Multifunctional radar working mode learning method and device and electronic equipment

Abstract

The invention discloses a multifunctional radar working mode learning method and device and electronic equipment, and the method comprises the steps: carrying out the feature extraction of an input radar pulse feature parameter through a preset feature extractor; radar features can be extracted more comprehensively and accurately, so that accurate identification and effective rejection of the radar working mode in a countermeasure scene with dynamic change of an electromagnetic environment can be realized. And then, based on a DBSCAN algorithm accelerated by a Kdtree data structure, carrying out clustering marking on radar pulse characteristic parameters corresponding to an unknown-class radar working mode, so that the clustering efficiency is remarkably improved. Finally, training a complementary learning model based on the data of the scene memory buffer area and the radar pulse characteristic parameters marked with the category labels; new radar pulse characteristic parameters are continuously added into the scene memory buffer area, so that the complementary learning model can keep the memory ability for a known radar working mode and can continuously learn an unknown radar working mode.

Description

Technical Field

[0001] The present invention relates to the technical field of information processing, and in particular, to a multi-functional radar working mode learning method, device and electronic device. Background Art

[0002] With the rapid development of technologies such as electronic information, microelectronics and game theory, radar systems are gradually transforming from single-function to multi-functional and multi-tasking directions. Modern multi-functional radars can flexibly switch between multiple working modes according to task requirements and environmental conditions, and have the ability of self-adaptation, being able to adjust parameters in real time to optimize performance, making their signal characteristics elusive, which brings great challenges to radar working mode recognition.

[0003] Current multi-functional radar working mode recognition models mostly obtain all category training samples before training for offline training. In actual applications after training, the model can only recognize known categories existing in the training set. Therefore, when unknown category data that does not exist in the training set is input into the model, it cannot be correctly recognized as an unknown category, but is forcibly recognized as a certain known category belonging to the training set. Thus, it can be seen that the current multi-functional radar working mode recognition model lacks the ability to recognize and update new working modes and new samples, and cannot continuously learn complex and changeable unknown radar working modes, making it difficult to adapt to dynamic incremental observation scenarios, resulting in a significant weakening of the information control ability for complex and changeable reconnaissance environments. Summary of the Invention

[0004] In order to solve the above problems existing in the prior art, the present invention provides a multi-functional radar working mode learning method, device and electronic device. The technical problems to be solved by the present invention are realized through the following technical solutions:

[0005] According to the first aspect of the embodiments of the present invention, a multi-functional radar working mode learning method is provided, and the method includes:

[0006] Performing feature extraction on input radar pulse feature parameters through a preset time-series multi-scale residual attention network feature extractor;

[0007] Identifying the radar working mode according to the feature extraction result, and marking the radar working mode as a known category or an unknown category;

[0008] Based on the DBSCAN algorithm accelerated by the Kdtree data structure, clustering and marking the radar pulse feature parameters corresponding to the unknown category radar working mode to obtain radar pulse feature parameters with marked category labels;

[0009] Train a complementary learning model based on the data in the episodic memory buffer and the radar pulse feature parameters of the labeled category tags to obtain a feature extractor for the next-time multi-scale residual attention network; the feature extractor for the next-time multi-scale residual attention network is used to extract features from the radar pulse feature parameters input next time;

[0010] Use the reservoir sampling algorithm to update the radar pulse feature parameters of the labeled category tags to the episodic memory buffer, and the updated episodic memory buffer is used for the next complementary learning model training.

[0011] In an embodiment of the present invention, the DBSCAN algorithm accelerated based on the Kdtree data structure is used to cluster and label the radar pulse feature parameters corresponding to the unknown-class radar operating modes to obtain the radar pulse feature parameters of the labeled category tags, including:

[0012] Take the radar pulse feature parameters corresponding to the unknown-class radar operating modes as the unknown-class data set, and according to the median of the current dimension in the Kdtree data structure, divide the data in the unknown-class data set that is less than the median into the left subtree, and the data that is greater than or equal to the median into the right subtree;

[0013] Recursively divide the left subtree and the right subtree based on the median until each tree node in the Kdtree data structure has only one data;

[0014] Traverse the unknown-class data set, and for each data in it, use the Kdtree neighborhood search function to find all the neighborhood point sets N(x i ) within the specified neighborhood radius of the current data x; if the number of neighborhood points satisfies |N(x i )| ≥ MinPts, then mark x i as a core point and add it to the core point set C i ; the MinPts is the number of the least sample points in the density neighborhood; core Set the status of all data to "unvisited", and traverse each core point x

[0015] in the core point set C core , if x i is not visited, then create a cluster C i and add x k to this cluster, and the neighborhood point set N(x i ) of x i is used as the expansion queue of C i ; k

[0016] Sequentially obtain the data x j from the expansion queue, if x j ​Not visited. Use Kdtree for neighborhood search to obtain x j The set of neighborhood points N(x within the specified neighborhood radius j ); If N(x j ) ≥ MinPts, add these neighborhood points to the expansion queue and mark x j as a core point and add it to the cluster C k , until the expansion queue of the current cluster is empty;

[0017] According to different radar working modes, label the cluster C k to obtain the radar pulse feature parameters with marked class labels.

[0018] In an embodiment of the present invention, the complementary learning model includes a generation working model, a plasticity model, and a stability model;

[0019] Train the complementary learning model with the data in the scenario memory buffer and the radar pulse feature parameters with marked class labels to obtain the next - time - series multi - scale residual attention network feature extractor, including:

[0020] Input the data in the scenario memory buffer and the radar pulse feature parameters with marked class labels into the working model to obtain a feature vector representing the radar working mode

[0021] Input the data in the scenario memory buffer into the plasticity model and the stability model respectively to obtain feature vectors z p 、z s ;

[0022] Based on the feature vectors z p 、z s , train the working model, plasticity model, and stability model to obtain the next - time - series multi - scale residual attention network feature extractor.

[0023] In an embodiment of the present invention, the training of the working model, plasticity model, and stability model based on the feature vectors z p 、z s to obtain the next - time - series multi - scale residual attention network feature extractor includes:

[0024] Initialize the plasticity model weight θ p 、the stability model weight θ s 、the working model weight θ w ;

[0025] Based on the total loss function Update the plastic model weight θ p and stabilize the model weight θ s and the working model weight θ w ;

[0026] Based on the updated plastic model weight θ p and the stabilized model weight θ s and the working model weight θ w , obtain the next time-series multi-scale residual attention network feature extractor.

[0027] In one embodiment of the present invention, the updating of the plastic model weight θ , the stabilized model weight θ p and the working model weight θ s based on the total loss function w includes:

[0028] Based on a preset learning rate η and the total loss function update the working model weight θ w : where represents the gradient of the total loss function with respect to θ w ;

[0029] Based on the working model weight θ w , a preset update rate r p and r s , perform exponential moving average updates on the plastic model weight θ p and the stabilized model weight θ s :

[0030]

[0031] where α s , α p is a number between 0 and 1, used to control the update probability.

[0032] In one embodiment of the present invention, the calculation formula of the total loss function is:

[0033]

[0034] where represents the mixed loss, represents the consistency loss, label represents the true class label, represents the output vector obtained using the data in the scenario memory buffer, represents the output vector obtained using the radar pulse feature parameters of the marked class label, λ is a hyperparameter, zbest The best semantic memory selected by confidence: z best = argmax c {z p , z s}.

[0035] In one embodiment of the present invention, the pre-training of the temporal multi-scale residual attention network feature extractor includes:

[0036] Obtain training samples, where the training samples are radar pulse feature parameters of known categories;

[0037] Through the temporal multi-scale residual attention network feature extractor, perform feature extraction on the training samples to obtain a feature vector representing the radar working mode;

[0038] Based on the feature vector representing the radar working mode, optimize the parameters of the temporal multi-scale residual attention network feature extractor by minimizing the hybrid loss function and the backpropagation algorithm;

[0039] The hybrid loss function Loss hybrid The calculation formula is:

[0040] Loss hybrid = L center + αL triplet + βL DWCE

[0041] where α and β are weight coefficients used to control the proportion of each sub-loss function; L center is the center loss, which is used to enhance the within-class aggregation by minimizing the distance between the training samples and the class center; L triplet is the triplet loss, which is used to pull closer the same-class samples and push away the different-class samples; L DWCE represents the distance cross-entropy, which is used to convert the distance between the training samples and the class center into a probability distribution.

[0042] According to the second aspect of the embodiments of the present invention, a multi-functional radar working mode learning device is provided, and the device includes:

[0043] An open-set recognition network, which includes a temporal multi-scale residual attention network feature extractor and an adaptive threshold rejection module; the temporal multi-scale residual attention network feature extractor is used to perform feature extraction on the input radar pulse feature parameters; the adaptive threshold rejection module is used to identify the radar working mode according to the feature extraction result and mark the radar working mode as a known class or an unknown class;

[0044] An unsupervised clustering module, which is used to perform clustering and labeling on the radar pulse feature parameters corresponding to unknown-class radar working modes based on the DBSCAN algorithm accelerated by the Kdtree data structure, and obtain the radar pulse feature parameters with labeled category labels;

[0045] A class incremental learning module, which is used to train a complementary learning model based on the data in the scenario memory buffer and the radar pulse feature parameters with the labeled category labels, and obtain the next-time multi-scale residual attention network feature extractor; the next-time multi-scale residual attention network feature extractor is used to extract features from the radar pulse feature parameters input next time; this module is also used to update the radar pulse feature parameters with the labeled category labels to the scenario memory buffer by using the reservoir sampling algorithm, and the updated scenario memory buffer is used for the next complementary learning model training.

[0046] In an embodiment of the present invention, the multi-scale residual attention network feature extractor includes a channel-time step dual-track feature aggregation module and a multi-scale efficient attention module;

[0047] The channel-time step dual-track feature aggregation module is specifically used to capture the non-linear relationship between the time step and the channel;

[0048] The multi-scale efficient attention module is specifically used to capture the relationships between space and channel and capture the dependencies between channels from multiple scales through a parallel multi-scale convolution structure and a cross-space learning mechanism.

[0049] According to the third aspect of the embodiments of the present invention, an electronic device is provided, and the device includes:

[0050] One or more processors;

[0051] A computer-readable medium configured to store one or more programs;

[0052] When the one or more programs are executed by the one or more processors, the one or more processors implement the multi-functional radar working mode learning method as described in any one of the first aspects.

[0053] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0054] The multi-functional radar working mode learning method, device and electronic device provided by the embodiments of the present invention extract features of the input radar pulse feature parameters through a preset time-series multi-scale residual attention network feature extractor; identify the radar working mode according to the feature extraction result, and mark the radar working mode as a known class or an unknown class; based on the DBSCAN algorithm accelerated by the Kdtree data structure, cluster and mark the radar pulse feature parameters corresponding to the unknown-class radar working mode to obtain radar pulse feature parameters with marked class labels; train a complementary learning model based on the data in the scenario memory buffer and the radar pulse feature parameters with the marked class labels to obtain the next time-series multi-scale residual attention network feature extractor; the next time-series multi-scale residual attention network feature extractor is used to extract features of the radar pulse feature parameters input next time; the reservoir sampling algorithm is used to update the radar pulse feature parameters with the marked class labels to the scenario memory buffer, and the updated scenario memory buffer is used for the next complementary learning model training. The present invention extracts features of the input radar pulse feature parameters through a preset time-series multi-scale residual attention network feature extractor; it can extract radar features more comprehensively and accurately to achieve accurate identification and effective rejection of the radar working mode in an adversarial scenario with dynamic changes in the electromagnetic environment. Then, based on the DBSCAN algorithm accelerated by the Kdtree data structure, cluster and mark the radar pulse feature parameters corresponding to the unknown-class radar working mode, which can significantly improve the clustering efficiency. Finally, based on the data in the scenario memory buffer and the radar pulse feature parameters with the marked class labels, train a complementary learning model to obtain the next time-series multi-scale residual attention network feature extractor, and use the reservoir sampling algorithm to update the radar pulse feature parameters with the marked class labels to the scenario memory buffer; it can be seen that new radar pulse feature parameters are continuously added to the scenario memory buffer, and scenario memory replay is used to help the model remember past tasks, so that the complementary learning model can continuously learn new classes without forgetting old classes, and can not only maintain the memory ability of the known radar working mode, but also continuously learn the unknown radar working mode. This method can effectively cope with the complex and changeable electromagnetic environment and task requirements.

[0055] The following will further describe the present invention in detail with reference to the drawings and embodiments. Description of the Drawings

[0056] Figure 1 It is a step diagram of a multi-functional radar working mode learning method provided by an embodiment of the present invention;

[0057] Figure 2 It is another flow chart of the multi-functional radar working mode learning method provided by an embodiment of the present invention;

[0058] Figure 3 Neural network model diagram of the temporal multi-scale residual attention network provided by an embodiment of the present invention;

[0059] Figure 4 Model diagram of the radar working mode continuous learning method for the dynamic incremental observation scenario provided by an embodiment of the present invention;

[0060] Figure 5 Recognition accuracy and loss change curve of the temporal multi-scale residual attention network provided by an embodiment of the present invention;

[0061] Figure 6 Test result diagram provided by an embodiment of the present invention;

[0062] Figure 7 A structural block diagram of a multi-functional radar working mode learning device provided by an embodiment of the present invention;

[0063] Figure 8 Another structural block diagram of the multi-functional radar working mode learning device provided by an embodiment of the present invention;

[0064] Figure 9 A schematic structural diagram of an electronic device provided by an embodiment of the present invention. Specific embodiments

[0065] The present invention will be further described in detail below in conjunction with specific embodiments, but the embodiments of the present invention are not limited thereto.

[0066] Embodiment 1

[0067] Referring to Figure 1 , a step flowchart of a multi-functional radar working mode learning method provided by Embodiment 1 of the present invention is shown.

[0068] The multi-functional radar working mode learning method of this embodiment includes the following steps:

[0069] Step 101: Extract features from the input radar pulse feature parameters through a preset temporal multi-scale residual attention network feature extractor.

[0070] Among them, the input radar pulse feature parameters can be PDW (Pulse Descriptor Word). In radar signal processing, PDW (Pulse Descriptor Word) is a core parameter set for describing radar pulse features, and is used for subsequent signal sorting, radiation source identification, and working mode analysis.

[0071] Step 102: Identify the radar working mode according to the feature extraction result, and mark the radar working mode as a known class or an unknown class.

[0072] Step 103: Use the DBSCAN algorithm accelerated by the Kdtree data structure to cluster and label the radar pulse feature parameters corresponding to the unknown-class radar working modes, and obtain the radar pulse feature parameters with labeled category labels.

[0073] The above DBSCAN (Density-Based Spatial Clustering of Applications with Noise) is a density-based spatial clustering algorithm that can divide regions with sufficient density into clusters and identify points in low-density regions as noise points; while the KD tree (K-dimensional tree) is a data structure used to efficiently process high-dimensional spatial data and can be used to quickly find nearest neighbor points. By applying the KD tree to the DBSCAN algorithm, its search efficiency can be significantly improved.

[0074] Step 104: Based on the data in the scenario memory buffer and the radar pulse feature parameters with labeled category labels, train a complementary learning model to obtain the next-time multi-scale residual attention network feature extractor.

[0075] Among them, the next-time multi-scale residual attention network feature extractor is used to extract features from the radar pulse feature parameters input next time.

[0076] In this embodiment, the complementary learning model is a self-constructed model that can learn the data in the scenario memory buffer and the radar pulse feature parameters with new labeled category labels.

[0077] Step 105: Use the reservoir sampling algorithm to update the radar pulse feature parameters with the above-mentioned labeled category labels to the scenario memory buffer.

[0078] Among them, the updated scenario memory buffer is used for the next complementary learning model training.

[0079] The reservoir sampling algorithm is an algorithm for random sampling in a data stream, used to randomly select a specified number of samples with equal probability from a data stream containing a large amount of data.

[0080] The multi-functional radar working mode learning method provided by the embodiment of the present invention extracts features of the input radar pulse feature parameters through a preset time-series multi-scale residual attention network feature extractor; identifies the radar working mode according to the feature extraction result, and marks the radar working mode as a known class or an unknown class; based on the DBSCAN algorithm accelerated by the Kdtree data structure, clusters and marks the radar pulse feature parameters corresponding to the unknown-class radar working mode to obtain the radar pulse feature parameters with marked class labels; trains a complementary learning model based on the data in the scenario memory buffer and the radar pulse feature parameters with the marked class labels to obtain the next time-series multi-scale residual attention network feature extractor; the next time-series multi-scale residual attention network feature extractor is used to extract features of the radar pulse feature parameters input next time; the reservoir sampling algorithm is used to update the radar pulse feature parameters with the marked class labels to the scenario memory buffer, and the updated scenario memory buffer is used for the training of the complementary learning model next time. Through the preset time-series multi-scale residual attention network feature extractor, the present invention extracts features of the input radar pulse feature parameters; can extract radar features more comprehensively and accurately to achieve accurate identification and effective rejection of the radar working mode in the confrontation scenario with dynamic changes in the electromagnetic environment. Then, based on the DBSCAN algorithm accelerated by the Kdtree data structure, the radar pulse feature parameters corresponding to the unknown-class radar working mode are clustered and marked, which can significantly improve the clustering efficiency. Finally, based on the data in the scenario memory buffer and the radar pulse feature parameters with the marked class labels, a complementary learning model is trained to obtain the next time-series multi-scale residual attention network feature extractor, and the reservoir sampling algorithm is used to update the radar pulse feature parameters with the marked class labels to the scenario memory buffer; it can be seen that new radar pulse feature parameters are continuously added to the scenario memory buffer, and scenario memory replay is used to help the model remember past tasks, so that the complementary learning model can continuously learn new classes without forgetting old classes, can not only maintain the memory ability of the known radar working mode, but also continuously learn the unknown radar working mode, and this method can effectively cope with the complex and changeable electromagnetic environment and task requirements.

[0081] Embodiment 2

[0082] The following will describe the multi-functional radar working mode learning method implemented by the present invention in more detail. As Figure 2 shown, Figure 2 is another flowchart of the multi-functional radar working mode learning method provided by the embodiment of the present application, and may include the following steps:

[0083] Step 201: Extract features of the input radar pulse feature parameters through a preset time-series multi-scale residual attention network feature extractor.

[0084] In this embodiment, a neural network based on a temporal multi-scale residual attention network is constructed, denoted as network M. This network M can be composed of a channel-time step dual-track feature aggregation module and a multi-scale efficient attention module. As Figure 3 shown, depthwise separable convolution can expand the receptive field and enhance the spatial-channel feature interaction ability without increasing the computational complexity. Therefore, it is added after the multi-head attention; the channel-time step dual-track feature aggregation module can capture deeper and more complex non-linear relationships between time steps and channels; the multi-scale efficient attention module aims to enable network M to capture richer spatial and channel relationships from multiple scales while ensuring efficient computation through a parallel multi-scale convolution structure and a cross-space learning mechanism, helping the network effectively capture the dependencies between each channel and strengthening the correlation between each position in the spatial dimension.

[0085] Specifically, the pre-training of the above-mentioned temporal multi-scale residual attention network feature extractor includes:

[0086] Obtain training samples The training samples are radar pulse feature parameters of known categories; through the temporal multi-scale residual attention network feature extractor, the training samples are feature-extracted to obtain a feature vector representing the radar working mode; based on the feature vector representing the radar working mode, the parameters of the temporal multi-scale residual attention network feature extractor are optimized by minimizing the hybrid loss function and the backpropagation algorithm.

[0087] Hybrid loss function Loss hybrid The calculation formula is:

[0088] Loss hybrid = L center + αL triplet + βL DWCE

[0089] where α and β are weight coefficients used to control the proportion of each sub-loss function; L center is the center loss, which is used to enhance the within-class aggregation by minimizing the distance between the training samples and the class center; L triplet is the triplet loss, which is used to pull closer the same-class samples and push away the different-class samples; L DWCE represents the distance cross-entropy, which is used to convert the distance between the training samples and the class center into a probability distribution.

[0090] Furthermore, the calculation formula of L center is:

[0091]

[0092] where N is the total number of training samples, f(xi ) is the output feature representation of training sample i (i.e., the feature vector representing the radar working mode), is the class y i 's class center vector.

[0093] L triplet By comparing the distances of Anchor (reference sample), Positive (samples belonging to the same class as Anchor), and Negative (samples belonging to different classes from Anchor) samples, the distance between Anchor and Negative is made greater than the distance between Anchor and Positive, pulling closer samples of the same class and pushing away samples of different classes. The calculation formula is:

[0094]

[0095] Among them, is the feature representation of Anchor, Positive, and Negative samples extracted by network M; is the square of the Euclidean distance, used to measure the similarity between samples; m is a preset hyperparameter, representing the minimum margin between samples, that is, ensuring that the distance between Anchor and Negative is at least greater than the distance between Anchor and Positive; i represents the i-th triple, and N represents the total number of training samples, which is a natural number.

[0096] L DWCE represents distance-weighted cross-entropy, which converts the distance between the training sample and the class center into a probability distribution, implicitly making the output feature vector (i.e., the output feature representation) close to the center of the correct class and far from the centers of other classes. The calculation formula is:

[0097]

[0098] Among them, y jk is the sign function, taking 1 if the true class of training sample j is equal to k, otherwise taking 0, and J is the number of samples.

[0099]

[0100] d(g(x),c i ) represents the distance between g(x) and the center vector c i of class i, K is the number of classes of known classes, and γ is a hyperparameter that adjusts the influence degree of the distance.

[0101] Reference Figure 4 , Figure 4For a radar working mode continuous learning method model for dynamic incremental observation scenarios, in this embodiment, during the training phase T1 of network M, a training set composed of all known category radar pulse feature parameters in this phase is used Open-set learning can be performed to obtain the trained network M 1 , at this time M 1 Can identify known classes and reject unknown classes. Then the network M 1 And the subsequent updated scenario memory buffer B1 are used in the first clustering marking phase. Such as Figure 4 , in the clustering marking and incremental learning phase T t (2 ≤ t ≤ N). It can be understood that the samples (That is, the above input radar pulse feature parameters) usually include data of known classes and unknown classes.

[0102] Step 202: Identify the radar working mode according to the feature extraction result, and mark the radar working mode as a known class or an unknown class.

[0103] In this embodiment, an adaptive threshold rejection module can be constructed. Specifically, the feature extraction result (that is, the feature vector that can represent the radar working mode) is obtained through network M, and the Euclidean distance distance between the sample and the prototype data of each known class is calculated. During the preliminary prediction, the sample will be classified as the class to which the prototype data with the closest distance distance min Belongs, and then distance min Is compared with the rejection threshold of this class. If distance min Does not exceed the rejection threshold of the corresponding class, the preliminary prediction result is maintained. Otherwise, it is rejected as an unknown class. The calculation formula is as follows:[[]]

[0104]

[0105] Among them, K + 1 is the label of the unknown radar working mode, K is the number of known classes, y pred Is the above preliminary prediction result,[[]] Is the rejection threshold.

[0106] After obtaining the unknown class radar working mode, a sample set can be selected to update the scenario memory buffer B.

[0107] First, based on the initial training set Add noise to obtain pseudo-samples Will Input into network M 1 Perform pseudo-update to obtain network

[0108] Secondly, will Input to network M 1 and respectively obtain output feature vectors z and z * ; Calculate all training sets The Euclidean distance ||z - z * ||2 and cosine similarity cos(z, z * ) of the samples in

[0109] Then, calculate the comprehensive weight w, and the calculation formula is:

[0110] w = α·||z - z * ||2 + (1 - α)·cos(z, z * )

[0111] where α is a hyperparameter used to balance the influence of Euclidean distance and cosine similarity.

[0112] Finally, in each category of the sample training set Select the k samples with the largest comprehensive weight w to form an exemplar set, and merge The exemplar sets of each category form Use Initialize the scenario memory buffer B to obtain B1. These samples are highly representative and can avoid catastrophic forgetting to the greatest extent.

[0113] Step 203: Use the radar pulse feature parameters corresponding to the unknown-class radar working mode as the unknown-class data set. According to the median of the current dimension in the Kdtree data structure, divide the data in the unknown-class data set that is less than the median into the left subtree, and the data that is greater than or equal to the median into the right subtree.

[0114] In this embodiment, the DBSCAN algorithm accelerated based on the Kdtree data structure is introduced as the unsupervised clustering module. Use the radar pulse feature parameters corresponding to the unknown-class radar working mode that has been rejected as the data set Set the neighborhood radius ε (defining the range of the density neighborhood) and the minimum number of samples MinPts (defining the minimum number of sample points in the density neighborhood). Then construct the Kdtree data structure to accelerate the nearest neighbor search: sequentially and cyclically select each dimension feature as the current division dimension; divide the data set into left and right parts according to the median of the current dimension, with the smaller part entering the left subtree and the larger part entering the right subtree.

[0115] Step 204: Recursively divide the left subtree and the right subtree based on the median until each tree node in the Kdtree data structure has only one data.

[0116] Repeat this step for the left and right subtrees by changing the dimension until all nodes are added to the Kdtree.

[0117] Step 205, traverse the unknown class dataset, and for each data in it, use the Kdtree neighborhood search function to find the current data x i All neighborhood point sets N(x i ) within the specified neighborhood radius; if the number of neighborhood points satisfies |N(x i )| ≥ MinPts, then mark x i as a core point and add it to the core point set C core .

[0118] Traverse the dataset For each data x i Use the neighborhood search function of the Kdtree to find all neighbor point sets N(x i ) within the radius ε; if the number of neighbor points satisfies |N(x i )| ≥ MinPts, then mark x i as a core point and add it to the core point set C i . core .

[0119] Step 206, set the status of all data to "unvisited", and traverse each core point x core in the core point set C i . If x i is not visited, then create a cluster C k and add x i to this cluster. The neighborhood point set N(x i ) of x i is used as the expansion queue of C k .

[0120] Initialize the clustering label, set the status of all data to "unvisited"; traverse each core point x core in the core point set C i . If x i is not visited, then create a new cluster C k and add x i to this cluster; the neighborhood point set N(x i ) of x i is used as the "expansion queue".

[0121] Step 207, sequentially obtain the data x j from the expansion queue. If x j is not visited, use the Kdtree to perform neighborhood search to obtain the neighborhood point set N(x j ) within the specified neighborhood radius; if N(x j )j ) ≥ MinPts, then add these neighborhood points to the expansion queue, and mark x j as a core point and add it to the cluster C k , until the expansion queue of the current cluster is empty.

[0122] Take out x from the "expansion queue" in sequence j , if x j has not been visited, use Kdtree for neighborhood search to obtain the set of neighbor points N(x i ) within the radius ε; if |N(x j )| ≥ MinPts, then add these neighbor points to the "expansion queue", and mark x j as a core point; repeat the above steps until the "expansion queue" of the current cluster is empty. j

[0123] Step 208, according to different radar working modes, label the cluster C k to obtain the radar pulse feature parameters with labeled category labels.

[0124] Traverse all the data in the dataset . If a certain data is not marked by any cluster and is not a core point, then mark it as a noise point. Use the obtained radar pulse feature parameters with labeled category labels as new label samples

[0125] Step 209, based on the data in the scenario memory buffer and the radar pulse feature parameters with the labeled category labels, train a complementary learning model to obtain the next temporal multi-scale residual attention network feature extractor.

[0126] Specifically, referring to Figure 4 , the complementary learning model includes a generative working model a plasticity model and a stability model

[0127] When training the complementary learning model based on the data in the scenario memory buffer and the radar pulse feature parameters with the labeled category labels to obtain the next temporal multi-scale residual attention network feature extractor, the data in the scenario memory buffer and the above-mentioned radar pulse feature parameters with the labeled category labels can be input into the working model to obtain a feature vector representing the radar working mode Input the data in the scenario memory buffer into the plasticity model and the stability model respectively to obtain feature vectors z p 、z s ; then based on the feature vectors z p 、z s ​, train the working model, the plasticity model, and the stability model to obtain the next-time multi-scale residual attention network feature extractor.

[0128] Further, based on the feature vectors z p and z s , when training the working model, the plasticity model, and the stability model to obtain the next-time multi-scale residual attention network feature extractor, the weights θ p of the plasticity model, the weights θ s of the stability model, and the weights θ w of the working model can be initialized first; then, based on the total loss function , update the weights θ p of the plasticity model, the weights θ s of the stability model, and the weights θ w of the working model; finally, based on the updated weights θ p of the plasticity model, the weights θ s of the stability model, and the weights θ w of the working model, obtain the next-time multi-scale residual attention network feature extractor.

[0129] The above total loss function is calculated as follows:

[0130]

[0131] Wherein, represents the hybrid loss, represents the consistency loss, label represents the true class label, represents the output vector obtained by using the data in the scenario memory buffer, represents the output vector obtained by using the radar pulse feature parameters of the marked class label, λ is a hyperparameter, z best is the best semantic memory selected by confidence: z best = argmax c {z p , z s}.

[0132] In this embodiment, the best semantic memory z best = argmax c {z p , z s} can be selected by confidence to calculate the hybrid loss and the consistency loss The calculation formula of the consistency loss is:

[0133]

[0134] Furthermore, when updating the plastic model weight θ , stabilizing the model weight θ p , and the working model weight θ s , based on the preset learning rate η and the total loss function w , the working model weight θ can be updated as follows: w : where represents the gradient of the total loss function with respect to θ w . Then, based on the working model weight θ w , the preset update rate r p and r s , the plastic model weight θ p and the stable model weight θ s are updated by exponential moving average:

[0135]

[0136] where α s , α p is a number between 0 and 1, used to control the update probability. Then, a new working model plastic model stable model is obtained. The previous feature extraction network M is replaced with the new working model t-1 to obtain M t , that is, a new time-series multi-scale residual attention network feature extractor is obtained, which is used to extract the feature parameters of the radar pulse features input next time.

[0137] Then, the reservoir sampling algorithm is used to add the radar pulse feature parameters with labeled class labels to the scenario memory buffer, and the updated scenario memory buffer is used for the next complementary learning model training.

[0138] The effects of the present invention can be further illustrated by the following experiments. Taking seven radar operating modes, namely Range While Search (RWS), Tracking While Search (TWS), Track and Search (TAS), Anti-Weather (AW), Burn Through (BT), Single Target Track (STT), and Guidance as examples.

[0139] First, let RWS, TWS, and TAS be known categories, and the anti-weather mode, burn-through mode, STT, and guidance mode be unknown categories. The unknown categories do not participate in the training process, and all samples are involved in the testing phase.

[0140] The accuracy and loss change curves of the open-set recognition of radar operating modes based on the temporal multi-scale residual attention network are as Figure 5 shown. Figure 5 They are the accuracy and loss change curves of the temporal multi-scale residual attention network. It can be seen that when the number of training rounds is greater than 15, the loss tends to be stable, and the final recognition accuracy is as high as 99.1%.

[0141] To further explore the incremental learning ability of the model, new categories are gradually introduced, and the changing rules of the rejection ability (AUS) of the network / model for new categories and the overall performance (NA, F1) while maintaining the classification ability (AKS) of known categories are analyzed. In terms of the dataset composition, the single-object tracking mode is the new category for the first increment; the anti-weather mode is the new category for the second increment; the burn-through mode and the guidance mode are the new categories for the third increment, and the buffer size is set to 1024. For each round of incremental learning, the test set contains samples of all categories, and the categories that have not been learned are regarded as the current unknown categories. In the third round of incremental learning, all categories have been learned, so there are no unknown categories. In this case, AUS and NA cannot be calculated and are recorded as " / " in the experimental results. The results are shown in Table 1 and Figure 6 (a~h). It can be seen that with the increase of new categories, various indicators decline, but the recognition of known and unknown categories can still be well performed. In the third increment, the confusion between the TAS mode and the STT mode is the most serious. This is because the TAS mode needs to switch between the search task and the tracking task according to the radar scheduling, and as an old category, the sample size is small, and it is greatly affected by catastrophic forgetting. Therefore, it is easy to be confused with the STT mode with relatively high similarity between categories.

[0142] Table 1 Recognition accuracy results

[0143] Initial Network After the First Round of Increment After the Second Round of Increment After the Third Round of Increment AKS 97.45％ 96.40％ 95.25％ 94.06％ AUS 96.78％ 95.34％ 94.41％ / NA 97.07％ 95.95％ 95.01％ / F1 Score 0.958 0.944 0.937 0.924

[0144] The multi-functional radar working mode learning method provided by the embodiments of the present invention is a radar working mode continuous learning method capable of facing dynamic incremental observation scenarios. The present invention extracts features of the input radar pulse feature parameters through the constructed time-series multi-scale residual attention network feature extractor, which can extract radar features more comprehensively and accurately, so as to accurately identify and effectively reject the radar working mode in the confrontation scenario of dynamic changes in the electromagnetic environment. The clustering module designs a DBSCAN clustering algorithm accelerated based on the Kdtree data structure to cluster and label the radar pulse feature parameters corresponding to the unknown-class radar working modes, significantly improving the clustering efficiency. Finally, based on the data in the scenario memory buffer and the radar pulse feature parameters with the labeled category tags, a complementary learning model is trained to obtain the next time-series multi-scale residual attention network feature extractor. Moreover, the reservoir sampling algorithm is used to update the radar pulse feature parameters with the labeled category tags to the scenario memory buffer. It can be seen that new radar pulse feature parameters are constantly added to the scenario memory buffer. Based on the complementary learning system theory, scenario memory replay is used to help the model remember past tasks, so that the complementary learning model can continuously learn new categories without forgetting old categories, and can not only maintain the memory ability of known radar working modes but also continuously learn unknown radar working modes. That is, the incremental learning module balances the rapid learning of new knowledge and the retention of old knowledge by integrating short-term and long-term semantic memories, enabling the network to structure the learned knowledge during the continuous learning process for sharing and reuse between different tasks and avoiding knowledge forgetting. Finally, experiments prove that this method can not only maintain the memory ability of known working modes but also continuously learn unknown working modes in the incremental recognition of multi-functional radar working modes. This method can effectively cope with the changing electromagnetic environment and task requirements, and has important theoretical and engineering significance for improving our reconnaissance effectiveness.

[0145] Embodiment III

[0146] The embodiments of the present invention also provide a multi-functional radar working mode learning device, as Figure 7 shown. The device 70 includes:

[0147] An open-set recognition network 700, which includes a time-series multi-scale residual attention network feature extractor 7001 and an adaptive threshold rejection module 7002. The time-series multi-scale residual attention network feature extractor 7001 is used to extract features of the input radar pulse feature parameters. The adaptive threshold rejection module 7002 is used to identify the radar working mode according to the feature extraction result and label the radar working mode as a known class or an unknown class.

[0148] The unsupervised clustering module 702 is used to cluster and label the radar pulse feature parameters corresponding to the unknown-class radar working modes based on the DBSCAN algorithm accelerated by the Kdtree data structure, and obtain the radar pulse feature parameters with labeled class tags.

[0149] The class incremental learning module 704 is used to train a complementary learning model based on the data in the scenario memory buffer and the radar pulse feature parameters with labeled class tags, and obtain the next-time multi-scale residual attention network feature extractor; the next-time multi-scale residual attention network feature extractor is used to extract features from the radar pulse feature parameters input next time; the class incremental learning module 704 is also used to update the radar pulse feature parameters with labeled class tags to the scenario memory buffer by using the reservoir sampling algorithm, and the updated scenario memory buffer is used for the next complementary learning model training.

[0150] Specifically, as Figure 8 shown, the multi-scale residual attention network feature extractor 7001 includes a channel-time step dual-track feature aggregation module 70011 and a multi-scale efficient attention module 70012; the channel-time step dual-track feature aggregation module 70011 is specifically used to capture the non-linear relationship between the time step and the channel; the multi-scale efficient attention module 70012 is specifically used to capture the relationships between space and channels and capture the dependencies between channels through a parallel multi-scale convolution structure and a cross-space learning mechanism.

[0151] Referring to Figure 3 , the specific structure of the multi-scale residual attention network feature extractor 7001 is: embedding layer → (positional encoding: to capture time dependence) → multi-head attention / masked multi-head attention → residual connection & layer normalization → feed-forward network (to enhance the model's feature extraction ability) → residual connection & layer normalization → depthwise separable convolution → gating module (to effectively fuse the correlation features extracted by the time step encoder and the channel encoder) → feature grouping (to strengthen the representation of important regions inside local features) → multi-scale feature extraction → cross-space learning → fully connected layer.

[0152] The multi-functional radar working mode learning device provided by the embodiment of the present invention includes an open-set recognition network 700, an unsupervised clustering module 702, and a class incremental learning module 704. The open-set recognition network 700 includes a feature extractor 7001 based on a temporal multi-scale residual attention network and an adaptive threshold rejection module 7002. The feature extractor 7001 based on the temporal multi-scale residual attention network further includes a channel-time step dual-track feature aggregation module 70011 and a multi-scale efficient attention module 70012. The depthwise separable convolution inside it can expand the receptive field and enhance the spatial-channel feature interaction ability without increasing the computational complexity, so it is added after the multi-head attention; the channel-time step dual-track feature aggregation module can capture deeper and more complex non-linear relationships between time steps and channels; the multi-scale efficient attention module aims to enable the network to capture richer spatial and channel relationships from multiple scales while ensuring efficient computation through a parallel multi-scale convolution structure and a cross-space learning mechanism, helping the network effectively capture the dependencies between each channel and strengthening the correlation between positions in the spatial dimension. The unsupervised clustering module 702 designs a DBSCAN clustering algorithm accelerated based on the Kdtree data structure to cluster and label the unknown classes after rejection, significantly improving the clustering efficiency; the class incremental learning module 704 designs a multi-functional radar working mode class incremental recognition method based on complementary experience replay, which can continuously update the existing model with new working modes, so that the new model can continuously learn new working modes without forgetting the old classes. Aiming at the challenges that training data is difficult to cover all sample classes and unknown classes continuously appear during the testing process, this method can accurately identify and effectively reject the radar working mode in the adversarial scenario of dynamic change of the electromagnetic environment, and continuously learn new classes without forgetting the old classes, which has important theoretical and technical support in the field of pattern recognition and has strong practical significance for improving our reconnaissance and confrontation capabilities.

[0153] Embodiment 4

[0154] The embodiment of the present invention also provides an electronic device, as Figure 9 shown, including a processor 301, a communication interface 302, a memory 303, and a communication bus 304. Among them, the processor 301, the communication interface 302, and the memory 303 complete communication with each other through the communication bus 304.

[0155] The memory 303 is used to store a computer program;

[0156] When the processor 301 is used to execute the program 305 stored on the memory 303, the following steps are implemented:

[0157] Extract the features of the input radar pulse feature parameters through a preset time-series multi-scale residual attention network feature extractor;

[0158] Identify the radar operating mode according to the feature extraction results, and label the radar operating mode as a known class or an unknown class;

[0159] Based on the DBSCAN algorithm accelerated by the Kdtree data structure, cluster and label the radar pulse feature parameters corresponding to the unknown-class radar operating mode to obtain the radar pulse feature parameters with labeled class labels;

[0160] Train a complementary learning model based on the data in the scenario memory buffer and the radar pulse feature parameters with the labeled class labels to obtain the next time-series multi-scale residual attention network feature extractor; the next time-series multi-scale residual attention network feature extractor is used to extract the features of the radar pulse feature parameters input next time;

[0161] Use the reservoir sampling algorithm to update the radar pulse feature parameters with the labeled class labels to the scenario memory buffer, and the updated scenario memory buffer is used for the next complementary learning model training.

[0162] The communication bus mentioned in the above electronic device can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into an address bus, a data bus, a control bus, etc. For the sake of simplicity, only a thick line is used in the figure to represent it, but it does not mean that there is only one bus or one type of bus.

[0163] The communication interface is used for communication between the above electronic device and other devices.

[0164] The memory can include a Random Access Memory (RAM), and can also include a Non-Volatile Memory (NVM), such as at least one disk memory. Optionally, the memory can also be at least one storage device located far from the aforementioned processor.

[0165] The above-mentioned processor may be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it may also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

[0166] The method provided by the embodiments of the present invention can be applied to an electronic device. Specifically, the electronic device may be: a desktop computer, a portable computer, a smart mobile terminal, a server, etc. There is no limitation here, and any electronic device that can implement the present invention belongs to the protection scope of the present invention.

[0167] For the embodiments of the electronic device / storage medium, since they are basically similar to the method embodiments, the description is relatively simple, and for the relevant parts, please refer to the partial description of the method embodiments.

[0168] It should be noted that the electronic device and storage medium of the embodiments of the present invention are respectively the electronic device and storage medium applying the above-mentioned multi-functional radar working mode learning method. Then, all embodiments of the above-mentioned multi-functional radar working mode learning method are applicable to the electronic device and storage medium, and can achieve the same or similar beneficial effects.

[0169] By applying the terminal device provided by the embodiments of the present invention, proper nouns and / or fixed phrases can be displayed for the user to select, thereby reducing the user input time and improving the user experience.

[0170] The terminal device exists in various forms, including but not limited to:

[0171] (1) Mobile communication device: The characteristic of this type of device is that it has mobile communication functions and mainly aims to provide voice and data communication. This type of terminal includes: smart phones (such as iPhone), multimedia phones, functional phones, and low-end phones, etc.

[0172] (2) Ultra-mobile personal computer device: This type of device belongs to the category of personal computers, has computing and processing functions, and generally also has the characteristic of mobile Internet access. This type of terminal includes: PDA, MID, and UMPC devices, etc., such as iPad.

[0173] (3) Portable entertainment devices: Such devices can display and play multimedia content. This type of device includes: audio and video players (such as iPods), handheld game consoles, e-books, as well as smart toys and portable in-vehicle navigation devices.

[0174] (4) Other electronic devices with data interaction functions.

[0175] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of such features. In the description of the present invention, "a plurality" means two or more unless otherwise specifically defined.

[0176] In the description of this specification, descriptions with reference to terms such as "an embodiment", "some embodiments", "example", "specific example", or "some examples", etc. mean that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples. In addition, those skilled in the art can combine and combine the different embodiments or examples described in this specification.

[0177] Although the present application has been described in conjunction with various embodiments herein, however, in the process of implementing the claimed present application, those skilled in the art can understand and achieve other variations of the disclosed embodiments by viewing the accompanying drawings, the disclosure content, and the appended claims. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality. A single processor or other unit can implement several functions recited in the claims. Certain measures are recited in mutually different dependent claims, but this does not mean that these measures cannot be combined to produce good results.

[0178] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, an apparatus (device), or a computer program product. Therefore, the present application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects, which are collectively referred to as "modules" or "systems" herein. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The computer program is stored / distributed in a suitable medium, provided together with other hardware or as part of the hardware, and can also be distributed in other forms, such as via the Internet or other wired or wireless telecommunication systems.

[0179] The present application is described with reference to the flowcharts and / or block diagrams of the methods, apparatuses (devices), and computer program products of the embodiments of the present application. It should be understood that each flow and / or block in the flowchart and / or block diagram, as well as the combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices generate a device for implementing the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0180] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory generate a manufactured article including an instruction device that implements the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0181] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0182] The above content is a further detailed description of the present invention in combination with specific preferred embodiments. It cannot be determined that the specific implementation of the present invention is only limited to these descriptions. For those of ordinary skill in the technical field to which the present invention pertains, without departing from the concept of the present invention, several simple deductions or substitutions can still be made, and all should be regarded as falling within the protection scope of the present invention.

Claims

1. A learning method for a multi-functional radar operating mode, characterized in that, The method includes: extracting features of the input radar pulse feature parameters through a preset time-series multi-scale residual attention network feature extractor; identifying the radar working mode according to the feature extraction result, and marking the radar working mode as a known class or an unknown class; clustering and marking the radar pulse feature parameters corresponding to the unknown-class radar working mode based on the DBSCAN algorithm accelerated by the Kdtree data structure to obtain the radar pulse feature parameters with marked class labels; training a complementary learning model based on the data in the scenario memory buffer and the radar pulse feature parameters with the marked class labels to obtain the next time-series multi-scale residual attention network feature extractor; the next time-series multi-scale residual attention network feature extractor is used to extract features of the radar pulse feature parameters input next time; updating the radar pulse feature parameters with the marked class labels to the scenario memory buffer by using the reservoir sampling algorithm, and the updated scenario memory buffer is used for the next complementary learning model training.

2. The method according to claim 1, characterized in that, The DBSCAN algorithm accelerated by the Kdtree data structure clustering and marking the radar pulse feature parameters corresponding to the unknown-class radar working mode to obtain the radar pulse feature parameters with marked class labels includes: taking the radar pulse feature parameters corresponding to the unknown-class radar working mode as an unknown-class data set, and dividing the data in the unknown-class data set that is less than the median in the current dimension of the Kdtree data structure into the left subtree, and dividing the data that is greater than or equal to the median into the right subtree; recursively dividing the left subtree and the right subtree based on the median until each tree node in the Kdtree data structure has only one data; Traverse the unknown class dataset, and for each data in it, use the Kdtree neighborhood search function to find the current data x i The set of all neighborhood points N(x within the specified neighborhood radius i ); If the number of neighborhood points satisfies |N(x i )|≥MinPts, then mark x i as a core point and add it to the core point set C core ; The MinPts is the number of the fewest sample points in the density neighborhood; Set the status of all data to "unvisited" and traverse each core point x in the core point set C core For each core point x i in it, if x i has not been visited, create a cluster C k and add x i to this cluster. Take the neighborhood point set N(x i ) of x as the expansion queue of C i . k ​ Obtain data x from the expansion queue in sequence j , if x j has not been visited, use Kdtree for neighborhood search to obtain the set of neighborhood points N(x j ) within the specified neighborhood radius; if N(x j ) ≥ MinPts, add these neighborhood points to the expansion queue and mark x j as a core point and add it to the cluster C j until the expansion queue of the current cluster is empty; k ​ According to different radar operating modes, the cluster C k is classified to obtain the radar pulse feature parameters marked with class labels.

3. The method according to claim 1, characterized in that, The complementary learning model includes a generation working model, a plasticity model, and a stability model; training the complementary learning model based on the data in the scenario memory buffer and the radar pulse feature parameters with the marked class labels to obtain the next time-series multi-scale residual attention network feature extractor includes: Input the data in the episodic memory buffer and the radar pulse feature parameters of the labeled category tags into the working model to obtain a feature vector representing the radar working mode. Input the data in the episodic memory buffer into the plasticity model and the stability model respectively to obtain the feature vectors z p and z s ; Based on the feature vector z p 、z s , train the working model, the plasticity model, and the stability model to obtain the next-time multi-scale residual attention network feature extractor.

4. The method according to claim 3, wherein Based on the feature vector z p and z s , train the working model, plastic model, and stability model to obtain the next-time-series multi-scale residual attention network feature extractor, including: Initialize the plastic model weights θ p 、Stabilize the model weights θ s 、Work with the model weights θ w ; Based on the total loss function Update the plastic model weight θ p , stabilize the model weight θ s , and the working model weight θ w ; Based on the updated plastic model weights θ p , the stable model weights θ s , and the working model weights θ w , the next time-series multi-scale residual attention network feature extractor is obtained.

5. The method according to claim 4, characterized in that, The total loss function-based updates the plasticity model weight θ p , stabilizes the model weight θ s , and the working model weight θ w , including: Based on a preset learning rate η and the total loss function Update the weights θ of the working model w : where represents the total loss function with respect to θ w gradient; Based on the working model weights θ w , the preset update rate r p and r s , perform exponential moving average updates on the plasticity model weights θ p and the stability model weights θ s : Among them, α s , α p is a number between 0 and 1, which is used to control the update probability.

6. The method according to claim 4, characterized in that The total loss function has the following calculation formula: Among them, represents the hybrid loss, represents the consistency loss, label represents the true class label, represents the output vector obtained by using the data in the scenario memory buffer, represents the output vector obtained by using the radar pulse feature parameters of the marked class label, λ is a hyperparameter, z best is the best semantic memory selected by confidence: z best = argmax c {z p , z s}.

7. The method according to claim 1, characterized in that, The pre-training of the time-series multi-scale residual attention network feature extractor includes: obtaining training samples, where the training samples are radar pulse feature parameters of known classes; extracting features of the training samples through the time-series multi-scale residual attention network feature extractor to obtain a feature vector representing the radar working mode; optimizing the parameters of the time-series multi-scale residual attention network feature extractor based on the feature vector representing the radar working mode by minimizing the mixed loss function and the backpropagation algorithm; The mixed loss function Loss hybrid The calculation formula is as follows: Loss hybrid = L center + αL triplet + βL DWCE Among them, α and β are weight coefficients used to control the proportion of each sub-loss function; L center is the center loss, which is used to enhance intra-class aggregation by minimizing the distance between training samples and class centers; L triplet is the triplet loss, which is used to pull closer samples of the same class and push away samples of different classes; L DWCE represents the distance cross-entropy, which is used to convert the distance between training samples and class centers into a probability distribution.

8. A multifunctional radar working mode learning device, characterized in that, The device includes: an open-set recognition network, where the open-set recognition network includes a time-series multi-scale residual attention network feature extractor and an adaptive threshold rejection module; the time-series multi-scale residual attention network feature extractor is used to extract features of the input radar pulse feature parameters; the adaptive threshold rejection module is used to identify the radar working mode according to the feature extraction result and mark the radar working mode as a known class or an unknown class; An unsupervised clustering module, which is used to cluster and label the radar pulse feature parameters corresponding to the unknown-class radar working modes based on the DBSCAN algorithm accelerated by the Kdtree data structure, and obtain the radar pulse feature parameters with labeled class tags; A class incremental learning module, which is used to train a complementary learning model based on the data in the scenario memory buffer and the radar pulse feature parameters with the labeled class tags, and obtain the next-time multi-scale residual attention network feature extractor; the next-time multi-scale residual attention network feature extractor is used to extract features from the radar pulse feature parameters input next time; this module is also used to update the radar pulse feature parameters with the labeled class tags to the scenario memory buffer by using the reservoir sampling algorithm, and the updated scenario memory buffer is used for the next complementary learning model training.

9. The device according to claim 8, wherein The multi-scale residual attention network feature extractor includes a channel-time step dual-track feature aggregation module and a multi-scale efficient attention module; The channel-time step dual-track feature aggregation module is specifically used to capture the non-linear relationship between the time step and the channel; The multi-scale efficient attention module is specifically used to capture the relationships between space and channel and capture the dependencies between channels through a parallel multi-scale convolution structure and a cross-space learning mechanism.

10. An electronic device, characterized in that, The device includes: One or more processors; A computer-readable medium configured to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the multi-functional radar working mode learning method according to any one of claims 1-7.

Images