Adaptive dynamic neuron electromagnetic target category incremental learning method and device

Abstract

This invention discloses an adaptive dynamic neuron-based method and apparatus for incremental learning of electromagnetic target categories, relating to the field of intelligent electromagnetic target classification technology. The invention first performs independent normalization on the multiple physical quantities of electromagnetic target samples; then, it uses old category samples to jointly train an autoencoder and a classifier, obtaining an original classification model with feature extraction and classification capabilities; during category increment, an adaptive aggregated dynamic neuron structure containing stable and incremental blocks is constructed in the classification network, a learnable scaling factor is introduced to allocate the weights of the two modules, and differential update amplitude is used to control parameter changes, while the autoencoder features complete incremental training. This invention can effectively retain old category classification knowledge, alleviate the catastrophic forgetting problem, reduce the computational load of incremental training, and is suitable for electromagnetic target classification scenarios with a small number of new category samples and frequent category increments.

Description

Technical Field

[0001] This invention relates to the field of intelligent classification technology for electromagnetic targets, specifically to an adaptive dynamic neuron-based method and apparatus for incremental learning of electromagnetic target categories. Background Technology

[0002] As the types of electromagnetic targets in open electromagnetic space continue to grow, the demand for continuous intelligent classification of electromagnetic targets is constantly increasing. Incremental learning of categories is the core technology for realizing the iterative classification model of electromagnetic targets in open environments.

[0003] Existing Class Incremental Learning (CIL) methods typically construct an example set based on existing data, augment the training by incorporating new category samples, and adjust the original classification network structure to accommodate the new category. In electromagnetic target classification scenarios, the sample sizes of the old and new categories are extremely imbalanced, with the new category usually having only a small number of samples. Furthermore, the feature distributions of electromagnetic signals across different categories vary widely, and category increments are frequent. Using existing methods for incremental training easily leads to overfitting to the new category and the curse of knowledge (forgetting) of the old category. Simultaneously, full-scale incremental training is computationally expensive, making it difficult to meet the engineering application requirements of frequent and rapid increments in electromagnetic target classification. Summary of the Invention

[0004] In view of the above problems, the present invention provides an electromagnetic target category incremental learning method and device with adaptive dynamic neurons. It retains old category feature knowledge through autoencoder pre-training and combines adaptive aggregation dynamic neuron structure to realize category increment under small sample conditions, alleviate the disaster forgetting problem, and reduce the computational load of incremental training.

[0005] In a first aspect, embodiments of the present invention provide an incremental learning method for electromagnetic target categories of adaptive dynamic neurons, the method comprising: Normalize the electromagnetic target sample data; Using the normalized old category electromagnetic target samples, the autoencoder and classifier are jointly trained to obtain the original classification model; When a new category of electromagnetic target sample is input, an adaptive aggregated dynamic neuron structure is constructed in the classification network of the original classification model. The new neurons are expanded and a scaling factor is introduced to allocate the weights of the original neurons and the new neurons, thus completing the incremental training of the categories.

[0006] In some embodiments, the normalization processing of the electromagnetic target sample data includes: For the three physical quantities of electromagnetic pulses—frequency, pulse width, and pulse interval—independent normalization calculations are performed for each type of physical quantity.

[0007] In some embodiments, the joint training of the autoencoder and classifier includes: The autoencoder adopts a symmetrical network structure between the encoder and decoder. The encoder outputs encoded features after being processed by convolutional layers, weight aggregation layers, and normalization layers in sequence. The decoder outputs reconstructed features with the same dimension as the input sample after being processed by convolutional layers, weight aggregation layers, and normalization layers in sequence. The classifier consists of multiple stacked one-dimensional convolutional layers and adaptive weight scaling layers, which perform classification calculations on the encoded features.

[0008] In some embodiments, the classifier further includes a Softmax layer; the encoded features are processed by a linear layer, an adaptive weight scaling layer, and a normalization layer, and then input into the Softmax layer to output the classification result.

[0009] In some embodiments, the adaptive aggregation dynamic neuron structure includes a stable block and an incremental block; The stable block corresponds to the original neurons in the original classification model, and retains the original weight parameters; The incremental block corresponds to the newly added neurons, which are used to adapt to the newly added class of samples.

[0010] In some embodiments, the expansion of the newly added neurons and the introduction of a scaling factor to allocate the weights of the original neurons and the newly added neurons includes: A learnable scaling factor is introduced, and the scaling factor is multiplied element-wise with the weights of the stable block and the incremental block to achieve weight allocation. The scaling factor satisfies the constraint that the summation is 1; The stable block uses a first update parameter to control the weight update magnitude, and the incremental block uses a second update parameter to control the weight update magnitude. The weight update magnitude corresponding to the first update parameter is less than the weight update magnitude corresponding to the second update parameter.

[0011] In some embodiments, during the process of completing the category incremental training, the parameters of the feature extraction module of the autoencoder in the original classification model are frozen, and the parameters of the adaptive aggregation dynamic neuron structure in the classification network are updated.

[0012] Secondly, embodiments of the present invention provide an electromagnetic target category incremental learning device for adaptive dynamic neurons, the device comprising: The normalization module is used to normalize the electromagnetic target sample data; The pre-training module is used to jointly train the autoencoder and classifier using normalized old category electromagnetic target samples to obtain the original classification model. The incremental training module is used to construct an adaptive aggregated dynamic neuron structure in the classification network of the original classification model when a new category of electromagnetic target sample is input. It expands the new neurons and introduces a scaling factor to allocate the weights of the original neurons and the new neurons, thereby completing the incremental training of the category.

[0013] Thirdly, embodiments of the present invention provide an electronic device, including a memory and a processor, wherein the memory stores program code executable on the processor, and when the program code is executed by the processor, it implements an adaptive dynamic neuron electromagnetic target category incremental learning method as described in any embodiment of the first aspect.

[0014] Fourthly, embodiments of the present invention provide a computer storage medium storing one or more programs, which can be executed by an electronic device as described in the third aspect to implement an adaptive dynamic neuron electromagnetic target category incremental learning method as described in any embodiment of the first aspect.

[0015] This invention utilizes an autoencoder to learn features from old samples, ensuring the retention and distillation of classification knowledge for old target categories and avoiding catastrophe forgetting. Simultaneously, after feature changes via the autoencoder, an adaptive aggregation network dynamic adjustment mechanism is designed within the classification network. This ensures that the main weights of feature extraction and the classification network simultaneously consider both the adaptability to new categories and the inheritance of knowledge from old categories, preventing drastic changes in the classification network caused by a small number of new category samples. This enables incremental learning of complex electromagnetic targets with frequently added categories and significant differences in sample size between old and new categories using small samples. Under relatively low computational requirements, the incrementally updated model ensures minimal degradation in classification performance for old categories while exhibiting excellent adaptability to new categories.

[0016] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0017] The invention will now be described in more detail with reference to embodiments and the accompanying drawings.

[0018] Figure 1 This diagram illustrates an exemplary method for incremental learning of electromagnetic target categories using adaptive dynamic neurons, as proposed in one embodiment of the present invention. Figure 2 A schematic diagram of an exemplary adaptive aggregation network structure proposed in one embodiment of the present invention is shown; Figure 3A schematic diagram of an exemplary autoencoder and classification joint network is shown in one embodiment of the present invention; Figure 4 This diagram illustrates a confusion matrix for 12 types of old-class data classification provided in an embodiment of the present invention. Figure 5 A schematic diagram of four new types of data classification confusion matrices provided in an embodiment of the present invention is shown; Figure 6 This diagram illustrates a structural block diagram of an exemplary adaptive dynamic neuron electromagnetic target category incremental learning device proposed in one embodiment of the present invention; Figure 7 A structural block diagram of an electronic device for performing an electromagnetic target category incremental learning method for an adaptive dynamic neuron according to an embodiment of the present invention is shown. Figure 8 A structural block diagram of a computer-readable storage medium for storing or carrying an electromagnetic target category incremental learning method for an adaptive dynamic neuron according to an embodiment of the present invention is shown. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0020] The adaptive dynamic neuron-based incremental learning method for electromagnetic target categories provided in this embodiment can be applied to scenarios such as radar signal recognition and electromagnetic spectrum monitoring, and is suitable for classification requirements in open environments where electromagnetic target categories are frequently added and the sample size of new categories is small. The specific implementation steps of the method are described in detail below with reference to the accompanying drawings.

[0021] This embodiment provides an adaptive dynamic neuron-based electromagnetic target category incremental learning method, which can be deployed on low-altitude aircraft ground stations or airborne safety control systems. The overall execution process is divided into two stages: offline configuration and online execution. The following is a detailed description with reference to the accompanying drawings.

[0022] Please see Figure 1 As shown, an adaptive dynamic neuron electromagnetic target category incremental learning method includes the following steps S110 to S130, applied to... Figure 3 The overall network structure shown is as follows: S110: Electromagnetic target sample normalization processing.

[0023] In this embodiment, the electromagnetic target sample consists of radar signal parameters of multiple different physical quantities. To eliminate numerical differences between variables with different dimensions, normalization calculations are performed separately for each type of physical quantity. In this embodiment, three characteristic physical quantities are selected: electromagnetic pulse frequency (rf), pulse width (pw), and pulse interval (pri). Independent normalization calculations are performed for each type of physical quantity. The normalization calculation is based on the average value of the physical quantity in the corresponding category of sample set, eliminating the influence of dimension and numerical range on subsequent feature extraction.

[0024] Since the data sample is described by multiple different physical quantities of radar signals, each variable is normalized separately in order to eliminate the differences between variables with different dimensions.

[0025] For example, this embodiment mainly uses three quantities of electromagnetic pulses: frequency rf, pulse width pw, and pulse interval pri. For each sample... All three quantities are normalized using the following expression:

[0026] In the formula, , , This represents the average value of the variable in each category of the sample set.

[0027] S120: Using the normalized old category electromagnetic target samples, the autoencoder and classifier are jointly trained to obtain the original classification model.

[0028] In this embodiment, let the normalized sample vector of the known K-type electromagnetic signal be... The original classification model is obtained by jointly training an autoencoder and a classifier. The model expression is as follows:

[0029] In the formula, and These represent the parameters of the learnable autoencoder and classifier, respectively.

[0030] The classifier consists of multiple stacked one-dimensional convolutional layers and custom adaptive weight scaling layers. The autoencoder employs a symmetrical network structure between the encoder and decoder ends to pre-train the intrinsic feature associations of impulse signals. The encoder end sequentially includes convolutional layers, weight aggregation layers, and normalization layers, and is input with samples from the old class. Output after encoding m 3D encoded features, the encoding process is represented as:

[0031] In the formula, , and represents the trainable weights of the convolutional, weight aggregation, and normalization layers of the classification joint network, respectively; the normalization layer in the autoencoder uses the BatchNorm layer.

[0032] The decoding and encoding ends have symmetrical structures, with convolutional layers, weight aggregation layers, and normalization layers arranged sequentially. The encoded feature Z is processed by the decoding end to output the reconstructed feature. The dimensions of the reconstructed features are the same as those of the input samples. Since the dimensions are consistent, the decoding process is represented as follows:

[0033] The classifier consists of multiple stacked one-dimensional convolutional layers and adaptive weight scaling layers, which perform classification calculations on the encoded feature Z. For the old-class K-class classification task, the encoded feature Z undergoes multiple channel scaling processes, successively passing through a linear layer, an adaptive weight scaling layer, and a normalization layer, before being input into the Softmax layer to output the K-class classification result. The classification process is represented as follows:

[0034] in, This represents the trainable weights of the normalized layer.

[0035] During joint training, the total loss function is the weighted sum of the reconstruction loss of the autoencoder and the classification loss of the classifier. All trainable parameters of the autoencoder and the classifier are updated synchronously through backpropagation, so that the original model has both robust feature extraction capability and old category classification capability.

[0036] S130: When a new category of electromagnetic target sample is input, an adaptive aggregated dynamic neuron structure is constructed in the classification network of the original classification model. The new neurons are expanded and a scaling factor is introduced to allocate the weights of the original neurons and the new neurons, thus completing the incremental training of the categories.

[0037] In this embodiment, when receiving new N classes of data to complete incremental training, the adaptive weight aggregation network needs to expand its neurons, accepting new class data while retaining the weights of the old neurons, and realizing a weighted redistribution of weights, as referenced. Figure 2 As shown.

[0038] When the original classification model receives N new electromagnetic target samples for incremental training, an adaptive aggregation dynamic neuron structure is constructed in the adaptive weight scaling layer of the classification network. This structure consists of two parts: a stable block and an incremental block. The incremental block inherits the weight parameters obtained from training the old classes from the original neurons in the original classification model, serving as a carrier of the old class classification knowledge; The newly added neurons are used to learn new category features and adapt to the new categories.

[0039] During the old category training phase, the network does not add incremental blocks, and the input features of the intermediate layers... The output features are obtained after stable block weight transformation. ,Right now:

[0040] Learnable weights Transform into It is used for adaptive feature distribution.

[0041] During the incremental learning training phase, stable blocks With incremental blocks Between them via a learnable scaling factor Weights are assigned. The scaling factor is multiplied element-wise by the weights of the stable block and the incremental block, transforming the input features. The intermediate layer outputs the features. Represented as:

[0042] in, Element-wise multiplication, scaling factor For learnable parameters, to ensure In mathematical consistency, the sum of all elements of the scaling factor is 1, a constraint. .

[0043] During incremental training, different update parameters are used to control the weight update magnitude for stable blocks and incremental blocks: the stable block uses the first update parameter to control the weight to update slowly with a small magnitude, so as to retain the old category classification knowledge to the greatest extent; the incremental block uses the second update parameter to control the weight to update quickly with a larger magnitude, so as to quickly fit the new category data features; the weight update magnitude corresponding to the first update parameter is smaller than the weight update magnitude corresponding to the second update parameter.

[0044] Furthermore, during incremental training, the parameters of the feature extraction module of the autoencoder in the original classification model are frozen, and only the adaptive aggregation dynamic neuron structure and classification output layer parameters in the classification network are updated. While retaining the general feature extraction capability of the old categories, the number of parameters to be updated during incremental training is greatly reduced, thus reducing computational overhead.

[0045] In summary, this invention designs an architecture that combines feature extraction and classification networks, and utilizes a large number of existing old category data samples for pre-training to form robust feature extraction and classification networks. This avoids catastrophic forgetting caused by drastic changes in model parameters after the addition of small samples of new category data. At the same time, while ensuring the classification performance of new and old targets, it effectively reduces the number of samples required for incremental training, thereby improving the adaptability of the model for engineering applications.

[0046] In addition, an adaptive aggregation dynamic neuron was designed in the classification model to dynamically adjust the balance between old network modules and new knowledge according to the new category of electromagnetic target samples, so that the main weights of the new classification network can simultaneously take into account the adaptability of the new class and the knowledge inheritance of the old class.

[0047] To verify the effectiveness of this method, comparative experiments were conducted using simulated radar signal samples from 16 categories. The samples simulated the signal characteristics of different types of radar under complex scenarios such as rapid movement, switching of working modes, switching of scanning cycles, and changes in signal strength and quality.

[0048] Experimental setup: Twelve radar target categories were randomly selected as the old categories, and the remaining four were selected as the new categories. Each category contained 2000 samples, each sample being 150,000 characters long, and was segmented into sample fragments of length 80 and overlap length 10, resulting in a total of approximately 24,000 samples for the old categories. The initial learning rate for model training was set to 0.0001, decreasing by 50% with each training epoch and validation set accuracy step. The pre-training phase consisted of 40 epochs. During the incremental training phase, the autoencoder feature extraction module was frozen, and only the parameters of the adaptive aggregation dynamic neuron structure of the classifier were updated.

[0049] Figure 4 This is the classification confusion matrix of the old category samples after incremental training. The average classification accuracy of the old category is 98.6%, which is very close to the classification performance of the pre-trained model. The degree of forgetting of old category knowledge is negligible. Figure 5 The classification confusion matrix for new class samples after incremental training shows that the classification accuracy of the four new classes is 90.05%, 99.87%, 100%, and 100%, respectively, demonstrating excellent adaptation of the new classes under small sample conditions.

[0050] To illustrate the effectiveness of using dynamic neurons in incremental training, Table 1 below shows the impact of whether or not dynamic neurons are used during the training process.

[0051] Comparison table of the impact of dynamic neuron mechanisms on incremental learning performance:

[0052] Table 1 Table 1 shows that after implementation, the accuracy of the old classes improved by 0.43%, 0.11%, and 1.15%, respectively, while the accuracy of the new classes improved by 0.31%, 0.84%, and 1.43%, respectively. The dynamic neural network mechanism has a greater impact when the proportion of new classes is high. In scenarios with a large volume of new classes, the dynamic neural network mechanism can better improve network performance and adapt to more complex incremental scenarios. As shown in Table 1 above, after introducing the adaptive aggregation dynamic neuron structure, the classification accuracy of both the old and new categories improved under different sample proportion scenarios; and the performance improvement was greater in scenarios with a higher proportion of new category samples, indicating that the dynamic neuron mechanism can better adapt to complex incremental scenarios and take into account the classification performance of both the old and new categories.

[0053] Please see Figure 6 , Figure 6 A structural block diagram of an adaptive dynamic neuron electromagnetic target category incremental learning device 300 provided by the present invention includes: a normalization module 310, a pre-training module 320, and an incremental training module 330, wherein: Normalization module 310 is used to normalize electromagnetic target sample data; The pre-training module 320 is used to jointly train the autoencoder and classifier using normalized old category electromagnetic target samples to obtain the original classification model. The incremental training module 330 is used to construct an adaptive aggregated dynamic neuron structure in the classification network of the original classification model when a new category of electromagnetic target sample is input. It expands the new neurons and introduces a scaling factor to allocate the weights of the original neurons and the new neurons, thus completing the incremental training of the category.

[0054] It should be noted that the present invention may also include other modules. The device embodiments in the present invention correspond to the foregoing method embodiments. The specific principles of the module embodiments in the device can be found in the foregoing method embodiments, and will not be repeated here.

[0055] In the several embodiments provided in this example, the coupling between modules can be electrical, mechanical, or other forms of coupling.

[0056] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0057] Please see Figure 7 , Figure 7The present invention provides a structural block diagram of an electronic device 200 capable of executing the above-described electromagnetic target category incremental learning method for an adaptive dynamic neuron. The electronic device 200 may be a tablet computer, a computer, or a portable computer.

[0058] The electronic device 200 also includes a processor 202 and a memory 204. The memory 204 stores programs that can execute the contents of the foregoing embodiments, and the processor 202 can execute the programs stored in the memory 204.

[0059] The processor 202 may include one or more cores for data processing and message matrix units. The processor 202 connects to various parts within the electronic device 200 using various interfaces and lines, and performs various functions and processes data by running or executing instructions, programs, code sets, or instruction sets stored in the memory 204, and by calling data stored in the memory 204. Optionally, the processor 202 may be implemented using at least one hardware form of Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). The processor 202 may integrate one or more of the following: Central Processing Unit (CPU), Graphics Processing Unit (GPU), and modem / decoder. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the displayed content; and the modem / decoder handles wireless communication. It is understood that the modem / decoder may also be implemented separately as a communication chip, without being integrated into the processor.

[0060] Memory 204 may include random access memory (RAM) or read-only memory (ROM). Memory 204 can be used to store instructions, programs, code, code sets, or instruction sets. Memory 204 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for implementing at least one function (e.g., instructions for a user to obtain random numbers), instructions for implementing the various method embodiments described below, etc. The data storage area may also store data (e.g., random numbers) created by the terminal during use.

[0061] Electronic device 200 may also include a network module and a screen. The network module is used to receive and transmit electromagnetic waves, converting electromagnetic waves into electrical signals, thereby enabling communication with communication networks or other devices, such as audio playback devices. The network module may include various existing circuit elements used to perform these functions, such as antennas, radio frequency transceivers, digital signal processors, encryption / decryption chips, SIM cards, memory, etc. The network module can communicate with various networks such as the Internet, corporate intranets, and wireless networks, or communicate with other devices via wireless networks. The aforementioned wireless networks may include cellular telephone networks, wireless local area networks, or metropolitan area networks. The screen can display interface content and facilitate data interaction.

[0062] Please refer to Figure 8 , Figure 8 A structural block diagram of a computer-readable storage medium provided in an embodiment of the present invention is shown. The computer-readable storage medium 400 stores program code 410, which can be called by a processor to execute the methods described in the above method embodiments.

[0063] The computer-readable storage medium 400 may be an electronic memory such as flash memory, EEPROM (Electrically Erasable Programmable Read-Only Memory), EPROM, hard disk, or ROM. Optionally, the computer-readable storage medium includes a non-transitory computer-readable storage medium. The computer-readable storage medium 400 has storage space for program code 410 that performs any of the method steps described above. This program code 410 can be read from or written to one or more computer program products. The program code 410 may be compressed, for example, in a suitable form.

[0064] This invention also provides a computer program product or computer program that includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform an adaptive dynamic neuron electromagnetic target category incremental learning method described in the various optional implementations above.

[0065] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An adaptive dynamic neuron-based incremental learning method for electromagnetic target categories, characterized in that, The method includes: Normalize the electromagnetic target sample data; Using the normalized old category electromagnetic target samples, the autoencoder and classifier are jointly trained to obtain the original classification model; When a new category of electromagnetic target sample is input, an adaptive aggregated dynamic neuron structure is constructed in the classification network of the original classification model. The new neurons are expanded and a scaling factor is introduced to allocate the weights of the original neurons and the new neurons, thus completing the incremental training of the categories.

2. The adaptive dynamic neuron electromagnetic target category incremental learning method according to claim 1, characterized in that, The normalization process for the electromagnetic target sample data includes: For the three physical quantities of electromagnetic pulses—frequency, pulse width, and pulse interval—independent normalization calculations are performed for each type of physical quantity.

3. The electromagnetic target category incremental learning method for adaptive dynamic neurons according to claim 1, characterized in that, The jointly trained autoencoder and classifier include: The autoencoder adopts a symmetrical network structure between the encoder and decoder. The encoder outputs encoded features after being processed by convolutional layers, weight aggregation layers, and normalization layers in sequence. The decoder outputs reconstructed features with the same dimension as the input sample after being processed by convolutional layers, weight aggregation layers, and normalization layers in sequence. The classifier consists of multiple stacked one-dimensional convolutional layers and adaptive weight scaling layers, which perform classification calculations on the encoded features.

4. The electromagnetic target category incremental learning method for adaptive dynamic neurons according to claim 3, characterized in that, The classifier also includes a Softmax layer; The encoded features are processed by a linear layer, an adaptive weight scaling layer, and a normalization layer, and then input into a Softmax layer to output the classification result.

5. The electromagnetic target category incremental learning method for adaptive dynamic neurons according to claim 1, characterized in that, The adaptive aggregation dynamic neuron structure includes a stable block and an incremental block; The stable block corresponds to the original neurons in the original classification model, and retains the original weight parameters; The incremental block corresponds to the newly added neurons, which are used to adapt to the newly added class of samples.

6. The method for incremental learning of electromagnetic target categories using adaptive dynamic neurons according to claim 5, characterized in that, The process of expanding and adding new neurons and introducing a scaling factor to allocate weights between existing and new neurons includes: A learnable scaling factor is introduced, and the scaling factor is multiplied element-wise with the weights of the stable block and the incremental block to achieve weight allocation. The scaling factor satisfies the constraint that the summation is 1; The stable block uses a first update parameter to control the weight update magnitude, and the incremental block uses a second update parameter to control the weight update magnitude. The weight update magnitude corresponding to the first update parameter is less than the weight update magnitude corresponding to the second update parameter.

7. The electromagnetic target category incremental learning method for adaptive dynamic neurons according to claim 1, characterized in that, During the process of completing the category incremental training, the parameters of the feature extraction module of the autoencoder in the original classification model are frozen, and the parameters of the adaptive aggregation dynamic neuron structure in the classification network are updated.

8. An adaptive dynamic neuron electromagnetic target category incremental learning device, characterized in that, include: The normalization module is used to normalize the electromagnetic target sample data; The pre-training module is used to jointly train the autoencoder and classifier using normalized old category electromagnetic target samples to obtain the original classification model. The incremental training module is used to construct an adaptive aggregated dynamic neuron structure in the classification network of the original classification model when a new category of electromagnetic target sample is input. It expands the new neurons and introduces a scaling factor to allocate the weights of the original neurons and the new neurons, thereby completing the incremental training of the category.

9. An electronic device, characterized in that, The electronic device includes a memory and a processor. The memory stores program code that can run on the processor. When the program code is executed by the processor, it implements an electromagnetic target category incremental learning method for an adaptive dynamic neuron as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program code that can be invoked by one or more processors to execute an electromagnetic target category incremental learning method for adaptive dynamic neurons as described in any one of claims 1-7.

Images