Distributed data collaboration method and device based on federal control

Abstract

The invention discloses a distributed data collaboration method and device based on federal control, and belongs to the field of computer technology and machine learning, and the method comprises the steps: obtaining network configuration parameters, and generating initial credit parameters for a plurality of entities; acquiring local data, performing feature extraction to generate feature data, generating a dynamic mask in combination with the initial credit parameter and the network state parameter acquired in real time, and processing the feature data to obtain mask feature data; and performing weighted aggregation on the mask feature data received from each entity according to the initial credit parameters, generating a global collaboration result, performing evaluation, generating an effect score, combining the obtained network performance parameters, and generating an updated collaboration strategy. According to the method, initial and joint credit is established, dynamic mask processing data is generated in combination with a network state, credit weighted aggregation is performed, and self-adaption, high robustness and high efficiency of distributed data collaboration in a complex environment are realized based on a collaboration effect closed-loop updating strategy.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of computer technology and machine learning, in particular to a distributed data collaboration method and device based on federal control. BACKGROUND

[0002] Distributed data collaboration is a key technology in the field of artificial intelligence and large-scale machine learning, which allows multiple geographically dispersed entities to jointly complete data analysis or model training tasks without directly sharing local raw data. This paradigm, especially under the framework of federated learning, reduces data privacy, security and data silos by performing calculations locally and uploading intermediate results to a central server for aggregation, and has broad application prospects in sensitive fields such as finance and medicine.

[0003] Existing distributed data collaboration methods usually adopt a relatively fixed collaboration mode. For example, in federated learning, the central server periodically distributes the global model to selected participants, each participant uploads the model update after local data training, and the server then simply averages or weights the aggregation according to the data volume. These methods can work effectively in ideal environments, but their collaboration strategies are often static, and the selection of participants and the way of data aggregation weight distribution are relatively single.

[0004] However, the above existing technical solutions have obvious technical defects in actual deployment. First, they generally lack differentiated consideration of the internal attributes and reliability of participating entities, treating all entities equally, which can lead to low-quality updates from entities with low computing power or unstable behavior, thus reducing the efficiency of global collaboration and the accuracy of the final results. Second, existing solutions lack awareness of dynamic changes in network environment, and cannot flexibly adjust data transmission and aggregation strategies according to real-time network latency, bandwidth, etc., which can easily lead to a decline in collaboration efficiency or task failure in unstable network conditions. Finally, existing technologies lack an effective closed-loop feedback mechanism to dynamically optimize and adjust subsequent collaboration strategies based on historical collaboration results. SUMMARY

[0005] To solve the above problems, the present application provides a distributed data collaboration method and device based on federal control, which adopts the technical solutions of establishing initial credit, joint credit and network state to generate dynamic mask processing data, credit weighted aggregation and closed-loop update of collaboration strategy based on collaboration effect, which can realize adaptive, high robustness and high efficiency of distributed data collaboration in complex environment.

[0006] The above objectives can be achieved by the following solutions:

[0007] A distributed data collaboration method and device based on federal control, comprising: acquiring network configuration parameters for defining a network topology, generating initial credit parameters for a plurality of entities based on the network configuration parameters; acquiring local data and performing feature extraction to generate feature data, generating a dynamic mask by combining the initial credit parameters and real-time acquired network state parameters, and processing the feature data using the dynamic mask to obtain mask feature data; weighting and aggregating the mask feature data received from each entity according to the initial credit parameters to generate a global collaboration result; evaluating the effectiveness of the global collaboration result to generate an effect score, and combining the effect score with acquired network performance parameters to generate an updated collaboration strategy for adjusting a preset dynamic mask generation strategy and a preset credit allocation rule.

[0008] Optionally, the generating initial credit parameters for a plurality of entities based on the network configuration parameters comprises: extracting entity attribute parameters of the plurality of entities from the network configuration parameters; performing credit quantification processing based on the entity attribute parameters to calculate initial credit values for each entity; and integrating the initial credit values of all entities to form an initial credit parameter.

[0009] Optionally, the combining the initial credit parameters with real-time acquired network state parameters to generate a dynamic mask comprises: real-time monitoring of communication link state to acquire the network state parameters including network delay and bandwidth parameters; performing joint processing based on the network state parameters and the initial credit parameters to calculate a strength coefficient of the dynamic mask; and performing random sampling based on the strength coefficient of the dynamic mask to generate a dynamic mask.

[0010] Optionally, the obtaining mask feature data comprises: performing numerical scaling of the feature data according to the strength coefficient of the dynamic mask to generate scaled feature data; and performing bitwise XOR operation on the scaled feature data and the dynamic mask to obtain mask feature data.

[0011] Optionally, the generating a global collaboration result comprises: assigning weights to the mask feature data of each entity according to the initial credit parameters to generate a weight set; and performing weighted fusion based on the mask feature data and the weight set to calculate a global collaboration result.

[0012] Optionally, the evaluation based on the effectiveness of the global collaboration result, the generation of the effect score, the combination of the effect score and the obtained network performance parameters, and the generation of the updated collaboration strategy for adjusting the preset dynamic mask generation strategy and the preset credit allocation rule include: continuously monitoring the communication state of the federal control network to obtain network performance parameters including delay, packet loss rate and throughput; performing consistency check and error analysis on the global collaboration result, and combining historical collaboration effect data to generate an effect score reflecting the effectiveness of the current collaboration; calculating a strategy adjustment amount according to the effect score and the network performance parameters through a preset multi-objective optimization function; modifying the complexity control parameter in the dynamic mask generation strategy according to the strategy adjustment amount, and synchronously updating the reward and punishment amplitude in the credit allocation rule to generate an updated collaboration strategy.

[0013] Optionally, the calculation of the intensity coefficient of the dynamic mask includes: performing quantitative analysis based on the initial credit parameter to determine the priority level of the entity; performing joint mapping based on the priority level and the network state parameter to obtain a mask reference value; and dynamically adjusting the mask reference value according to the network state parameter to generate the intensity coefficient of the dynamic mask.

[0014] Optionally, the generation of the weight set includes: calculating the credit weight of each entity based on the initial credit parameter; obtaining network state parameters for quantifying communication quality, and adjusting the credit weight of each entity according to the network state parameters to generate adaptive weight values; and collecting the adaptive weight values of all entities to form a weight set.

[0015] Optionally, the quantitative analysis based on the initial credit parameter to determine the priority level of the entity includes: inputting the initial credit parameter into a credit grading unit to divide a plurality of credit intervals according to the size of the credit value; matching the corresponding priority label according to the credit interval to which each entity belongs; obtaining historical collaboration behavior data, and fine-tuning the priority label in combination with the historical collaboration behavior data to form a priority level.

[0016] Based on the same inventive concept, the application also provides a distributed data collaboration device based on federal control, comprising: a network credit initialization module for obtaining network configuration parameters for defining network topology, generating initial credit parameters for multiple entities based on the network configuration parameters; a dynamic mask generation module for obtaining local data and performing feature extraction to generate feature data, generating a dynamic mask by combining the initial credit parameters with real-time obtained network state parameters, and processing the feature data using the dynamic mask to obtain masked feature data; a weighted aggregation collaboration module for weighting and aggregating the masked feature data received from each entity according to the initial credit parameters, generating a global collaboration result; a strategy evaluation update module for evaluating based on the effectiveness of the global collaboration result, generating an effect score, and generating an updated collaboration strategy for adjusting the preset dynamic mask generation strategy and the preset credit allocation rule in combination with the obtained network performance parameters.

[0017] Compared with the prior art, the application has the following advantages:

[0018] 1. The application improves the high adaptability and robustness of the system by constructing a closed-loop feedback collaboration strategy update mechanism. This method can continuously evaluate the effectiveness of the global collaboration result and monitor the network performance, dynamically adjust the dynamic mask generation strategy and the credit allocation rule according to the evaluation result, so that the system can actively adapt to the dynamic changes of the participating entity behavior and the network environment, and maintain the stability and efficiency of the collaboration task in complex and variable environment.

[0019] 2. The application improves the accuracy and reliability of the global collaboration result by introducing a dynamic weighting mechanism based on entity attributes and network state. This method does not treat all participating entities equally, but weights and aggregates the data contributed by each entity based on the initial credit parameters, and dynamically adjusts the weight in combination with the real-time network state, ensuring that high-credit and high-communication-quality entities play a leading role in collaboration, reducing the negative impact of low-quality data and network jitter on the final result.

[0020] 3. The application realizes fine and secure control of data contribution by designing a dynamic mask generation and application method combined with initial credit and network state. This dynamic mask can adaptively scale and perturb the uploaded feature data according to the credibility of each entity and the network environment it is in, which not only dynamically adjusts the amount of information transmission according to the network bandwidth on the premise of ensuring the collaboration effect, but also provides a certain degree of privacy protection for the data sharing process, enhancing the security of the entire distributed collaboration system.

[0021] Other features and advantages of the present application will be set forth in the descriptions that follow, and in part will be apparent from the description, or can be learned by practice of the application. The purposes and other advantages of the present application will be realized and attained by the structures particularly pointed out in the written description and claims hereof as well as the appended drawings. BRIEF DESCRIPTION OF DRAWINGS

[0022] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings described below are only some embodiments of the present application, and the person of ordinary skill in the art can obtain other drawings according to these drawings without any creative effort.

[0023] Figure 1 is a flowchart of a distributed data collaboration method based on federal control according to an embodiment of the present application.

[0024] Figure 2 is a schematic diagram of the relationship between dynamic mask strength coefficients and influence factors according to an embodiment of the present application.

[0025] Figure 3 is a strategy evaluation mechanism diagram according to an embodiment of the present application.

[0026] Figure 4 is an entity priority level determination schematic diagram according to an embodiment of the present application.

[0027] Figure 5 is a structural schematic diagram of a distributed data collaboration device based on federal control according to an embodiment of the present application. DETAILED DESCRIPTION

[0028] In order to make the purposes, technical solutions and advantages of the embodiments of the present application more clear, the following will combine the drawings in the embodiments of the present application to clearly and completely explain the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by the person of ordinary skill in the art without any creative effort are within the protection scope of the present application.

[0029] With reference to Figure 1 , one embodiment of the present application proposes a distributed data collaboration method based on federal control, which adopts the technical solutions of establishing initial credit, generating dynamic mask processing data by joint credit and network state, performing credit weighted aggregation, and updating collaboration strategy based on collaboration effect closed loop, and can realize self-adaptation, high robustness and high efficiency of distributed data collaboration in complex environment.

[0030] The method according to the embodiment specifically includes:

[0031] obtaining network configuration parameters for defining a network topology, generating initial credit parameters for a plurality of entities based on the network configuration parameters;

[0032] obtaining local data and performing feature extraction to generate feature data, generating a dynamic mask by combining the initial credit parameters and real-time obtained network state parameters, and processing the feature data using the dynamic mask to obtain masked feature data;

[0033] performing weighted aggregation on the masked feature data received from each entity according to the initial credit parameters to generate a global collaborative result;

[0034] performing evaluation based on the effectiveness of the global collaborative result to generate an effect score, and generating an updated collaborative strategy for adjusting a preset dynamic mask generation strategy and a preset credit allocation rule by combining the effect score and obtained network performance parameters.

[0035] Specifically, first, an initial credit system is established by analyzing the inherent configuration parameters of network entities, and each participant is given a basic initial credit parameter. In the collaborative process, instead of directly sharing raw features, each entity applies a dynamically generated mask for processing; this dynamic mask depends on both the entity's static initial credit and the dynamically changing network state parameters, thereby achieving intelligent adjustment of data contribution. In the data aggregation stage, the initial credit parameters are used to weight and fuse the masked feature data submitted by each party, ensuring that high-credit entities have greater influence on the final result. Most importantly, this method introduces a feedback control loop that evaluates the effectiveness of the global collaborative result after each round of collaboration and combines the effect score with network performance parameters to drive a strategy update mechanism, thereby dynamically adjusting the generation strategy of future dynamic masks and the allocation rules of the credit system, enabling the entire collaborative system to continuously optimize and evolve. By introducing a federal control mechanism, this method improves the intelligence, security, and efficiency of the distributed data collaboration process. The dynamic mask generated based on the initial credit and real-time network state, as well as the subsequent weighted aggregation, improves the accuracy and robustness of the global collaborative result, effectively reducing the interference of low-quality nodes and adverse network environments on the collaborative task. More importantly, through strategy module updates, the entire collaborative framework is given the ability to learn and continuously optimize, dynamically adapting to complex and changing network environments and participant behavior, automatically finding and maintaining the optimal balance point between collaboration effectiveness and data security, thereby ensuring the stability, efficiency, and reliability of distributed data collaboration in the long run.

[0036] Optionally, the generating initial credit parameters for a plurality of entities based on the network configuration parameters comprises:

[0037] Extract entity attribute parameters of multiple entities from the network configuration parameters;

[0038] Credit quantification is performed based on the entity attribute parameters to calculate an initial credit value for each entity.

[0039] The initial credit values ​​of all entities are combined to form initial credit parameters.

[0040] Specifically, firstly, entity attribute parameters related to each entity are extracted from the network configuration parameters used to define the network topology. These entity attribute parameters are a collection describing the entity's static capabilities, including but not limited to hardware computing capabilities such as CPU performance and memory capacity, network connectivity capabilities such as nominal bandwidth, and security authentication levels and historical operational stability records. Next, credit quantification processing is performed on the extracted entity attribute parameters to calculate an initial credit value for each entity. This quantification process aims to transform multi-dimensional, heterogeneous attributes into a unified credit scalar. For calculating the first... Initial credit value of an entity ,have:

[0041] ;

[0042] in, Representing the The entity in the first The quantized values ​​of the entity attribute parameters are obtained by extracting the original attribute values ​​of the entity from the network configuration parameters and converting these heterogeneous attributes into values ​​with uniform dimensions by performing max-min normalization. This represents the total number of entity attribute parameters. Representing the The weight coefficients corresponding to the attribute parameters of each entity are set based on the analysis of the contribution of different attributes to the collaborative effect in historical collaborative tasks, such as regression analysis or feature importance assessment. This reflects the importance of different attributes in the credit assessment system and satisfies the condition that the sum of all weight coefficients is 1. Through this calculation, each entity obtains an initial credit value that reflects its comprehensive capabilities. Finally, the initial credit values ​​of all entities are integrated to form a structured set of initial credit parameters, such as a credit vector or mapping table. These initial credit parameters will serve as the objective basis for subsequent steps such as dynamic mask generation and weighted aggregation collaboration. The initial credit parameters generated in this way provide an objective and differentiated starting point for distributed data collaboration in federated control. By conducting initial trust assessments on the inherent and verifiable attributes of entities, the rationality of resource allocation and collaborative weight distribution is improved. Entities with strong computing power, good network conditions, and high security levels can be identified and preferentially selected from the outset, laying the foundation for subsequent collaborative tasks and improving the quality of the overall collaborative results.

[0043] Optionally, the joint processing of the initial credit parameter and the real-time acquired network state parameter to generate a dynamic mask comprises:

[0044] The real-time monitoring of the communication link state acquires the network state parameter containing network delay and bandwidth parameters;

[0045] Based on the joint processing of the network state parameter and the initial credit parameter, the intensity coefficient of the dynamic mask is calculated;

[0046] Based on the intensity coefficient of the dynamic mask, random sampling is performed to generate a dynamic mask;

[0047] Specifically, first, through the built-in network monitoring unit, the communication link state between each entity and other nodes is monitored in real time, which is usually achieved by sending probe packets or analyzing feedback information of data transmission protocol, so as to acquire real-time network state parameters, the key indicators including network delay reflecting data transmission delay and network bandwidth parameter representing link transmission capacity. After acquiring these parameters, the network state parameter is jointly processed with the initial credit parameter to calculate an intensity coefficient of a dynamic mask. The intensity coefficient of the dynamic mask is a quantitative indicator that integrates the credibility of the entity and the current network communication quality. After calculating the intensity coefficient of the dynamic mask, random sampling is performed to generate the final dynamic mask. The dynamic mask is a data structure with the same dimension as the feature data, and the generation probability of its elements is related to the intensity coefficient of the dynamic mask. For example, each bit or element value in the dynamic mask can be drawn from a Bernoulli distribution, and the probability of taking the value 1 is equal to the intensity coefficient of the dynamic mask. The generation of the dynamic mask is no longer static or unchangeable, but can respond in real time to the changes of the intrinsic properties of the participating entities and the external network environment. For example, Figure 2 As shown in the figure, the three-dimensional scatter plot shows how the dynamic mask intensity coefficient is jointly affected by the initial credit parameter of the entity and the real-time network quality index. Each point in the figure represents the state of an entity at a certain time, and the darker the color, the higher the mask intensity coefficient. As can be seen from the figure, when the initial credit parameter is high and the network quality index is also high, i.e. in the front right area of the figure, the dynamic mask intensity coefficient tends to be low. Conversely, when the credit is low or the network quality is poor, i.e. in the left rear area of the figure, the dynamic mask intensity coefficient is relatively high. When a high-credit entity is in an excellent network environment, a low-intensity mask is generated, allowing it to contribute higher-fidelity feature data, thereby improving the accuracy of the global collaborative result. Conversely, when the entity credit is low or the network condition deteriorates, a higher-intensity mask is generated, which can not only reduce the negative impact of low-quality data on the global result, but also adapt to limited bandwidth by reducing the amount of effective information, improving the timeliness of data transmission and enhancing the adaptability and robustness of the entire distributed data collaboration in complex and variable environments.

[0048] Optionally, the obtaining the mask feature data comprises:

[0049] numerically scaling the feature data according to the intensity coefficient of the dynamic mask, to generate scaled feature data;

[0050] performing a bitwise XOR operation between the scaled feature data and the dynamic mask, to obtain mask feature data;

[0051] Specifically, first, the feature data extracted locally is numerically scaled according to the intensity coefficient of the dynamic mask obtained above. Local data is obtained from a storage unit or a real-time data acquisition interface of the entity, which includes but is not limited to sensor time series data, image data, text logs or model update gradients. A feature extraction algorithm is used to process the local data to extract a feature vector representing the essential features of the data, as the feature data. For example, principal component analysis can be used for dimensionality reduction, or a convolutional neural network can be used for deep feature extraction of image data. The numerical scaling operation is to multiply the intensity coefficient of the dynamic mask with each element in the feature data, to generate scaled feature data. This operation adjusts the overall magnitude of the feature data in numerical value according to real-time trust and network evaluation. Subsequently, a bitwise XOR operation is performed between the scaled feature data and the dynamic mask also generated in the previous step. Since the bitwise XOR operation requires the operands to be in binary format, the scaled feature data in floating-point or integer type needs to be converted into a binary bit string through quantization or encoding before the operation, and the dynamic mask itself is a binary bit string or matrix. Finally, the converted scaled feature data and the dynamic mask are subjected to a bitwise XOR operation to obtain the final mask feature data. By combining the numerical scaling and bitwise XOR operations, the original feature data is subjected to double processing, and the numerical scaling step adjusts the contribution weight of the feature data according to the intensity coefficient of the dynamic mask, which is equivalent to a pre-weighting before data uploading, so that the data contributed by entities with high credit and good network conditions has greater numerical influence, and vice versa, thus optimizing the quality of data fusion. The subsequent bitwise XOR operation, without completely destroying the data structure, performs a controllable disturbance on the data, which plays a certain role in privacy protection and data confusion. This double processing mechanism not only improves the accuracy and robustness of the collaborative result, but also takes into account the security of the data in a distributed environment.

[0052] Optionally, the generating the global collaborative result comprises:

[0053] assigning a weight to the mask feature data of each entity according to the initial credit parameter, to generate a weight set;

[0054] performing weighted fusion based on the mask feature data and the weight set, to calculate a global collaborative result.

[0055] Specifically, firstly, based on the initial credit parameters, a corresponding weight is assigned to each piece of mask feature data. All these weights together form a weight set. The logic for weight allocation is that the higher the initial credit parameters of an entity, the greater the proportion of its contributed data in the aggregation process. For example, the... The weights of each entity can be calculated by normalizing its initial credit value to the total credit value of all participating entities. Next, a weighted fusion operation is performed, multiplying the masked feature data of each entity by its corresponding weight, and then summing all the products to calculate the final global collaborative result. (The calculation of the global collaborative result is described in the original text.) and weight set assigned to the first Weight of each entity ,have:

[0056] ;

[0057] in, Then from the first The mask feature data received by each entity; This is the sum of the initial credit scores of all participating entities. This calculation allows data scattered across various entities to be organically combined according to their credibility, forming a unified and representative collaborative output. By introducing a weighting mechanism based on initial credit parameters, the data aggregation process is no longer a simple, undifferentiated averaging, but an intelligent and biased fusion. This plays a more decisive role in the formation of the global collaborative result, improving its accuracy and reliability. Simultaneously, it suppresses the negative interference of low-quality or unreliable data from low-credit entities on the final result, enhancing the robustness of the entire distributed collaboration and the value of the final output.

[0058] Optionally, the evaluation of the effectiveness of the global collaboration results, the generation of an effectiveness score, and the combination of the effectiveness score with the obtained network performance parameters to generate an updated collaboration strategy for adjusting the preset dynamic mask generation strategy and the preset credit allocation rules include:

[0059] Continuously monitor the communication status of the federal control network and obtain network performance parameters including latency, packet loss rate, and throughput;

[0060] The global collaboration results are subjected to consistency verification and error analysis. Combined with historical collaboration effect data, an effect score reflecting the effectiveness of this collaboration is generated.

[0061] Based on the performance score and the network performance parameters, the strategy adjustment amount is calculated using a preset multi-objective optimization function;

[0062] According to the policy adjustment amount, the complexity control parameter in the dynamic mask generation strategy is modified, and the reward and punishment amplitude in the credit allocation rule is synchronously updated, to generate an updated cooperative strategy.

[0063] Specifically, first, the communication state is continuously monitored by the monitoring agent deployed in the federal control network, so as to obtain a set of network performance parameters in real time. The set of network performance parameters is multidimensional, and specifically includes delay for measuring round-trip time of data, packet loss rate for reflecting data transmission reliability, and throughput for characterizing effective data transmission rate. At the same time, the effectiveness of the global cooperative result generated by the last round of cooperation is deeply evaluated. This evaluation includes two levels. One is to compare the global cooperative result with the preset benchmark or theoretical optimal solution through consistency check and error analysis, to quantify the accuracy. The other is to analyze the performance fluctuation trend by longitudinal comparison with historical cooperative effect data. Comprehensive analysis of the two levels generates a quantitative effect score for this cooperative generation. Next, the network performance parameters just obtained and the generated effect score are jointly sent into a preset multi-objective optimization function, to calculate a policy adjustment amount. The multi-objective optimization function aims to find a balance point, which maximizes the effect score of cooperation and minimizes the consumption of network resources. The specific form of the multi-objective optimization function can be a trained reinforcement learning model, which aims to output the optimal policy adjustment instruction according to the current effect and network state. The effect score and network performance parameters are taken as state inputs, and the policy parameter adjustment is taken as action output. The model such as deep Q network is trained through historical cooperative data, so that it can learn the policy adjustment decision that maximizes the long-term cumulative reward in a complex environment. Finally, according to the calculated policy adjustment amount, the two core strategies are synchronously modified. On the one hand, the complexity control parameter in the dynamic mask generation strategy is modified, such as adjusting the weights of credit and network state in the dynamic mask intensity calculation, or directly adjusting the sparsity of the dynamic mask, to adapt to the current network bearing capacity. On the other hand, the reward and punishment amplitude in the credit allocation rule is synchronously updated, that is, according to the effect of this cooperation, the credit reward for entities contributing high-quality data is increased, or the credit punishment for entities performing poorly is increased. After the two adjustments, an updated cooperative strategy that adapts to the current state is generated, which is used to guide the next round of data cooperation. Figure 3 As shown in the heat map, the adjustment of different strategy parameters in multiple rounds of cooperation is intuitively displayed, where the color depth represents the adjustment amplitude, and the positive and negative values represent enhancement or weakening. The core operation strategy can be adjusted in real time and automatically according to the actual cooperative effect and network environment. This closed-loop feedback mechanism enables the balance between cooperative effect and resource consumption to be continuously sought under changing network conditions, so as to maintain stability and robustness in the long-term operation, and finally improve the global cooperative performance of continuous optimization.

[0064] Optionally, the calculation of the strength coefficient of the dynamic mask includes:

[0065] Based on the initial credit parameters, a quantitative analysis is performed to determine the priority level of the entity;

[0066] A mask reference value is obtained by jointly mapping the priority level with the network state parameters.

[0067] The mask reference value is dynamically adjusted based on the network state parameters to generate the strength coefficient of the dynamic mask.

[0068] Specifically, the process begins with quantitative analysis based on an entity's initial credit parameters to determine its priority level. This involves comparing each entity's initial credit value with a set of preset credit thresholds, classifying it into different priority levels such as "high," "medium," or "low" according to its credit range. This classification provides a foundational, macro-level categorization for subsequent processing. After determining the priority level, a joint mapping is performed between this priority level and real-time acquired network state parameters to obtain a mask baseline value. This joint mapping can be understood as a two-dimensional lookup table or a decision function, with discrete priority levels and quantified network state parameters as inputs, and an initial mask strength baseline as the output. For example, a "high-priority" entity will be mapped to a lower mask baseline value when the network state is good; conversely, a "low-priority" entity may be mapped to a relatively high mask baseline value even when the network state is good. This mask baseline value establishes the basic level of mask strength under specific trust and macro-level network conditions. Finally, the mask baseline value is dynamically fine-tuned based on real-time, continuously changing network state parameters, ultimately generating the strength coefficients of the dynamic mask. For calculating the... Dynamic mask strength coefficients generated by each entity ,have:

[0069] ;

[0070] in, It is the mask reference value obtained in the previous step through joint mapping; It is based on real-time network performance parameters The calculated adjustment factor It includes continuously changing metrics such as network latency and bandwidth. (Function) The design makes the output value positive when the network condition deteriorates, such as increased delay or reduced bandwidth, thereby increasing the final strength coefficient; on the contrary, when the network condition improves, the output value is negative or tends to zero, thereby reducing or maintaining the strength coefficient. The function form can be determined by multi-objective optimization modeling and simulation parameter adjustment. Through these steps, a dynamic mask strength coefficient is obtained, which not only reflects the inherent credibility of the entity, but also responds to the fluctuation of the network environment in real time. By introducing priority levels and mask reference values as two intermediate levels, the method clearly separates the static, trust-based policy setting from the dynamic, network environment-based real-time adjustment. This hierarchical control not only makes the formulation and adjustment of the policy clearer and more interpretable, but also reduces the non-linear effects and parameter sensitivity problems that may be caused by directly and simply mixing credit and network state parameters.

[0071] Optionally, the generating the weight set comprises:

[0072] calculating, based on the initial credit parameter, a credit weight of each entity;

[0073] obtaining a network state parameter for quantifying the communication quality, and adjusting the credit weight of each entity according to the network state parameter to generate an adaptive weight value;

[0074] collecting the adaptive weight values of all entities to form a weight set.

[0075] Specifically, first, the credit weight of each entity is calculated based on the predetermined initial credit parameter. The calculation is usually done by normalizing the initial credit parameters of all entities, ensuring that high-credit entities obtain a relatively high basic weight, and the sum of the credit weights of each entity is 1. This step establishes a static benchmark for weight distribution, reflecting the long-term and inherent reliability of the entity. Then, the network state parameters for quantifying the communication quality are obtained in real time, which usually include the bandwidth, delay and packet loss rate of the link, etc. These heterogeneous network state parameters are integrated into a unified, dimensionless communication quality adjustment factor. Then, the credit weight of each entity is dynamically adjusted using this adjustment factor to generate adaptive weight values. The adjustment process can multiply the credit weight of each entity by its corresponding communication quality adjustment factor to obtain the adaptive weight value of the entity. Specifically, the credit weight of each entity is derived from the normalization calculation of the initial credit parameter of the entity, reflecting the long-term credibility of the entity; while the communication quality adjustment factor is a positive number dynamically calculated based on real-time network state parameters, which reflects the current transmission conditions of the entity; when the network quality is better than the set benchmark, the communication quality adjustment factor is greater than 1, which plays a role in promoting the weight; when the network quality decreases, the communication quality adjustment factor is less than 1, which correspondingly reduces the weight; the calculation of the communication quality adjustment factor is based on the normalized network parameters, so that the bandwidth, delay and other different dimensional indicators can be comprehensively evaluated. After calculating the adaptive weight values of all entities, normalization is performed again to ensure that the sum of the final adaptive weight values is 1. Finally, the adaptive weight values of all entities after the final normalization are collected to form a complete weight set that can be directly used for weighted aggregation. By upgrading the weight allocation of data aggregation from a single-dimensional static evaluation to a multi-dimensional and dynamically perceived decision-making process, not only the historical credibility of the entity is reflected, but also the current data transmission capability of the entity is immediately embodied. When a high-credit entity encounters temporary network congestion or unstable connection, its weight will be dynamically adjusted downward, thereby reducing the negative impact of its possible delayed or incomplete data on the global collaborative result; on the contrary, entities with good network conditions can obtain higher weights, and their high-quality data contributions will be fully utilized. This adaptive weight adjustment enhances the robustness of distributed data collaboration to network environment fluctuations and improves the overall timeliness and accuracy of the final result.

[0076] Optionally, the quantitatively analyzing based on the initial credit parameter to determine the priority level of the entity comprises:

[0077] inputting the initial credit parameter into a credit grading unit, and dividing a plurality of credit intervals according to the credit value;

[0078] According to the credit interval to which each entity belongs, a corresponding priority label is matched;

[0079] Historical collaborative behavior data is obtained, and the priority label is fine-tuned in combination with the historical collaborative behavior data to form a priority level.

[0080] Specifically, first, the initial credit parameter of an entity, that is, a quantitative credit score value, is input into a special credit grading unit. The unit internally presets multiple credit intervals, which divide the complete credit score range through a series of thresholds, for example, 0 to 60 points, 61 to 85 points, and 86 to 100 points. The credit score value of the entity is matched with these credit intervals, and according to the specific interval it falls into, an initial priority label is assigned to it, such as three levels of “low”, “medium”, and “high”. This step completes the preliminary classification based on static attributes. However, the priority label is not the final result, and historical collaborative behavior data of the entity will be further obtained and analyzed. These data record the actual performance of the entity in past collaborative tasks, including but not limited to the timeliness of data contribution, the quality assessment of submitted data, and the record of compliance with collaborative agreements. Based on these historical collaborative behavior data, the initially assigned priority label is dynamically fine-tuned. For example, an entity with an initial label of “high” may have its label temporarily downgraded if its recent behavior data shows frequent delayed submission or decreased data quality. Conversely, an entity with a label of “medium” but long-term excellent performance may have its label upgraded. Through this fine-tuning combined with historical behavior, the initial, static priority label is calibrated into a final priority level that more comprehensively and dynamically reflects the current comprehensive state of the entity. Figure 4 As shown in the figure, the horizontal axis represents entity ID, and the vertical axis represents the initial credit value; the initial credit parameter is input into a credit grading unit, and the initial credit value is divided into multiple credit intervals such as “low”, “medium”, and “high” through preset credit thresholds, that is, the “low / medium threshold” and the “medium / high threshold” shown in the figure. Each entity is matched with a priority according to the interval in which its initial credit value falls. By combining the initial credit evaluation based on inherent attributes and the behavior history based on actual contribution, the system reduces the lag and one-sidedness that may be caused by relying solely on static credit. This dual evaluation mechanism enables the priority level to not only reflect the basic ability of an entity, but also embody its real attitude and long-term value in the collaborative network, enabling timely identification and reward of entities that consistently make high-quality contributions, while also enabling discovery and limitation of entities that perform poorly or are potentially malicious, thereby improving the security and collaborative efficiency of the entire federation control system.

[0081] Based on the same inventive concept, as Figure 5As shown, the application also provides a distributed data collaboration device based on federal control, which comprises:

[0082] a network credit initialization module, configured to obtain network configuration parameters for defining a network topology, and generate initial credit parameters for a plurality of entities based on the network configuration parameters;

[0083] a dynamic mask generation module, configured to obtain local data and perform feature extraction to generate feature data, generate a dynamic mask by combining the initial credit parameters and real-time obtained network state parameters, and process the feature data using the dynamic mask to obtain mask feature data;

[0084] a weighted aggregation collaboration module, configured to perform weighted aggregation on the mask feature data received from each entity according to the initial credit parameters, and generate a global collaboration result;

[0085] a policy evaluation update module, configured to evaluate based on the effectiveness of the global collaboration result, generate an effect score, and combine the effect score with obtained network performance parameters to generate an updated collaboration strategy for adjusting a preset dynamic mask generation strategy and a preset credit allocation rule.

[0086] In order to verify the feasibility of the application in implementation, the application is applied to a set of distributed industrial Internet of Things predictive maintenance system of a manufacturing enterprise. The enterprise hopes to construct a high-precision global fault prediction model by collaboratively analyzing the operation data of a plurality of numerical control machine tools distributed throughout the factory, so as to discover potential equipment faults in advance.

[0087] The traditional collaboration method usually performs simple average aggregation on the data submitted by all edge computing nodes, ignoring the huge differences in hardware capabilities, network environments and historical data quality of each node, which leads to low precision of the global model and low collaboration efficiency in network congestion, and is difficult to meet the requirements of high-precision and high-timeliness of industrial production.

[0088] In this embodiment, the enterprise deploys the distributed data collaboration system based on federal control proposed by the present application. The system first generates an initial credit parameter for each node through the network credit initialization module according to the hardware configuration of each edge computing node, such as CPU performance, memory capacity, and network access capability, such as nominal bandwidth, and other network configuration parameters. In the data collaboration process, each node uses the dynamic mask generation module to extract features from the collected machine tool vibration and temperature data locally, and generates a dynamic mask by combining the initial credit parameter and the real-time monitored network state parameter such as network delay, bandwidth, and then generates the mask feature data. Subsequently, the weighted aggregation collaboration module of the central collaboration server generates a global fault prediction model by weighting and aggregating all the mask feature data received according to the initial credit parameters of each node. Finally, the strategy evaluation and update module evaluates the prediction accuracy of the model to generate an effect score, and combines the real-time acquired network performance parameters such as delay and packet loss rate to perform closed-loop feedback and adaptive update on the dynamic mask generation strategy and credit allocation rules.

[0089] In order to verify the beneficial effects of the present application, we conducted a 6-month comparative test. The experimental group used the method of the present application, and selected three representative edge computing nodes, namely nodes A, B, and C, for intensive observation; the control group used the traditional indiscriminate average aggregation method. During the test period, we recorded in detail the credit rating of each node, the change of dynamic mask intensity, the prediction accuracy of the global model, and the network resource consumption.

[0090] In the system initialization phase, the network credit initialization module evaluates each node. For example, node A is deployed on a high-performance server and accesses through a gigabit wired network, and its initial credit parameter is rated as 95; node B is deployed on a standard industrial computer and accesses through industrial WiFi, and its initial credit parameter is 70; and node C is an old device with unstable performance and network connection, and its initial credit parameter is only 55.

[0091] In the execution of the collaboration task, the role of the dynamic mask generation module is highlighted. In a collaboration task at 10:30 on April 15, 2025, serious interference was detected in the wireless network environment of node B, with network delay increasing to 180 ms and bandwidth decreasing to 30% of the available value. The system obtains this network state parameter in real time and, combined with the initial credit parameter of node B, calculates a dynamic mask intensity coefficient of 0.88, which performs strong mask processing on the feature data uploaded by the node to reduce the negative impact of its low-quality data on the global result. At the same time, node A with stable network environment has a delay of 12 ms, and its dynamic mask intensity coefficient is only 0.25, thereby strengthening the contribution of its high-quality data.

[0092] In the weighted aggregation coordination phase, the system assigns weights to each node according to the initial credit parameters. For example, the weight of the data contributed by node A is set to 0.22, node B to 0.11, and node C to only 0.06. Through this trust-based weighted aggregation, the global model generated by the present application achieves a failure prediction accuracy of 91.5% in the first round of testing, while the control group using simple average aggregation has a model accuracy of only 82.3%.

[0093] The system's policy evaluation update module demonstrates its closed-loop adaptive capability. After the first month of operation, the system evaluates the effectiveness score of the global coordination result as 92, while monitoring that the overall packet loss rate of the plant network has increased somewhat due to the busy production tasks. Based on the effectiveness score and network performance parameters, the system calculates the policy adjustment amount through a multi-objective optimization function and generates an updated coordination strategy. This strategy adjusts the complexity control parameter in the dynamic mask generation strategy upwards to adapt to the current poor network environment; at the same time, since node A has been consistently contributing high-quality data in the past month, its credit is rewarded and increased to 98, while node C, which has been underperforming, is punished and reduced to 51. After this round of policy update, the system's model accuracy is further improved in the coordination task of the next month, and the network bandwidth occupancy is lower than last month.

[0094] Through six months of continuous operation and optimization, the experimental group using the present application ultimately achieves a global model failure prediction accuracy of more than 96%, which is superior to the control group's accuracy of about 85%. At the same time, the adaptive mechanism of the system effectively balances the coordination effect and resource consumption, and the overall network bandwidth consumption is also reduced compared to the transmission method without strategy control.

[0095] It should be noted that the electrical connection between the above-mentioned units does not necessarily represent a direct connection, and an indirect connection mode can also be used as long as the purpose of the present application is achieved. The above-described embodiments are only exemplary embodiments of the present application, and cannot limit the scope of the present application.

[0096] That is, any equivalent changes and modifications made in accordance with the teachings of the present application are still within the scope of the present application. Other embodiments of the present application will be readily apparent to those skilled in the art upon considering the description set forth in the specification and the teachings of the disclosure. The present application is intended to cover any variations, uses, or adaptive changes to the present application that follow the general principles of the present application and include common knowledge or conventional technical means in the art not described in the present application.

Claims

1. A distributed data collaboration method based on federal control, characterized in that: obtaining network configuration parameters for defining network topology, and generating initial credit parameters for multiple entities based on the network configuration parameters; obtaining local data and performing feature extraction to generate feature data, generating a dynamic mask by combining the initial credit parameters and real-time obtained network state parameters, and processing the feature data using the dynamic mask to obtain mask feature data; weighting and aggregating the mask feature data received from each entity according to the initial credit parameters to generate a global collaboration result; evaluating the effectiveness of the global collaboration result to generate an effect score, and combining the effect score with obtained network performance parameters to generate an updated collaboration strategy for adjusting the preset dynamic mask generation strategy and the preset credit allocation rule.

2. The distributed data collaboration method based on federal control according to claim 1, characterized in that, The generation of initial credit parameters for multiple entities based on the network configuration parameters includes: extracting entity attribute parameters of multiple entities from the network configuration parameters; performing credit quantification processing based on the entity attribute parameters to calculate an initial credit value for each entity; integrating the initial credit values of all entities to form an initial credit parameter.

3. The distributed data collaboration method based on federal control according to claim 1, characterized in that, The generation of a dynamic mask by combining the initial credit parameters and real-time obtained network state parameters includes: real-time monitoring of communication link state to obtain network state parameters including network delay and bandwidth parameters; jointly processing the network state parameters and the initial credit parameters to calculate the intensity coefficient of the dynamic mask; random sampling based on the intensity coefficient of the dynamic mask to generate a dynamic mask.

4. The distributed data collaboration method based on federal control according to claim 3, characterized in that, The generation of mask feature data includes: performing numerical scaling on the feature data according to the intensity coefficient of the dynamic mask to generate scaled feature data; performing bitwise XOR operation on the scaled feature data and the dynamic mask to obtain mask feature data.

5. The method of claim 1, wherein the method is a federated control based distributed data collaboration method. The generation of a global collaboration result includes: assigning weights to the mask feature data of each entity according to the initial credit parameters to generate a weight set; performing weighted fusion based on the mask feature data and the weight set to calculate a global collaboration result.

6. The method of claim 1, wherein the method is a federated control based distributed data collaboration method. The evaluation of the effectiveness of the global collaboration result to generate an effect score, and the combination of the effect score with obtained network performance parameters to generate an updated collaboration strategy for adjusting the preset dynamic mask generation strategy and the preset credit allocation rule includes: continuously monitoring the communication state of the federal control network to obtain network performance parameters including delay, packet loss rate and throughput; performing consistency checking and error analysis on the global collaboration result, and combining historical collaboration effect data to generate an effect score reflecting the effectiveness of this collaboration; calculating the strategy adjustment amount through a preset multi-objective optimization function according to the effect score and the network performance parameters; modifying the complexity control parameter in the dynamic mask generation strategy and synchronously updating the reward and punishment amplitude in the credit allocation rule according to the strategy adjustment amount to generate an updated collaboration strategy.

7. The distributed data collaboration method based on federal control according to claim 3, characterized in that, The calculation of the intensity coefficient of the dynamic mask includes: Perform quantitative analysis based on the initial credit parameters to determine the priority level of the entity; Jointly map the priority level and the network state parameters to obtain a mask reference value; Adjust the mask reference value based on the network state parameters to generate the intensity coefficient of the dynamic mask.

8. The distributed data collaboration method based on federal control according to claim 5, characterized in that, The generation of the weight set includes: Calculate the credit weight of each entity based on the initial credit parameters; Obtain network state parameters for quantifying communication quality, and adjust the credit weight of each entity based on the network state parameters to generate adaptive weight values; Collect the adaptive weight values of all entities to form a weight set.

9. The distributed data collaboration method based on federal control according to claim 7, characterized in that, The quantitative analysis based on the initial credit parameters to determine the priority level of the entity includes: Input the initial credit parameters into a credit grading unit, and divide multiple credit intervals according to the size of the credit value; According to the credit interval to which each entity belongs, match the corresponding priority label; Obtain historical cooperative behavior data, and fine-tune the priority label based on the historical cooperative behavior data to form a priority level.

10. A distributed data collaboration device based on federal control, applied to a distributed data collaboration method based on federal control according to any one of claims 1-9, characterized in that, The device includes: A network credit initialization module for obtaining network configuration parameters for defining network topology, and generating initial credit parameters for multiple entities based on the network configuration parameters; A dynamic mask generation module for obtaining local data and performing feature extraction to generate feature data, generating a dynamic mask based on the initial credit parameters and real-time obtained network state parameters, and processing the feature data using the dynamic mask to obtain mask feature data; A weighted aggregation cooperation module for weighting and aggregating the mask feature data received from each entity based on the initial credit parameters to generate a global cooperation result; A strategy evaluation update module for evaluating based on the effectiveness of the global cooperation result to generate an effect score, and generating an updated cooperation strategy for adjusting the preset dynamic mask generation strategy and the preset credit allocation rule based on the effect score and the obtained network performance parameters.

Images