Distributed learning and training method for electromagnetic signal intelligent model based on adaptive collaborative optimization

Abstract

The present invention discloses a distributed learning and training method for electromagnetic signal intelligent models with adaptive collaborative optimization, which belongs to the field of radio signal processing technology and solves the problem of poor adaptive collaborative optimization capabilities of existing intelligent models. The method comprises: a scheduling node assigns a training node to each training task, and distributes a training sample set and the same electromagnetic intelligent model to be trained to each training node performing the same training task; the training node uses the training sample set to perform multiple rounds of distributed training on the electromagnetic intelligent model to be trained; each time a round of training is performed, the scheduling node generates a current round aggregation parameter based on the current round final training parameters of all training nodes performing the same training task, which serves as the initial training parameter for the corresponding training node to perform the next round of training until the preset rounds of training are completed; the scheduling node selects the optimal aggregation parameter from each round aggregation parameter of each training task to obtain a trained electromagnetic intelligent model for the corresponding training task.

Description

Technical Field

[0001] The present invention relates to the technical field of radio signal processing, and in particular to a distributed learning and training method for an electromagnetic signal intelligent model with adaptive collaborative optimization. Background Art

[0002] In the era of big data, achieving breakthroughs in the ultimate performance of intelligent models often requires the infusion of massive amounts of training data. This greatly increases the pressure on traditional centralized intelligent model training, leading to a series of problems such as low model training efficiency and inefficient use of computing resources. Currently, distributed learning and training of intelligent models is an effective means to address the problems caused by the explosive growth of training data. It distributes data and models to different training nodes, and the scheduling node coordinates the computing resources of the training nodes to achieve distributed collaborative training among multiple nodes. Existing distributed training methods for intelligent models generally require that the data on each training node be consistently divided and evenly distributed in terms of quantity and category.

[0003] However, in reality, when scheduling nodes allocate data to training nodes, it's difficult to ensure that each training node receives the same amount of data. This can lead to a "short board" effect for categorized tasks, hindering model collaboration. Furthermore, traditional methods cannot adaptively select optimization strategies based on different task types, nor can they adaptively allocate computing resources.

[0004] Therefore, how to study a distributed learning and training method for electromagnetic signal intelligent models that can achieve adaptive collaborative optimization is a technical problem that needs to be solved urgently. Summary of the Invention

[0005] In view of the above analysis, an embodiment of the present invention aims to provide a distributed learning and training method for an electromagnetic signal intelligent model with adaptive collaborative optimization, so as to solve the problem of poor adaptive collaborative optimization capability of existing intelligent models.

[0006] The present invention discloses a distributed learning and training method for an electromagnetic signal intelligent model with adaptive collaborative optimization, the distributed learning and training method comprising:

[0007] The scheduling node allocates a training node for each training task, and distributes a training sample set and the same electromagnetic intelligent model to be trained to each training node that performs the same training task;

[0008] The training nodes use the training sample set to perform multiple rounds of distributed training on the electromagnetic intelligent model to be trained. During each round of training, the scheduling node generates the current round aggregate parameters based on the current round final training parameters of all training nodes performing the same training task, and uses them as the initial training parameters for the next round of training performed by the corresponding training nodes until the preset rounds of training are completed.

[0009] The scheduling node selects the optimal aggregation parameters from each round of aggregation parameters of each training task, and constructs an electromagnetic intelligent model using the optimal aggregation parameters of each training task as the model parameters, which serves as the trained electromagnetic intelligent model of the corresponding training task.

[0010] On the basis of the above solution, the present invention also makes the following improvements:

[0011] Furthermore, after each round of training, the scheduling node also updates the training sample set of the training node by performing the following operations:

[0012] The scheduling node updates the training sample set of each training node executing the corresponding training task based on the time when all training nodes executing the same training task feedback the final training parameters of the current round, and distributes the current round aggregation parameters and updated training sample set of each training task to the corresponding training node.

[0013] Furthermore, the scheduling node updates the training sample set of the training node by performing the following operations:

[0014] Based on the time it takes for all training nodes executing the same training task to feedback the final training parameters of the current round, calculate the average time it takes for all training nodes executing the same training task to feedback the final training parameters of the current round, as well as the time difference between the earliest feedback and the latest feedback;

[0015] Calculate the average number of training samples that each training node should obtain when performing the same training task under the average distribution method;

[0016] Recalculate the updated training sample size of each training node performing the same training task based on the time when each training node feeds back the final training parameters of the current round, the average time, the time difference, and the average training sample size;

[0017] According to the updated training sample size of each training node, the electromagnetic signal training set of the corresponding training task is re-divided to obtain the updated training sample set of each training node.

[0018] Furthermore, the updated training sample size of the training node k performing the same training task is ;in,

[0019] Indicates the average number of training samples that each training node should obtain when performing the same training task under the average distribution method. It represents the average time it takes for all training nodes that execute the same training task to feedback the final training parameters of the current round. Indicates the time difference between the earliest feedback and the latest feedback, Represents a training node Feedback time of the final training parameters of the current round.

[0020] Furthermore, the scheduling node selects the optimal aggregation parameters from each round of aggregation parameters of each training task and executes:

[0021] The scheduling node obtains the prediction accuracy of the electromagnetic intelligent model using the aggregation parameters of each round as model parameters on the electromagnetic signal verification set, and takes the aggregation parameter with the highest prediction accuracy as the optimal aggregation parameter.

[0022] Furthermore, the distributed learning training method further includes:

[0023] The scheduling node cleans up the training nodes corresponding to each training task and closes the training task.

[0024] Furthermore, after each round of training, the training node obtains the final training parameters of the current round in the following way:

[0025] The training node initializes the training parameters of the electromagnetic intelligent model using the initial training parameters of the current round, trains the electromagnetic intelligent model multiple times based on the training sample set, obtains the final training parameters of the current round, and sends the final training parameters of the current round to the scheduling node.

[0026] Furthermore, the electromagnetic intelligent model is trained multiple times based on the training sample set to obtain the final training parameters of the current round, including:

[0027] The training node trains the electromagnetic intelligent model multiple times based on the training sample set. Each training session calculates the loss function of the corresponding training task and optimizes the training parameters of the electromagnetic intelligent model based on the calculation results of the loss function until the number of training rounds is reached and the final training parameters of the current round are obtained.

[0028] Furthermore, each time a round of training is performed, the scheduling node generates the aggregate parameters for the current round in the following way:

[0029] The scheduling node performs weighted average aggregation on the current round final training parameters of all training nodes executing the same training task to generate the current round aggregate parameters of the corresponding training task.

[0030] Furthermore, the current round aggregation parameters are generated in the following way:

[0031] For each training task, the scheduling node verifies the prediction accuracy of the electromagnetic intelligent model using the final training parameters of the current round of each training node as the model parameters on the electromagnetic signal verification set, and determines the weighted ratio of the final training parameters of the current round of each training node based on the prediction accuracy;

[0032] And according to the determined weighted ratio, the final training parameters of the current round of all training nodes performing the same training task are weighted averaged and aggregated.

[0033] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0034] An adaptive collaborative optimization distributed learning and training method for an electromagnetic signal intelligent model in an embodiment of the present invention designs an adaptive training resource allocation mechanism, and also designs adaptive training optimization strategies and model parameter aggregation strategies according to different training task types. It solves the performance bottlenecks caused by inconsistent data allocation, the inability to adaptively select optimization strategies based on task types, and the inability to adaptively allocate training resources in traditional distributed training methods. It can autonomously optimize collaborative strategies, improve training efficiency, and break through the model performance bottleneck under distributed training.

[0035] Therefore, this method can address the performance bottlenecks caused by inconsistent data distribution, the inability to adaptively select optimization strategies based on task types, and the inability to adaptively allocate training resources in existing methods, so as to realize distributed adaptive collaborative optimization training of intelligent models.

[0036] In the present invention, the above-mentioned technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of the present invention will be described in the following description, and some advantages will become apparent from the description or be learned through practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the contents particularly pointed out in the description and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The accompanying drawings are only used for the purpose of illustrating specific embodiments and are not to be considered as limiting the present invention. Throughout the drawings, the same reference symbols denote the same components.

[0038] Figure 1 A flowchart of a distributed learning and training method for an electromagnetic signal intelligent model with adaptive collaborative optimization provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0039] The preferred embodiments of the present invention will be described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, and are not used to limit the scope of the present invention.

[0040] A specific embodiment of the present invention discloses a distributed learning and training method for an electromagnetic signal intelligent model with adaptive collaborative optimization, as shown in the flowchart. Figure 1 shown.

[0041] Step S1: The scheduling node allocates a training node for each training task, and distributes a training sample set and the same electromagnetic intelligent model to be trained to each training node executing the same training task.

[0042] In this embodiment, the user can create several training tasks according to the training requirements of the electromagnetic intelligent model and import them into the scheduling node. Each training task includes a task type, an electromagnetic signal training set, an electromagnetic signal verification set, and an electromagnetic intelligent model to be trained. At present, in the field of electromagnetic signal processing, the task types of training tasks can be divided into classification tasks, regression tasks, and reconstruction tasks according to the different task objectives. If other task types that can use the distributed training method in this embodiment appear in subsequent academic research, they should also be included in the scope of protection of this application. Specifically, the classification tasks in the field of electromagnetic signal processing can realize signal source identification, modulation method identification, and signal frequency band classification. The regression tasks in the field of electromagnetic signal processing can realize signal strength prediction, signal propagation delay estimation, and signal frequency offset estimation. The reconstruction tasks in the field of electromagnetic signal processing can realize signal denoising, signal missing data recovery, signal compression and reconstruction, etc. It should be emphasized that the training tasks for the different task types described above in the field of electromagnetic signal processing are all common training tasks in this field. The technical solution provided in this embodiment is intended to train the model parameters of an electromagnetic intelligent model to be trained given a known task type, electromagnetic signal training set, electromagnetic signal verification set, and the model structure of the electromagnetic intelligent model to be trained. The datasets and model structures for each training task can be implemented using existing methods, and this embodiment does not limit this. Furthermore, it should be noted that the labels of the electromagnetic signal training set and electromagnetic signal verification set differ for training tasks of different task types. The neural network structures of the electromagnetic intelligent model to be trained also differ for training tasks of different task types. Specifically, the labels of training samples for classification tasks are the classification categories and must be integers; the labels of training samples for regression tasks are the next numerical value of the training sample or other numerical values, which can be decimals; and the labels of training samples for reconstruction tasks depend on the model optimization process.

[0043] Preferably, in this embodiment, considering the different urgency of training tasks of different priority levels, the training tasks may also include priorities, which are pre-configured by the user to ensure that the scheduling node preferentially executes high-priority training tasks. The specific implementation process of step S1 is described as follows.

[0044] Step S11: The scheduling node receives the training task and determines whether the idle training nodes can execute all the training tasks. If so, the corresponding training nodes are assigned to each training task respectively; otherwise, the training nodes are assigned to the high-priority training tasks first, and the low-priority training tasks enter the waiting queue at the same time until the number of idle training nodes can meet the needs of the low-priority training tasks.

[0045] During implementation, after receiving a user-created training task, the scheduling node identifies the index of each training task. The scheduling node can directly allocate training nodes based on the training task. Furthermore, if there are sufficient idle nodes, the scheduling node can also consider providing a list of idle training nodes to the user, who can then allocate training nodes and then provide the allocation results back to the scheduling node. Whether training node allocation is handled by the scheduling node or the user, the allocation strategy is the same. The difference is that, if there are sufficient idle training nodes, the user can independently select the appropriate number of training nodes based on training needs and hand them over to the scheduling node for allocation. Specifically, if the number of idle training nodes is insufficient to meet the user's training task requirements, the scheduling node can prioritize training nodes for high-priority training tasks. If the number of idle training nodes is still insufficient to meet the needs of high-priority training tasks, the scheduling node can consider stopping some training nodes currently executing low-priority training tasks and expanding them to idle training nodes. During implementation, the available computing power of the training node (computing device) can be used to determine whether a training node is idle.

[0046] Preferably, the present embodiment can also allocate training nodes in the following manner. The scheduling node calculates the proportion of the training sample volume of the electromagnetic signal training set of each training task to the total number of training sample volumes of the electromagnetic signal training set of all training tasks, and uses this as the allocation ratio of training nodes for different training tasks, allocates training nodes to each training task, and ensures that each training task is allocated to at least one training node. Assume that the number of idle training nodes is K, i represents the index of the training task, and the i-th training task The corresponding training sample size of the electromagnetic signal training set is ;in, Indicates the task type, , Indicates the total number of task types. For example, in this embodiment, Take 3, “1” represents classification task, “2” represents regression task, and “3” represents reconstruction task. Therefore, the initial allocation process of training nodes can be expressed as:

[0047] (1)

[0048] in, Indicates that the task type is , training task with index i The number of training nodes to allocate. , This is a floor operation.

[0049] Step S12: Divide the electromagnetic signal training set in each training task, and distribute different training sample sets to each training node executing the same training task.

[0050] Preferably, in this embodiment, different training sample sets are distributed to each training node performing the same training task, so that the data in the training sample set allocated to each training node has different distribution characteristics, thereby enriching data diversity and effectively enhancing the generalization ability of the electromagnetic model to be trained.

[0051] For example, for each training task, if the training task is assigned to a unique training node, the electromagnetic signal training set in the training task can be distributed to the corresponding training node as a training sample set; if the training task is assigned to two or more training nodes, the electromagnetic signal training set in the training task can be evenly divided according to the number of training nodes assigned to the training task, and the training samples obtained by the division can be distributed to the corresponding training nodes respectively.

[0052] In addition, it is necessary to distribute the same electromagnetic intelligent model to be trained to each training node that performs the same training task to achieve distributed training. The specific implementation process of distributed training is described in the subsequent instructions.

[0053] Step S2: The training node uses the training sample set to perform multiple rounds of distributed training on the electromagnetic intelligent model to be trained; each time a round of training is performed, the scheduling node generates the current round aggregate parameters based on the current round final training parameters of all training nodes performing the same training task, and uses them as the initial training parameters for the corresponding training nodes to perform the next round of training until the preset rounds of training are completed.

[0054] Step S21: Each time the training node performs a round of training, it initializes the training parameters of the electromagnetic intelligent model using the initial training parameters of the current round, trains the electromagnetic intelligent model multiple times based on the training sample set, obtains the final training parameters of the current round, and sends the final training parameters of the current round to the scheduling node.

[0055] Specifically, it should be noted that when performing the first round of training, each training node can use existing methods to determine the first round's initial training parameters and use these initial training parameters to initialize the electromagnetic intelligent model's training parameters. Starting from the second round of training, the aggregated parameters of the same training task in the previous round are used as the initial training parameters for the current round, and the training parameters of the electromagnetic intelligent model are initialized using these initial training parameters. That is, starting from the second round, the training parameters of the electromagnetic intelligent model are initialized to the aggregated parameters of the same training task in the previous round. This means that the starting point of each round of training is the combined training results of all training nodes in the previous round, allowing the training parameters of the electromagnetic intelligent model to be updated and optimized during each round of training.

[0056] Specifically, during each round of training, the training node initializes the training parameters of the electromagnetic intelligent model using the initial training parameters of the current round, and then trains the electromagnetic intelligent model multiple times based on the training sample set. Each training session calculates the loss function of the corresponding training task, and optimizes the training parameters of the electromagnetic intelligent model based on the calculation results of the loss function until the number of training rounds is reached.

[0057] Preferably, in this embodiment, the model training optimization method for classification tasks is improved, as described below. For classification tasks, the first round of training uses the cross-entropy loss function to train and optimize the electromagnetic intelligence model; starting from the second round of training, each round of training uses a loss function that combines cross-entropy loss and adaptive constraint loss to train and optimize the electromagnetic intelligence model.

[0058] Loss function that combines cross entropy loss and adaptive constraint loss Expressed as:

[0059] (2)

[0060] in, represents the loss weighted ratio, , As a hyperparameter, it can be preset manually; represents the cross entropy loss function, represents the adaptive constraint loss function.

[0061] For regression tasks, the conventional regression loss function is used to train and optimize the electromagnetic intelligent model in each round of training. Expressed as:

[0062] (3)

[0063] in, The parameterized function representing the electromagnetic intelligence model for regression tasks, For training samples The corresponding true label; Represents the training parameters of the electromagnetic intelligence model for regression tasks.

[0064] For reconstruction tasks, the conventional reconstruction loss function is used to train and optimize the electromagnetic intelligent model in each round of training. The reconstruction loss function Expressed as:

[0065] (4)

[0066] in, Represents the training parameters of the electromagnetic intelligence model for reconstruction tasks.

[0067] As can be seen above, for regression and reconstruction tasks, each training node can initialize the electromagnetic intelligent model's training parameters using existing methods during the first round of training. The electromagnetic intelligent model is then trained multiple times in the first round, calculating the corresponding loss function for each training session. The model's training parameters are optimized based on the loss function's calculation results until the current round of training ends (reaching the full number of training rounds), obtaining the final training parameters for the first round. Starting with the second round of training, each training round initializes the electromagnetic intelligent model's training parameters to the aggregated parameters from the previous round (i.e., the aggregated parameters for the same training task in the previous round) fed back by the scheduling node. The electromagnetic intelligent model is then trained multiple times in the current round using the same method to obtain the final training parameters for the current round.

[0068] For classification tasks, during the first round of training, each training node first initializes the training parameters of the electromagnetic intelligent model using the existing method. Then, during the multiple trainings of the electromagnetic intelligent model in the first round, the corresponding loss function (cross entropy loss function) is calculated for each training, and the training parameters of the electromagnetic intelligent model are optimized based on the calculation results of the loss function until the first round of training is completed (the number of trainings in one round is reached), and the final training parameters of the first round are obtained. Starting from the second round of training, each round of training initializes the training parameters of the electromagnetic intelligent model to the aggregated parameters of the same training task in the previous round, and during the multiple trainings in the current round, each training calculates the loss function of the comprehensive cross entropy loss and the adaptive constraint loss, and optimizes the training parameters of the electromagnetic intelligent model based on the calculation results of the loss function of the comprehensive cross entropy loss and the adaptive constraint loss, until the end of the current round of training, and the final training parameters of the current round are obtained. That is, starting from the second round of training, each training except for calculating the cross entropy loss function In addition, it is necessary to calculate the adaptive constraint loss function based on the aggregation parameters and the training parameters of this training , thereby calculating the loss function based on comprehensive cross entropy loss and adaptive constraint loss , and according to the loss function Update the training parameters of the electromagnetic intelligent model when performing the next training.

[0069] It should be noted that this embodiment creatively proposes an adaptive constraint loss function based on the existing cross entropy loss function to solve the performance bottleneck problem caused by inconsistent data distribution during the training of electromagnetic intelligent models for classification tasks. , to optimize model performance. Adaptive constraint loss function It is related to the aggregate parameters (distributed training model parameters) obtained by the training node in the previous round of training and the training parameters of each training performed by the training node in the current round (second round and subsequent) of distributed training. It should be emphasized that the cross entropy loss function Labels are required; the adaptive constraint function No labels are required, and constraints are based on the training output of the electromagnetic intelligent model. The details are as follows.

[0070] In this embodiment, the electromagnetic intelligence model for the classification task includes an input layer, a hidden layer, and a softmax layer in sequence. The input layer and the hidden layer are collectively referred to as the pre-order layer. Starting from the second round of training, the training parameters of the electromagnetic intelligent model are initialized to the aggregated parameters of the same training task in the previous round. At this time, the training node performs the first training of the current round and obtains the electromagnetic intelligent model for the training sample Pre-process Output of layer processing , represents the training sample, represents the current aggregation parameters of the electromagnetic smart model, Indicates the preamble The parameterization function of the layer, vector The dimension of is equal to the dimension of the classification category. At the same time, when the training node performs each training of the current round, it also obtains the electromagnetic intelligent model for the training sample. Pre-process Output of layer processing , It should be noted that starting from the second round of training, the first training of each round of training will initialize the training parameters of the electromagnetic intelligent model. Aggregate parameters for the same training task in the previous round , therefore, the first training obtained and At the same time, since each training is to optimize and update the training parameters, the training node needs to obtain the electromagnetic intelligent model for the training samples in each subsequent training process. Pre-process Output of layer processing , and based on 、 Calculate the adaptive constraint loss function of this training to obtain the loss function of the comprehensive cross entropy loss and the adaptive constraint loss, and update the training parameters for the next training accordingly.

[0071] Then, respectively 、 Selective softening is performed, and the softening formula is:

[0072] (5)

[0073] (6)

[0074] in, Represents a vector Middle The elements corresponding to the classification categories, express The softening result, Indicates the softening temperature, Indicates the classification category set with unbalanced sample size in the training sample set of the training node. Selective softening is also performed to obtain .

[0075] It should be noted that conventional softmax is a process performed on all classification categories and has no softening temperature parameter. Selective softening is essentially softmax processing, but with the addition of a softening temperature hyperparameter. Therefore, softmax processing is only performed on the set of classification categories with an imbalanced sample size in the training sample set. In other words, selectivity is reflected in the need to select the imbalanced sample size of the classification categories output by the electromagnetic intelligence model. For example, if there are 10 classification categories in total, and the classification category labels are {0, 1, 2, ..., 9}, if the number of samples in classification categories 2 and 4 is small, that is, unbalanced, then ={2,4}. Therefore, Represents a set of classification categories with an unbalanced number of training samples assigned to the training node.

[0076] Specifically, in the classification task, when the scheduling node allocates the electromagnetic signal training set to the training node, if the number of training samples of category a assigned on training node 1 is relatively large, and the number of training samples of category a assigned on training node 2 is relatively small, then the electromagnetic intelligent models on training node 1 and training node 2 have different classification capabilities for category a during their respective training processes. Training node 1 performs better on the classification task of category a because it processes more training samples of category a. In contrast, training node 2 has slightly inferior classification performance for category a due to the insufficient number of training samples of category a. Therefore, the training parameters obtained by training node 2 may not reach the optimal state. In this way, the weighted averaging of parameters on the scheduling node will definitely be affected, so it is necessary to use the adaptive constraint loss function proposed in this embodiment. Limit such losses.

[0077] Therefore, if a training node processes a small number of training samples of a certain category, its classification performance for the corresponding category will be worse than that of other training nodes. This performance difference will be reflected in the vector output by the model. In the example, it is specifically represented by the vector output by the corresponding training node The values ​​of the elements of the corresponding classification category in may not be as reliable as the corresponding structures of the other training nodes. If the number of training samples of different categories is balanced, the adaptive constraint loss function finally calculated It will degenerate into the cross entropy loss function. This is because the cross entropy loss function can effectively measure the difference between the probability distribution predicted by the model and the probability distribution of the true label, thereby guiding the training and optimization of the model.

[0078] Finally, by and Calculating selectivity constraint loss :

[0079] (7)

[0080] Step S22: The scheduling node performs weighted average aggregation on the current round final training parameters of all training nodes executing the same training task to generate the current round aggregated parameters of the corresponding training task.

[0081] Preferably, for each training task, the scheduling node separately verifies the prediction accuracy of the electromagnetic intelligent model using the final training parameters of each training node in the current round as model parameters on the electromagnetic signal validation set, determines the weighted ratio of the final training parameters of each training node in the current round based on the prediction accuracy, and then, based on the determined weighted ratio, performs a weighted average aggregation of the final training parameters of the current round for all training nodes performing the same training task. During specific implementation, the weighted average aggregation strategy for different training tasks can be independently selected based on the task type. For example, for classification tasks, the scheduling node can separately verify the classification accuracy of the electromagnetic intelligent model using the final training parameters of each training node in the current round as model parameters on the electromagnetic signal validation set, and determine the weighted ratio of the final training parameters of each training node in the current round. For regression or reconstruction tasks, the scheduling node can separately verify the mean squared error of the final training parameters of the current round fed back by each training node in the electromagnetic signal validation set, and determine the weighted ratio of the training parameters of each training node.

[0082] Specifically, both training parameters and aggregation parameters are sets of model parameters for the electromagnetic intelligent model (including weight parameters and bias parameters for each layer). Because multiple training nodes performing the same training task are assigned the same electromagnetic intelligent model, the aggregation parameters obtained by weighted averaging the model parameters for each layer also comprise the set of model parameters for the electromagnetic intelligent model.

[0083] The scheduling node receives the training task Corresponding After the final training parameters of the current round are fed back by the training nodes, the weighted average aggregation can be performed according to the following formula to obtain the aggregated parameters: :

[0084] (8)

[0085] in, Indicates the execution of training tasks Training nodes The weighted ratio of the final training parameters of the current round, The final training parameters of the current round fed back by training node k.

[0086] Based on the above description, we can see that the weighted ratio According to task type For classification tasks ( ), , It represents the recognition accuracy of the electromagnetic intelligent model fed back by training node k on the electromagnetic signal verification set.

[0087] For regression or reconstruction tasks ( ), , Represents a training node Mean square error of the feedback electromagnetic intelligent model on the electromagnetic signal validation set.

[0088] Step S23: The scheduling node determines whether each training task has reached the preset round of training. If not, the scheduling node distributes the current round aggregation parameters of each training task to the corresponding training node and jumps to the next round of training; otherwise, the training ends.

[0089] Preferably, to optimize the training effect, the scheduling node can also update the training sample set for the next round of training for each training task while generating the aggregation parameters for each round of training. The specific implementation process is described as follows.

[0090] Step S24: The scheduling node updates the training sample set of each training node that performs the corresponding training task based on the time when all training nodes that perform the same training task feedback the final training parameters of the current round, and distributes the current round aggregation parameters and the updated training sample set of each training task to the corresponding training node.

[0091] Specifically, the scheduling node calculates the average time for all training nodes that perform the same training task to feedback the final training parameters of the current round based on the time for all training nodes that perform the same training task to feedback the final training parameters of the current round. , and the time difference between the earliest feedback and the latest feedback .

[0092] Assume that the training node Feedback time of the final training parameters of the current round At this time, the average time for all training nodes executing the same training task to feedback the final training parameters of the current round is Expressed as:

[0093] (9)

[0094] The scheduling node calculates the average number of training samples that each training node should obtain when executing the same training task under the average distribution method :

[0095] (10)

[0096] The scheduling node is based on the time of each training node feedback of the final training parameters of the current round, 、 and , recalculate the updated training sample size of each training node that performs the same training task. Specifically, the updated training sample size of training node k is Finally, based on the updated training sample size of each training node, the electromagnetic signal training set of the corresponding training task is re-divided to obtain the updated training sample set of each training node, and the updated training sample set and aggregation parameters are distributed to the corresponding training nodes.

[0097] It should be noted that due to differences in training node computing power, the training speed of different training nodes varies. If a training node trained slowly in the previous round of training, meaning that it fed the model back to the scheduling node too late, the scheduling node will allocate less data to it in the next round of training, thereby improving overall training speed. Data is redistributed in each round to achieve dynamic sample updates and adjustments, enhancing the model's adaptability to different samples.

[0098] Step S3: The scheduling node selects the optimal aggregation parameters from each round of aggregation parameters of each training task, and constructs an electromagnetic intelligent model using the optimal aggregation parameters of each training task as the model parameters, as the trained electromagnetic intelligent model of the corresponding training task.

[0099] Specifically, the scheduling node obtains the prediction accuracy of the electromagnetic intelligent model using the aggregation parameters of each round as model parameters on the electromagnetic signal verification set, and takes the aggregation parameter with the highest prediction accuracy as the optimal aggregation parameter.

[0100] During the specific implementation process, the prediction process of the electromagnetic intelligent model using the aggregation parameters of each round as the model parameters on the electromagnetic signal verification set can be performed after generating the aggregation parameters of each round, or after obtaining the aggregation parameters of all preset rounds. The specific description is as follows.

[0101] For example, the scheduling node can generate aggregation parameters for each round, verify the prediction accuracy of the electromagnetic intelligent model with the current round aggregation parameters as model parameters on the electromagnetic signal verification set. If the prediction accuracy is higher than the optimal aggregation parameters saved in the previous round, the current round aggregation parameters are updated to the optimal aggregation parameters, otherwise they are not updated. Specifically, the aggregation parameters are tested on the electromagnetic signal verification set and compared. and the optimal aggregation parameters saved in the historical rounds The performance on the electromagnetic signal validation set is When , the recognition accuracy is compared. If it is higher than the historical round, Updated to and save, otherwise not save; when When , the mean square error is compared. If it is lower than the historical round, then Updated to Alternatively, after obtaining the aggregation parameters of all preset rounds, the scheduling node obtains the prediction accuracy of the electromagnetic intelligent model using the aggregation parameters of each round as model parameters on the electromagnetic signal verification set, and uses the aggregation parameter with the highest prediction accuracy as the optimal aggregation parameter.

[0102] Preferably, the method may further include the following steps.

[0103] Step S4: The scheduling node clears the training nodes corresponding to each training task and closes the training task.

[0104] In summary, this embodiment provides a distributed learning and training method for electromagnetic signal intelligent models with adaptive collaborative optimization. The method includes: a scheduling node in a distributed training system adaptively assigns training nodes to tasks based on task priority and the number of idle training nodes, and then distributes electromagnetic signal intelligent models and electromagnetic signal training samples to the training nodes; after receiving the model and training samples, each training node adaptively selects an optimization strategy based on the type of task to which the node belongs; when the training on the node reaches a specified number of rounds, the model is fed back to the scheduling node; the scheduling node adaptively selects an aggregation strategy based on the type of task to which the feedback model belongs, adaptively redistributes the sample size based on the speed of the node feedback model to control the training speed, and then redistributes the aggregated parameters and newly allocated samples to the training nodes, repeating this process until the task reaches the specified number of distributed training rounds; the scheduling node outputs the saved task model, which is the trained electromagnetic signal intelligent model. The solution of the embodiment of the present invention can adaptively select optimization and aggregation strategies and adaptively allocate training resources based on the training task type, providing a new technical approach for distributed learning and training of electromagnetic signal intelligent models.

[0105] Those skilled in the art will appreciate that all or part of the process steps of the above-described embodiments can be implemented by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium, such as a magnetic disk, an optical disk, a read-only memory, or a random access memory.

[0106] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by any technician familiar with this technical field within the technical scope disclosed by the present invention should be covered by the scope of protection of the present invention.

Claims

1. A distributed learning and training method for electromagnetic signal intelligent model with adaptive collaborative optimization, characterized in that: The distributed learning training method includes: The scheduling node allocates a training node for each training task, and distributes a training sample set and the same electromagnetic intelligent model to be trained to each training node that performs the same training task; The training nodes use the training sample set to perform multiple rounds of distributed training on the electromagnetic intelligent model to be trained. During each round of training, the scheduling node generates the current round aggregate parameters based on the current round final training parameters of all training nodes performing the same training task, and uses them as the initial training parameters for the next round of training performed by the corresponding training nodes until the preset rounds of training are completed. The scheduling node selects the optimal aggregation parameters from each round of aggregation parameters of each training task, and constructs an electromagnetic intelligent model using the optimal aggregation parameters of each training task as the model parameters, which serves as the trained electromagnetic intelligent model of the corresponding training task; Each time a training round is performed, the scheduling node updates the training sample set of the training node by performing the following operations: Based on the time it takes for all training nodes executing the same training task to feedback the final training parameters of the current round, calculate the average time it takes for all training nodes executing the same training task to feedback the final training parameters of the current round, as well as the time difference between the earliest feedback and the latest feedback; Calculate the average number of training samples that each training node should obtain when performing the same training task under the average distribution method; Recalculate the updated training sample size of each training node performing the same training task based on the time when each training node feeds back the final training parameters of the current round, the average time, the time difference, and the average training sample size; According to the updated training sample size of each training node, the electromagnetic signal training set of the corresponding training task is re-divided to obtain the updated training sample set of each training node; The current round aggregation parameters and updated training sample set of each training task are distributed to the corresponding training nodes respectively.

2. The adaptive collaborative optimization electromagnetic signal intelligent model distributed learning and training method according to claim 1 is characterized in that: The updated training sample size of the training node k performing the same training task is ;in, Indicates the average number of training samples that each training node should obtain when performing the same training task under the average distribution method. It represents the average time it takes for all training nodes that execute the same training task to feedback the final training parameters of the current round. Indicates the time difference between the earliest feedback and the latest feedback, Represents a training node Feedback time of the final training parameters of the current round.

3. The adaptive collaborative optimization electromagnetic signal intelligent model distributed learning and training method according to claim 1 or 2, characterized in that: The scheduling node selects the optimal aggregation parameters from each round of aggregation parameters for each training task and executes: The scheduling node obtains the prediction accuracy of the electromagnetic intelligent model using the aggregation parameters of each round as model parameters on the electromagnetic signal verification set, and takes the aggregation parameter with the highest prediction accuracy as the optimal aggregation parameter.

4. The adaptive collaborative optimization electromagnetic signal intelligent model distributed learning and training method according to claim 3 is characterized in that: The distributed learning training method further includes: The scheduling node cleans up the training nodes corresponding to each training task and closes the training task.

5. The adaptive collaborative optimization electromagnetic signal intelligent model distributed learning and training method according to claim 3 is characterized in that: Each time a training round is performed, the training node obtains the final training parameters of the current round in the following way: The training node initializes the training parameters of the electromagnetic intelligent model using the initial training parameters of the current round, trains the electromagnetic intelligent model multiple times based on the training sample set, obtains the final training parameters of the current round, and sends the final training parameters of the current round to the scheduling node.

6. The adaptive collaborative optimization electromagnetic signal intelligent model distributed learning and training method according to claim 5 is characterized in that: The electromagnetic intelligent model is trained multiple times based on the training sample set to obtain the final training parameters of the current round, including: The training node trains the electromagnetic intelligent model multiple times based on the training sample set. Each training session calculates the loss function of the corresponding training task and optimizes the training parameters of the electromagnetic intelligent model based on the calculation results of the loss function until the number of training rounds is reached and the final training parameters of the current round are obtained.

7. The adaptive collaborative optimization electromagnetic signal intelligent model distributed learning and training method according to claim 6 is characterized in that: Each time a round of training is performed, the scheduling node generates the aggregate parameters for the current round in the following way: The scheduling node performs weighted average aggregation on the current round final training parameters of all training nodes executing the same training task to generate the current round aggregate parameters of the corresponding training task.

8. The adaptive collaborative optimization electromagnetic signal intelligent model distributed learning and training method according to claim 7 is characterized in that: The current round aggregation parameters are generated by: For each training task, the scheduling node verifies the prediction accuracy of the electromagnetic intelligent model using the final training parameters of the current round of each training node as the model parameters on the electromagnetic signal verification set, and determines the weighted ratio of the final training parameters of the current round of each training node based on the prediction accuracy; And according to the determined weighted ratio, the final training parameters of the current round of all training nodes performing the same training task are weighted averaged and aggregated.

Images