Dynamic pseudo-label based data value-aware federated continual learning method and system

Abstract

The application provides a kind of data value perception federated continuous learning method and system based on dynamic pseudo label, belongs to federated learning technical field.The application generates mechanism by dynamic threshold pseudo label, enables edge device to be able to according to training phase, model confidence, new and old data distribution difference and class sample number adaptive screening unlabeled sample, to improve the participation ratio of credible sample in the early stage of new data, and enhance the reliability of pseudo label after model gradually stabilizes, reduce the risk of new and old class confusion and confirmation bias accumulation caused by false pseudo label;Through new and old data value evaluation mechanism, the contribution of different edge data to the current global model update is described from the aspects of sample freshness, distribution characteristics and information increment, so that the system can identify recent samples with new environment information, and also consider the preservation value of old classes, rare classes and historical scenarios, to improve the utilization efficiency of high-value unlabeled data.

Description

Technical Field

[0001] This invention relates to the field of federated learning technology, specifically to a data value-aware federated continuous learning method and system based on dynamic pseudo-labels. Background Technology

[0002] With the large-scale deployment of the Internet of Things (IoT), edge intelligent devices, and mobile terminals, Federated Learning (FL) has become an important technical approach that balances data privacy protection and distributed model training. In real-world applications, the environments in which devices such as autonomous vehicles, industrial terminals, smart sensors, and robots operate are constantly changing, causing data distribution to evolve over time. To enable models to maintain effective adaptability in such continuously changing task scenarios, Federated Continual Learning (FCL) has emerged. However, due to the significant spatiotemporal heterogeneity of data distribution, models are prone to overwriting previously acquired knowledge when learning new scene features, leading to catastrophic forgetting.

[0003] In cloud-edge collaborative architectures, servers typically only receive a small amount of labeled data, while edge devices distributed across various locations continuously generate massive streams of unlabeled data. To extract value from this unlabeled data, semi-supervised learning (SSL) usually generates pseudo-labels for unlabeled samples through the model and then uses these pseudo-labels as training signals for optimization. Introducing semi-supervised learning mechanisms into federated continuous learning scenarios allows the system to utilize continuously collected unlabeled data from edge devices without directly sharing the original data, but it also introduces new risks of error accumulation. Because the model has not fully learned the relevant features in the early stages of encountering new distributed data, directly using newly arrived unlabeled data for training can easily lead to incorrectly classifying new samples into old categories, thus causing confirmation bias. This leads to the accumulation of incorrect pseudo-labels in subsequent iterations, ultimately resulting in a decline in model performance.

[0004] Existing federated continuous learning methods in scenarios with scarce labels still face two key technical bottlenecks. Firstly, methods with fixed pseudo-label thresholds lack awareness of the value differences between new and old samples. If a uniformly high threshold is adopted, only a very small number of samples can participate in training in the initial stages of encountering new data distributions. The model struggles to absorb new features in a timely manner and is more prone to misclassifying new samples as belonging to old categories. Secondly, under conditions of limited communication bandwidth and computing resources, servers often can only select a portion of edge devices for aggregation. However, existing device selection mechanisms typically focus more on operational status such as computing power, channel capacity, and load, ignoring the comprehensive contribution of edge data in terms of value, heterogeneity, and timeliness. Consequently, it is difficult to achieve rapid convergence under limited resources.

[0005] In the fields of federated continuous learning and federated semi-supervised learning, existing technologies mainly revolve around continuous data utilization, unlabeled sample training, dynamic environment adaptation, device scheduling, and communication efficiency optimization. Current federated continuous learning methods largely focus on online incremental training, cross-environment adaptation, or buffer mechanisms, lacking dynamic pseudo-label selection strategies for unlabeled data streams. If pseudo-labels are generated directly using a fixed threshold before the model has adapted to the new distribution, it can easily introduce erroneous supervision signals, leading to accumulated confirmation bias and exacerbating confusion between old and new classes and catastrophic forgetting.

[0006] While existing federated semi-supervised learning methods can improve the utilization of unlabeled data by using pseudo-labels, teacher-student models, or parameter decomposition, they often struggle to adaptively adjust the pseudo-label threshold based on the training stage, sample freshness, distribution differences, and class scarcity. This results in underutilization of new samples or the incorrect inclusion of low-quality pseudo-labels in the training.

[0007] Furthermore, existing device scheduling methods often focus on communication conditions, computing resources, training time, or the number of devices, while neglecting the freshness of the data held by the devices, their ability to complete categories, and the incremental information. This makes it difficult to prioritize high-value devices that truly help the model adapt to new distributions. Therefore, in scenarios with scarce annotations, constantly changing environments, and limited resources, existing technologies still struggle to simultaneously address new sample utilization, false label suppression, key device selection, and the model's continuous adaptability.

[0008] Federated continuous learning, as a distributed collaborative training method for dynamic data environments, can fully utilize the continuously collected data resources and local computing power of each edge device without directly sharing the raw data, supporting the continuous updating and evolution of the model in dynamic environments. In this architecture, the server may only hold a small amount of initial labeled data and gradually receive newly added reliable labeled samples as the system runs; each edge device continuously collects a large amount of unlabeled data in different geographical locations, time points, and application scenarios. To utilize the new environmental information contained in the unlabeled data, the system can use a semi-supervised learning mechanism to generate pseudo-labels for the unlabeled samples and incorporate credible pseudo-label samples into the local training process. Its basic architecture is as follows: Figure 1 As shown.

[0009] However, in real-world edge intelligence scenarios, the environments, acquisition times, and task scenarios of different devices vary significantly, leading to substantial spatiotemporal heterogeneity in edge data distribution. Due to the scarcity of labeled samples at the edge and the fact that models are not yet fully adapted to the new data distribution in the early stages of a new environment, directly using a fixed threshold to generate pseudo-labels for training can easily reinforce erroneous predictions as supervisory signals, leading to confirmation bias, error accumulation, and model performance degradation. Therefore, in the face of continuously changing environments, dynamically adjusting the pseudo-label selection threshold, improving the effective utilization of new samples, and selecting more contributing devices for updates based on the difference in value between new and old data have become key issues in improving the convergence efficiency, generalization ability, and environmental adaptability of federated continuous learning. Summary of the Invention

[0010] The purpose of this invention is to provide a data value-aware federated continuous learning method and system based on dynamic pseudo-labels. A dynamic pseudo-label screening mechanism is established for continuously arriving unlabeled data to improve the effective utilization rate of new samples and reduce confirmation bias. Based on this, a dynamic joint evaluation mechanism oriented towards data value is constructed, enabling the system to prioritize the use of edge data and model updates with greater learning value under conditions of limited communication and computing resources. This improves the system's convergence efficiency, model generalization ability, and continuous adaptability, thereby solving at least one of the technical problems existing in the aforementioned background art.

[0011] To achieve the above objectives, the present invention adopts the following technical solution:

[0012] In a first aspect, the present invention provides a data value-aware federated continuous learning method based on dynamic pseudo-labels, comprising:

[0013] The global model is pre-trained using the currently available labeled data obtained from the server, so that the global model has basic category recognition ability and initial feature expression ability, reducing the risk of initial misjudgment in the subsequent process of generating pseudo-labels for unlabeled samples.

[0014] The pre-trained global model is used to predict unlabeled data acquired by edge devices, and the category probability distribution and prediction confidence of each unlabeled sample are obtained. The high-confidence pseudo-label samples are selected to form a local training dataset for subsequent local semi-supervised training.

[0015] The pre-trained global model is semi-supervised by using the local training set composed of the obtained pseudo-label samples; in this process, the model is updated with gradients for several rounds using the selected high-confidence pseudo-label samples to obtain the updated local model parameters or model parameter differences.

[0016] After completing local semi-supervised training, a comprehensive evaluation is performed on the data used in the current training round, the local training process, and the operating status of the edge device itself, and a data utility report and an edge device self-performance report are generated. The data utility report describes the potential contribution of local data to the current global model update; the self-performance report describes the resource status of the device after completing local training and the feasibility of uploading model updates.

[0017] Based on the data utility report and the self-performance report, edge devices that can improve the model's adaptability and have the feasibility of uploading are selected to form the model update upload device set for the current round. A model update upload instruction is sent to the selected devices, which is used to notify the devices to upload the model update parameters obtained after local training.

[0018] The obtained model update parameters are aggregated according to the federated aggregation algorithm to generate a global model. The global model is then fine-tuned and calibrated using labeled data to obtain the final calibrated global model. This model is then redistributed to all edge devices as the initial model for the next round of training.

[0019] As a further limitation of the first aspect of the present invention, the server's preheating training of the global model using labeled data includes:

[0020] The server retrieves the existing labeled dataset for the current frame:

[0021] ;

[0022] in, Indicates as of the date The frame server has accumulated labeled data. Indicates the first The frame was updated and confirmed to contain reliable labeled data;

[0023] The server obtains a labeled dataset. Randomly select training batches:

[0024] ;

[0025] in, Indicates the current training batch. This represents the input samples and their corresponding true category labels in the batch;

[0026] The server will send the sample Input the current global model, obtain the class prediction probabilities through forward propagation, and construct the server-side supervised cross-entropy loss function:

[0027] ;

[0028] in, Represents global model parameters. Indicates the number of samples in the training batch. The model parameters are: Time Sample Belongs to the real category The predicted probability;

[0029] The server uses the final global model from the previous frame as the initial model for preheating and training the current frame.

[0030] ;

[0031] in, Indicates the first Global model parameters after frame training is complete.

[0032] The server uses the stochastic gradient descent algorithm to execute... Secondary parameter update:

[0033] ;

[0034] in, This represents the server-side learning rate. This indicates that the supervised loss function has respect to the model parameters. gradient

[0035] After completing the warm-up training, the server obtains the initial global model for the current frame:

[0036] .

[0037] As a further limitation of the first aspect of the present invention, after the edge device receives the trained global model from the server, it uses the global model to predict local unlabeled data, specifically including the following steps:

[0038] Get the local unlabeled dataset available for the current round:

[0039] ;

[0040] in, Indicates edge device Unlabeled data accumulated or retained during historical periods, Indicates edge device In the Newly acquired unlabeled data in each data evolution frame;

[0041] Load global model parameters issued by the server and for unlabeled samples Weak data augmentation is performed to obtain the processed samples. ;

[0042] Processed sample Input the global model, and through the model's feature extraction and classification process, obtain the predicted probability distribution of each sample belonging to each category:

[0043] ;

[0044] in, Indicates as of the date The total number of categories that appear cumulatively in each data evolution frame system. Indicates edge device In the The first data evolution frame Wheel of Samples Predicted as category The probability of;

[0045] The category corresponding to the highest predicted probability is selected as the unlabeled sample. The candidate pseudo-labels are selected, and the maximum predicted probability is used as the prediction confidence of the sample:

[0046]

[0047]

[0048] in, Indicates sample Candidate pseudo-tags, This indicates the prediction confidence level corresponding to the candidate pseudo-label;

[0049] Prediction confidence Dynamic pseudo-label threshold corresponding to the current training round The comparison is performed; when the prediction confidence is higher than the dynamic threshold, the sample is considered to have high credibility, and the category corresponding to the highest probability is used as the pseudo-label of the sample; when the prediction confidence is lower than the dynamic threshold, the sample is considered to have insufficient credibility at present, and the sample will not be used in this round of training.

[0050] As a further limitation of the first aspect of the present invention, during the local training process, the edge device calculates the local semi-supervised training loss using the selected high-confidence pseudo-label samples, and performs several gradient updates on the local model in the following manner:

[0051] ;

[0052] in, Indicates edge device In the Round completed Model parameters updated locally. Indicates the local learning rate. This represents the gradient of the local semi-supervised training loss with respect to the model parameters. This indicates the total number of iterations during local training.

[0053] As a further limitation of the first aspect of the present invention, the data utility report is used to describe the potential contribution of local data to the current global model update, and includes at least three types of information: sample freshness, sample scarcity, and information increment. Sample freshness is used to measure the proportion of newly collected samples in the current data of the edge device; sample scarcity identifies local long-tail categories and class distribution imbalance; information increment is calculated based on the relative decrease in pseudo-label training loss before and after local training.

[0054] The self-performance report includes:

[0055] Computational resources used to describe the device’s current processor, graphics processing unit, memory, storage space, and remaining computing power available for subsequent training or model uploading;

[0056] Communication status used to describe the device's current uplink rate, communication latency, channel stability, network congestion, and the feasibility of uploading model updates;

[0057] Energy status is used to describe the device's current battery level, energy budget, and ability to continue performing communication uploads or subsequent training tasks;

[0058] This describes the status of training results, including whether local training on the device is complete, changes in training loss, the magnitude of model updates, and the stability of the local training process.

[0059] As a further limitation of the first aspect of the present invention, the server processes the local model updates uploaded by the edge device, including:

[0060] The server performs weighted aggregation on the local model parameters uploaded by the selected edge devices to obtain the aggregated global model:

[0061] ;

[0062] in, This represents the global model parameters after federated aggregation. Indicates the first Frame number A collection of edge devices participating in the aggregation. Indicates edge device The corresponding aggregate weights, Indicates edge device Model parameters after local training is complete;

[0063] The server uses its labeled data to fine-tune the aggregated global model, calculates supervised loss based on the true labels, and performs several gradient updates on the model parameters:

[0064] ;

[0065] in, This indicates that the server has completed the first... The global model parameters after the second fine-tuning This indicates that the learning rate is fine-tuned on the server side. This represents the gradient of the loss with respect to the model parameters. This indicates the number of fine-tuning iterations.

[0066] Secondly, this invention provides a data value-aware federated continuous learning system based on dynamic pseudo-labels, comprising:

[0067] The preheating training module is used to preheat the global model using the currently available labeled data obtained from the server, so that the global model has basic category recognition ability and initial feature expression ability, reducing the risk of initial misjudgment in the subsequent process of generating pseudo-labels for unlabeled samples.

[0068] The prediction module is used to predict unlabeled data acquired by edge devices using a pre-trained global model, obtain the category probability distribution and prediction confidence of each unlabeled sample, and select high-confidence pseudo-label samples to form a local training dataset for subsequent local semi-supervised training.

[0069] The semi-supervised training module is used to perform semi-supervised training on the pre-trained global model using a local training set composed of acquired pseudo-label samples. Specifically, the model is updated with gradients for several rounds using selected high-confidence pseudo-label samples to obtain the updated local model parameters or model parameter differences.

[0070] The evaluation module is used to comprehensively evaluate the data used in the current training round, the local training process, and the operating status of the edge device itself after the local semi-supervised training is completed, and to generate a data utility report and an edge device self-performance report. The data utility report is used to describe the potential contribution of local data to the current global model update. The self-performance report is used to describe the resource status of the device after completing local training and the feasibility of uploading the model update.

[0071] The filtering module is used to select edge devices that can improve the model's adaptability and have the feasibility of uploading based on the data utility report and the self-performance report, forming a set of model update upload devices for the current round, and sending model update upload instructions to the selected devices. The instructions are used to notify the devices to upload the model update parameters obtained after local training.

[0072] The aggregation and calibration module is used to aggregate the obtained model update parameters according to the federated aggregation algorithm to generate a global model. The global model is then fine-tuned and calibrated using labeled data to obtain the final calibrated global model, which serves as the initial model for the next round of training.

[0073] Thirdly, the present invention provides a non-transitory computer-readable storage medium for storing computer instructions, which, when executed by a processor, implement the data value-aware federated continuous learning method based on dynamic pseudo-labels as described in the first aspect.

[0074] Fourthly, the present invention provides a computer device including a memory and a processor, wherein the processor and the memory communicate with each other, the memory stores program instructions executable by the processor, and the processor invokes the program instructions to execute the data value-aware federated continuous learning method based on dynamic pseudo-labels as described in the first aspect.

[0075] Fifthly, the present invention provides an electronic device, comprising: a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to cause the electronic device to execute instructions for implementing the data value-aware federated continuous learning method based on dynamic pseudo-labels as described in the first aspect.

[0076] The beneficial effects of this invention are as follows: Through a dynamic threshold pseudo-label generation mechanism, edge devices can adaptively filter unlabeled samples based on the training stage, model confidence, differences in the distribution of new and old data, and the number of samples in each category. This increases the participation rate of reliable samples in the early stages of new data appearance and enhances the reliability of pseudo-labels as the model gradually stabilizes, reducing the risk of confusion between new and old categories and the accumulation of confirmation bias caused by erroneous pseudo-labels. Through a new and old data value assessment mechanism, the contribution of different edge data to the current global model update is characterized from aspects such as sample freshness, distribution characteristics, and information increment. This allows the system to identify recent samples with new environmental information while also preserving the value of old categories, rare categories, and historical scenarios, improving the utilization efficiency of high-value unlabeled data. Through a data value-aware collaborative scheduling mechanism, the server can combine device data value, training result status, communication conditions, computing power, and energy consumption status to prioritize high-value, reliable devices with upload capabilities for federated aggregation, avoiding the waste of communication and computing resources caused by indiscriminate device selection, and improving the quality of model updates and the utilization rate of training resources. While protecting the privacy of raw data at the edge, it can improve the convergence efficiency, generalization ability and stability of federated continuous learning in scenarios where labeled samples are scarce, unlabeled data continues to arrive and the environment is dynamically changing, and can provide more practical technical support for applications such as autonomous driving, smart healthcare and industrial monitoring.

[0077] The advantages of additional aspects of the invention will be set forth more clearly in the following description or will be learned by practice of the invention. Attached Figure Description

[0078] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0079] Figure 1 This is a schematic diagram of an existing technology-based federated continuous semi-supervised learning architecture.

[0080] Figure 2 This is a flowchart illustrating the overall process of federated continuous semi-supervised learning based on dynamic threshold driving, as described in an embodiment of the present invention.

[0081] Figure 3 This is a schematic diagram of information transmission in federated continuous semi-supervised learning based on dynamic threshold driving, as described in an embodiment of the present invention. Detailed Implementation

[0082] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0083] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0084] It should also be understood that terms such as those defined in general dictionaries should be understood to have meanings consistent with their meanings in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as here.

[0085] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, and / or groups thereof.

[0086] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0087] To facilitate understanding of the present invention, the present invention will be further explained and described below with reference to the accompanying drawings and specific embodiments. However, the specific embodiments do not constitute a limitation on the embodiments of the present invention.

[0088] Those skilled in the art should understand that the accompanying drawings are merely schematic diagrams of embodiments, and the components in the drawings are not necessarily essential for implementing the present invention.

[0089] Example 1

[0090] This embodiment first provides a data value-aware federated continuous learning system based on dynamic pseudo-labels, including: a preheating training module (e.g., located in a server), used to preheat the global model using currently available labeled data obtained from the server, enabling the global model to have basic category recognition capabilities and initial feature expression capabilities, reducing the risk of initial misjudgment in the subsequent generation of pseudo-labels for unlabeled samples; a prediction module (e.g., located in an edge device, where the server distributes the preheated global model to the edge device), used to predict the unlabeled data obtained by the edge device using the preheated global model, obtaining the category probability distribution and prediction confidence of each unlabeled sample, and selecting high-confidence pseudo-label samples to form a local training dataset for subsequent local semi-supervised training; and a semi-supervised training module (located in the edge device), used to perform semi-supervised training on the preheated global model using the local training set composed of the obtained pseudo-label samples; wherein, the selected high-confidence pseudo-label samples are used to perform several rounds of gradient updates on the model to obtain updated local model parameters or model parameters. The evaluation module (located in the edge device) comprehensively evaluates the data used in the current training round, the local training process, and the edge device's own operating status after completing local semi-supervised training, and generates a data utility report and an edge device self-performance report. The data utility report describes the potential contribution of local data to the current global model update; the self-performance report describes the resource status of the device after completing local training and the feasibility of uploading model updates. The filtering module (located in the server) selects edge devices that can improve model adaptability and have upload feasibility based on the data utility report and the self-performance report, forming a set of model update upload devices for the current round, and issues model update upload instructions to the selected devices, which are used to notify the devices to upload the model update parameters obtained after local training. The aggregation and calibration module (located in the server) aggregates the obtained model update parameters according to the federated aggregation algorithm to generate a global model, and fine-tunes and calibrates the global model using labeled data to obtain the final calibrated global model, which serves as the initial model for the next round of training.

[0091] The above system was used to implement a data value-aware federated continuous learning method based on dynamic pseudo-labels, including: pre-training the global model using currently available labeled data obtained from the server, enabling the global model to have basic category recognition capabilities and initial feature representation capabilities, reducing the risk of initial misjudgment in the subsequent pseudo-label generation process for unlabeled samples; using the pre-trained global model to predict unlabeled data obtained from edge devices, obtaining the category probability distribution and prediction confidence of each unlabeled sample, and selecting high-confidence pseudo-label samples to form a local training dataset for subsequent local semi-supervised training; using the local training set composed of the obtained pseudo-label samples to perform semi-supervised training on the pre-trained global model; wherein, using the selected high-confidence pseudo-label samples to perform several rounds of gradient updates on the model, obtaining the updated local model parameters or model parameter differences; after completing the local semi-supervised training, the current round of training participants... The system comprehensively evaluates the data, local training process, and the operating status of the edge devices, generating a data utility report and an edge device performance report. The data utility report describes the potential contribution of local data to the current global model update. The performance report describes the resource status of the device after completing local training and the feasibility of uploading model updates. Based on the data utility report and the performance report, edge devices that can improve model adaptability and have upload feasibility are selected to form the current round of model update upload device set. A model update upload instruction is issued to the selected devices, which is used to notify the devices to upload the model update parameters obtained after local training. The obtained model update parameters are aggregated according to the federated aggregation algorithm to generate a global model. The global model is fine-tuned and calibrated using labeled data to obtain the final calibrated global model, which is then redistributed to all edge devices as the initial model for the next round of training.

[0092] The data value-aware federated continuous learning method based on dynamic pseudo-labels implemented in this embodiment takes the actual contribution of marginal data to the continuous updating of the model as the core basis, and constructs a closed-loop learning mechanism that combines dynamic threshold pseudo-label generation, evaluation of the value of new and old data, and multi-device collaborative scheduling. First, by dynamically adjusting the pseudo-label screening threshold, the utilization rate of credible unlabeled samples is improved and the confirmation bias is reduced. On this basis, the value of new and old data is evaluated from aspects such as sample freshness, contribution of category distribution, and information increment. Furthermore, considering the communication, computing, and training status of devices, high-value devices with upload capabilities are prioritized to participate in federated aggregation, thereby improving the model's convergence efficiency, generalization ability, and continuous adaptability.

[0093] like Figure 2As shown, the overall workflow of the data value-aware federated continuous learning method based on dynamic pseudo-labels described in this embodiment includes steps such as data acquisition, server warm-up, pseudo-label generation, local semi-supervised training, data value assessment, collaborative scheduling, federated aggregation, and server fine-tuning, forming a closed-loop optimization process oriented towards dynamic environments. The system includes one server and multiple edge devices. The server is mainly responsible for labeled data management, model warm-up and training, device utility assessment, model update selection, federated aggregation, and model fine-tuning and calibration; the edge devices are mainly responsible for local data acquisition and aggregation, unlabeled sample determination, dynamic threshold pseudo-label generation, data value quantification, local model training, and model update uploading. The edge devices communicate with the server through the network, receive the global model issued by the server, and after completing pseudo-label generation, semi-supervised training, and data value assessment locally, upload a utility report containing sample freshness, distribution characteristics, information increment, and device operating status; the server forms a set of devices for model update uploading based on the utility report, and the selected edge devices then upload the updated parameters of the trained model.

[0094] The specific process of the data value-aware federated continuous learning method based on dynamic pseudo-labels described in this embodiment includes the following steps:

[0095] Step 1: Data Acquisition and Server Warm-up Preparation, including: Edge devices continuously collect local environmental data during the data acquisition phase. This data primarily consists of unlabeled samples, but may also include a small number of labeled samples obtained through manual verification or other methods. Before each training round, the devices aggregate the local data to form a usable dataset for the current round, providing a foundation for subsequent pseudo-label generation and data value assessment. The server receives and manages the currently available labeled data, including initial labeled samples accumulated historically on the server side, as well as newly added and confirmed reliable labeled samples during system operation. The server uses this labeled data to warm up and train the global model, enabling the model to possess basic category recognition capabilities and initial feature expression capabilities, reducing the initial misjudgment risk during the subsequent generation of pseudo-labels for unlabeled samples.

[0096] The server distributes the pre-warmed global model to all available edge devices. Server-side pre-training avoids the model's initial reliance on low-quality pseudo-labels, thus providing a more stable model foundation for subsequent semi-supervised learning. Edge devices continuously collect local environmental data, helping the system to promptly perceive changes in the environment of different devices, providing data support for the subsequent collaborative scheduling of new and old data value assessment and data value perception.

[0097] Specifically, the process of the server using the labeled data to pre-train the global model includes the following steps:

[0098] The first step is for the server to obtain the existing labeled dataset for the current frame:

[0099] ;

[0100] in, Indicates as of the date The frame server has accumulated labeled data. Indicates the first The frame adds and confirms reliable labeled data.

[0101] The second step involves the server accessing the labeled dataset. Randomly select training batches:

[0102] ;

[0103] in, Indicates the current training batch. This represents the input samples and their corresponding true category labels in the batch.

[0104] Third step, the server will send the sample Input the current global model, obtain the class prediction probabilities through forward propagation, and construct the server-side supervised cross-entropy loss function:

[0105] ;

[0106] in, Represents global model parameters. Indicates the number of samples in the training batch. The model parameters are: Time Sample Belongs to the real category The predicted probability.

[0107] The fourth step is to use the final global model from the previous frame as the initial model for preheating and training the current frame:

[0108] ;

[0109] in, Indicates the first Global model parameters after frame training is complete.

[0110] The server uses the stochastic gradient descent algorithm to execute... Secondary parameter update:

[0111] ;

[0112] in, This represents the server-side learning rate. This indicates that the supervised loss function has respect to the model parameters. The gradient.

[0113] Fifth step: After completing the warm-up training, the server obtains the initial global model for the current frame:

[0114] ;

[0115] The model is then sent to various edge devices for subsequent unlabeled sample prediction, pseudo-label generation, and local semi-supervised training.

[0116] Step 2: Dynamic threshold pseudo-label generation for edge devices, including: after receiving the global model from the server, the edge device uses the model to predict the unlabeled data aggregated in this round, obtains the category probability distribution and prediction confidence of each unlabeled sample, and determines whether the sample currently has the credible conditions to participate in training.

[0117] Specifically, after receiving the trained global model from the server, the edge device uses the global model to predict local unlabeled data, which includes the following steps:

[0118] The first step is for the edge device to obtain the local unlabeled dataset available for the current round:

[0119] ;

[0120] in, Indicates edge device Unlabeled data accumulated or retained during historical periods, Indicates edge device In the Each data evolution frame contains newly acquired unlabeled data.

[0121] The second step is for the edge device to load the global model parameters issued by the server. and for unlabeled samples Weak data augmentation is performed to obtain the processed samples. .

[0122] The third step is for the edge device to process the sample. Input the global model, and through the model's feature extraction and classification process, obtain the predicted probability distribution of each sample belonging to each category:

[0123] ;

[0124] in, Indicates as of the date The total number of categories that appear cumulatively in each data evolution frame system. Indicates edge device In the The first data evolution frame Wheel of Samples Predicted as category The probability of.

[0125] Fourth, the edge device selects the category corresponding to the highest predicted probability as the unlabeled sample. The candidate pseudo-labels are selected, and the maximum predicted probability is used as the prediction confidence of the sample:

[0126] ;

[0127] ;

[0128] in, Indicates sample Candidate pseudo-tags, This indicates the prediction confidence level corresponding to the candidate pseudo-label.

[0129] Fifth, the edge device will predict the confidence level. Dynamic pseudo-label threshold corresponding to the current training round The process involves comparison. When the prediction confidence is higher than the dynamic threshold, the sample is considered to have high credibility, and the category corresponding to the highest probability is used as the pseudo-label for the sample. When the prediction confidence is lower than the dynamic threshold, the sample is considered to have insufficient credibility and will not be used in this round of training. The sample selection process is represented as follows: ;in, Indicates the sample selection results. Indicates an indicator function. When At that time, the sample Selected for this round of pseudo-label training dataset; when At that time, the sample They will not participate in this round of training for the time being.

[0130] The dynamic threshold can be adjusted based on factors such as training rounds, model fit, prediction stability, differences in the distribution of new and old samples, and the number of samples in each category. In the early stages of training, the system can appropriately lower the threshold to increase the chances of newly arriving samples participating in training; as the model gradually adapts to the new data distribution, the system can gradually increase the threshold to enhance the reliability of pseudo-labeled samples.

[0131] The edge device uses the selected high-confidence pseudo-label samples to form a local training dataset for subsequent local semi-supervised training.

[0132] By employing a dynamic threshold pseudo-label generation mechanism, the system avoids the problem of fixed thresholds becoming too rigid in dynamic environments. On one hand, this mechanism can improve the utilization rate of potentially effective samples in the early stages of new data, enabling the model to absorb new environmental features more quickly; on the other hand, it can also improve the quality of pseudo-labels as the model gradually stabilizes, reducing confirmation bias caused by erroneous pseudo-labels.

[0133] Step 3: The edge device performs local semi-supervised training, including: after completing the screening of unlabeled samples and the generation of pseudo-labels, the edge device performs local semi-supervised training based on the global model issued by the server, local high-confidence pseudo-label samples, and the device's local preset training strategy.

[0134] During local training, the edge device uses the selected high-confidence pseudo-label samples to perform several rounds of gradient updates on the model, obtaining the updated local model parameters or model parameter differences.

[0135] During local training, the edge device uses the selected high-confidence pseudo-label samples to calculate the local semi-supervised training loss, and performs several gradient updates on the local model in the following manner:

[0136] ;

[0137] in, Indicates edge device In the Round completed Model parameters updated locally. Indicates the local learning rate. This represents the gradient of the local semi-supervised training loss with respect to the model parameters. This represents the total number of iterations during local training. After local training is complete, the edge device receives the updated local model parameters. Alternatively, one can obtain the parameter difference relative to the initial global model. .

[0138] The local preset training strategy can be determined by the edge device itself based on its own computing resources, energy consumption, local data scale, and current task requirements, including but not limited to the number of local training rounds, batch size, training time window, and optimizer parameters. The server does not send specific training configuration parameters to the device.

[0139] During local training, edge devices not only need to absorb the new distribution information contained in the newly collected data, but also need to retain the model's ability to recognize existing categories and historical scenes as much as possible, so as to avoid forgetting old knowledge due to over-adaptation to new data.

[0140] Local semi-supervised training enables edge devices to update their models using dynamically selected, reliable pseudo-labeled samples without uploading the original data. This process improves the efficiency of unlabeled data utilization and provides training results as a basis for subsequently evaluating the actual contribution of local data to global model updates.

[0141] Step 4: The edge device generates a utility report for evaluating the value of both new and old data. This includes: after completing local semi-supervised training, the edge device comprehensively evaluates the data used in the current training round, the local training process, and the device's own operating status, and generates a data utility report and a self-performance report.

[0142] The data utility report describes the potential contribution of local data to the current global model update, and includes at least three types of information: sample freshness, sample scarcity, and information increment. Specifically, it includes:

[0143] Since edge devices only hold unlabeled data, sample category information is statistically analyzed based on high-confidence pseudo-labels obtained after dynamic threshold filtering. Sample freshness, sample scarcity, and information increment can be obtained as follows:

[0144] Sample freshness measures the proportion of newly collected samples in the current data of edge devices, and can be expressed as: ;in, Indicates edge device In the Sample freshness of each data evolution frame This represents the set of newly acquired unlabeled samples in the current frame. This represents the complete set of unlabeled samples currently available. The larger the value, the more accurately the device's data reflects current environmental changes.

[0145] Sample scarcity is used to identify local long-tail categories and imbalanced category distribution. Edge devices first count the number of valid samples for each category based on high-confidence pseudo-labels:

[0146] ;

[0147] Furthermore, categories The scarcity of samples can be expressed as: ;

[0148] in, Indicates device In high-confidence pseudo-label samples, the categories The number of samples, This indicates the sample screening results. Indicates sample pseudo-tags Indicates an indicator function, Indicates as of the date The total number of categories that appear cumulatively in each data evolution frame. This indicates a positive number set to prevent the denominator from being zero. The larger the value, the higher the category. The scarcer the samples in the device, the better it can be used to identify local long-tail categories and imbalanced category distribution.

[0149] The information increment is calculated based on the relative decrease in pseudo-label training loss before and after local training:

[0150] ;

[0151] in, Indicates edge device In the Information increment of the wheel, This represents the pseudo-label training loss before local training begins. This represents the pseudo-label training loss after local training is completed. The larger the value, the more significant the reduction in model loss can be achieved with the device's local data, and the greater its potential contribution to the current global model update.

[0152] The self-performance report is used to describe the resource status of the device after local training and the feasibility of uploading model updates, specifically including:

[0153] Computing resources. Computing resources describe the device's current processor, graphics processing unit, memory, storage space, and remaining computing power available for subsequent training or model uploading.

[0154] Communication status. Communication status describes the device's current uplink speed, communication latency, channel stability, network congestion, and the feasibility of uploading model updates.

[0155] Power Status. Power status describes the device's current battery level, power budget, and ability to continue performing communication uploads or subsequent training tasks.

[0156] Training result status. The training result status describes whether local training on the device is complete, the changes in training loss, the magnitude of model updates, and whether the local training process is stable.

[0157] Edge devices upload data utility reports and their own performance reports to the server, while temporarily storing the model update parameters obtained from local training, awaiting subsequent scheduling instructions from the server.

[0158] By evaluating the value of both new and old data, the server can not only determine whether the device has completed local training, but also further assess the value of the device's data for the current model update. This mechanism can simultaneously characterize the timely contribution of new data, the value of preserving old categories, and the ability to complete rare categories, avoiding device selection based solely on the amount of data or communication status.

[0159] Step 5: The server performs collaborative scheduling based on data value perception, including: after receiving data utility reports and its own performance reports uploaded by multiple edge devices, the server summarizes and analyzes the data freshness, category coverage, information increment, training result status, computing resources and communication status of each device.

[0160] The server combines the current global model's performance status to identify the model's weaknesses under different categories, scenarios, or data distributions, i.e., the distributional shortcomings that the current model needs to focus on strengthening.

[0161] Based on the data utility reports of each device, the server determines whether the model updates of candidate devices can provide the key gains required by the current global model. For example, when some devices hold more recently collected samples, rare class samples, or samples that differ significantly from the current global distribution, the local model updates corresponding to those devices can be assigned higher aggregation value.

[0162] Based on each device's own performance report, the server determines whether the candidate device has the resource conditions to upload local model updates and participate in this round of aggregation. For example, if some devices have unstable communication links, insufficient power consumption, or abnormal local training results, their priority may be reduced or they may not be included in this round of aggregation, even if they have some data value.

[0163] During the device selection process, the server comprehensively considers the following factors and selects several devices from the candidate device set to upload model updates and participate in this round of federated aggregation: device data value; sample freshness; contribution to class balance; information increment; local training result status; communication conditions and computing power; energy consumption status; historical participation; system resource constraints.

[0164] The server weighs the benefits of model updates against the overhead of communication resources, prioritizing devices that can improve model adaptability and are feasible for uploading, thus forming a set of devices for model update uploads in the current round. The server sends model update upload instructions to the selected devices. These instructions are used to instruct the devices to upload the model update parameters obtained after local training, excluding specific training configuration parameters such as learning rate, number of local training rounds, and batch size.

[0165] Through data-value-aware collaborative scheduling, the system avoids the waste of communication resources caused by all devices indiscriminately uploading model updates. It prioritizes limited communication resources for devices holding high-value data, with valid training results, and capable of uploading. This mechanism helps improve the quality of model aggregation and enhances the model's learning ability for new distributions, new categories, and long-tail scenarios.

[0166] Step 6: Edge devices upload model updates, and the server performs federated aggregation and fine-tuning calibration. This includes: After receiving the model update upload instruction from the server, the selected edge devices upload the updated model parameters obtained from local training to the server. The updated model parameters can be local model parameters, model parameter differences, or gradient information. After receiving model updates uploaded by multiple devices, the server aggregates these updates according to the federated aggregation algorithm to generate a global model. During the aggregation process, the server can assign different aggregation weights to different devices based on factors such as device data volume, data quality, utility score, category contribution, local training result status, or historical participation contribution, so that the data contributions of high-value and highly reliable devices can be more fully reflected in the global model update.

[0167] After completing the federated aggregation, the server uses the labeled data it holds to fine-tune and calibrate the global model, so as to use real labels to correct the bias that may occur during the training of unlabeled pseudo-labels and suppress the accumulation and reinforcement of erroneous pseudo-labels in multiple rounds of training.

[0168] Specifically, the server processes the local model updates uploaded by the edge device, which includes the following two steps:

[0169] The first step is for the server to perform weighted aggregation of the local model parameters uploaded by the selected edge devices to obtain the aggregated global model:

[0170] ;

[0171] in, This represents the global model parameters after federated aggregation. Indicates the first Frame number A collection of edge devices participating in the aggregation. Indicates edge device The corresponding aggregate weights, Indicates edge device Model parameters after local training is complete.

[0172] The second step involves the server using its labeled data to fine-tune the aggregated global model, calculating the supervised loss based on the true labels, and performing several gradient updates on the model parameters.

[0173] ;

[0174] in, This indicates that the server has completed the first... The global model parameters after the second fine-tuning This indicates that the learning rate is fine-tuned on the server side. This represents the gradient of the loss with respect to the model parameters. This indicates the number of fine-tuning iterations.

[0175] After fine-tuning, the server obtains the final calibrated global model and redistributes it to all edge devices as the initial model for the next round of training.

[0176] Federated aggregation allows the server to aggregate local model updates from multiple edge devices, fully leveraging data features from different devices to improve the generalization ability of the global model. Server-side fine-tuning and calibration with labeled data further mitigates confirmation bias caused by false label errors, enhancing the model's stability and reliability in dynamic environments.

[0177] Combination Figure 3 As shown, the key information exchange steps of the method described in this embodiment are as follows:

[0178] Model distribution: After the server completes the model warm-up training using labeled data, it distributes the warmed-up global model to all available edge devices, providing an initial model foundation for the edge devices to perform unlabeled sample judgment, dynamic threshold pseudo-label generation, and local semi-supervised training.

[0179] Report Upload: After completing unlabeled sample screening, dynamic threshold pseudo-label generation, and local semi-supervised training, the edge device generates and uploads a data utility report and a self-performance report. The data utility report supports the evaluation of the value of new and old data, including sample freshness, distribution characteristics, and information increment; the self-performance report includes computing resources, communication status, energy consumption status, and training result status.

[0180] Upload command issuance: Based on the data utility reports of each device, its own performance reports, and the current global model performance status, the server performs data value-aware collaborative scheduling, prioritizing devices that hold high-value data, have reliable training results, and meet the upload conditions to participate in this round of model update upload and federated aggregation, and issues model update upload commands to the selected devices.

[0181] Model Update Upload: After receiving the upload command, the selected devices upload the model update parameters obtained from local semi-supervised training to the server. The uploaded model update parameters reflect the combined contribution of dynamically selected reliable pseudo-label samples, the value of new and old data, and the local training effect to the model update.

[0182] Model distribution: After receiving the model update uploaded by the selected devices, the server completes federated aggregation and fine-tuning calibration of labeled data on the server side, and redistributes the final calibrated global model to all edge devices as the initial model for the next round of continuous learning, providing an updated model foundation for the next round of dynamic threshold pseudo-label generation, new and old data value assessment and collaborative scheduling.

[0183] Example 2

[0184] In this second embodiment, taking traffic sign image recognition in a vehicle-to-everything (V2X) scenario as an example, the data value perception federated continuous learning method based on dynamic pseudo-labels described in the first embodiment is further explained.

[0185] The system in this embodiment includes a base station server and multiple vehicles. The base station server can be an edge server deployed in a cellular base station, roadside unit, or mobile edge computing node, responsible for global model warm-up, vehicle utility evaluation, participation in vehicle selection, model aggregation, and fine-tuning calibration; the vehicles, as edge devices, are equipped with onboard cameras, onboard computing units, and wireless communication modules to continuously collect road environment images during driving and perform inference and training of traffic sign image recognition models locally.

[0186] The image data collected by the vehicle can include speed limit signs, stop signs, no-entry signs, turn signs, warning signs, instruction signs, and other road traffic signs. Due to differences in road conditions, weather, lighting, shooting angle, and driving area, the data collected by different vehicles varies significantly; at the same time, new traffic sign categories, special road conditions, and traffic signs with different appearances may continuously appear as the vehicle travels.

[0187] The base station server holds a small number of manually labeled traffic sign images. Each labeled image includes a traffic scene image and a corresponding traffic sign category label. The base station server uses these labeled images to pre-train a traffic sign recognition model, enabling the model to possess basic image feature extraction and traffic sign category recognition capabilities. The traffic sign recognition model can employ convolutional neural networks, residual neural networks, or other neural network models suitable for image classification or object recognition tasks. The specific training process adopts the server-side pre-training method described in Example 1.

[0188] After completing the preheating training, the base station server distributes the global traffic sign recognition model to each vehicle. The vehicles use their onboard cameras to acquire unlabeled road images collected during the current driving process, and perform preprocessing operations such as resizing, normalization, cropping, and data augmentation on the images.

[0189] The vehicle inputs the preprocessed unlabeled image into the global model sent by the base station server to obtain the predicted probability and prediction confidence of the traffic sign category corresponding to the image. Following the dynamic pseudo-label generation method described in Example 1, the vehicle uses the traffic sign category corresponding to the highest predicted probability as a candidate pseudo-label and determines whether the image meets the confidence condition based on the dynamic threshold of the current training round.

[0190] When the prediction confidence of an unlabeled image is not lower than the dynamic threshold, the vehicle uses the candidate category as the effective pseudo-label for that image and adds it to the local high-confidence pseudo-label sample set. When the prediction confidence is lower than the dynamic threshold, the vehicle temporarily does not use the image for this round of training. By dynamically adjusting the pseudo-label threshold, the utilization rate of new images can be improved in the early stages of new road environments or new traffic signs, and the reliability of pseudo-labels can be improved as the model gradually adapts.

[0191] The vehicle performs local semi-supervised training using the high-confidence traffic sign pseudo-label images obtained through screening, and performs several rounds of gradient updates on the traffic sign recognition model according to the local model update method described in Example 1, to obtain the updated local model parameters or the parameter differences relative to the initial global model.

[0192] After local training is completed, the vehicle calculates the sample freshness based on the proportion of newly acquired images to the total number of local images; it counts the number of samples for each traffic sign category based on high-confidence pseudo-labels and determines the sample scarcity of different traffic sign categories; and it calculates the information increment based on the change in training loss before and after local training.

[0193] For example, when a vehicle passes through a school, construction zone, or special speed limit area, the data it collects may contain more warning signs, temporary traffic signs, or special speed limit signs that are less frequently collected by other vehicles. The base station server can determine whether the data from that vehicle can supplement the current global model's learning ability for scarce traffic sign categories based on the scarcity information of the samples uploaded by the vehicle.

[0194] The vehicle also generates its own performance report based on the load of the onboard computing unit, remaining battery power, wireless channel quality, uplink transmission rate, local training completion status, and model update stability. The vehicle uploads the data utility report and its own performance report to the base station server, while temporarily storing local model updates locally on the vehicle.

[0195] After receiving reports uploaded by multiple vehicles, the base station server comprehensively analyzes the sample freshness, traffic sign category scarcity, information increment, on-board computing resources, communication conditions, and energy consumption status of each vehicle, and selects vehicles with high data value and the conditions for model upload from the candidate vehicles to participate in this round of federated aggregation.

[0196] After receiving the model update upload instruction from the base station server, the selected vehicles upload their local traffic sign recognition model parameters or model parameter differences to the base station server. The base station server performs weighted aggregation on the model updates uploaded by multiple selected vehicles to obtain the aggregated global traffic sign recognition model.

[0197] After federated aggregation is complete, the base station server uses its held real-labeled traffic sign images to fine-tune and calibrate the aggregated global model. This corrects potential class biases that may arise during vehicle-side pseudo-label training and suppresses the accumulation of erroneous pseudo-labels over multiple training rounds. Once fine-tuned, the base station server redistributes the final calibrated global model to each vehicle.

[0198] The vehicle uses the updated global traffic sign recognition model to identify subsequently acquired road images, outputs the traffic sign category and the corresponding recognition result, and uses the recognition result for driving assistance, road environment perception, vehicle warning or autonomous driving decision-making.

[0199] Through the above process, each vehicle can collaboratively update the traffic sign recognition model using unlabeled traffic images continuously collected by different vehicles without uploading the original road images to the base station server. This improves the model's ability to continuously adapt to different road areas, weather conditions, lighting environments, newly added traffic signs, and scarce traffic sign categories.

[0200] Since the traffic sign image recognition in this embodiment and Embodiment 1 both belong to multi-class classification tasks, the loss function and model update formula used in global model warm-up, unlabeled image prediction, dynamic pseudo-label generation, local semi-supervised training, federated aggregation, and server fine-tuning calibration are the same as in Embodiment 1, and will not be repeated in this embodiment. Embodiment 2 only specifically limits the input samples in Embodiment 1 to traffic sign images collected by vehicles, the edge devices to vehicles, and the server to base station servers.

[0201] Example 3

[0202] This embodiment 3 provides a non-transitory computer-readable storage medium for storing computer instructions. When the computer instructions are executed by a processor, they implement the data value-aware federated continuous learning method based on dynamic pseudo-labels as described above. The method includes:

[0203] The global model is pre-trained using the currently available labeled data obtained from the server, so that the global model has basic category recognition ability and initial feature expression ability, reducing the risk of initial misjudgment in the subsequent process of generating pseudo-labels for unlabeled samples.

[0204] The pre-trained global model is used to predict unlabeled data acquired by edge devices, and the category probability distribution and prediction confidence of each unlabeled sample are obtained. The high-confidence pseudo-label samples are selected to form a local training dataset for subsequent local semi-supervised training.

[0205] The pre-trained global model is semi-supervised by using the local training set composed of the obtained pseudo-label samples; in this process, the model is updated with gradients for several rounds using the selected high-confidence pseudo-label samples to obtain the updated local model parameters or model parameter differences.

[0206] After completing local semi-supervised training, a comprehensive evaluation is performed on the data used in the current training round, the local training process, and the operating status of the edge device itself, and a data utility report and an edge device self-performance report are generated. The data utility report describes the potential contribution of local data to the current global model update; the self-performance report describes the resource status of the device after completing local training and the feasibility of uploading model updates.

[0207] Based on the data utility report and the self-performance report, edge devices that can improve the model's adaptability and have the feasibility of uploading are selected to form the model update upload device set for the current round. A model update upload instruction is sent to the selected devices, which is used to notify the devices to upload the model update parameters obtained after local training.

[0208] The obtained model update parameters are aggregated according to the federated aggregation algorithm to generate a global model. The global model is then fine-tuned and calibrated using labeled data to obtain the final calibrated global model. This model is then redistributed to all edge devices as the initial model for the next round of training.

[0209] Example 4

[0210] This embodiment 4 provides a computer device, including a memory and a processor, wherein the processor and the memory communicate with each other, and the memory stores program instructions that can be executed by the processor. The processor calls the program instructions to execute the data value-aware federated continuous learning method based on dynamic pseudo-labels as described above, the method including:

[0211] The global model is pre-trained using the currently available labeled data obtained from the server, so that the global model has basic category recognition ability and initial feature expression ability, reducing the risk of initial misjudgment in the subsequent process of generating pseudo-labels for unlabeled samples.

[0212] The pre-trained global model is used to predict unlabeled data acquired by edge devices, and the category probability distribution and prediction confidence of each unlabeled sample are obtained. The high-confidence pseudo-label samples are selected to form a local training dataset for subsequent local semi-supervised training.

[0213] The pre-trained global model is semi-supervised by using the local training set composed of the obtained pseudo-label samples; in this process, the model is updated with gradients for several rounds using the selected high-confidence pseudo-label samples to obtain the updated local model parameters or model parameter differences.

[0214] After completing local semi-supervised training, a comprehensive evaluation is performed on the data used in the current training round, the local training process, and the operating status of the edge device itself, and a data utility report and an edge device self-performance report are generated. The data utility report describes the potential contribution of local data to the current global model update; the self-performance report describes the resource status of the device after completing local training and the feasibility of uploading model updates.

[0215] Based on the data utility report and the self-performance report, edge devices that can improve the model's adaptability and have the feasibility of uploading are selected to form the model update upload device set for the current round. A model update upload instruction is sent to the selected devices, which is used to notify the devices to upload the model update parameters obtained after local training.

[0216] The obtained model update parameters are aggregated according to the federated aggregation algorithm to generate a global model. The global model is then fine-tuned and calibrated using labeled data to obtain the final calibrated global model. This model is then redistributed to all edge devices as the initial model for the next round of training.

[0217] Example 5

[0218] This embodiment 5 provides an electronic device, including: a processor, a memory, and a computer program; wherein, the processor is connected to the memory, and the computer program is stored in the memory. When the electronic device is running, the processor executes the computer program stored in the memory to cause the electronic device to execute instructions for implementing the data value-aware federated continuous learning method based on dynamic pseudo-labels as described above. The method includes:

[0219] The global model is pre-trained using the currently available labeled data obtained from the server, so that the global model has basic category recognition ability and initial feature expression ability, reducing the risk of initial misjudgment in the subsequent process of generating pseudo-labels for unlabeled samples.

[0220] The pre-trained global model is used to predict unlabeled data acquired by edge devices, and the category probability distribution and prediction confidence of each unlabeled sample are obtained. The high-confidence pseudo-label samples are selected to form a local training dataset for subsequent local semi-supervised training.

[0221] The pre-trained global model is semi-supervised by using the local training set composed of the obtained pseudo-label samples; in this process, the model is updated with gradients for several rounds using the selected high-confidence pseudo-label samples to obtain the updated local model parameters or model parameter differences.

[0222] After completing local semi-supervised training, a comprehensive evaluation is performed on the data used in the current training round, the local training process, and the operating status of the edge device itself, and a data utility report and an edge device self-performance report are generated. The data utility report describes the potential contribution of local data to the current global model update; the self-performance report describes the resource status of the device after completing local training and the feasibility of uploading model updates.

[0223] Based on the data utility report and the self-performance report, edge devices that can improve the model's adaptability and have the feasibility of uploading are selected to form the model update upload device set for the current round. A model update upload instruction is sent to the selected devices, which is used to notify the devices to upload the model update parameters obtained after local training.

[0224] The obtained model update parameters are aggregated according to the federated aggregation algorithm to generate a global model. The global model is then fine-tuned and calibrated using labeled data to obtain the final calibrated global model. This model is then redistributed to all edge devices as the initial model for the next round of training.

[0225] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0226] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0227] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0228] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment, whereby a series of operational steps are performed to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0229] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that, based on the technical solutions disclosed in the present invention, various modifications or variations that can be made by those skilled in the art without creative effort should be included within the scope of protection of the present invention.

Claims

1. A data value-aware federated continuous learning method based on dynamic pseudo-labels, characterized in that, include: The global model is pre-trained using the currently available labeled data obtained from the server, which enables the global model to have basic category recognition capabilities and initial feature expression capabilities, thereby reducing the risk of initial misjudgment in the subsequent process of generating pseudo-labels for unlabeled samples. The pre-trained global model is used to predict unlabeled data acquired by edge devices, and the category probability distribution and prediction confidence of each unlabeled sample are obtained. The high-confidence pseudo-label samples are selected to form a local training dataset for subsequent local semi-supervised training. The pre-trained global model is semi-supervised by using the local training set composed of the obtained pseudo-label samples; among them, the model is updated several times by using the selected high-confidence pseudo-label samples to obtain the updated local model parameters or model parameter differences. After completing local semi-supervised training, a comprehensive evaluation is performed on the data used in the current training round, the local training process, and the operating status of the edge device itself, and a data utility report and an edge device self-performance report are generated. The data utility report describes the potential contribution of local data to the current global model update; the self-performance report describes the resource status of the device after completing local training and the feasibility of uploading model updates. Based on the data utility report and the self-performance report, edge devices that can improve the model's adaptability and have the feasibility of uploading are selected to form the model update upload device set for the current round. A model update upload instruction is sent to the selected devices, which is used to notify the devices to upload the model update parameters obtained after local training. The obtained model update parameters are aggregated according to the federated aggregation algorithm to generate a global model. The global model is then fine-tuned and calibrated using labeled data to obtain the final calibrated global model. This model is then redistributed to all edge devices as the initial model for the next round of training.

2. The data value-aware federated continuous learning method based on dynamic pseudo-labels according to claim 1, characterized in that, The server uses labeled data to perform pre-training of the global model, including: The server retrieves the existing labeled dataset for the current frame: ； in, Indicates as of the date The frame server has accumulated labeled data. Indicates the first The frame was updated and confirmed to contain reliable labeled data; The server obtains a labeled dataset. Randomly select training batches: ； in, Indicates the current training batch. This represents the input samples and their corresponding true category labels in the batch; The server will send the sample Input the current global model, obtain the class prediction probabilities through forward propagation, and construct the server-side supervised cross-entropy loss function: ； in, Represents global model parameters. Indicates the number of samples in the training batch. The model parameters are: Time Sample Belongs to the real category The predicted probability; The server uses the final global model from the previous frame as the initial model for preheating and training the current frame. ； in, Indicates the first Global model parameters after frame training is complete. The server uses the stochastic gradient descent algorithm to execute... Secondary parameter update: ； in, This represents the server-side learning rate. This indicates that the supervised loss function has respect to the model parameters. The gradient; After completing the warm-up training, the server obtains the initial global model for the current frame: 。 3. The data value-aware federated continuous learning method based on dynamic pseudo-labels according to claim 1, characterized in that, After receiving the trained global model from the server, the edge device uses the global model to predict local unlabeled data, specifically including the following steps: Get the local unlabeled dataset available for the current round: ； in, Indicates edge device Unlabeled data accumulated or retained during historical periods, Indicates edge device In the Newly acquired unlabeled data in each data evolution frame; Load global model parameters issued by the server and for unlabeled samples Weak data augmentation is performed to obtain the processed samples. ; Processed sample Input the global model, and through the model's feature extraction and classification process, obtain the predicted probability distribution of each sample belonging to each category: ； in, Indicates as of the date The total number of categories that appear cumulatively in each data evolution frame system. Indicates edge device In the The first data evolution frame Wheel of Samples Predicted as category The probability of; The category corresponding to the highest predicted probability is selected as the unlabeled sample. The candidate pseudo-labels are selected, and the maximum predicted probability is used as the prediction confidence of the sample: ； ； in, Indicates sample Candidate pseudo-tags, This indicates the prediction confidence level corresponding to the candidate pseudo-label; Prediction confidence Dynamic pseudo-label threshold corresponding to the current training round The comparison is performed; when the prediction confidence is higher than the dynamic threshold, the sample is considered to have high credibility, and the category corresponding to the highest probability is used as the pseudo-label of the sample; when the prediction confidence is lower than the dynamic threshold, the sample is considered to have insufficient credibility at present, and the sample will not be used in this round of training.

4. The data value-aware federated continuous learning method based on dynamic pseudo-labels according to claim 3, characterized in that, During local training, the edge device uses the selected high-confidence pseudo-label samples to calculate the local semi-supervised training loss, and performs several gradient updates on the local model in the following manner: ； in, Indicates edge device In the Round completed The model parameters after the next local update. Indicates the local learning rate. This represents the gradient of the local semi-supervised training loss with respect to the model parameters. This indicates the total number of iterations during local training.

5. The data value-aware federated continuous learning method based on dynamic pseudo-labels according to claim 1, characterized in that, The data utility report is used to describe the potential contribution of local data to the current global model update, and includes at least three types of information: sample freshness, sample scarcity, and information increment. Sample freshness is used to measure the proportion of newly collected samples in the current data of edge devices. Sample scarcity identifies local long-tail categories and imbalanced category distribution; The information increment is calculated based on the relative decrease in pseudo-label training loss before and after local training; The self-performance report includes: Computational resources used to describe the device’s current processor, graphics processing unit, memory, storage space, and remaining computing power available for subsequent training or model uploading; Communication status used to describe the device's current uplink rate, communication latency, channel stability, network congestion, and the feasibility of uploading model updates; Energy status is used to describe the device's current battery level, energy budget, and ability to continue performing communication uploads or subsequent training tasks; This describes the status of training results, including whether local training on the device is complete, changes in training loss, the magnitude of model updates, and the stability of the local training process.

6. The data value-aware federated continuous learning method based on dynamic pseudo-labels according to claim 1, characterized in that, The server processes local model updates uploaded by edge devices, including: The server performs weighted aggregation on the local model parameters uploaded by the selected edge devices to obtain the aggregated global model: ； in, This represents the global model parameters after federated aggregation. Indicates the first Frame number A collection of edge devices participating in the aggregation. Indicates edge device The corresponding aggregate weights, Indicates edge device Model parameters after local training is complete; The server uses its labeled data to fine-tune the aggregated global model, calculates supervised loss based on the true labels, and performs several gradient updates on the model parameters: ； in, This indicates that the server has completed the first... The global model parameters after this fine-tuning. This indicates that the learning rate is fine-tuned on the server side. This represents the gradient of the loss with respect to the model parameters. This indicates the number of fine-tuning iterations.

7. A data value-aware federated continuous learning system based on dynamic pseudo-labels, characterized in that, include: The preheating training module is used to preheat the global model using the currently available labeled data obtained from the server, so that the global model has basic category recognition ability and initial feature expression ability, reducing the risk of initial misjudgment in the subsequent process of generating pseudo-labels for unlabeled samples. The prediction module is used to predict unlabeled data acquired by edge devices using a pre-trained global model, obtain the category probability distribution and prediction confidence of each unlabeled sample, and select high-confidence pseudo-label samples to form a local training dataset for subsequent local semi-supervised training. The semi-supervised training module is used to perform semi-supervised training on the pre-trained global model using a local training set composed of acquired pseudo-label samples. Specifically, the model is updated with gradients for several rounds using selected high-confidence pseudo-label samples to obtain the updated local model parameters or model parameter differences. The evaluation module is used to comprehensively evaluate the data used in the current training round, the local training process, and the operating status of the edge device itself after the local semi-supervised training is completed, and to generate a data utility report and an edge device self-performance report. The data utility report is used to describe the potential contribution of local data to the current global model update. The self-performance report is used to describe the resource status of the device after completing local training and the feasibility of uploading the model update. The filtering module is used to select edge devices that can improve the model's adaptability and have the feasibility of uploading based on the data utility report and the self-performance report, forming a set of model update upload devices for the current round, and sending model update upload instructions to the selected devices. The instructions are used to notify the devices to upload the model update parameters obtained after local training. The aggregation and calibration module is used to aggregate the obtained model update parameters according to the federated aggregation algorithm to generate a global model. The global model is then fine-tuned and calibrated using labeled data to obtain the final calibrated global model, which serves as the initial model for the next round of training.

8. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium is used to store computer instructions, which, when executed by a processor, implement the data value-aware federated continuous learning method based on dynamic pseudo-labels as described in any one of claims 1-6.

9. A computer device, characterized in that, The system includes a memory and a processor, the processor and the memory communicating with each other, the memory storing program instructions executable by the processor, and the processor calling the program instructions to execute the data value-aware federated continuous learning method based on dynamic pseudo-labels as described in any one of claims 1-6.

10. An electronic device, characterized in that, include: The device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to cause the electronic device to execute instructions that implement the data value-aware federated continuous learning method based on dynamic pseudo-labels as described in any one of claims 1-6.

Images