Multi-station radar signal sorting method based on federal learning
By combining federated learning and LSTM, the problems of accuracy and data sharing in multi-station radar signal sorting under complex electromagnetic environments are solved, achieving efficient and accurate multi-station radar signal sorting, which is suitable for real-time reconnaissance and electronic warfare missions in complex battlefield environments.
Patent Information
- Application Number
- CN202511373198.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-24
- Publication Date
- 2026-01-09
AI Technical Summary
Existing multi-station radar signal sorting methods have low accuracy in complex electromagnetic environments, making it difficult to effectively utilize the potential of multi-station collaboration, and there are also problems with data sharing and transmission.
A multi-station radar signal sorting method based on federated learning is adopted. By training local models at each reconnaissance station and sharing model update parameters, and combining Long Short-Term Memory (LSTM) network for temporal modeling, a multi-station collaborative sorting framework is constructed. The temporal and fuzzy characteristics of pulse signals are utilized to achieve pulse sorting.
It improves the accuracy and robustness of radar signal sorting, reduces data transmission risks, adapts to complex battlefield environments, and has efficient and accurate pulse sorting capabilities.
Smart Images
Figure CN121302158A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of electronic reconnaissance and situation awareness, and particularly relates to a multi-station radar signal sorting method based on federated learning. BACKGROUND
[0002] Radar signal sorting is a key technology in electronic reconnaissance and battlefield situation awareness, and its accuracy directly affects subsequent combat decision-making and planning. In a complex electromagnetic environment, the reconnaissance system intercepts radar pulse streams from multiple radar emitters interlaced together. Accurate sorting of different types of radar signals is a prerequisite for subsequent processing such as emitter identification and situation awareness.
[0003] Electronic reconnaissance systems can be divided into single-station reconnaissance systems and multi-station reconnaissance systems according to the number of reconnaissance stations. Single-station reconnaissance systems are limited by external factors such as line of sight, target obstruction, and multipath effects, and cannot completely cover the monitoring area. It is also difficult to effectively handle signal distortion and interference in complex electromagnetic environments. Unlike this, multi-station reconnaissance systems take advantage of the spatial diversity characteristics of the equipment, and through cooperative observation and data sharing, can significantly enhance the sorting performance.
[0004] Existing multi-station sorting methods mainly use the Time Difference of Arrival (TDOA) information in the multi-station system to achieve pulse matching by introducing the time difference of the same pulse at multiple stations, thereby completing the pulse sorting between multiple stations. Typical methods include time difference histogram, extended time difference histogram, and TDOA clustering. Among them, the sorting method based on time difference histogram has good sorting effect for low repetition frequency pulses, but due to the time difference ambiguity characteristic, it is difficult to effectively sort high repetition frequency pulses. Moreover, this method relies on the hyperparameters in the histogram construction, making it difficult to achieve stable and effective sorting in complex electromagnetic environments. The TDOA clustering method is similar to the single-station clustering method, although it can handle time difference information, most clustering algorithms require iteration or search, resulting in high time complexity and making it difficult to apply to actual sorting tasks.
[0005] Intelligent algorithms based on deep learning have achieved excellent performance in certain sorting and identification tasks compared to traditional algorithms. However, existing intelligent algorithms are mostly based on centralized learning processing mode for single stations, with little research on multi-station scenarios. The main problems are data barriers between multiple stations, difficulties in data transmission, and strong information confidentiality. In practical applications, due to different security levels and data confidentiality requirements between different reconnaissance stations, data sharing and transmission are often strictly limited, making it difficult to sort multiple stations based on centralized learning methods. In addition, the transmission process of cross-station data may be affected by network bandwidth and time delay, etc., resulting in constraints on the real-time performance and accuracy of the sorting process. SUMMARY
[0006] The purpose of the present application is to provide a multi-station radar signal sorting method based on federated learning, so as to overcome the problem that the existing multi-station radar signal sorting method fails to fully utilize the potential of multi-station cooperation, and thus has a low sorting accuracy in a complex electromagnetic environment. Through multi-station cooperative learning and information sharing, the sorting precision and robustness are improved.
[0007] In order to achieve the above-mentioned task, the present application adopts the following technical solutions: A multi-station radar signal sorting method based on federated learning, comprising: For each reconnaissance station participating in reconnaissance, a radar pulse parameter sequence of the reconnaissance station is obtained to form a pulse sorting data set of the reconnaissance station; the time of arrival sequence in the pulse sorting data set is replaced by a time difference of arrival sequence, and each parameter sequence is standardized to obtain a pulse sorting data set; A group of parameters corresponding to each pulse in the pulse sorting data set is taken as a sorting sample, and a sorting category is taken as a label of the sorting sample; the sorting category includes a radar category and a stray pulse; A federated sorting processing framework of multiple reconnaissance stations is constructed and initialized; a local model is allocated to each reconnaissance station, and the local model is trained and tested by using the pulse sorting data set; during the training process, the class weight of each sorting category is calculated by dynamic weighting, and a weighted loss function during the training process is constructed based on this; The parameters of the local model of each reconnaissance station are uploaded to a central server; the central server aggregates the parameters of the models of each reconnaissance station by using a parameter averaging strategy to form the parameters of a new global model; the parameters of the global model are synchronized and distributed to the local models of each reconnaissance station, and the local models continue the next round of local training based on the distributed parameters; For a group of radar pulse parameter sequences to be identified, the time of arrival is replaced by the time difference of arrival and input into the trained local model to obtain a sorting result.
[0008] Further, the parameter types in the parameter sequence include the time of arrival, the carrier frequency, the pulse width, and the pulse amplitude.
[0009] Further, the sorting category includes a radar category and a stray pulse.
[0010] Further, the local model adopts an LSTM model, and the structure of the LSTM model is: An input layer Input is used to receive a sorting sample at each time step; A fully connected layer Linear is used to perform linear transformation and feature dimension mapping on the sorting sample to obtain a 256-dimensional feature; a layer normalization layer LayerNorm, an activation function layer Tanh, and a dropout layer Dropout are arranged after the fully connected layer; The first LSTM layer has an input feature dimension of 256, 256 hidden units, and is unidirectional. The second to fourth LSTM layers each have an input dimension of 256 and a hidden unit count of 256, and are unidirectional; each LSTM layer is followed by a Dropout layer with a preset probability. The bidirectional layer concatenates the output of the unidirectional LSTM with the output of the inverse LSTM to obtain the output of the bidirectional LSTM layer. The fully connected Linear layer performs a linear transformation on the output of the bidirectional LSTM layer, mapping the feature dimension from 512 to 128; after the fully connected Linear layer, the ReLU activation function layer and the Dropout layer are set. The fully connected Linear layer performs a linear transformation on the mapped features, mapping the feature dimension from 128 to 6, to obtain the final classification result; The Output layer outputs the classification probability distribution for each time step, which is used to predict the label.
[0011] Furthermore, the class weights for each sorting category are calculated through dynamic weighting, and a weighted loss function is constructed based on this for the training process, including:
[0012] in, For category weights, It is the number of sorted samples in the training set. It is the first The number of sorted samples for each sorting category in the current batch. It is the number of categories sorted in the training set;
[0013] in, This represents the weighted loss function. Labels for sorted samples, The predicted label for the sorted sample is the first... The probability of each sorting category. Category weights The normalized result.
[0014] Furthermore, the central server uses a parameter averaging strategy to aggregate the parameters of the LSTM models of each reconnaissance station, forming the parameters of a new global model, including: From all the local LSTM models of the reconnaissance sites, all parameters of an arbitrarily selected local LSTM model are taken as an initial template; each parameter in the initial template is sequentially added to the tensor of the corresponding parameter in the local LSTM model of the other reconnaissance sites; after the addition of each parameter is completed, the sum is divided by the number of reconnaissance sites to obtain the average value of the parameter, which is used to update the parameters of the aggregated global LSTM model.
[0015] A terminal device comprising a processor, a memory, and a computer program stored in the memory; when the processor executes the computer program, the federated learning-based multi-station radar signal sorting method is implemented.
[0016] A computer-readable storage medium, the medium storing a computer program; when the computer program is executed by a processor, the federated learning-based multi-station radar signal sorting method is implemented.
[0017] Compared with the prior art, the present application has the following technical features: The present application provides a radar pulse sorting method based on multi-station cooperation and federated learning, which can realize joint intelligent processing of multi-station pulse data by local model training and global model aggregation while maintaining the privacy of each site data. The method utilizes the timing and ambiguity characteristics of pulse signals, combined with the powerful timing modeling capability of long short-term memory network (LSTM), effectively improving the accuracy and robustness of pulse sorting. By training locally at each site and fusing the model parameters of each site in a federated learning manner, the risk of data transmission and storage is significantly reduced, while the performance of the model is continuously optimized.
[0018] Compared with the prior art, the present application can realize efficient and accurate pulse sorting in a distributed environment, adapt to multi-source heterogeneous radar environment, has the advantages of fast training speed, low resource consumption, good scalability, strong fault tolerance to pulse data disturbance and mismatch, etc., and is particularly suitable for real-time reconnaissance and electronic countermeasure tasks in complex battlefield environments. BRIEF DESCRIPTION OF DRAWINGS
[0019] Figure 1 The flowchart for local model training in the method of the present application is shown in the figure; Figure 2 The distribution diagram of the carrier frequency parameter sequence of the simulation data is shown in the figure; Figure 3 The distribution diagram of the pulse width parameter sequence of the simulation data is shown in the figure; Figure 4 The distribution diagram of the pulse amplitude parameter sequence of the simulation data is shown in the figure; Figure 5 The loss function iteration diagram is shown in the figure; Figure 6An iteration schematic diagram for testing data accuracy of the present application; Figure 7 A confusion matrix for sorting precision of the present application; Figure 8 A confusion matrix for sorting recall rate of the present application; Figure 9 A confusion matrix for local LSTM sorting precision; Figure 10 A confusion matrix for local LSTM sorting recall rate. DETAILED DESCRIPTION
[0020] The present application proposes a method combining federated learning and long short-term memory network (LSTM). Federated learning avoids direct data transmission by training models locally at each site and only sharing model update parameters, thus overcoming data privacy and transmission difficulties. Each site only needs to share the model weights generated during the training process, without having to transmit raw data, which ensures data security and reduces the burden of cross-site data transmission. At the same time, the LSTM network can effectively handle the timing characteristics of pulse signals, capture dependencies over long time spans, and improve the accuracy and robustness of multi-site sorting. By combining the advantages of federated learning and LSTM, the method can work collaboratively at multiple sites, preserving local features at each site while optimizing the global model through the federated learning framework, thus improving overall sorting performance. In addition, due to the advantages of LSTM in time series modeling, it can effectively deal with signal interference and time-varying characteristics in complex electromagnetic environments, ensuring the model's adaptability in dynamic environments. Therefore, the multi-site sorting method of the present application combining federated learning and LSTM not only breaks down data barriers and solves information sharing difficulties, but also fully leverages the advantages of time series modeling, providing a more efficient and robust solution for multi-site radar signal sorting in complex electromagnetic environments.
[0021] The present application proposes a method based on federated learning and long short-term memory network (LSTM), which deeply analyzes the characteristics of each parameter of pulse description word (PDW). Specifically, the present application introduces a federated learning framework to fully utilize distributed data for collaborative training, effectively avoiding data privacy issues and improving the model's generalization ability. At the same time, the LSTM network model is used to model the timing characteristics of pulse signals, which can capture dependencies over long time spans, thus improving sorting performance in high-repetition pulse and complex signal environments. By combining the advantages of federated learning and LSTM, the present application not only improves sorting accuracy, but also enhances the model's ability to recognize complex modulation, making it adaptable to dynamic changes and signal interference in a multi-site collaborative environment, thus providing more accurate data support for subsequent emitter identification, situation awareness analysis, etc.
[0022] The method of the present invention includes the following steps: Step 1: For each reconnaissance station participating in the reconnaissance, extract the four parameter sequences of radar pulses: Time of Arrival (TOA), Carrier Frequency (CF), Pulse Width (PW), and Pulse Amplitude (PA), to form the pulse sorting dataset for that reconnaissance station.
[0023] Step 2: Calculate the corresponding Time Difference of Arrival (DTOA) for the TOA parameters in the pulse dataset of each reconnaissance station.
[0024]
[0025] in, Represents the first in the pulse dataset One TOA parameter; Indicates the first The arrival time difference corresponding to each TOA parameter; due to the calculated arrival time difference The number of arrival times is less than the number of TOA parameters in the pulse dataset, therefore 0 is taken as the first arrival time difference. This ensures that the arrival time difference corresponds one-to-one with the TOA parameter.
[0026] Step 3: Replace the corresponding TOA parameter in the pulse dataset with the Time Difference of Arrival (DTOA) to obtain a new pulse dataset; standardize the sequence of each parameter (DTOA, CF, PW, PA) in the new pulse dataset to obtain the processed pulse sorting dataset.
[0027] Standardization eliminates the influence of different units of measurement and is a key step in intelligent algorithm processing; the formula for maximum and minimum standardization is as follows.
[0028]
[0029] Taking the Time Difference of Arrival (DTOA) sequence as an example, This represents the DTOA parameter in the DTOA sequence. for The result after standardization; These represent the minimum and maximum values of the DTOA sequence, respectively.
[0030] Step 4, the pulse sorting dataset is divided into a training set and a test set in a ratio of 5:5, and is labeled; wherein a group of DTOA, CF, PW, and PA parameters corresponding to the same time for each pulse in the pulse sorting dataset is taken as a sorting sample, and the sorting category is taken as the label of the sorting sample; the sorting category includes a radar category and a stray pulse.
[0031] Step 5, a federated sorting processing framework of multiple reconnaissance sites is constructed and initialized; a local model copy is assigned to each reconnaissance site, and the same initial parameter setting is adopted. The local model uses an LSTM model to adapt to the timing characteristics and ambiguity of the pulse sequence; wherein the structure of the LSTM model is as follows: an input layer Input for receiving a sorting sample at each time step; a fully connected layer Linear for linear transformation and feature dimension mapping of the sorting sample to obtain a 256-dimensional feature; a layer normalization layer LayerNorm, an activation function layer Tanh, and a dropout layer Dropout are arranged after the fully connected layer, wherein: the layer normalization layer is used to normalize the 256-dimensional feature after dimension mapping of the sorting sample; the activation function layer Tanh maps the normalized feature to the interval (-1, 1); the dropout layer Dropout is used to discard the output of part of the neurons of the fully connected layer with a preset probability; a first LSTM layer with an input feature dimension of 256, a hidden unit number of 256, and unidirectional; a second LSTM layer to a fourth LSTM layer, each with an input dimension of 256, a hidden unit number of 256, and unidirectional; each LSTM layer has a dropout layer Dropout with a preset probability after it, which is used to discard part of the output neurons; a bidirectional layer Bidirectional for splicing the unidirectional LSTM output and the reverse LSTM output to obtain the output of the bidirectional LSTM layer; a fully connected layer Linear for linear transformation of the output of the bidirectional LSTM layer to map the feature dimension from 512 to 128; an activation function layer ReLU and a dropout layer Dropout are arranged after the fully connected layer Linear; wherein the activation function layer ReLU is used to map the linearly transformed feature to a non-negative interval; the dropout layer Dropout discards the output of part of the neurons of the fully connected layer with a preset probability; a fully connected layer Linear for linear transformation of the mapped feature to map the feature dimension from 128 to 6 to obtain a final classification result; an output layer Output for outputting a classification probability distribution at each time step, which is used to predict the label.
[0032] Table 1 LSTM model parameter table
[0033] Step 6, each reconnaissance site trains a local LSTM model using the respective training set, and the local LSTM model updates the weights locally without uploading the training data to protect the privacy and security of the data; after a certain number of local training rounds, the number of sorting categories is imbalanced due to the different PRI of each radar, which makes the model insufficient in distinguishing the sorting categories with fewer sorting samples; therefore, the application improves the model's ability to distinguish sorting categories with fewer sorting samples through dynamic weighting:
[0034] wherein, is the category weight, is the number of sorting samples in the training set, is the number of sorting samples of the i-th sorting category in the current batch, is the number of sorting samples of the i-th sorting category in the current batch, is the number of sorting categories in the training set.
[0035] In order to further ensure the stability of the training process, the category weight is normalized as follows:
[0036] Therefore, the weighted loss function of the final LSTM model during training is as follows:
[0037] wherein, is the label of the sorting sample, is the predicted label of the sorting sample, and is the probability of the i-th sorting category.
[0038] Step 7, each reconnaissance site uploads the parameters of the local LSTM model to the central server; the server uses the parameter averaging strategy to aggregate the parameters of the LSTM model of each reconnaissance site to form the parameters of the new global model. First, from all the local LSTM models of all reconnaissance sites, randomly select all parameters (such as weights, biases, etc. of each layer of the network) of a local LSTM model as an initial template, and make a deep copy to prevent side effects on the initial template during the subsequent aggregation process; traverse each parameter in the initial template, and add the corresponding parameter tensors in the local LSTM models of the other reconnaissance sites to the parameters in the initial template one by one; after completing the addition of each parameter, divide the sum by the number of reconnaissance sites to obtain the average value of the parameter, which is used to update the parameters of the aggregated global LSTM model.
[0039] Step 8, the parameters of the global LSTM model are synchronized and issued to the local LSTM model of each reconnaissance site, and the local LSTM model continues the next round of local training based on the issued parameters; steps 6 to 8 are executed in a loop until a preset training round or a convergence condition is reached.
[0040] Step 9, the final local LSTM model evaluation is performed using the test set data, and the sorting accuracy, recall rate and F1 score of each sorting category are output; the test results of each reconnaissance site are comprehensively analyzed to evaluate the overall sorting performance.
[0041] Step 10, save the trained local LSTM model; in actual application, after obtaining a new set of radar pulse sequences, calculate the DTOA according to the method of step 2; input the DTOA, CF, PW and PA parameters of each pulse into the trained local LSTM model to obtain the sorting result.
[0042] Embodiment: The data of the embodiment of the application comes from simulation, and the simulation process is close to the complex pulse modulation mode in the actual environment. Five types of radar radiation sources are simulated, and 10% of stray pulses are added. The simulation data is shown in Table 2: Table 2 Simulation data parameter table
[0043] The distribution diagrams of the CF, PW and PA parameters of the pulses received by one of the stations are shown in Figures 2, 3 and 4; the proportion of stray pulses is 10%, the PRI jitter range is ±30%, the parameter physical jitter range is ±5%, and the CF agile conforms to a uniform distribution. The pulse loss probability of each station is not more than 50%, the time distribution of the lost pulses is independent of each other, and the arrival time of the received pulses at each station is determined by the relative positions of the stations.
[0044] This embodiment is implemented based on three reconnaissance sites, and the Time of Arrival (TOA), Carrier Frequency (CF), Pulse Width (PW) and Pulse Amplitude (PA) four parameter sequences of the received pulses of each participating site are extracted to form the pulse sorting data set of each station. The number of pulse sequences of each station is 29668, 30044 and 29560 respectively.
[0045] In this example, in the sorting category, the radar category is marked as 0~4, and the stray pulse is marked as 5. The local training set is processed into a format suitable for processing by the local LSTM model, and the time sequence length is 32; during local training, the training round is 8, and the batch size is 32; the total round of federated learning is 100.
[0046] The test set is used for final model evaluation, and the sorting accuracy, recall rate and F1 score of each category pulse are output. The test results of each station are comprehensively analyzed to evaluate the overall sorting performance. The LSTM algorithm trained locally without federated learning is compared (not compared with the LSTM prototype algorithm, which is meaningless), and the model is trained with the data of station 0, and the sorting effect of each station pulse is evaluated with the model.
[0047] Table 3 F1 score of sorting
[0048] The multi-station radar pulse sorting method based on federated learning proposed in the application shows significant technical progress in the classification task of five typical radar signals, and the key performance indicators are superior to the local LSTM model. The specific technical breakthroughs are: As shown in FIGS. 5 and 6, the model converges quickly, and the loss reduction and accuracy increase are relatively stable, and the performance is relatively ideal after about 15 rounds of federated learning; As shown in Table 3, the average F1 score is improved from 90.60% to 95.37%, with a relative optimization rate of 5.26%, and the model's overall sorting ability is greatly improved; As shown in FIGS. 7 and 9, the precision is improved from 92.82% to 96.37%, an increase of 3.55 percentage points; As shown in FIGS. 8 and 10, while maintaining an average precision of >96%, the recall rate of each category is higher than that of the locally trained LSTM model, with the most significant improvement in category 2, reaching a recall rate gain of more than 10%, indicating that the application is more robust in relatively difficult categories; The worst category performance: the F1 score of category 2 is improved from 82.43% to 91.54%; The application greatly improves the overall recall ability and practical reliability of the system while maintaining high precision, and has higher engineering application value.
[0049] The above embodiments are only used to illustrate the technical solutions of the present application, but not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can modify the technical solutions described in the foregoing embodiments, or make equivalent replacements for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.
Claims
1. A federated learning based multi-station radar signal sorting method, characterized in that, The method comprises the following steps: For each reconnaissance site participating in reconnaissance, a radar pulse parameter sequence of the reconnaissance site is obtained to form a pulse sorting dataset of the reconnaissance site; The time of arrival sequence in the pulse sorting dataset is replaced by a time difference of arrival sequence, and each parameter sequence is standardized to obtain a pulse sorting dataset; A group of parameters corresponding to each pulse in the pulse sorting dataset is taken as a sorting sample, and a sorting category is taken as a label of the sorting sample; the sorting category includes a radar category and a stray pulse; A federal sorting processing framework of multiple reconnaissance sites is constructed and initialized; A local model is assigned to each reconnaissance site, and the local model is trained and tested by using the pulse sorting dataset; During the training process, the class weights of each sorting category are calculated by dynamic weighting, and a weighted loss function during the training process is constructed based on the class weights; The parameters of the local model of each reconnaissance site are uploaded to a central server; The central server aggregates the parameters of the models of each reconnaissance site by using a parameter averaging strategy to form parameters of a new global model; The parameters of the global model are synchronized and distributed to the local models of each reconnaissance site, and the local models continue the next round of local training based on the distributed parameters; For a group of radar pulse parameter sequences to be identified, the time of arrival is replaced by the time difference of arrival, and then the sorting result is obtained by inputting the time difference of arrival into the trained local model.
2. The federated learning based multi-station radar signal sorting method according to claim 1, wherein, The parameter types in the parameter sequence include the time of arrival, the carrier frequency, the pulse width, and the pulse amplitude.
3. The federated learning based multi-station radar signal sorting method according to claim 1, wherein, The sorting category includes the radar category and the stray pulse.
4. The federated learning based multi-site radar signal sorting method according to claim 1, wherein, The local model adopts an LSTM model, and the structure of the LSTM model is as follows: An input layer Input is used to receive a sorting sample at each time step; A fully connected layer Linear is used to perform linear transformation and feature dimension mapping on the sorting sample to obtain a 256-dimensional feature; a layer normalization layer LayerNorm, an activation function layer Tanh, and a dropout layer Dropout are arranged after the fully connected layer; A first LSTM layer has an input feature dimension of 256, a hidden unit number of 256, and is unidirectional; A second LSTM layer to a fourth LSTM layer each have an input dimension of 256, a hidden unit number of 256, and are unidirectional; a dropout layer Dropout with a preset probability is arranged after each LSTM layer; A bidirectional layer Bidirectional is used to splice the output of the unidirectional LSTM and the output of the reverse LSTM to obtain the output of the bidirectional LSTM layer; A fully connected layer Linear is used to perform linear transformation on the output of the bidirectional LSTM layer to map the feature dimension from 512 to 128; an activation function layer ReLU and a dropout layer Dropout are arranged after the fully connected layer Linear; A fully connected layer Linear is used to perform linear transformation on the mapped feature to map the feature dimension from 128 to 6 to obtain a final classification result; An output layer Output is used to output a classification probability distribution at each time step for predicting a label.
5. The federated learning based multi-site radar signal sorting method according to claim 1, wherein, The class weights of each sorting category are calculated by dynamic weighting, and a weighted loss function during the training process is constructed based on the class weights, which comprises: wherein, is the class weight, is the number of sorted samples in the training set, is the number of sorted samples in the current batch for the th sorted class, is the number of sorted classes in the training set; in, This represents the weighted loss function. Labels for sorted samples, The predicted label for the sorted sample is the first... The probability of each sorting category. Category weights The normalized result.
6. The federated learning based multi-site radar signal sorting method according to claim 1, wherein, The central server adopts a parameter averaging strategy to aggregate the parameters of the LSTM models of each reconnaissance site to form the parameters of a new global model, including: From all the local LSTM models of the reconnaissance sites, all the parameters of an arbitrarily selected local LSTM model are taken as an initial template; each parameter in the initial template is sequentially added to the corresponding parameter tensor in the local LSTM model of the other reconnaissance sites; after the addition of each parameter is completed, the sum is divided by the number of reconnaissance sites to obtain the average value of the parameter, which is used to update the parameters of the aggregated global LSTM model. 7.A terminal device, comprising a processor, a memory, and a computer program stored in the memory; characterized in that, The processor executes the computer program to implement the multi-site radar signal sorting method based on federated learning according to any one of claims 1-6.
8. A computer readable storage medium having stored therein a computer program; characterized in that, The computer program is executed by the processor to implement the multi-site radar signal sorting method based on federated learning according to any one of claims 1-6.