Incremental Recognition Method for Multifunctional Radar Operating Modes Based on Deep Feature Expansion

Abstract

The present invention discloses a method for incrementally identifying multifunctional radar working modes based on deep feature expansion, including: inputting the sample data of the radar to be measured obtained into an incremental identification model, and the incremental identification model is trained according to the extended data set composed of new sample data and old sample data under the iteration of the hybrid loss function to obtain a trained model; then inputting the sample data into the trained model to obtain an identification result. According to the method provided by the present invention, by inputting the sample data into the incremental model for training, a trained model is obtained; then inputting the sample data into the trained model to obtain the identification result of the radar working mode category; since the extended feature extractor in the model is trained under the iteration of the hybrid loss function, the model can maintain the memory of old knowledge while learning new types, thereby enhancing the incremental identification effect in a dynamic environment, improving the identification accuracy, and expanding the application scenario.

Description

Technical Field

[0001] The present invention belongs to the technical field of signal recognition, and particularly relates to a multi-functional radar operating mode incremental recognition method based on deep feature expansion. Background Art

[0002] The processing of radar reconnaissance signals is a key task in modern electronic warfare, and radar operating mode recognition is an important part of radar reconnaissance signal processing. With the development of phased array radar technology, modern multi-functional radars have more flexible time and space resource allocation capabilities and stronger scene adaptability. Multi-functional radars have multiple operating mode categories, such as airspace surveillance, search, and target tracking. Under the constraints of the total radar system time and energy resources, when performing tasks of different operating mode categories, the radar can adjust the intra-pulse or inter-pulse parameters in real time according to different task categories to achieve the switching of operating modes.

[0003] Currently, there are roughly three categories of methods for identifying radar operating modes. One is the traditional identification method, such as the template matching method, statistical histogram method, etc., which can only be applied to conventional radars with fixed parameters. Due to the stronger dynamics and higher complexity of multi-functional radars, the identification effects of the above simple identification methods are greatly affected.

[0004] One is the model-based identification method, such as the two types of models based on the Hidden Markov Model (HMM) and the syntactic model. The identification method based on the Hidden Markov Model models the sequential pulse sequence of the radar, and then uses the classical algorithms of HMM for model training, signal identification, state decoding, and prediction. However, due to the large number of operating mode categories of multi-functional radars and the complex system state transitions, the model complexity of the identification method based on the HMM model is relatively high, the computational amount is large, and the usage scenarios are limited. The identification method based on the syntactic model equates the complex mapping relationship from the radar pulse sequence to the operating mode to the semantic relationship between "radar words", "radar phrases", and "radar sentences", and then conducts radar operating mode recognition through semantic analysis. The identification method based on the syntactic model is relatively dependent on the prior knowledge of the multi-functional radar to be identified. In practical applications, it is generally difficult to obtain the prior knowledge of the enemy radar, and during the actual detection process, the detected radar signals will contain a large number of non-ideal situations such as false pulses, missing pulses, and measurement errors, resulting in the destruction of the original pulse signal pattern, making it often difficult to extract radar words. This leads to poor identification effects of the identification method based on the syntactic model.

[0005] One is the recognition method based on deep neural network. Although this method does not require a large amount of prior knowledge about multifunctional radar, the recognition method based on deep neural network is trained and tested in a static observation scenario where data of all working mode categories can be obtained simultaneously. However, in practice, it is difficult to obtain data of all working mode categories of the enemy radar through a limited number of reconnaissances, and the working modes of most radars are only used in special situations, which results in limited data for training and testing the deep neural network. This method has poor effect in recognizing multifunctional radars with rapidly changing working modes in dynamic scenarios. Summary of the Invention

[0006] An embodiment of the present invention provides a method for incremental recognition of working modes of multifunctional radar based on deep feature expansion, which can solve the problems of low recognition accuracy and limited application scenarios of the current recognition methods.

[0007] In a first aspect, an embodiment of the present invention provides a method for incremental recognition of working modes of multifunctional radar based on deep feature expansion, the method comprising:

[0008] Obtain sample data of the radar to be measured, where the sample data includes samples of new working mode categories, and the new working mode categories are different from the data types of the old working mode categories stored in the incremental recognition model;

[0009] Input the sample data into the trained incremental recognition model, and the incremental recognition model is trained according to the extended data set to obtain the trained incremental recognition model,

[0010] wherein, the incremental recognition model includes an extended feature extractor, a sample management module and a working mode classification module. The extended feature extractor is used to extract the features of the sample data. The working mode classification module is used to determine the working mode category to which the sample data belongs according to the distance between the sample data and the average example vector of each working mode category. The sample management module is used to determine and store representative samples and / or delete redundant samples. The extended feature extractor is trained according to the extended data set under the iteration of the hybrid loss function, and the dynamic adjustment coefficient of the hybrid loss function decreases as the number of incremental recognition tasks increases.

[0011] Input the sample data into the trained incremental recognition model to obtain the recognition result of the category to which the radar working mode belongs.

[0012] According to the method provided by the present invention, by inputting sample data including new working mode categories into an incremental model for training, a trained incremental recognition model is obtained; then the sample data is input into the trained incremental recognition model to obtain the recognition result of the working mode category to which the sample data belongs; since the incremental recognition model trained under the iteration of the hybrid loss function can maintain the memory of old knowledge while learning new types, the recognition effect can be enhanced, the recognition accuracy can be improved, and the application scenario can be expanded.

[0013] In a possible implementation manner of the first aspect, the hybrid loss function satisfies the following formula:

[0014]

[0015]

[0016] Wherein, is the main loss function, is the auxiliary loss function, is the hybrid loss function, is the number of new working mode categories, is the number of old working mode categories, is the dynamic adjustment coefficient.

[0017] In a possible implementation manner of the first aspect, the main loss function may satisfy the following formula:

[0018]

[0019] Wherein, is the main loss function, is the working mode included in the sample data, is the predicted label corresponding to the working mode category, is the prediction result of the main classifier, is the number of samples in the extended dataset.

[0020] Exemplarily, the prediction result of the main classifier includes: the probabilities that the sample data in the extended dataset belong to each new working mode category respectively, and the probabilities that the sample data in the extended dataset belong to each old working mode category respectively.

[0021] Exemplarily, the prediction result of the main classifier is determined according to the features of the sample data.

[0022] In a possible implementation manner of the first aspect, the incremental recognition model may further include a main classifier, and the main classifier may be used to determine the prediction result of the main classifier according to the features of the sample data.

[0023] Exemplarily, the prediction result of the main classifier satisfies the following formula:

[0024] ;

[0025] Among them, is the working mode included in the sample data, is the predicted label corresponding to the working mode category, is the prediction result of the main classifier, is the feature of the sample data.

[0026] In a possible implementation of the first aspect, the auxiliary loss function may satisfy the following formula:

[0027]

[0028] Among them, is the auxiliary loss function, is the working mode included in the sample data, is the predicted label corresponding to the working mode category, is the prediction result of the auxiliary classifier, is the number of samples in the extended dataset.

[0029] Exemplarily, the prediction result of the auxiliary classifier may include: the probability that the sample data in the extended dataset belongs to the old working mode category, and the probabilities that the sample data in the extended dataset belongs to each new working mode category respectively.

[0030] Exemplarily, the prediction result of the auxiliary classifier is determined according to the features of the sample data.

[0031] In a possible implementation of the first aspect, the incremental recognition model may further include an auxiliary classifier, and the auxiliary classifier is used to determine the prediction result of the auxiliary classifier according to the features of the sample data.

[0032] Exemplarily, the prediction result of the auxiliary classifier satisfies the following formula:

[0033]

[0034] Among them, is the working mode included in the sample data, is the predicted label corresponding to the working mode category, is the prediction result of the auxiliary classifier, is the feature of the sample data.

[0035] In a possible implementation of the first aspect, the sample management module may specifically be used to sort the operating modes included in the operating mode category in descending order according to the output result of the Softmax layer in the sample data, so as to obtain a first sequence of the operating mode category. Determine the first N operating modes in the first sequence as the representative samples of the operating mode category, and then store the representative samples.

[0036] Exemplarily, the output result of the Softmax layer is the probability that the sample data belongs to the operating mode category;

[0037] Exemplarily, the output result of the Softmax layer is determined according to the prediction results of the main classifier and the auxiliary classifier.

[0038] Exemplarily, N is a positive integer.

[0039] In a possible implementation of the first aspect, the sample management module may also be used to: delete redundant sample data, where the redundant sample data is the sample data of the last m operating modes in the first sequence, and m is a positive integer.

[0040] In a possible implementation of the first aspect, the current extended feature extractor is determined according to the previous extended feature extractor and the basic feature extractor.

[0041] In a possible implementation of the first aspect, before obtaining the sample data, the original data may be obtained first; then according to the original data, a pulse parameter sequence may be determined; and then the pulse parameter sequence is processed such as abnormal pulse elimination, PRI parameter calculation, normalization, etc. to obtain the sample data.

[0042] Exemplarily, the original data includes: the carrier frequency, pulse width, pulse amplitude, and pulse arrival time of the radar transmitted pulse.

[0043] Exemplarily, the pulse parameter sequence includes: the carrier frequency, pulse width, pulse amplitude, and pulse repetition time of the radar transmitted pulse.

[0044] The beneficial effects of the embodiments of the present invention compared with the prior art are as follows: According to the method provided by the present invention, by inputting the sample data into the incremental model for training, a trained model is obtained; then the sample data is input into the trained model to obtain the recognition result of the radar operating mode category; since the extended feature extractor in the model is trained under the iteration of the mixed loss function, the model can maintain the memory of old knowledge while learning new types, so as to enhance the incremental recognition effect in a dynamic environment, improve the recognition accuracy, and expand the application scenario. Description of the Drawings

[0045] Figure 1 It is a schematic structural diagram of an incremental recognition model provided by an embodiment of the present invention;

[0046] Figure 2 Schematic flow chart of a multi-functional radar working mode incremental recognition method based on deep feature expansion provided by an embodiment of the present invention;

[0047] Figure 3 Schematic comparison diagram of an identification accuracy rate provided by an embodiment of the present invention;

[0048] Figure 4 Another schematic comparison diagram of an identification accuracy rate provided by an embodiment of the present invention;

[0049] Figure 5 Another schematic comparison diagram of an identification accuracy rate provided by an embodiment of the present invention. Detailed implementation manners

[0050] In the following description, specific details such as specific system structures and technologies are presented for the purpose of illustration rather than limitation, so as to thoroughly understand the embodiments of the present invention. However, those skilled in the art should clearly understand that the present invention can also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to avoid unnecessary details from obstructing the description of the present invention.

[0051] It should be understood that when used in the specification and appended claims of the present invention, the term "comprising" indicates the presence of the described features, wholes, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components, and / or their combinations.

[0052] It should also be understood that the term "and / or" as used in the specification and appended claims of the present invention refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0053] As used in the specification and appended claims of the present invention, the term "if" can be interpreted as "when", "once", "in response to determining", or "in response to detecting" according to the context. Similarly, the phrase "if determined" or "if [the described condition or event] is detected" can be interpreted as meaning "once determined", "in response to determining", "once [the described condition or event] is detected", or "in response to detecting [the described condition or event]" according to the context.

[0054] In addition, in the description of the specification and appended claims of the present invention, the terms "first", "second", "third", etc. are only used for distinguishing descriptions and cannot be understood as indicating or implying relative importance.

[0055] References to "one embodiment" or "some embodiments" etc. described in the specification of the present invention mean that specific features, structures or characteristics described in connection with that embodiment are included in one or more embodiments of the present invention. Thus, statements such as "in one embodiment", "in some embodiments", "in other some embodiments", "in still other embodiments" etc. that appear in different places in this specification do not necessarily all refer to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized. The terms "comprise", "include", "have" and their variants all mean "including but not limited to", unless otherwise specifically emphasized.

[0056] Figure 1 Shown is a schematic structural diagram of an incremental recognition model 100 provided by an embodiment of the present invention. By way of example and not limitation, the model 100 may include an extended feature extractor 110, a sample management module 120, and a working mode classification module 130.

[0057] The extended feature extractor 110 may be used to extract features of sample data in an extended data set.

[0058] Exemplarily, the extended data set includes data of new working mode categories and data of old working mode categories.

[0059] In some embodiments, the extended feature extractor for the t-th incremental learning task may extract features in the sample data through the following formula:

[0060]

[0061]

[0062] where, is the extended feature extractor for the t-th incremental learning task, is the extended feature extractor for the (t - 1)-th incremental learning task, is the basic feature extractor, is the sample data, is the feature of the sample data.

[0063] Exemplarily, the extended feature extractor of the current incremental learning task is obtained based on the basic feature extractor and the extended feature extractor of the previous incremental learning task.

[0064] Exemplarily, the basic feature extractor may be a Temporal Convolutional Network (TCN).

[0065] For example, the number of convolutional layers of the basic feature extractor can be 2, the size of the convolutional kernel can be 4, the number of input channels can be 4, and the number of output channels can be 128.

[0066] Exemplarily, the core structure of the basic feature extractor includes dilated causal convolution and residual modules.

[0067] Exemplarily, the sequence At The dilated convolution satisfies the following formula:

[0068]

[0069] Where Is the dilated convolution of the sequence s at The dilation factor d grows in the form of an exponential power of 2, Represents the size of the convolutional kernel, Is the filter.

[0070] Exemplarily, the extended feature extractor 110 is trained according to the extended dataset under the iteration of the hybrid loss function.

[0071] Exemplarily, the dynamic adjustment coefficient of the hybrid loss function decreases as the incremental task increases. In this way, the extended feature extractor can retain the memory of old knowledge while learning new knowledge.

[0072] The sample management module 120 can be used to determine and store representative samples and / or delete redundant samples according to the sample data.

[0073] In some embodiments, the sample management module can determine representative samples in each working mode category and delete redundant samples. And when performing a new incremental learning task, it will combine the stored old category radar working mode data and new category radar working mode data into an extended dataset for training.

[0074] In a possible implementation, the sample management module can sort the output results of the softmax layer in descending order to form a list, and obtain the first sequence corresponding to the working mode category. Then, the sample data of the first N working modes in the first sequence, where N is a positive integer, are determined as the representative samples of the working mode category.

[0075] For example, there are 50 working modes with a predicted label of Y3, and these 50 working modes are respectively denoted as: S1, S2, S3... S50. The probabilities that each working mode output by the softmax layer belongs to the working mode category Y3 are successively: 0.95, 0.92, 0.9, 0.98,.... The first sequence corresponding to the working mode Y3 is: S4, S1, S2, S3.... If N = 4, then the working modes S4, S1, S2, and S3 can be determined as representative samples of the working mode category Y3.

[0076] In another possible view mode, when there is insufficient space in the storage space, the sample management module can delete the last M sample data at the end of the first sequence, where M is a positive integer.

[0077] Exemplarily, the number of radar working mode samples that the storage space for storing sample data in the model 100 can store satisfies the following formula:

[0078]

[0079] Among them, is the capacity of the storage space, and m is the number of radar working mode samples that the storage space can store.

[0080] The working mode classification module 130 can be used to determine the working mode category to which the sample data belongs according to the distance between the sample data feature and the average example vector of each working mode category.

[0081] In some embodiments, the working mode classification module 130 can determine the average example vector of each working mode category according to the true label corresponding to the sample data of the representative sample.

[0082] Exemplarily, the working mode classification module 130 can be the Nearest-Mean-of-Exemplars Classification (NME).

[0083] Exemplarily, the average example vector of the working mode category can satisfy the following formula:

[0084]

[0085] Among them, is all the working mode samples under this working mode category, is the average example vector of this working mode category, is the sum of the features of all the working modes in this working mode category.

[0086] In one example, the working mode classification module 130 can determine the average example vector that is closest to the sample data, and then determine the prediction label of the working mode category corresponding to the average example vector as the prediction label of the sample data. The working mode category corresponding to the prediction label is the working mode category to which the sample data belongs.

[0087] Exemplarily, the working mode category to which the sample data belongs can be determined by the following formula:

[0088]

[0089] Where, is the prediction label of the sample data.

[0090] Optionally, the model 100 may further include a main classifier 141, and the main classifier is used to determine the prediction result of the sample data according to the features of the sample data.

[0091] Exemplarily, the main classifier may be a softmax classifier.

[0092] Exemplarily, the prediction result of the main classifier satisfies the following formula:

[0093]

[0094] Where, is the working mode included in the sample data of the current incremental learning task, is the prediction label corresponding to the working mode category, is the prediction result of the main classifier, is the feature of the sample data.

[0095] Optionally, the model 100 may further include an auxiliary classifier 141, and the auxiliary classifier is used to determine the prediction result of the sample data according to the features of the sample data.

[0096] Exemplarily, the auxiliary classifier may be a softmax classifier.

[0097] Exemplarily, the prediction result of the auxiliary classifier satisfies the following formula:

[0098]

[0099] Where, is the working mode included in the sample data of the current incremental learning task, is the prediction label corresponding to the working mode category, is the prediction result of the auxiliary classifier, is the feature of the sample data.

[0100] Exemplarily, the prediction results of the main classifier and the auxiliary classifier can be used to determine the hybrid loss function and perform iterative optimization.

[0101] The multi-functional radar working mode incremental recognition method based on deep feature expansion provided by the embodiments of the present invention can be used in processing devices such as radars, computers, and supercomputers. The embodiments of the present invention do not impose any restrictions on the specific types of processing devices.

[0102] Figure 2 Shown is a multi-functional radar working mode incremental recognition method 200 based on deep feature expansion. The method 2OO can be applied to the above-mentioned processing device. By way of example and not limitation, the method 200 may include steps S201-S203. The following explains steps S201-S203.

[0103] S201, obtaining sample data of the radar to be measured.

[0104] In some embodiments, generally the directly obtained is the raw data of the radar, such as the carrier frequency, pulse width, pulse amplitude, and pulse arrival time of the radar transmitted pulse. Therefore, before processing the sample data, the sample data can be determined according to the raw data first.

[0105] Exemplarily, the sample data may include pulse data of at least one working mode category.

[0106] Exemplarily, each working mode category may include at least one working mode.

[0107] In one possible implementation, the pulse parameter sequence can be determined according to the raw data, and then the pulse parameter sequence is normalized to obtain the sample data.

[0108] Exemplarily, the pulse parameter sequence includes the carrier frequency, pulse width, pulse amplitude, and pulse repetition time of the pulse.

[0109] In one example, the pulses below the first pulse width threshold in the raw data can be filtered first to obtain the filtered raw data. Then, a first-order difference operation is performed on the pulse arrival time in the filtered raw data to obtain the pulse repetition period, thereby obtaining the pulse parameter sequence.

[0110] Exemplarily, the first pulse width threshold may be 5 us.

[0111] In one example, the following formula can be used to normalize each parameter in the pulse parameter sequence, so as to exclude the influence of different dimensions of each parameter in the pulse parameter sequence:

[0112]

[0113] Wherein, is the parameter vector of each parameter in the pulse parameter sequence, such as the pulse width vector.

[0114] S202: Input the sample data into the incremental recognition model. The incremental recognition model is trained according to the extended data set to obtain the trained incremental recognition model.

[0115] Exemplarily, the sample data includes samples of new working mode categories.

[0116] Exemplarily, the samples of the new working mode category have a different data type from the data of the old working mode category stored in the incremental recognition model.

[0117] Exemplarily, the incremental recognition model can be the above-mentioned incremental recognition model 100.

[0118] Exemplarily, the incremental recognition model includes an extended feature extractor, a sample management module, and a working mode classification module. The extended feature extractor is used to extract the features of the sample data. The working mode classification module is used to determine the working mode category to which the sample data belongs according to the distance between the sample data features and the average example vector of each working mode category. The sample management module is used to determine and store representative samples and / or delete redundant samples. The extended feature extractor is trained according to the extended data set under the iteration of the hybrid loss function, and the dynamic adjustment coefficient of the hybrid loss function decreases as the number of incremental recognition tasks increases.

[0119] In some embodiments, the hybrid loss function satisfies the following formula:

[0120]

[0121]

[0122] Wherein, is the main loss function, is the auxiliary loss function, is the hybrid loss function, is the number of new working mode categories, is the number of old working mode categories, is the dynamic adjustment coefficient.

[0123] In a possible implementation manner, the main loss function can be represented by the following formula according to the cross entropy:

[0124]

[0125] Wherein, is the main loss function, is the working mode included in the sample data, is the predicted label corresponding to the working mode category, is the prediction result of the main classifier, is the number of samples in the extended dataset.

[0126] Exemplarily, the prediction result of the main classifier includes: the probabilities that the sample data belong to each new working mode category respectively, and the probabilities that the sample data belong to each old working mode category respectively.

[0127] Optionally, the prediction result of the main classifier can be determined by the main classifier in the incremental recognition model according to the sample data.

[0128] For example, among the sample data of the current incremental learning task, there are 5 pieces of data of radar working modes, which are respectively denoted as S1, S2, S3, S4, and S5. There are 3 new working mode categories in the sample data, and their corresponding prediction labels are Y1, Y2, and Y3 respectively. There are 100 old working mode categories stored in the incremental learning model 100, and their corresponding prediction labels are Y4, Y5, Y6... Y103 respectively. Then, after inputting the sample data into the main classifier, the prediction results that the 5 working modes in the sample data belong to the above 103 working mode categories respectively can be obtained, and a total of 515 prediction result values are obtained.

[0129] In a possible implementation manner, the auxiliary loss function can be expressed by the following formula according to the cross-entropy:

[0130]

[0131] where, is the auxiliary loss function, is the working mode included in the sample data, is the prediction label corresponding to the working mode category, is the prediction result of the auxiliary classifier, is the number of samples in the extended dataset.

[0132] Exemplarily, the prediction result of the auxiliary classifier includes the probability that the sample data belong to the old working mode category and the probabilities that the sample data belong to each new working mode category respectively.

[0133] Optionally, the prediction result of the auxiliary classifier can be determined by the auxiliary classifier of the incremental recognition model according to the features of the sample data.

[0134] Exemplarily, in the auxiliary classifier, the sum of all old working mode categories is regarded as a large working mode category, so as to increase the influence of the new working mode category when determining the auxiliary loss function.

[0135] For example, inputting the features of the sample data of the current incremental task into the auxiliary classifier can obtain the probabilities that the working modes S1, S2, S3, S4, and S5 belong to the new working mode categories Y1, Y2, and Y3 respectively, as well as the probabilities that the working modes S1, S2, S3, S4, and S5 belong to the old working mode categories Y4 - Y103. A total of 20 prediction result values are obtained.

[0136] In some embodiments, initial setting parameters can be obtained and input into the incremental recognition model. Then, the incremental recognition model is trained under the iteration of the hybrid loss function based on the initial setting parameters and the extended dataset, that is, the incremental recognition model starts incremental learning. Then, representative samples of each working mode category are determined and saved. Then, according to the number of working mode categories, sample data of multiple working modes are randomly selected as the extended dataset for the next incremental learning task for learning. This continues until all the data in the training set data have been selected, at which point the incremental learning ends, and the currently trained radar working mode recognition model is obtained.

[0137] Optionally, the predicted labels of the sample data can also be compared with the true labels to obtain the accuracy rate of incremental recognition.

[0138] Exemplarily, the initial setting parameters can include the maximum number of iterations for each incremental learning task, the initial number of incremental training categories, and the number of newly added categories each time.

[0139] For example, if the maximum number of iterations for each incremental learning task is times, the initial number of incremental training categories is , and the number of newly added categories each time is ; then sample data of working modes can be randomly selected to perform iterative training on the radar working mode recognition model. Then, the first sequence of each working category is determined, and representative samples of each working mode category are saved according to the first sequence. Then, sample data of working modes are randomly selected again to perform iterative training on the radar working mode recognition model... This continues until all the data in the training set data have been selected, at which point the training ends, and the trained radar working mode recognition model is obtained.

[0140] S203, input the sample data into the trained incremental recognition model to obtain the recognition result of the category to which the radar working mode belongs.

[0141] Inputting the sample data into the trained incremental model can obtain the working mode category to which the sample data belongs.

[0142] Through the following experiments, the beneficial effects of the radar working mode recognition model construction method provided according to the present invention can be better illustrated.

[0143] Exemplarily, the experimental data set includes pulse data of 20 different working mode categories, and 30,000 pulses of the intercepted radar are selected for each working mode category. Each working mode category includes 625 working modes. Then each working mode under each working mode category includes 48 pulse data. Then these pulse data can be randomly divided into training set data and test set data according to a certain ratio, such as 8:2.

[0144] Exemplarily, during the model training process, the Adam optimizer is adopted, and a non-uniform step size adjustment strategy is used to dynamically adjust the learning rate so that the model can better converge to the optimal solution. The initial learning rate in the initial and subsequent incremental training processes is set to 5e -4 , the number of training rounds in each incremental task is 200 times, and the batch size during training is 128. The old category sample retention method adopts a fixed memory component size, which is set to 400. The number of samples of each working mode category saved in the memory component decreases as the number of working mode types increases during the incremental learning process. Specifically, when the number of working mode types in the initial incremental task is 4, the number of each working mode saved in the memory component is 400 / 4 = 100. When the number of working mode types in the next incremental task is also 4, the number of samples of each radar working mode saved in the memory component at this time is 400 / 8 = 50. And so on until the training ends.

[0145] Figure 3 Shown is a comparison schematic diagram of the recognition accuracy provided by an embodiment of the present invention

[0146] Exemplarily, the method provided according to the present invention, the recognition method based on the LWF (learning without forgetting) incremental learning model, the recognition method based on the ICaRL (Incremental Classifier and Representation Learning) incremental learning model, the recognition method based on the EWC (Elastic weight consolidation) incremental learning model, the recognition method based on the DER (dynamically expandable representation) incremental learning model, and the recognition method based on the BIC (Bias Correction) incremental learning model are compared using the same sample data. The number of initial incremental training categories is 4, the number of newly added categories each time is 4 categories, and the number of samples saved in the memory component is set to 400.

[0147] SeeFigure 3 , where 301 is the recognition accuracy rate obtained by the method according to the present invention, 302 is an incremental learning model based on LWF (learning without forgetting), 303 is an identification method of an incremental learning model based on ICaRL (Incremental Classifier and Representation Learning), 304 is an identification method of an incremental learning model based on EWC (Elastic weight consolidation), 305 is an identification method of an incremental learning model based on BIC (Bias Correction), and 306 is an identification method of an incremental learning model based on DER (dynamically expandable representation). It can be seen that in the initial recognition task, all methods achieved a recognition accuracy rate of over 90%, but as the incremental task progressed, the recognition accuracy rates of all methods showed an obvious downward trend. And the accuracy rate of the method provided by the present invention is the highest.

[0148] Figure 4 Shown is another comparison schematic diagram of the recognition accuracy rate provided by the embodiment of the present invention.

[0149] Exemplarily, using the same sample data, the number of reserved samples is set to 400, the initial number of incremental categories is set to 4, and the number of incremental categories each time is 4.

[0150] See Figure 4 , Figure 4 , where 401 is the accuracy rate of using the method provided by the present invention, and 402 is the accuracy rate of the incremental learning model obtained by replacing the sample management module in the incremental learning model of the method provided by the present invention with an existing sample management method. It can be seen that the accuracy rate of the incremental learning model of the method provided by the present invention is 12.24% higher than that of the existing one.

[0151] Figure 5 Shown is yet another comparison schematic diagram of the recognition accuracy rate provided by the embodiment of the present invention.

[0152] Exemplarily, using the same sample data, the number of samples stored in the memory component is set to 400, the initial number of incremental categories is set to 4, and the number of incremental categories each time is 4.

[0153] See Figure 5 , Figure 5Among them, 501 is the accuracy rate of using the method provided by the present invention, and 502 is the accuracy rate of the incremental learning model obtained by replacing the dynamic adjustment coefficient of the mixed loss function of the incremental recognition model in the present invention with a fixed adjustment coefficient. It can be seen that in each incremental task, the recognition accuracy rate of using the dynamic adjustment coefficient is increased by 4.0%, 3.6%, 1.0%, and 0.24% in turn compared with that of using the fixed adjustment coefficient, and the overall average recognition accuracy rate is increased by 2.21%.

[0154] Thus, according to the method provided by the present invention, by inputting sample data into an incremental model for training, a trained model is obtained; then the sample data is input into the trained model to obtain the recognition result of the radar working mode category; since the extended feature extractor in the model is trained under the iteration of the mixed loss function, the model can maintain the memory of old knowledge while learning new types, so as to enhance the incremental recognition effect in a dynamic environment, improve the recognition accuracy rate, and expand the application scenarios. Further, since the sample management module in the incremental model can determine the samples with more distinct features as representative samples for comparison and recognition with the sample data, the method provided by the present invention can further enhance the recognition accuracy rate.

[0155] In the above embodiments, the descriptions of the respective embodiments have their own focuses. For parts not detailed or recorded in a certain embodiment, reference may be made to the relevant descriptions of other embodiments.

[0156] Those of ordinary skill in the art can realize that the units and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present invention.

Claims

1. A method for incrementally identifying multifunctional radar working modes based on deep feature expansion, characterized in that Including: Obtaining sample data of a radar to be measured, where the sample data includes samples of a new working mode category, and the new working mode category is different from the data type of the old working mode category stored in the incremental recognition model; Inputting the sample data into the incremental recognition model, and the incremental recognition model is trained according to an extended data set to obtain a trained incremental recognition model, wherein the incremental recognition model includes an extended feature extractor, a sample management module, and a working mode classification module. The extended feature extractor is used to extract features of the sample data in the extended data set. The working mode classification module is used to determine the working mode category to which the sample data belongs according to the distance between the features of the sample data and the average example vector of each working mode category. The sample management module is used to determine and store representative samples and / or delete redundant samples. The extended feature extractor is trained according to the extended data set under the iteration of a mixed loss function, and the dynamic adjustment coefficient of the mixed loss function decreases as the number of incremental recognition tasks increases. The extended data set includes the sample data and the sample data of the old working mode category; wherein the sample management module is specifically used for: sorting the working modes included in the working mode category in descending order according to the output result of the Softmax layer of the sample data in the extended data set to obtain a first sequence of the working mode category. The output result of the Softmax layer is determined according to the prediction results of the main classifier and the auxiliary classifier, and the output result of the Softmax layer is a probability sequence of the sample data belonging to the working mode category; determining the first N working modes in the first sequence as the representative samples of the working mode category, where N is a positive integer; storing the representative samples; Inputting the sample data into the trained incremental recognition model to obtain an identification result of the category to which the radar working mode belongs.

2. The method according to claim 1, characterized in that The mixed loss function satisfies the following formula: Among them, is the main loss function, is the auxiliary loss function, is the said hybrid loss function, is the number of the said new working mode categories, is the number of the said old working mode categories, is the said dynamic adjustment coefficient.

3. The method according to claim 2, wherein The main loss function satisfies the following formula: wherein, is the main loss function, is the working mode included in the sample data, is the prediction label corresponding to the working mode category, is the prediction result of the main classifier, is the number of samples in the extended dataset, and the prediction result of the main classifier includes the probabilities that the sample data in the extended dataset belong to each of the new working mode categories and the probabilities that the sample data in the extended dataset belong to each of the old working mode categories.

4. The method according to claim 3, wherein The incremental recognition model further includes a main classifier, and the main classifier is used to determine the prediction result of the main classifier according to the features of the sample data. The prediction result of the main classifier satisfies the following formula: ； wherein, is the working mode included in the sample data, is the prediction label corresponding to the working mode category, is the prediction result of the main classifier, is the feature of the sample data.

5. The method according to claim 3 or 4, characterized in that, The auxiliary loss function satisfies the following formula: Among them, is the auxiliary loss function, is the working mode included in the sample data, is the predicted label corresponding to the working mode category, is the prediction result of the auxiliary classifier, is the number of samples in the extended dataset. The prediction result of the auxiliary classifier includes the probability that the sample data in the extended dataset belongs to the old working mode category, and the probability that the sample data in the extended dataset belongs to each new working mode category respectively.

6. The method according to claim 5, wherein The incremental recognition model further includes an auxiliary classifier, and the auxiliary classifier is used to determine the prediction result of the auxiliary classifier according to the features of the sample data. The prediction result of the auxiliary classifier satisfies the following formula: Wherein, is the working mode included in the sample data, is the prediction label corresponding to the working mode category, is the prediction result of the auxiliary classifier, is the feature of the sample data.

7. The method according to claim 1, wherein The sample management module is further used for: Deleting redundant sample data, where the redundant sample data is the sample data of the last M working modes in the first sequence, and M is a positive integer.

8. The method according to claim 1, wherein The current extended feature extractor is determined according to the previous extended feature extractor and the basic feature extractor.

9. The method according to claim 1, wherein Before obtaining the sample data of the radar to be measured, the method further includes: Determine a pulse parameter sequence according to the original data, where the original data includes: the carrier frequency, pulse width, pulse amplitude, and pulse arrival time of the radar transmitted pulse, and the pulse parameter sequence includes: the carrier frequency, pulse width, pulse amplitude, and pulse repetition time of the radar transmitted pulse; Perform normalization processing on the pulse parameter sequence to obtain the sample data.

Images