A space target multi-dimensional multi-domain information data level fusion identification method

Abstract

The application discloses a kind of space target multi-dimensional multi-domain information data level fusion identification methods, comprising the following steps: obtaining the one-dimensional RCS sequence of space target, one-dimensional high-resolution range image and two-dimensional ISAR image;Obtain the one-dimensional luminosity sequence of space target, one-dimensional spectral image and two-dimensional optical image;Get corresponding segmented aggregate after approximation one-dimensional RCS sequence and segmented aggregate after approximation one-dimensional luminosity sequence;Space target one-dimensional high-resolution range image average image and one-dimensional spectral image average image;Gram angle field conversion is carried out unified modeling, is transformed into corresponding gram matrix and is saved as two-dimensional gray scale image;Construct space target multi-dimensional multi-domain image dataset;Get trained space target multi-dimensional multi-domain data fusion intelligent identification model;Get the classification recognition result of space target.The application carries out data level fusion intelligent identification to the multi-dimensional information of space target radar domain and optical domain, and the robustness of recognition result is better.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of signal processing, in particular to a multi-dimensional and multi-domain information data level fusion recognition method for space targets. BACKGROUND

[0002] Multi-dimensional and multi-domain data fusion recognition of space targets has very important significance for improving the performance of space target recognition and enhancing the ability of space situation awareness. Space targets generally include satellites and space debris, etc. At present, the observation of space targets mainly adopts radar observation and optical observation. Radar observation method belongs to active detection, which can obtain the effective scattering cross section (Radar Cross Section, RCS) of space targets, one-dimensional high-resolution range profile (High Resolution Range Profile, HRRP) and two-dimensional inverse synthetic aperture radar (Inverse Synthetic Aperture Radar, ISAR) image. Optical observation belongs to passive observation, which can obtain the visible light brightness (luminosity) of space targets, one-dimensional radiation spectrum information (spectrum) and two-dimensional optical image. Therefore, through the joint active and passive detection of photoelectric sensors, one-dimensional and two-dimensional data of space targets in radar domain and optical domain can be obtained, i.e. multi-dimensional and multi-domain data of space targets. Based on this, the research on multi-dimensional and multi-domain data fusion recognition of space targets is of great significance for improving the performance of space target recognition and has broad application prospects.

[0003] In the prior art, the fusion recognition of space targets mainly focuses on the fusion recognition of multi-dimensional data from a single sensor, such as the fusion recognition of only one-dimensional high-resolution range profile and two-dimensional high-quality ISAR image observed by radar, or the fusion recognition of only one-dimensional spectrum and two-dimensional optical image observed by optical equipment. However, the description of space target characteristics by a single sensor is limited and one-sided, and the utilization of space target information is insufficient, so there is still a problem of poor recognition performance in complex scenes. In addition, the current multi-dimensional data fusion recognition method of space targets by a single sensor mainly includes feature level fusion recognition and decision level fusion recognition. Feature level fusion recognition is to manually extract some basic features from multi-dimensional data of space targets, then obtain higher-dimensional fusion features through splicing or summation operation rules, and finally perform recognition and classification. Decision level fusion recognition is to obtain recognition results based on feature extraction of data in different dimensions, and then analyze and operate based on D-S evidence theory (Dempster-Shafer-Theory) and other artificially designed rules to obtain more accurate and reliable recognition results. However, the current feature level and decision level fusion recognition processes both rely on manual participation, and face serious challenges in processing complex scene space target recognition tasks. In summary, the data level fusion intelligent recognition of multi-dimensional and multi-domain information of space targets is a difficult problem to be solved. SUMMARY

[0004] To solve the above problems, the application provides a space target multi-dimensional multi-domain information data level fusion recognition method, comprising the following steps:

[0005] Step 1, by processing the wide and narrow band echo signals of the space target received by the radar observation equipment respectively, obtaining the one-dimensional RCS sequence, one-dimensional high-resolution range image and two-dimensional ISAR image of the space target; by processing the passive detection information of the space target obtained by the optical observation equipment respectively, obtaining the one-dimensional luminosity sequence, one-dimensional spectral image and two-dimensional optical image of the space target;

[0006] Step 2, using the segmented aggregation approximation method to transform the sequence length to the corresponding segmented aggregation approximation one-dimensional RCS sequence and segmented aggregation approximation one-dimensional luminosity sequence for the one-dimensional RCS sequence and one-dimensional luminosity sequence corresponding to the ISAR imaging time period of the space target;

[0007] Step 3, performing average processing on the multiple one-dimensional high-resolution range images and one-dimensional spectral images corresponding to the ISAR imaging time period of the space target to obtain the one-dimensional high-resolution range image average image and one-dimensional spectral image average image of the space target;

[0008] Step 4, the one-dimensional RCS sequence, one-dimensional luminosity sequence, one-dimensional high-resolution range image average image and one-dimensional spectral image average image of the space target obtained after steps 2 and 3 are respectively subjected to Gram angle field conversion for unified modeling, transformed into corresponding Gram matrix-like matrix and saved as two-dimensional gray scale image;

[0009] Step 5, combining the ISAR image, optical image obtained in step 1 and the multiple Gram angle field transformation two-dimensional gray scale images obtained after step 4, labeling according to the category of the space target, and expanding the image library through data augmentation operation, and then constructing a space target multi-dimensional multi-domain image dataset;

[0010] Step 6, according to the space target multi-dimensional multi-domain image dataset labeled and constructed in step 5, training the pre-designed deep neural network model to obtain a trained space target multi-dimensional multi-domain data fusion intelligent recognition model;

[0011] Step 7, receiving the space target multi-dimensional multi-domain data to be identified, inputting into the space target multi-dimensional multi-domain data fusion intelligent recognition model trained in step 6 to obtain the classification recognition result of the space target.

[0012] Further, step 1 specifically comprises the following steps:

[0013] 1.1 The wide and narrow band echo signals of a space target are recorded by an ISAR radar observation system, the one-dimensional RCS sequence of the space target is obtained by pulse compression processing of the narrow band echo signals, denoted as x rcs,t ; the one-dimensional high-resolution range image of the space target is obtained by pulse compression processing of the one-dimensional range echo signals obtained by the wideband radar, denoted as x hrrp,t ; the two-dimensional ISAR image of the space target is obtained by distance-Doppler algorithm processing of the wideband one-dimensional range echo signals within a certain time period, denoted as x ISAR ;

[0014] 1.2 The passive detection information of the space target obtained by the optical observation system is processed to obtain the one-dimensional luminosity sequence, the one-dimensional spectral image and the two-dimensional optical image of the space target, denoted as x phm,t , x spect,t , x optical .

[0015] Further, step 2 specifically includes the following steps:

[0016] 2.1 The specific formula for size transformation of the one-dimensional RCS sequence corresponding to the ISAR imaging time period of the space target is:

[0017]

[0018] Where t = 1, 2, …, T represents the original time sequence, l = 1, 2, …, N represents the time sequence after segmentation and aggregation approximation, x rcs,t represents the RCS value of the space target at the original t time, x rcs,l represents the average RCS value of the space target in the lth segment after segmentation and aggregation approximation.

[0019] 2.2 The specific formula for size transformation of the one-dimensional luminosity sequence corresponding to the ISAR imaging time period of the space target is:

[0020]

[0021] Where t = 1, 2, …, T represents the original time sequence, l = 1, 2, …, N represents the time sequence after segmentation and aggregation approximation, x phm,t represents the luminosity value of the space target at the original t time, x phm,l represents the average luminosity value of the space target in the lth segment after segmentation and aggregation approximation.

[0022] Further, step 3 specifically includes the following steps:

[0023] 3.1 The specific implementation formula for average processing of the multiple one-dimensional high-resolution range images corresponding to the ISAR imaging time period of the space target is:

[0024]

[0025] where x hrrp,t represents the one-dimensional high-resolution range profile at time t, x hrrp represents the one-dimensional high-resolution range profile average image.

[0026] 3.2 The specific formula for averaging the multiple one-dimensional spectral images corresponding to the ISAR imaging time period of the spatial target is:

[0027]

[0028] where x spec,t represents the one-dimensional spectral image at time t, x spec represents the one-dimensional spectral image average image.

[0029] Further, step 4 specifically includes the following steps:

[0030] 4.1 Scale the one-dimensional RCS sequence, the one-dimensional luminosity sequence, the one-dimensional high-resolution range profile average image, and the one-dimensional spectral image average image obtained after steps 2 and 3, respectively, to scale the data range to [0, 1], and the formula is as follows:

[0031]

[0032]

[0033]

[0034]

[0035] where X rcs = [x rcs,1 , x rcs,2 ,..., x rcs,N ] represents the one-dimensional RCS sequence after piecewise aggregation approximation, X phm = [x phm,1 , x phm,2 ,..., x phm,N ] represents the one-dimensional luminosity sequence after piecewise aggregation approximation, max() represents the maximum value in the sample sequence, and min() represents the minimum value in the sample sequence.

[0036] 4.2 Encode the data after scaling and normalization, respectively, and map them to the polar coordinate system:

[0037]

[0038] where l = 1, 2,..., L, L is a positive integer, φ l represents the angle in polar coordinates, and rl representing the radius in polar coordinates, x l representing the input time series or one-dimensional vector, i.e. representing the scaled and normalized data x in step 4.1 rcs,l , x phm,l , x hrrp , x spec ; the above data are respectively substituted into the above formula to obtain the angle and radius φ rcs,l , r rcs,l , φ phm,l , r phm,l , φ hrrp,l , r hrrp,l , φ spec,l , r spec,l ;

[0039] 4.3 The data φ rcs,l , r rcs,l , φ phm,l , r phm,l , φ hrrp,l , r hrrp,l , φ spec,l , r spec,l processed in step 4.2 are respectively unified modeling by Gram angle field conversion, transformed into corresponding Gram-like matrix and saved as a two-dimensional gray image, specifically:

[0040]

[0041]

[0042] The one-dimensional angle vector φ rcs,l , φ phm,l , φ hrrp,l , φ spec,l under the polar coordinate system is respectively substituted into the above formula to obtain the Gram-like matrix and saved as a corresponding two-dimensional gray image, denoted as X rcs,gaf , X hrrp,gaf , X phm,gaf , X spec,gaf .

[0043] Further, step 5 specifically includes the following steps:

[0044] 5.1 The two-dimensional ISAR image, the two-dimensional optical image of the space target obtained in step 1 and the multiple two-dimensional gray images after Gram angle field conversion obtained after step 4 are combined, if the image sizes are inconsistent, the image is zero-filled to make the image size consistent, and manual labeling is performed according to the category of the space target;

[0045] 5.2 The image is augmented, including image translation, image flipping, image rotation, image skewing, image cropping, image scaling, image noise adding, image occlusion, and combined transformation, to construct a spatial target multi-dimensional multi-domain image dataset.

[0046] Further, in step 6:

[0047] The spatial target multi-dimensional multi-domain image dataset includes a training set, a validation set, and a test set;

[0048] The pre-designed deep neural network model is a deep neural network model structure with multiple input channels, including two modules: a feature extraction module and a category prediction module; The formula of the category prediction module is specifically:

[0049] L FocalLoss =-α c (1-p c ) γ log(p c )

[0050] Wherein, L FocalLoss represents the Focal Loss loss function, p c represents the predicted probability value of the cth spatial target, and alpha c and gamma represent hyperparameters.

[0051] Further, step 7 specifically includes the following steps:

[0052] 7.1 Receive the spatial target multi-dimensional multi-domain data to be identified recorded by the radar observation equipment and the optical observation equipment, and uniformly model the spatial target multi-dimensional multi-domain data using the Gram angle field transformation according to the method of steps 1 to 4, and convert it into multiple gray scale images to be identified representing multiple dimensional information;

[0053] 7.2 Input the multiple gray scale images to be identified into the trained spatial target multi-dimensional multi-domain data fusion intelligent identification model to obtain the classification and identification result of the spatial target, and complete the identification task.

[0054] Compared with the prior art, the beneficial effects of the present application are:

[0055] (1) The spatial target multi-dimensional multi-domain data obtained by the radar observation equipment and the optical observation equipment is uniformly modeled for the first time, converted into the same modal data for processing and analysis, that is, data-level fusion intelligent identification;

[0056] (2) The spatial target is classified and recognized by using the spatial target multi-dimensional multi-domain data fusion intelligent recognition model with multiple input channels completed by pre-designed training; the multi-dimensional information of the spatial target in the radar domain and the optical domain can be data-level fusion intelligent recognized, the description of the characteristics of the spatial target is more comprehensive, the utilization of the information of the spatial target is more sufficient, the spatial target feature extraction and recognition process do not need human participation, the spatial target recognition precision in the complex scene is higher, and the robustness of the recognition result is better. BRIEF DESCRIPTION OF DRAWINGS

[0057] Figure 1 is the implementation flowchart of the embodiment of the present application.

[0058] Figure 2 is the one-dimensional RCS sequence, the one-dimensional high-resolution range image, and the two-dimensional ISAR image obtained after processing the spatial target echo data observed by the radar equipment of the embodiment of the present application;

[0059] Figure 3 is the two-dimensional optical image obtained after processing the passive detection information of the spatial target observed by the optical equipment of the embodiment of the present application;

[0060] Figure 4 is the one-dimensional data transformation process flowchart of the embodiment of the present application;

[0061] Figure 5 is the polar coordinate representation of the one-dimensional high-resolution range image of the spatial target and the two-dimensional image after the Gram angle field conversion of the embodiment of the present application.

[0062] Figure 6 is the deep neural network model structure schematic diagram designed by the embodiment of the present application.

[0063] Figure 7 is the confusion matrix obtained by using the trained spatial target multi-dimensional multi-domain data fusion intelligent recognition model to recognize the data in the test set in the embodiment of the present application; wherein (a) corresponds to the test set data in the radar multi-dimensional data, (b) corresponds to the test set data in the optical multi-dimensional data, and (c) corresponds to the test set data in the optical radar multi-dimensional multi-domain data.

[0064] Figure 8 is the feature visualization result graph extracted by using the trained spatial target multi-dimensional multi-domain data fusion intelligent recognition model to recognize the data in the test set in the embodiment of the present application; wherein (a) corresponds to the test set data in the radar multi-dimensional data, (b) corresponds to the test set data in the optical multi-dimensional data, and (c) corresponds to the test set data in the optical radar multi-dimensional multi-domain data. DETAILED DESCRIPTION

[0065] The following detailed description of a spatial target multi-dimensional multi-domain information data level fusion recognition method of the application is made with reference to the accompanying drawings.

[0066] Referring to Figure 1 , the embodiment discloses a spatial target multi-dimensional multi-domain information data level fusion recognition method, and the specific implementation steps are as follows:

[0067] Step 1, by processing the wideband and narrowband echo signals of the spatial target received by the radar observation equipment respectively, the one-dimensional RCS sequence, the one-dimensional high-resolution range image and the two-dimensional ISAR image of the spatial target are obtained; by processing the passive detection information of the spatial target obtained by the optical observation equipment respectively, the one-dimensional luminosity sequence, the one-dimensional spectral image and the two-dimensional optical image of the spatial target are obtained, as shown in Figure 2 and Figure 3 .

[0068] The specific sub-steps of step 1 are as follows:

[0069] 1.1 By recording the wideband and narrowband echo signals of the spatial target through the ISAR radar observation system, the one-dimensional RCS sequence of the spatial target is obtained by pulse compression processing of the narrowband echo signal, denoted as x rcs,t ; the one-dimensional high-resolution range image of the spatial target is obtained by pulse compression processing of the one-dimensional range echo signal obtained by the wideband radar, denoted as x hrrp,t ; the two-dimensional ISAR image of the spatial target is obtained by distance-Doppler algorithm processing of the wideband one-dimensional range echo signal within a certain time period, denoted as x ISAR .

[0070] 1.2 By processing the passive detection information of the spatial target obtained by the optical observation system, the one-dimensional luminosity sequence, the one-dimensional spectral image and the two-dimensional optical image of the spatial target are obtained by processing the data of each optical sensor respectively, denoted as x phm,t , x spect,t , x optical .

[0071] Step 2, by using the piecewise aggregation approximation method, the sequence length of the one-dimensional RCS sequence and the one-dimensional luminosity sequence corresponding to the ISAR imaging time period of the spatial target is transformed to a suitable size, and the corresponding one-dimensional RCS sequence after piecewise aggregation approximation and the one-dimensional luminosity sequence after piecewise aggregation approximation are obtained.

[0072] The specific sub-steps of step 2 are as follows:

[0073] 2.1 The one-dimensional RCS sequence corresponding to the space target ISAR imaging time period is transformed to a suitable size by using the segmented aggregation approximation method. Due to the large amount of RCS data, the RCS sequence length is too long, there is certain redundant information, and after transformation, the image dimension is too high, which is not suitable for directly using a deep neural network model for feature extraction. Therefore, the RCS sequence must be segmented and aggregated, that is, the sequence is first segmented, then the subsequence in each segment is compressed into a value by averaging, thereby reducing the sequence length, and the specific implementation method can be realized by the following formula:

[0074]

[0075] where t = 1, 2, …, T represents the original time sequence, l = 1, 2, …, N (N is a positive integer) represents the time sequence after segmented aggregation approximation, X rcs,t represents the RCS value of the space target at the original time t, and X rcs,l represents the average RCS value of the space target in the lth segment after segmented aggregation approximation.

[0076] 2.2 The one-dimensional photometric sequence corresponding to the space target optical imaging time period is transformed to a suitable size by using the segmented aggregation approximation method. Since the photometric sequence has redundant information, and after transformation, the image dimension is too high, it is not suitable for directly using a deep neural network model for feature extraction. Therefore, the photometric sequence must also be transformed to a suitable size by using the segmented aggregation approximation method

[0077]

[0078] where t = 1, 2, …, T represents the original time sequence, l = 1, 2, …, N represents the time sequence after segmented aggregation approximation, x phm,t represents the photometric value of the space target at the original time t, and x phm,l represents the average photometric value of the space target in the lth segment after segmented aggregation approximation.

[0079] Step 3, average processing of multiple one-dimensional high-resolution range images and one-dimensional spectral images corresponding to the space target ISAR imaging time period to obtain one-dimensional high-resolution range image average image and one-dimensional spectral image average image of the space target.

[0080] The specific sub-step of step 3 is:

[0081] 3.1 Average processing of multiple one-dimensional high-resolution range images corresponding to the space target ISAR imaging time period to obtain one-dimensional high-resolution range image average image of the space target, so as to alleviate the azimuth sensitivity between the HRRP range images corresponding to the ISAR image, and the scattering point distribution in the one-dimensional high-resolution range image average image is more stable, and the specific implementation method is:

[0082]

[0083] where x hrrp,t represents the one-dimensional high-resolution range image at time t, x hrrp represents the one-dimensional high-resolution range image average image.

[0084] 3.2 The one-dimensional spectral image average image of the spatial target is obtained by averaging a plurality of one-dimensional spectral images corresponding to the optical imaging time period of the spatial target, specifically:

[0085]

[0086] where x spec,t represents the one-dimensional spectral image at time t, x spec represents the one-dimensional spectral image average image.

[0087] Step 4: The one-dimensional RCS sequence, the one-dimensional luminosity sequence, the one-dimensional high-resolution range image average image, and the one-dimensional spectral image average image of the spatial target obtained after steps 2 and 3 are processed are respectively subjected to Gram angle field conversion for unified modeling, transformed into corresponding Gram-like matrices, and saved as two-dimensional gray images. The implementation results are shown in Figure 4 and Figure 5 .

[0088] The specific sub-steps of step 4 are:

[0089] 4.1 The one-dimensional RCS sequence, the one-dimensional luminosity sequence, the one-dimensional high-resolution range image average image, and the one-dimensional spectral image average image obtained after steps 2 and 3 are processed are respectively scaled, and the data range is scaled to [0, 1], and the formula is as follows:

[0090]

[0091]

[0092]

[0093]

[0094] where X rcs = [x rcs,1 , x rcs,2 ,..., x rcs,N ] represents the one-dimensional RCS sequence after piecewise aggregation approximation, X phm = [x phm,1 , x phm,2 ,..., x phm,N ] represents the one-dimensional luminosity sequence after piecewise aggregation approximation, max() represents the maximum value in the sample sequence, and min() represents the minimum value in the sample sequence.

[0095] 4.2 The scaled and normalized data are respectively encoded and mapped to the polar coordinate system:

[0096]

[0097] where l = 1, 2,..., L (L is a positive integer), φ l represents the angle in polar coordinates, r l represents the radius in polar coordinates, x l represents the input time series or one-dimensional vector, which here represents the scaled and normalized data x rcs,l , x phm,l , x hrrp , x spec . By substituting the above data into the above formula, the angle and radius φ rcs,l , r rcs,l , φ phm,l , r phm,l , φ hrrp,l , r hrrp,l , φ spec,l , x spec,l in the polar coordinate system can be obtained.

[0098] 4.3 The data in the polar coordinate system processed in step 4.2 are respectively unified modeled by Gramian Angular Summation Field (GASF) or Gramian Angular Difference Field (GADF), transformed into corresponding Gramian-like matrix and saved as a two-dimensional gray image, which is specifically:

[0099]

[0100]

[0101] The one-dimensional angle vector φ rcs,l , φ phm,l , φ hrrp,l , φ spec,l in the polar coordinate system is respectively substituted into the above formula for calculation to obtain the Gramian-like matrix and save it as a corresponding two-dimensional gray image, denoted as X rcs,gaf , X hrrp,gaf , X phm,gaf , X spec,gafThe advantage of converting one-dimensional sequence data into two-dimensional images by using the Gram angular field transformation method is that not only the integrity of the original data information is preserved, but also the dependence of one-dimensional sequence data on time is preserved through the positional relationship during the mapping process, as the corresponding mapping part of the Gram-like matrix moves from the upper left corner to the lower right corner with the passage of time.

[0102] Step 5: Combine the two-dimensional ISAR image of the spatial target obtained in step 1, the two-dimensional optical image, and the multiple two-dimensional gray-scale images after Gram angular field transformation processed in step 4, label according to the category of the spatial target, and expand the image library through data augmentation operation, and then construct a multi-dimensional and multi-domain image dataset of the spatial target.

[0103] The specific sub-steps of step 5 are:

[0104] 5.1 Combine the ISAR image, optical image obtained in step 1, and multiple two-dimensional gray-scale images after Gram angular field transformation processed in step 4, if the image sizes are inconsistent, perform zero padding operation to make the image width and height consistent, and manually label according to the category of the spatial target.

[0105] 5.2 Augment the image, including image translation, image flipping, image rotation, image scaling, image noise addition, image occlusion, and combination transformation, to construct a multi-dimensional and multi-domain image dataset of the spatial target.

[0106] Step 6: Train the pre-designed deep neural network model according to the multi-dimensional and multi-domain image dataset of the spatial target labeled and constructed in step 5, obtain the trained multi-dimensional and multi-domain data fusion intelligent recognition model, and the implementation process is shown in Figure 6 .

[0107] The specific sub-steps of step 6 are:

[0108] 6.1 Divide the multi-dimensional and multi-domain image dataset of the spatial target labeled and constructed in step 5 into training set, validation set and test set, with the proportion of 60%, 20% and 20% respectively.

[0109] 6.2 The structure of the deep neural network model with multiple input channels is shown in Figure 6 , mainly including two modules: feature extraction module and category prediction module.

[0110] a) The feature extraction module is used to complete the high-dimensional feature extraction of the multi-dimensional and multi-domain image data of the space target, which facilitates subsequent category prediction. Commonly used network modules for feature extraction include VGG network using repeated elements, NiN network in network, GoogleNet containing parallel connections, ResNet, DenseNet with dense connections, and various deep neural network model variants derived therefrom. In this example, a ResNet-50 deep neural network model is used as the space target multi-dimensional and multi-domain image data feature extraction network module.

[0111] b) The category prediction module is used to complete the classification and identification of the space target. Considering the uneven distribution of sample quantities and the inconsistent difficulty levels of sample classification of various categories of space targets in actual scenarios, this example uses a Focal Loss loss function for classification prediction. Specifically:

[0112] L FocalLoss = -α c (1-p c ) γ log(p c )

[0113] where p c represents the predicted probability value of the cth category of space target, α c and γ represent hyperparameters, and in this example, α c = 0.3 and γ = 2.

[0114] 6.3 Based on the design of the feature extraction module of the deep neural network model and the definition of the loss function of the category prediction module, the deep neural network model is trained, and the Adam optimizer and adaptive learning rate adjustment are used. Finally, a trained space target multi-dimensional and multi-domain data fusion intelligent identification model is obtained.

[0115] Step 7, receiving the space target multi-dimensional and multi-domain data to be identified, inputting into the trained space target multi-dimensional and multi-domain data fusion intelligent identification model in step 6 to obtain the classification and identification result of the space target to be identified.

[0116] The specific sub-steps of step 7 are:

[0117] 7.1 Receive the space target multi-dimensional and multi-domain data to be identified recorded by the radar observation equipment and the optical observation equipment, and use the Gram angle field transformation to uniformly model the space target multi-dimensional and multi-domain data to be identified according to the method of steps 1 to 4, and convert it into multiple gray scale images representing multiple dimensional information.

[0118] 7.2 inputting multiple gray scale images to be identified into the trained spatial target multi-dimensional multi-domain data fusion intelligent recognition model with multiple input channels trained in step 6, obtaining a classification recognition result of the spatial target, and completing the recognition task.

[0119] The simulation data set in the embodiment of the application has seven types of spatial targets, the optical multi-dimensional data of the spatial target is obtained by an optical simulation software such as Blender, the radar multi-dimensional data is obtained by an electromagnetic simulation software FEKO, a total of 74,000 samples are obtained, and the data set is divided into a training set and a test set according to a ratio of 8:2, and 20% of the training set is further divided as a verification set. Based on the deep neural network model shown in the embodiment of the application Figure 6 , the model in the embodiment is trained by using the pre-training weight of the ResNet-50 deep neural network model and Fine-Tuning, and the data in the test set is tested. Through the recognition and classification of the test set, the confusion matrix obtained is as shown in Figure 7 . As can be seen from the figure, the classification recognition accuracy of the spatial target by only relying on the multi-dimensional data of the optical sensor or the radar sensor is about 85%, and the recognition performance in a complex scene may be poor. The classification recognition accuracy of the spatial target radar optical multi-dimensional multi-domain information data level fusion recognition method proposed in the application is 100%, which indicates the feasibility of the method proposed in the application in the field of spatial target multi-source information fusion recognition. At the same time, based on the spatial target multi-dimensional multi-domain data fusion intelligent recognition model in the embodiment of the application, the features extracted from the data in the test set are two-dimensional plane visualized by using the t-SNE dimension reduction technology, as shown in Figure 8 . As can be seen from the figure, the features extracted by the data level fusion recognition of the radar optical multi-dimensional multi-domain information have obvious spatial separability and more dense spatial clustering, which can greatly improve the recognition accuracy. Finally, Table 1 shows the accuracy, recall rate and F1 value of the classification recognition of the data in the test set by the spatial target multi-dimensional multi-domain data fusion intelligent recognition model in the embodiment of the application. As can be seen from Table 1, the data level fusion recognition of the radar optical multi-dimensional multi-domain information of the spatial target achieves the best effect in the above three indicators.

[0120] Table 1

[0121] Test dataset Accuracy Recall F1 value Radar multidimensional data 83.8％ 83.4％ 83.6％ Optical multidimensional data 85.3％ 85.1％ 85.2％ Optical radar multidimensional multidomain data 100.0％ 100.0％ 100.0％

[0122] The application firstly unifies the spatial target multi-dimensional multi-domain data obtained by the radar observation equipment and the optical observation equipment, converts the data into the same modal data for processing and analysis, that is, data level fusion intelligent recognition. Then, the spatial target is classified and recognized by using the spatial target multi-dimensional multi-domain data fusion intelligent recognition model with multiple input channels trained in advance.

[0123] Compared with the prior art mainly focusing on single-sensor multi-dimensional data feature level fusion recognition or decision level fusion recognition, the method can perform data level fusion intelligent recognition on multi-dimensional information of the space target radar domain and the optical domain, more comprehensively describes the characteristics of the space target, more fully utilizes the space target information, and does not require manual participation in the space target feature extraction and recognition process, so that the space target recognition precision is higher in a complex scene, and the robustness of the recognition result is better.

[0124] The above merely describes preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for multi-dimensional and multi-domain information data level fusion recognition of space targets, characterized in that, Comprise the following steps: Step 1, by processing the wide and narrow band echo signals of the space target received by the radar observation equipment respectively, obtain the one-dimensional RCS sequence, one-dimensional high-resolution range image and two-dimensional ISAR image of the space target; by processing the passive detection information of the space target obtained by the optical observation equipment respectively, obtain the one-dimensional luminosity sequence, one-dimensional spectral image and two-dimensional optical image of the space target; Step 2, the one-dimensional RCS sequence and the one-dimensional luminosity sequence corresponding to the ISAR imaging time period of the space target are transformed in size by using the piecewise aggregation approximation method, and the corresponding one-dimensional RCS sequence after piecewise aggregation approximation and the one-dimensional luminosity sequence after piecewise aggregation approximation are obtained; Step 3, the multiple one-dimensional high-resolution range images and one-dimensional spectral images corresponding to the ISAR imaging time period of the space target are averaged to obtain the one-dimensional high-resolution range image average image and one-dimensional spectral image average image of the space target; Step 4, the one-dimensional RCS sequence, one-dimensional luminosity sequence, one-dimensional high-resolution range image average image and one-dimensional spectral image average image of the space target obtained after steps 2 and 3 are respectively subjected to Gram angle field conversion for unified modeling, transformed into corresponding Gram matrix and saved as two-dimensional gray image; Step 5, the two-dimensional ISAR image, two-dimensional optical image of the space target obtained in step 1 and the multiple Gram angle field conversion two-dimensional gray images obtained after step 4 are combined, labeled according to the category of the space target, and the image library is expanded through data augmentation operation, and then a space target multi-dimensional multi-domain image data set is constructed; Step 6, according to the space target multi-dimensional multi-domain image data set labeled in step 5, a pre-designed deep neural network model is trained to obtain a trained space target multi-dimensional multi-domain data fusion intelligent recognition model; Step 7, receiving the space target multi-dimensional multi-domain data to be identified, inputting into the space target multi-dimensional multi-domain data fusion intelligent recognition model trained in step 6 to obtain the classification recognition result of the space target.

2. The method according to claim 1, wherein, Step 1 specifically comprises the following steps: 1.1 The wide and narrow band echo signals of a space target are recorded by an ISAR radar observation system, the one-dimensional RCS sequence of the space target is obtained by pulse compression processing of the narrow band echo signals, denoted as ; the one-dimensional high-resolution range image of the space target is obtained by pulse compression processing of the one-dimensional range direction echo signals obtained by the wide band radar, denoted as ; the two-dimensional ISAR image of the space target is obtained by range-Doppler algorithm processing of the wide band one-dimensional range direction echo signals in a certain time period, denoted as ; 1.2 The passive detection information of the space target obtained by the optical observation system is processed respectively to obtain the one-dimensional luminosity sequence, one-dimensional spectral image and two-dimensional optical image of the space target, which are respectively denoted as , , .

3. The method of claim 2, wherein, Step 2 specifically comprises the following steps: 2.1 The specific formula for size transformation of the one-dimensional RCS sequence corresponding to the ISAR imaging time period of the space target is: wherein represents the original time series, represents the time series after piecewise aggregate approximation, represents the original spatial target RCS value at time instant, represents the spatial target average RCS value of the segment after piecewise aggregate approximation; 2.2 The specific formula for size transformation of the one-dimensional luminosity sequence corresponding to the ISAR imaging time period of the space target is: wherein represents the original time series, represents the time series after piecewise aggregate approximation, represents the original spatial target luminance value at time instant t, represents the spatial target average luminance value of the segment after piecewise aggregate approximation.

4. The method of claim 3, wherein, Step 3 specifically comprises the following steps: 3.1 The specific implementation formula for average processing of the multiple one-dimensional high-resolution range images corresponding to the ISAR imaging time period of the space target is: wherein, represent a one-dimensional high resolution range profile at a time instant, represent a one-dimensional high resolution range profile average image; 3.2 The specific formula for average processing of the multiple one-dimensional spectral images corresponding to the ISAR imaging time period of the space target is: wherein, represent one-dimensional spectral image at a time instant, represent an average image of one-dimensional spectral images.

5. The method of claim 4, wherein, Step 4 specifically comprises the following steps: 4.1 The one-dimensional RCS sequence, one-dimensional luminosity sequence, one-dimensional high-resolution range image average image and one-dimensional spectral image average image of the space target obtained after steps 2 and 3 are respectively scaled, and the data range is scaled to [0, 1], the formula is as follows: wherein denotes the one-dimensional RCS sequence after piecewise aggregation approximation, denotes the one-dimensional photometric sequence after piecewise aggregation approximation, denotes the maximum value in the sample sequence, denotes the minimum value in the sample sequence; 4.2 The data after scaling and normalization processing is respectively encoded and mapped to the polar coordinate system: wherein L is a positive integer, represents the angle in polar coordinates, represents the radius in polar coordinates, represents the input time series or one-dimensional vector, i.e. the scaled and normalized data from step 4.1 , , , ; the above data are put into the above equations, respectively, to obtain the angle and the radius in polar coordinates , , , , , , , ; 4.3 Data in polar coordinate system after step 4.2 is processed 、 、 、 、 、 、 、 Respectively, through the unified modeling of Gram angle field conversion, it is transformed into the corresponding Gram-like matrix and saved as a two-dimensional gray image, which is: One-dimensional angle vector in polar coordinate system , , , Respectively, the calculation is carried out by bringing into the above formula, get the class gram matrix and save as the corresponding two-dimensional gray image, recorded as 、 、 、 .

6. The method of claim 5, wherein, Step 5 specifically comprises the following steps: 5.1 Combine the two-dimensional ISAR image of the space target obtained in step 1, the two-dimensional optical image, and the multiple two-dimensional gray images after Gram angle field transformation obtained after step 4 processing. If the image sizes are inconsistent, perform zero padding operation on the images to make the image width and height sizes consistent, and manually label according to the category of the space target; 5.2 Perform augmentation processing on the images, including image translation, image flipping, image rotation, image skewing, image cropping, image scaling, image noise adding, image occlusion, and combination transformation operation, to construct a space target multi-dimensional multi-domain image dataset.

7. The method of claim 6, wherein, In step 6: The space target multi-dimensional multi-domain image dataset includes a training set, a validation set, and a test set; The pre-designed deep neural network model is a deep neural network model structure with multiple input channels, including two modules: a feature extraction module and a category prediction module. The formula of the category prediction module is specifically: wherein, denotes the Focal Loss loss function, denotes the prediction probability value of the i-th class space object, denotes the prediction probability value of the i-th class space object, denotes a hyper-parameter.

8. The method of claim 7, wherein, Step 7 specifically includes the following steps: 7.1 Receive the space target multi-dimensional multi-domain data to be identified recorded by the radar observation equipment and the optical observation equipment, and uniformly model the space target multi-dimensional multi-domain data using Gram angle field transformation according to the method of steps 1 to 4, converting it into multiple gray images to be identified representing multiple dimensional information; 7.2 Input the multiple gray images to be identified into the trained space target multi-dimensional multi-domain data fusion intelligent recognition model to obtain the classification recognition result of the space target, completing the recognition task.

Images