Radar working mode recognition system and semi-supervised training method

Abstract

The invention relates to a radar working mode recognition system and a semi-supervised training method. The system comprises a supervising module and a self-supervising module which are updated alternately. During training, the supervision module adopts annotation data to drive feature extraction and classifier optimization, and captures high-order feature differences by fusing polynomial loss, so that the model discrimination capability is improved; the self-supervision module innovatively introduces a dynamic time slice division mechanism, adjusts a time slice division proportion in real time according to a verification set index, and forms an elastic training strategy. The two modules share the weight of a feature extraction network, and bidirectional gradient transmission is realized in iteration: feature parameters updated by supervision training are synchronized to the self-supervision module, and the self-supervision module optimizes a generated pseudo-label relation loss feedback feature space by constructing positive / negative samples. According to the collaborative mechanism, labeled data and unlabeled data are aligned in a unified feature space, and the error accumulation problem of a traditional training method is effectively relieved. According to the method, the radar working mode recognition capability in a complex electromagnetic environment can be effectively improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar data analysis, and particularly relates to a radar working mode recognition system and a semi-supervised training method. Background Art

[0002] Currently, the working mode recognition methods for multifunctional radars can be mainly divided into three categories: statistical feature-based, behavior reasoning-based, and feature extraction-based. The statistical feature-based working mode recognition algorithm determines the mode type by setting the parameter range of each working mode and matching these parameters, but the number of recognizable working modes is small, and a threshold still needs to be set manually. The behavior reasoning-based working mode recognition algorithm mainly models the dynamic laws of the system state and signals, and infers the dynamic changes of the working mode using state transition and observation probability, but the finite state is difficult to effectively depict the numerous state numbers and complex conversion laws of the multifunctional radar system. The feature extraction-based pattern recognition algorithm extracts the key features of PDW through a feature extraction model and then uses the differences in features under different working modes to recognize the working mode, but it is too dependent on training data, and a large number of labeled samples are difficult to obtain in the actual environment.

[0003] The semi-supervised learning recognition algorithm provides a new idea to solve this problem. As the name implies, semi-supervised learning is a learning method that lies between supervised learning and unsupervised learning. In supervised learning, the model is trained with input and corresponding labeled data to predict the labels of unseen samples; while unsupervised learning only relies on unlabeled data to find the potential structure or pattern of the data. According to the design of the algorithm, semi-supervised learning can be roughly divided into three main types: consistency regularization, graph-based semi-supervised learning, and pseudo-label method. Consistency regularization is a method based on the assumption of data invariance of the model. This method assumes that when the input data undergoes a small perturbation, the output of the model should remain consistent. By adding noise or perturbation to the input data, the model is forced to learn a robust and consistent representation, thereby improving its performance on unlabeled data. The graph-based semi-supervised learning method regards data points as nodes in a graph, and the edges between nodes represent the similarity between data points. This type of method infers the labels of unlabeled data by embedding labeled data and unlabeled data into a graph and using information transfer on the graph structure. The pseudo-label method is a common technique in semi-supervised learning. The core idea is to first train a preliminary model with a small amount of labeled data, then use this model to predict unlabeled data, and the prediction results are used as pseudo-labels, and the model is trained together with the original labeled data using these pseudo-labels.

[0004] However, when the existing semi-supervised training methods for target recognition models process data carrying time series, there are problems such as a cumbersome and complex training process and serious error accumulation, which seriously affect the recognition accuracy in later use. Summary of the Invention

[0005] To solve the above problems existing in the prior art, the present invention provides a radar working mode recognition system and a semi-supervised training method. The technical problems to be solved by the present invention are realized through the following technical solutions:

[0006] In a first aspect, the present invention provides a semi-supervised training method for a radar working mode recognition system, the radar working mode recognition system including: a supervision module and a self-supervision module, the supervision module including a first feature extraction unit and a classifier unit, the self-supervision module including a second feature extraction unit and a predictor unit, and the model parameters in the first feature extraction unit and the second feature extraction unit being the same; the training method including:

[0007] S1: In the i-th iterative training, input a plurality of labeled radar data carrying time series into the (i - 1)-th supervision module to obtain the i-th first loss value, and update the (i - 1)-th supervision module using the i-th first loss value to obtain the i-th supervision module and the corresponding i-th feature extraction parameters;

[0008] S2: Input a plurality of unlabeled radar data carrying time series into the (i - 1)-th self-supervision module, and divide the plurality of unlabeled radar data into a first prediction segment and a second prediction segment using the i-th time segment division ratio value, determine the i-th second loss value using the first prediction segment, the second prediction segment, and the i-th feature extraction parameters, and update the (i - 1)-th self-supervision module using the i-th second loss value to obtain the i-th self-supervision module;

[0009] S3: Judge whether i is equal to I; if so, execute S4, if not, execute S5;

[0010] S4: Stop the iterative training to obtain the trained radar working mode recognition system;

[0011] S5: Replace i in S1 to S3 with i + 1 and continue the next iterative training.

[0012] Preferably, S1 includes: using the (i-1)-th first feature extraction unit to extract the first embedded feature vectors of each labeled radar data, obtaining a plurality of first embedded feature vectors; using the (i-1)-th classifier unit to predict the class distributions of the plurality of first embedded feature vectors, obtaining a plurality of class distribution results; using the fusion polynomial loss function and the triplet loss function to calculate the plurality of class distribution results, respectively obtaining the i-th fusion polynomial loss value and the i-th triplet loss value; summing the i-th fusion polynomial loss value and the i-th triplet loss value to obtain the i-th first loss value; using the i-th first loss value to update the (i-1)-th first feature extraction unit and the (i-1)-th classifier unit, obtaining the i-th supervised module and the i-th feature extraction parameters.

[0013] Preferably, S2 includes: using the i-th time segment division ratio value to divide the plurality of unlabeled radar data in the dataset into a first prediction segment and a second prediction segment; using the first prediction segment and the second prediction segment to construct a plurality of positive sample pairs and a plurality of negative sample pairs; using the i-th feature extraction parameters to update the (i-1)-th second feature extraction unit to obtain the i-th second feature extraction unit, and respectively extracting features from the plurality of positive sample pairs and the plurality of negative sample pairs to obtain a plurality of positive embedded feature vectors and a plurality of negative embedded feature vectors; using the (i-1)-th predictor unit to respectively predict the plurality of positive embedded feature vectors and the plurality of negative embedded feature vectors to obtain a plurality of positive prediction results and a plurality of negative prediction results; using the binary cross-entropy loss function to calculate the plurality of positive prediction results and the plurality of negative prediction results to obtain the i-th second loss value; using the i-th second loss value to update the i-th second feature extraction unit and the (i-1)-th predictor unit to obtain the i-th self-supervised module.

[0014] Preferably, the i-th time segment division ratio value is calculated using the (i-1)-th time segment division ratio value, the evaluation index of the (i-1)-th time, and the optimal historical index in the previous (i-1) times. Wherein, when i takes the value of 1, the value of the first time segment division ratio value is a preset value.

[0015] Preferably, before executing S3, the semi-supervised training method further includes: inputting the validation set into the i-th self-supervised module to obtain the evaluation index of the i-th time.

[0016] Preferably, the semi-supervised training method further includes: before the start of the i-th iterative training, summarizing the evaluation indexes of the previous (i-1) times, and screening out the maximum value from the evaluation indexes of the previous (i-1) times as the optimal historical index in the previous (i-1) times.

[0017] Optimally, constructing a plurality of positive sample pairs and a plurality of negative sample pairs by using the first prediction segment and the second prediction segment includes: constructing a plurality of positive sample pairs by using a plurality of unlabeled radar data in the first prediction segment and a plurality of unlabeled radar data in the second prediction segment, each sample pair including two unlabeled radar data in each prediction segment, and each sample pair corresponding to a positive relationship label; respectively constructing a plurality of negative sample pairs by using one unlabeled radar data in the first prediction segment and the second prediction segment, each negative sample pair corresponding to a negative relationship label.

[0018] Optimally, before performing S1, the semi-supervised training method further includes: normalizing all the plurality of unlabeled radar data and all the labeled radar data.

[0019] Optimally, the plurality of labeled radar data are obtained by randomly selecting some of the plurality of unlabeled radar data and assigning corresponding label information.

[0020] In a second aspect, the present invention further provides a radar working mode recognition system, which is trained by using the semi-supervised training method of the radar working mode recognition system described above. The system is used to output a corresponding radar working mode recognition result according to the input radar data to be recognized. The radar working mode recognition result at least includes: ranging while searching mode, tracking while searching mode, search plus tracking mode, anti-weather mode, burn-through mode, single target tracking mode, and guidance mode.

[0021] Compared with the prior art, the beneficial effects of the present invention are:

[0022] In view of the problems existing in the existing semi-supervised training methods when dealing with time series, such as the cumbersome and complex training process, serious error accumulation, and low recognition accuracy of radar working modes, the present invention proposes a radar working mode recognition system and a semi-supervised training method, which achieve performance breakthroughs through the collaborative optimization of supervised learning and self-supervised learning. The supervised module uses labeled data to drive the optimization of the classification boundary, and combines the high-order feature difference loss to enhance the discriminability of the mode; the self-supervised module adopts a dynamic time segment division strategy, adaptively adjusts the prediction segment span according to the verification index, reduces the division ratio to strengthen the long-range time series modeling when the model performance improves, and expands the ratio to reduce the learning difficulty when the performance fluctuates. The two achieve parameter synchronization updates through a shared feature extraction network, avoiding the problem of feature space offset. The composite loss function fuses supervised classification and self-supervised relationship prediction, constrains the semantic consistency of time series segments, so that the features of the same type of mode are highly aggregated and the boundaries between different types are clear. The dynamic adjustment mechanism effectively deals with interferences such as pulse loss and parameter errors, maintains stable recognition ability in complex electromagnetic environments, and accurately distinguishes modes such as "search and track" and "search while tracking" with similar carrier frequency and pulse width parameters, providing reliable technical support for the real-time and accurate recognition of radar working modes in complex electromagnetic environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 FIG. is a schematic structural diagram of a radar working mode recognition system provided by an embodiment of the present invention;

[0024] Figure 2 FIG. is a schematic structural diagram of a feature extraction unit provided by an embodiment of the present invention;

[0025] Figure 3 FIG. is a schematic structural diagram of a classifier unit provided by an embodiment of the present invention;

[0026] Figure 4 FIG. is a schematic flow diagram of a semi-supervised training method for a radar working mode recognition system provided by an embodiment of the present invention;

[0027] Figure 5 FIG. is a training schematic diagram of a radar working mode recognition system provided by an embodiment of the present invention;

[0028] Figure 6 FIG. is an example diagram of a confusion matrix corresponding to the case of different proportions of labeled radar data;

[0029] Figure 7 FIG. is an example diagram of a T-SNE distribution diagram corresponding to the case of different proportions of labeled radar data;

[0030] Figure 8 FIG. is the radar working mode recognition performance of the system under different interference scenarios. DETAILED DESCRIPTION OF THE INVENTION

[0031] The present invention will be further described in detail below in conjunction with specific embodiments, but the implementation manners of the present invention are not limited thereto.

[0032] In the description of the present invention, the terms "first" and "second" are only used for descriptive purposes and cannot be construed as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, the 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 of" means two or more unless otherwise specifically defined.

[0033] In the description of this specification, the description with reference to terms such as "an embodiment", "some embodiments", "example", "specific example", or "some examples" means 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 any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine the different embodiments or examples described in this specification.

[0034] Although the present invention has been described herein in connection with various embodiments, however, in the process of implementing the claimed invention, those skilled in the art can understand and achieve other variations of the disclosed embodiments by viewing the drawings, the disclosure, and the appended claims. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "one" does not exclude a plurality of cases. 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.

[0035] Now, in conjunction with the drawings, a semi-supervised training method for a radar working mode recognition system proposed by the present invention will be described in detail.

[0036] Before describing the semi-supervised training method for the radar working mode recognition system, the architecture of the radar working mode recognition system will be described first.

[0037] Figure 1 is a schematic structural diagram of a radar working mode recognition system provided by an embodiment of the present invention. As Figure 1As shown in the figure, the radar working mode recognition system includes: a supervised module and a self-supervised module. The supervised module includes a first feature extraction unit and a classifier unit, and the self-supervised module includes a second feature extraction unit and a predictor unit. The model parameters in the first feature extraction unit and the second feature extraction unit are the same.

[0038] Here, the structures and model parameters of the first feature extraction unit and the second feature extraction unit are exactly the same, and during the training process, the first feature extraction unit and the second feature extraction unit will be updated synchronously.

[0039] For a clear understanding, Figure 2 is a schematic structural diagram of the feature extraction unit provided by an embodiment of the present invention. As Figure 2 shown, the feature extraction unit f θ adopts a dual-track mechanism, and the specific structure is: embedding layer → (positional encoding) → multi-head attention / masked multi-head attention → residual connection & layer normalization → feed-forward network → residual connection & layer normalization → depthwise separable convolution → gated module → feature grouping → multi-scale feature extraction → cross-space learning → fully connected layer. This feature extraction unit is responsible for mapping the input time series t i ={t i,1 ,t i,2 ,…,t i,T} to a high-dimensional feature space to generate an effective feature representation z i =f θ (t i ).

[0040] Figure 3 is a schematic structural diagram of the classifier unit provided by an embodiment of the present invention. As Figure 3 shown, this unit adopts a multi-layer perceptron structure and consists of three fully connected layers and a non-linear activation function. Here, the classifier unit h μ is responsible for converting the high-dimensional feature representation generated by the first feature extraction unit into a class probability distribution where C is the number of classes. The specific structure of the classifier unit h μ is as follows:

[0041] The first fully connected layer and the second fully connected layer are used to perform a linear transformation on the high-dimensional feature representation output by the feature extraction network, and then perform non-linear processing through the ReLU activation function. The specific formula is: h i =ReLU(W (2) (ReLU(W (1) z i +b (1) ))+b (2) ), where W (1) ,W (2) are weight matrices, b (1) ,b(2) is the bias vector; the fully connected layer 3 is used to map the feature representations output by the previous two layers to the class space to generate class scores (unnormalized scores), and the specific formula is: where W (3) is the weight matrix, b (3) is the bias vector, and the output is the class score vector; the Softmax normalization layer is used to normalize the class scores into a probability distribution using the Softmax function, and the specific formula is: where p i,j is the predicted probability of the j-th class, satisfying The classifier network completes the mapping from the feature representation to the class probability distribution through the above process, supporting the optimization of the overall supervised classification task of the network.

[0042] It should be noted that the structure of the predictor unit is the same as that of the classifier unit h μ , and the main difference between the two lies in the parameter weights and the positions where they play roles in the model. The predictor unit h φ plays a role in the self-supervised module and is responsible for converting the sample pair features generated by the feature extraction network in the self-supervised module into the relationship prediction probability between two samples.

[0043] Now, in combination with the above Figures 1-3 , and Figure 4 a semi-supervised training method for the radar working mode recognition system provided by the embodiments of the present invention will be introduced. It should be understood that in the present invention, the "supervised module" can also be called the "supervised training module", and the "self-supervised module" can also be called the "self-supervised training module".

[0044] Figure 4 is a schematic flowchart of the semi-supervised training method for the radar working mode recognition system provided by the embodiments of the present invention. Figure 5 is a schematic diagram of the training of the radar working mode recognition system provided by the embodiments of the present invention.

[0045] As Figure 4 and Figure 5 shown, the method includes:

[0046] S1: In the i-th iterative training, input multiple labeled radar data carrying time series into the (i - 1)-th supervised module to obtain the i-th first loss value, and use the i-th first loss value to update the (i - 1)-th supervised module to obtain the i-th supervised module and the corresponding i-th feature extraction parameters.

[0047] Here, before executing S1, the semi-supervised training method further includes: normalizing all the multiple unlabeled radar data and all the labeled radar data. Subsequently, the normalized data is divided into a training set, a validation set, and a test set according to the ratio of 7:1:2. The training set is used for data training; the validation set is used to evaluate the training quality during the training process; the test set is used to test the network model after the training is completed. Among them, the normalized data t i ′ can be expressed as: where t i is the value of the i-th sample, min(t) is the minimum value of the sample, and max(t) is the maximum value of the sample.

[0048] Here, the multiple labeled radar data are obtained by randomly selecting some of the unlabeled radar data from the multiple unlabeled radar data and assigning corresponding label information. That is to say, the entire dataset D U are all unlabeled radar data. By selecting a part of the data according to the selection ratio and assigning label information to it, it is used as the labeled training set D L .

[0049] Specifically, S1 includes: using the (i - 1)-th first feature extraction unit to extract the first embedded feature vectors of each labeled radar data, obtaining multiple first embedded feature vectors; using the (i - 1)-th classifier unit to predict the class distribution of the multiple first embedded feature vectors, obtaining multiple class distribution results; using the fused polynomial loss function and the triplet loss function to calculate the multiple class distribution results, respectively obtaining the i-th fused polynomial loss value and the i-th triplet loss value; summing the i-th fused polynomial loss value and the i-th triplet loss value to obtain the i-th first loss value; using the i-th first loss value to update the (i - 1)-th first feature extraction unit and the (i - 1)-th classifier unit, obtaining the i-th supervised module and the i-th feature extraction parameters.

[0050] In a possible implementation manner, the expression of the i-th first embedded feature vector z i satisfies: where y i is the label of the sample t i . And, after being processed by the classifier unit for the i-th first embedded feature vector z i , the obtained class distribution result can be expressed as: p i = h μ (z i ).

[0051] In a possible implementation manner, each first loss value can be expressed as:

[0052] L label= L Poly + λL Triplet ;

[0053]

[0054] where λ ranges from 0 to 1, and are the feature representations of Anchor (reference sample), Positive (samples belonging to the same category as Anchor), and Negative (samples belonging to different categories from Anchor) samples extracted through the network. is the square of the Euclidean distance, used to measure the similarity between samples. α is a hyperparameter, representing the minimum margin between samples, that is, the distance between Anchor and Negative should be at least greater than α times the distance between Anchor and Positive, P t represents the predicted probability of the model for the true category, α j is the weight coefficient, used to control the influence of the j-th order polynomial term.

[0055] where, when truncating the polynomial in the fused polynomial loss function to the N-th order, the fused polynomial loss function is transformed into:

[0056]

[0057] where N is the highest order of the polynomial, ∈ j represents the offset of each order polynomial term relative to the cross-entropy loss. When ∈ j = 0, the fused polynomial loss is completely equivalent to the cross-entropy loss.

[0058] Furthermore, after obtaining the i-th first loss value, update the (i - 1)-th first feature extraction unit and the (i - 1)-th classifier unit. The feature extraction parameters in the i-th first feature extraction unit can be expressed as: The weight parameters in the classifier unit can be expressed as: η is the preset learning rate.

[0059] It should be understood that after obtaining the i-th first feature extraction unit in S1, the second feature extraction unit in the self-supervised module will also be updated synchronously.

[0060] S2: Input multiple unlabeled radar data with time series into the (i - 1)-th self-supervised module, divide the multiple unlabeled radar data into a first prediction segment and a second prediction segment by using the i-th time segment division ratio value, determine the i-th second loss value by using the first prediction segment, the second prediction segment and the i-th feature extraction parameter, and update the (i - 1)-th self-supervised module by using the i-th second loss value to obtain the i-th self-supervised module.

[0061] Here, S2 includes: dividing multiple unlabeled radar data in the dataset into a first prediction segment and a second prediction segment by using the i-th time segment division ratio value; constructing multiple positive sample pairs and multiple negative sample pairs by using the first prediction segment and the second prediction segment; respectively extracting features from the multiple positive sample pairs and the multiple negative sample pairs by using the i-th feature extraction parameter for the i-th second feature extraction unit updated from the (i - 1)-th second feature extraction unit to obtain multiple positive embedded feature vectors and multiple negative embedded feature vectors; respectively predicting the multiple positive embedded feature vectors and the multiple negative embedded feature vectors by using the (i - 1)-th predictor unit to obtain multiple positive prediction results and multiple negative prediction results; calculating the multiple positive prediction results and the multiple negative prediction results by using the binary cross-entropy loss function to obtain the i-th second loss value; updating the i-th second feature extraction unit and the (i - 1)-th predictor unit by using the i-th second loss value to obtain the i-th self-supervised module.

[0062] Here, before executing S3, the semi-supervised training method further includes: inputting the validation set into the i-th self-supervised module to obtain the i-th evaluation metric. Exemplarily, during the (i - 1)-th iterative training, inputting the validation set into the (i - 1)-th self-supervised module can obtain the (i - 1)-th evaluation metric. And, the semi-supervised training method further includes: before the start of the i-th iterative training, aggregating the evaluation metrics of the previous i - 1 times and screening out the maximum value from the evaluation metrics of the previous i - 1 times as the optimal historical metric among the previous i - 1 times. Further, the i-th time segment division ratio value is calculated by using the (i - 1)-th time segment division ratio value, the (i - 1)-th evaluation metric and the optimal historical metric among the previous i - 1 times, where when i takes the value of 1, the value of the 1st time segment division ratio value is a preset value (for example, 0.5).

[0063] It should be noted that the time segment division ratio value α is an adjustable parameter. If the model performance improves, the prediction difficulty is gradually increased (α is decreased), and if the performance decreases, the prediction difficulty is decreased (α is increased).

[0064] Here, the time segment division ratio value α can be expressed as: where α i is the division ratio of the i-th time, and M i is the evaluation metric of the t-th validation set. is the historical indicator that is the best up to the t-th time, i.e., Δα is a fixed adjustment step size (which can take 0.05), and clip(·) is a boundary constraint function that constrains the division ratio within the range of 0.3 to 0.7.

[0065] It should be understood that when then the current α t is in the favorable direction, and continue to search for a better solution along the negative gradient direction; when then escape from the local optimum along the positive gradient direction.

[0066] Here, using the first prediction segment and the second prediction segment, multiple positive sample pairs and multiple negative sample pairs are constructed, including: using multiple unlabeled radar data in the first prediction segment and multiple unlabeled radar data in the second prediction segment to construct multiple positive sample pairs, each sample pair contains two unlabeled radar data in each prediction segment, and each sample pair corresponds to a positive relationship label; respectively using one unlabeled radar data in the first prediction segment and one unlabeled radar data in the second prediction segment to construct multiple negative sample pairs, and each negative sample pair corresponds to a negative relationship label.

[0067] Exemplarily, from the same prediction segment, obtain s i,α and as a positive sample pair and its corresponding positive relationship label is 1, obtain an unlabeled radar data s from the first prediction segment i,α , and obtain an unlabeled radar data from the second prediction segment as a negative sample pair and the corresponding negative relationship label is 0.

[0068] In a possible implementation, the i-th positive embedding feature vector and the i-th negative embedding feature vector can be expressed as: Furthermore, the i-th positive prediction result can be expressed as: and the i-th negative prediction result can be expressed as: It should be understood that the expressions of each positive embedding feature vector, each negative embedding feature vector, each positive prediction result, and each negative prediction result are the same.

[0069] In a possible implementation, each second loss value can be expressed as: where represents the positive relationship, represents the negative relationship. p i refers to the value of the i-th positive prediction result or the i-th negative prediction result.

[0070] In a possible implementation, the feature extraction parameters in the (i + 1)-th second feature extraction unit can be expressed as: The weight parameters in the i-th predictor unit can be expressed as:

[0071] It should be understood that the first prediction segment can also be referred to as the "past time relationship segment", and the second prediction segment can also be referred to as the "future time relationship segment".

[0072] S3: Determine whether i is equal to I; if so, execute S4, otherwise execute S5.

[0073] S4: Stop the iterative training and obtain a trained radar working mode recognition system;

[0074] S5: Replace i in S1 to S3 with i + 1 and continue the next iterative training.

[0075] For the above-mentioned semi-supervised training method of the radar working mode recognition system, the present invention also provides a radar working mode recognition system, which is trained by using the above semi-supervised training method. The system is used to output the corresponding radar working mode recognition result according to the input radar data to be recognized. The radar working mode recognition result at least includes: ranging while searching mode, tracking while searching mode, search plus tracking mode, anti-weather mode, burn-through mode, single target tracking mode and guidance mode.

[0076] To verify the recognition accuracy of the above system, a simulation software is used to verify the system. PDWs are generated according to the ranging while searching mode, tracking while searching mode, search plus tracking mode, anti-weather mode, burn-through mode, single target tracking mode and guidance mode to obtain 28,000 pulse data. The specific parameters are shown in Table 1, and the hyperparameters in the experiment are shown in Table 2; each working mode includes 4,000 pulse description words. During the experiment, the overall pulse sequence is divided as required to obtain the sample to be recognized with the dimension of L×C, where C is the dimension of the PDW and L is the length of the time step. The data is divided into a training set, a validation set and a test set according to the ratio of 7:1:2. The labeled training set D L is composed of random samples selected from the training set, and the selection ratios are 40%, 20%, 10% and 5% in sequence. The unlabeled training set D U is composed of the entire training set. The validation set is used to evaluate the training quality during the training process and determine the size of the time segment division ratio value α. The test set is used to test the network model after the training is completed.

[0077] Table 1

[0078]

[0079] Table 2

[0080]

[0081] Figure 6 It is an example diagram of the confusion matrix corresponding to the case where the proportion of labeled radar data is different. Figure 7 It is an example diagram of the T-SNE distribution map when the proportion of labeled radar data is different. Among them, Figure 6 (a) in Figure 7 and (a) in Figure 6 are the cases when the proportion of labeled samples is 40%, Figure 7 and (b) in Figure 6 and (b) in Figure 7 are the cases when the proportion of labeled samples is 20%, Figure 6 and (c) in Figure 7 and (c) in Figure 6 and Figure 7 are the cases when the proportion of labeled samples is 10%,

[0082] To evaluate the robustness of the system in complex scenarios, five types of scenarios were designed in this experiment: scenarios that only contain false pulses, only contain missing pulses, only contain pulse measurement errors, scenarios that contain both false and missing pulses, and scenarios that contain both false and missing pulses as well as pulse measurement errors to evaluate the system's performance in radar working mode recognition. Figure 8 It is the system's performance in radar working mode recognition under different interference scenarios. Combining Figure 8 the parameter changes in

[0083] In the missing pulse scenario, based on a preset ratio of 5% - 50%, random pulse deletion processing was performed on the original pulse sequence to simulate the pulse loss phenomenon during the radar signal reception process.

[0084] In the false pulse scenario, false pulses were randomly inserted into the original pulse sequence, and the insertion ratio was set to 5% - 50% to represent false targets in the signal environment.

[0085] In the parameter measurement error scenario, the following maximum error thresholds were set with reference to the actual engineering level: carrier frequency measurement error: Δf c≤5000 kHz; Pulse width measurement error: ΔPW ≤ 5%·PW + 1 μs; Pulse repetition interval measurement error: ΔPRI ≤ 5 μs; Bandwidth measurement error: ΔBW ≤ 5%·BW + 200 kHz. In the experiment, random perturbations within the range of 0 - 100% maximum error are generated and superimposed on the PDW parameters to simulate measurement errors.

[0086] In the scenario of missing and misfiring pulses, valid pulses in the pulse sequence are randomly deleted at a ratio of 5% - 50% respectively, and error pulses accounting for 5% - 50% of the total number of pulses are randomly added. On the basis of the former, in the scenario of parameter measurement error superposition, errors are added to the PDW with a maximum measurement error of 5% - 50% respectively.

[0087] In the composite interference scenario, the algorithm performance shows significant non - linear decay characteristics. When the ratio of missing and misfiring pulses is greater than 20%, and the ratio of missing and misfiring pulses + parameter measurement error is greater than 15%, the recognition accuracy drops rapidly, indicating that the synergistic effect of deletion and insertion operations amplifies the feature distortion. When the ratio of missing pulses is less than 35%, the ratio of misfiring pulses is less than 25%, and the parameter measurement error is less than 40% of the maximum measurement error, the recognition accuracy of the semi - supervised learning algorithm proposed in this paper is still above 80%, indicating that the present invention can still play a role in complex electromagnetic environments and has engineering application value.

[0088] Aiming at the problems existing in the existing semi - supervised training methods when dealing with time series, such as the cumbersome and complex training process, serious error accumulation, and low recognition accuracy of radar working mode, the present invention proposes a radar working mode recognition system and a semi - supervised training method. Through the collaborative optimization mechanism of supervised and self - supervised learning, the dependence on labeled data in the radar working mode recognition task is reduced, and the model robustness is significantly improved. Its technical effects are mainly reflected in the following three aspects:

[0089] (1) Aiming at the problem of feature extraction of time - series data, through the dual - track feature extraction module, the multi - head attention mechanism and multi - scale feature extraction technology are fused, and the pattern representation ability is significantly improved under the condition of a small amount of labeled data. Experiments show that this method can still maintain high - precision recognition in extremely low - label scenarios, breaking through the strong dependence of traditional supervised learning on labeled data.

[0090] (2) The innovative design of the dynamic time - segment division strategy optimizes the utilization efficiency of unlabeled data. This method realizes the dynamic balance between the model training difficulty and the data representation ability through the time - segment ratio adaptive adjustment mechanism driven by verification indicators. Compared with the fixed division strategy, this design shows stronger robustness in complex electromagnetic interference scenarios, effectively alleviating the error accumulation problem in the pseudo - label method, and providing a new idea for the feature consistency learning of unlabeled time - series data.

[0091] (3) The design of the composite loss function realizes the deep coupling of supervised learning and self-supervised constraints. By fusing high-order feature difference modeling and feature space geometric constraints, this method demonstrates significant advantages in the task of differentiating similar working modes. The analysis of the training process shows that this loss combination effectively smooths the optimization path in semi-supervised learning and significantly accelerates the model convergence speed, providing an innovative technical path to address the practical engineering challenges of high annotation costs and dynamically changing electromagnetic environments.

[0092] The above content is a further detailed description of the present invention in combination with specific preferred embodiments, and 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 belonging to the protection scope of the present invention.

Claims

1. A semi-supervised training method for a radar working mode recognition system, characterized in that, The described radar working mode recognition system includes: a supervised module and a self-supervised module. The supervised module includes a first feature extraction unit and a classifier unit. The self-supervised module includes a second feature extraction unit and a predictor unit. The model parameters in the first feature extraction unit and the second feature extraction unit are the same. This semi-supervised training method includes: S1: In the i-th iterative training, input multiple labeled radar data with time series into the (i - 1)-th supervised module to obtain the i-th first loss value. Use the i-th first loss value to update the (i - 1)-th supervised module to obtain the i-th supervised module and the corresponding i-th feature extraction parameters; S2: Input multiple unlabeled radar data with time series into the (i - 1)-th self-supervised module. Use the i-th time segment division ratio value to divide the multiple unlabeled radar data into a first prediction segment and a second prediction segment. Use the first prediction segment, the second prediction segment, and the i-th feature extraction parameters to determine the i-th second loss value. Use the i-th second loss value to update the (i - 1)-th self-supervised module to obtain the i-th self-supervised module; S3: Judge whether i is equal to I; if so, execute S4, if not, execute S5; S4: Stop the iterative training to obtain the trained radar working mode recognition system; S5: Replace i in S1 to S3 with i + 1 and continue the next iterative training.

2. The semi-supervised training method of the radar working mode recognition system according to claim 1, characterized in that The S1 includes: Use the (i - 1)-th first feature extraction unit to extract the first embedded feature vectors of each labeled radar data to obtain multiple first embedded feature vectors; Use the (i - 1)-th classifier unit to predict the class distribution of the multiple first embedded feature vectors to obtain multiple class distribution results; Use the fusion polynomial loss function and the triplet loss function to calculate the multiple class distribution results to obtain the i-th fusion polynomial loss value and the i-th triplet loss value respectively; Sum the i-th fusion polynomial loss value and the i-th triplet loss value to obtain the i-th first loss value; Use the i-th first loss value to update the (i - 1)-th first feature extraction unit and the (i - 1)-th classifier unit to obtain the i-th supervised module and the i-th feature extraction parameters.

3. The semi-supervised training method of the radar working mode recognition system according to claim 2, characterized in that, The S2 includes: Use the i-th time segment division ratio value to divide the multiple unlabeled radar data in the dataset into a first prediction segment and a second prediction segment; Use the first prediction segment and the second prediction segment to construct multiple positive sample pairs and multiple negative sample pairs; Use the i-th feature extraction parameters to update the (i - 1)-th second feature extraction unit to obtain the i-th second feature extraction unit, and respectively extract features from the multiple positive sample pairs and the multiple negative sample pairs to obtain multiple positive embedded feature vectors and multiple negative embedded feature vectors; Use the (i - 1)-th predictor unit to respectively predict the multiple positive embedded feature vectors and the multiple negative embedded feature vectors to obtain multiple positive prediction results and multiple negative prediction results; Using the binary cross-entropy loss function, calculate the multiple positive prediction results and multiple negative prediction results to obtain the i-th second loss value; Update the i-th second feature extraction unit and the (i - 1)-th predictor unit using the i-th second loss value to obtain the i-th self-supervised module.

4. The semi-supervised training method of the radar working mode recognition system according to claim 3, characterized in that The i-th time segment division ratio value is calculated using the (i - 1)-th time segment division ratio value, the evaluation index of the (i - 1)-th time, and the optimal historical index among the previous (i - 1) times. Wherein, when i takes the value of 1, the value of the first time segment division ratio value is a preset value.

5. The semi-supervised training method of the radar working mode recognition system according to claim 3, characterized in that Before executing S3, the semi-supervised training method further includes: Input the validation set into the i-th self-supervised module to obtain the evaluation index of the i-th time.

6. The semi-supervised training method of the radar working mode recognition system according to claim 4, characterized in that, The semi-supervised training method further includes: Before the start of the i-th iterative training, summarize the evaluation indexes of the previous (i - 1) times, and select the maximum value from the evaluation indexes of the previous (i - 1) times as the optimal historical index among the previous (i - 1) times.

7. The semi-supervised training method of the radar working mode recognition system according to claim 3, characterized in that, The constructing of multiple positive sample pairs and multiple negative sample pairs using the first prediction segment and the second prediction segment includes: Using the multiple unlabeled radar data in the first prediction segment and the multiple unlabeled radar data in the second prediction segment to construct multiple positive sample pairs. Each sample pair contains two unlabeled radar data in each prediction segment, and each sample pair corresponds to a positive relationship label; Respectively use one unlabeled radar data in the first prediction segment and the second prediction segment to construct multiple negative sample pairs. Each negative sample pair corresponds to a negative relationship label.

8. The semi-supervised training method of the radar working mode recognition system according to claim 1, characterized in that, Before executing S1, the semi-supervised training method further includes: Perform normalization processing on all the multiple unlabeled radar data and all the labeled radar data.

9. The semi-supervised training method of the radar working mode recognition system according to claim 4, characterized in that The multiple labeled radar data are obtained by randomly selecting some of the multiple unlabeled radar data and assigning one-to-one corresponding label information.

10. A radar operating mode recognition system, characterized in that, This system is trained using the semi-supervised training method of the radar working mode recognition system according to any one of claims 1 to 9 above. The system is used to output the corresponding radar working mode recognition result according to the input radar data to be recognized. The radar working mode recognition result at least includes: ranging while searching mode, tracking while searching mode, search and track mode, anti-weather mode, burn-through mode, single target tracking mode, and guidance mode.

Images