of the adaptive threshold. If AT i If the value is greater than the threshold, the long window time-frequency peak filtering method is used to remove the noise in the heading state of the mixed data in the middle and high frequency bands of the classification results; if AT i If the value is less than the threshold, the short window time-frequency peak filtering method is used to remove the noise in the heading state of the mixed data in the middle and high frequency bands of the classification results.

Images