Digital twin driven federated learning fair contribution evaluation and individualized reward method

Abstract

This invention provides a digital twin-driven federated learning fair contribution evaluation and personalized reward method, belonging to the field of contribution evaluation technology, for jointly training multiple equipment failure prediction models for predictive maintenance of large industrial equipment such as wind turbines and oil drilling platforms. This invention combines blockchain and digital twin technology to select safe and stable clients to participate in training, minimizing the negative impact of unstable clients; it effectively reduces computational costs by evaluating the Shapley value of the client group rather than the Shapley value of an individual client. The model is updated personalized based on each client's contribution, thereby improving fairness and efficiency in the collaborative process. This invention ensures the credibility of client identities and the integrity of data while effectively selecting stable participants, improving the security and stability of the training process. It reduces computational complexity while improving training efficiency and fairness, optimizing overall performance.

Description

Technical Field

[0001] This invention belongs to the field of contribution assessment technology, and in particular relates to a digital twin-driven method for fair contribution assessment and personalized rewards in federated learning. Background Technology

[0002] Federated Learning (FL) is a distributed machine learning paradigm where clients do not upload raw data but train locally, only uploading model updates such as gradients or parameters to the server for aggregation. This enables collaborative modeling without centralized data. In this process, client contribution evaluation quantifies the marginal value of each participant's contribution to the global model improvement. Based on fair evaluation results, the system can establish a matching reward mechanism to incentivize clients to continuously provide high-quality updates, thereby improving overall collaboration efficiency and global model performance. However, in open or semi-open environments, malicious or semi-honest clients may inject harmful updates or tamper with and falsify training data locally, leading to risks such as poisoning attacks, model theft, and data pollution. Therefore, additional trust safeguards are needed.

[0003] In federated learning scenarios with dynamic client participation, although numerous methods exist for evaluating client contributions and guiding participation selection and incentive allocation, practical applications still face multiple challenges that limit the fairness, security, and stability of federated learning in real-world environments. First, there are security issues with client data and identities. In federated learning, sensor data and client identities can be tampered with by malicious clients, leading to false or inaccurate results. Malicious clients may not only tamper with data but also forge identities to bypass security verification, submitting false training updates. This behavior not only affects the training quality of the model, leading to incorrect predictions and decisions, but may also undermine the overall credibility and stability of the system, especially in critical application areas such as industrial control, medical monitoring, and intelligent transportation systems, potentially causing serious consequences. Second, there are issues with the long-term stability and availability of clients. Clients may experience disconnections, delays, or packet loss due to network instability, equipment failure, or other factors. This not only causes other clients to wait for updates but may also affect the completion of local training and the timeliness of updates, further impacting the convergence speed and accuracy of the global model. Finally, there are issues with the computational complexity of contribution evaluation and the problem of non-independent identically distributed (Non-IID) data. Existing contribution assessment methods based on Shapley values ​​suffer from high computational complexity and unfair assessment in environments with large-scale participants and non-IID data. The calculation of Shapley values ​​increases exponentially with the number of participants, and in non-IID data scenarios, some useful clients may be misclassified as negative contributors, leading to unfair assessment results. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a digital twin-driven method for fair contribution assessment and personalized rewards in federated learning, comprising the following steps:

[0005] The physical entities of industrial equipment participating in the joint training of the equipment failure prediction model are used as clients. Client registration and identity verification are completed on the blockchain. The system reviews whether the client is allowed to join the global training and builds a digital twin for each client that passes the review.

[0006] In each global training round, the first status index of the approved clients is continuously collected. Based on the first status index, the selected clients are selected and the candidate set for the current global training round is generated.

[0007] Within the current global training round, several rounds of local training are performed on each selected client based on the initial global model and the local data of the selected clients, generating the gradient of each selected client in the current global training round and uploading it to the server;

[0008] Based on the gradients uploaded by the selected clients, the selected clients are preprocessed and clustered to obtain a set of strategic equivalence groups;

[0009] The group-level contribution assessment is performed on the strategic equivalence group set to obtain the individual contribution weight of each selected client;

[0010] Based on the gradient and individual contribution weight of each selected client, the final global model for the current global training round is generated; based on the individual contribution weight of each selected client, the personalized local model for each selected client is generated and distributed to the corresponding selected client.

[0011] Based on the individual contribution weight and stability score of the selected clients, the reward and penalty settlement of the selected clients is executed and notarized on the blockchain. The digital twins of the selected clients are updated, the current global training round is completed and the next global training round is entered, until the final global model obtained after several global training rounds satisfies the convergence condition and the final global model with stable performance is obtained.

[0012] The final global model, whose performance has reached a stable state, is deployed on the server side for centralized fault reasoning and prediction based on the collected device operation data.

[0013] Furthermore, the specific method for completing client registration and identity verification on the blockchain is as follows:

[0014] The client generates a key pair, including a public key and a private key, using an encryption algorithm; the client submits registration information to the identity management contract and generates a digital signature of the registration information using the private key to prove the client's identity;

[0015] The identity management contract uses the client's public key to verify the digital signature of the registration information. After successful identity verification, the identity management contract sets the client's registration status to available and enables the client's access permissions and settlement rules.

[0016] Furthermore, the specific method for reviewing whether a client is allowed to participate in global training is as follows:

[0017] When the client requests to join the During each global training round, the client sends a join request based on the client's private key. A digital signature for the join request is generated and submitted to the identity management contract for review. The identity management contract verifies the client's registration status and access permissions based on the digital signature. If the client has completed registration and possesses valid access permissions, the review is approved, and the identity management contract allows the client to join the [new client's] ... Each global training round; hashing the client's request to join the global training and writing it to the blockchain to obtain an auditable record;

[0018] The server creates an initial packet, calculates its hash value, generates a server signature using its private key, and distributes the initial packet, its hash value, and the server signature to several approved clients. Including the initial global model Training hyperparameters Privacy and security configuration The system distributes the hash of the initial package and the server signature to approved clients.

[0019] Furthermore, the first status indicator includes online rate, latency jitter, failure rate, energy consumption, and bandwidth usage;

[0020] The specific method for selecting clients and generating the candidate set for the current global training round based on the first state index is as follows:

[0021] In the Within each global training round, the first state indicators of each approved client are continuously collected, and the first state indicators are summarized according to preset weights to obtain the stability score of each approved client.

[0022] The first status indicator of each approved client is compared with a preset security threshold. If any of the first status indicators of the approved client exceeds the preset security threshold, the approved client is determined to be in an abnormal state, and a risk label is generated for the approved client. The risk label is used to indicate the risk status of the approved client, including high risk and low risk.

[0023] Clients are selected based on a preset stability threshold and the risk label of approved clients. When the stability score of an approved client is greater than or equal to the preset stability threshold and the risk label of the approved client is low risk, the approved client is added to the candidate set of the current global training round.

[0024] If the number of clients that have passed the review in the current global training round candidate set exceeds the concurrency limit. All approved clients were sorted in descending order of stability score, and the top-performing clients were selected. One client that passes the review will be selected as the selected client.

[0025] Furthermore, the specific method for performing several rounds of local training on each selected client based on the initial global model and the local data of the selected clients is as follows:

[0026] Selected clients from the current global training round candidate set In the In the first global training round, the first round... The final global model obtained after one global training round As the initial global model trained locally, it is selected by the client. Perform several local training iterations on local data to obtain the selected clients. No. Local model in each global training round ;

[0027] Among them, in the first In each global training epoch, the initial global model is used. As the initial global model for local training; in the... In the first global training round, the first... The final global model obtained after one global training round As the initial global model for local training;

[0028] The specific method for generating the gradient for each selected client in the current global training round is as follows:

[0029] Based on selected clients No. Local model in each global training round and the The initial global model in each global training round Calculate the selected clients No. Gradient of each global training epoch .

[0030] Furthermore, the specific methods for preprocessing and clustering the selected clients are as follows:

[0031] For each selected client's gradient, retain the gradient with the largest absolute value. One component is set to zero in each position, and the rest are set to zero to obtain the gradient sparse vector of each selected client.

[0032] L2 normalization is performed on the gradient sparse vector of each selected client to obtain the normalized sparse vector of each selected client; a set of normalized sparse vectors is generated based on the normalized sparse vector of each selected client.

[0033] A top-down bisecting K-means method is used to cluster the normalized sparse vector set to generate a set of strategic equivalence groups.

[0034] Furthermore, the specific method for clustering the normalized sparse vector set using the top-down K-means method is as follows:

[0035] Clustering is performed using a normalized sparse vector set as the initial cluster. The proportion of common non-zero dimensions in the initial cluster is calculated. If the proportion of common non-zero dimensions in the initial cluster is not lower than a preset threshold, then the cluster is considered a cluster. If the initial cluster is strategically equivalent, the selected clients in the initial cluster are determined to be strategically equivalent, clustering is stopped, and the initial cluster is treated as a strategically equivalent group; otherwise, the initial cluster is split into two sub-clusters, the proportion of common non-zero dimensions of the two sub-clusters is calculated, and the proportion of common non-zero dimensions of the two sub-clusters is compared with the threshold. The size of the subcluster is used to determine whether to split it into subclusters, until the proportion of the common non-zero dimension of all subclusters is not less than the threshold. Alternatively, the total number of all clusters may reach the target number of clusters. After clustering is completed, each cluster is treated as a strategic equivalence group, and a set of strategic equivalence groups is generated based on all strategic equivalence groups, including several strategic equivalence groups.

[0036] Furthermore, the specific method for evaluating the group-level contribution of the strategic equivalence group set is as follows:

[0037] Based on the gradients uploaded by all selected clients, calculate the gradients for all client groups. Group model, client group It is a subset of the current global training round candidate set;

[0038] Based on the Initial global model and client group for each global training round The group model establishes feature functions;

[0039] Calculate the cooperation Shapley value for each selected client in the set of strategic equivalence groups;

[0040] The cooperation Shapley value of each selected client in the strategic equivalence set is normalized to obtain the individual contribution weight of each selected client.

[0041] Furthermore, the specific method for generating the final global model for the current global training round is as follows:

[0042] The gradients received from each selected client are weighted and aggregated based on the individual contribution weight of each selected client, and the aggregation results are used to analyze the gradients of the selected clients. The initial global model of each global training round is updated to obtain the final global model of the current global training round;

[0043] The specific method for generating personalized local models for selected clients is as follows:

[0044] Based on the individual contribution weight of each selected client, a personalized model for each selected client is generated by combining the local model of each selected client.

[0045] The beneficial effects of adopting the above technical solution are as follows: The present invention provides a digital twin-driven federated learning fair contribution evaluation and personalized reward method, which combines blockchain and digital twin technology to ensure reliable verification of client identities, protection of data integrity, and selection of safe and stable clients to participate in training, minimizing the negative impact of unstable clients. It effectively reduces computational costs by evaluating the Shapley value of the client group, rather than the Shapley value of an individual client. Furthermore, based on each participant's contribution, the present invention can update their model individually, thereby improving fairness and efficiency in the collaborative process. The digital twin-driven federated learning fair contribution evaluation and personalized reward method provided by the present invention can ensure the credibility of client identities and the integrity of data, while effectively selecting stable participants, significantly improving the security and stability of the training process. While reducing the computational complexity of Shapley value evaluation, it improves training efficiency and fairness through client group evaluation and personalized model updates, optimizing overall performance. Attached Figure Description

[0046] Figure 1 Flowchart of the digital twin-driven federated learning fair contribution assessment and personalized reward method provided in Embodiment 1 of this invention;

[0047] Figure 2 A schematic diagram of federated learning training provided in Embodiment 1 of the present invention. Detailed Implementation

[0048] The specific implementation methods of this application will be further described in detail below with reference to the accompanying drawings and embodiments.

[0049] Example 1:

[0050] Existing contribution evaluation methods typically require training a global model separately for each possible combination of participant data and calculating the marginal contribution of each participant based on the model's performance on the validation dataset. This approach is computationally extremely expensive, with computational complexity increasing exponentially with the number of participants, making it difficult to apply to real-world scenarios with large-scale participation. Furthermore, existing Shapley value-based methods do not adequately consider the dynamic changes of participants, especially in open federated learning environments where frequent client joining and leaving make traditional methods difficult to adapt, thus limiting their effectiveness in real-time applications.

[0051] Blockchain is a decentralized, immutable, and traceable distributed ledger. Through a consensus mechanism, it establishes a shared state among multiple parties without requiring complete trust. Combined with smart contracts, it enables programmable rule execution and automatic settlement. When used in federated learning, blockchain provides a trusted distributed ledger for client registration and identity management. Smart contracts automatically verify identities and access requests on-chain according to rules, and write metadata such as model update hashes and access logs to the chain. The immutability of blockchain ensures that records are traceable and auditable. However, relying solely on on-chain records is insufficient to depict the long-term evolution and real-time state of clients; therefore, a digital twin that sustainably maps and updates their state is needed.

[0052] A digital twin (DT) is a dynamic, high-fidelity virtual mapping of a physical entity or system built in the digital space. Through continuous data acquisition and two-way interaction, it enables the monitoring, simulation, analysis, and optimization of the entity. Utilizing sensor data collection and network transmission combined with model computation, the virtual and physical ends can be aligned in near real-time. Based on this mirror relationship, the digital twin presents the client's operational status and trajectory online, while continuously accumulating and updating its long-term state profile to assess stability and availability, thereby supporting stability-oriented selection and decision-making.

[0053] Against this backdrop, this embodiment provides a digital twin-driven federated learning fair contribution assessment and personalized reward method for the industrial Internet of Things (IIoT) field, used for cross-enterprise joint training of equipment fault prediction models to predict and maintain large industrial equipment such as wind turbines and oil drilling platforms.

[0054] Because different enterprises use different types and batches of equipment, and the operating environments and failure modes of these devices vary, the data exhibits significant heterogeneity and dynamism. To address this issue, this embodiment establishes a digital twin for each device or factory, mapping and managing its latest contribution status, data quality, and historical performance in real time. Blockchain and digital twins work together to ensure the security and stability of clients in federated learning, conduct fair client contribution assessments in a dynamically participating environment, and distribute personalized models as incentives, thereby achieving fair collaboration.

[0055] In joint training, this embodiment utilizes strategic equivalence group clustering to handle device type differences and uses an improved Shapley value for fair contribution assessment. Smart contracts are liquidated based on the assessment results and the digital twin state: high-contributing enterprises are rewarded with resource allocation; enterprises providing low-quality data are penalized by differentially affecting the quality of their next round of local models. This mechanism achieves accurate assessment of contributions and fair distribution of rewards, significantly improving the reliability of industrial collaboration models and the operational efficiency of enterprises.

[0056] This embodiment presents a digital twin-driven federated learning fair contribution evaluation and personalized reward method for jointly training multiple device fault prediction models, such as... Figure 1 As shown, it includes the following steps:

[0057] Step 1: The physical entities of industrial equipment participating in the joint training of the equipment failure prediction model are treated as clients. Client registration and identity verification are completed on the blockchain. The approval of whether the client is allowed to join the global training is reviewed, and a digital twin is built for each approved client.

[0058] In this embodiment, the physical entities of industrial equipment participating in the joint training of the equipment fault prediction model are treated as clients. Client registration and identity verification are completed on the blockchain. Based on the identity management contract, client access permissions and settlement rules are determined, and the server reviews whether to allow a client to join the training. The server synchronously sends an initial package, its hash value, and the server signature to several approved clients. The initial package includes the initial global model, training hyperparameters, and privacy / security configuration. A schematic diagram of the federated learning training of multiple equipment fault prediction models is shown below. Figure 2 As shown;

[0059] Step 1.1: Using the physical entities of industrial equipment participating in the joint training of the equipment failure prediction model as clients, generate client key pairs using encryption algorithms. ,in, For public key, The client submits registration information to the identity management contract using the private key. A digital signature is generated for the registration information to prove the client's identity;

[0060] The identity management contract uses the client's public key. The digital signature of the registration information is verified for identity. After the identity verification is successful, the identity management contract sets the client's registration status to available and sets the client's access permissions and settlement rules to valid.

[0061] Step 1.2: When the client requests to join the... During each global training round, the client sends a join request based on the client's private key. A digital signature for the join request is generated and submitted to the identity management contract for review. The identity management contract verifies the client's registration status and access permissions based on the digital signature. If the client has completed registration and possesses valid access permissions, the review is approved, and the identity management contract allows the client to join the [new client's] ... One global training round; simultaneously, the join request sent by the client is hashed and written to the blockchain to form an auditable trace record;

[0062] Step 1.3: The server creates an initial packet, calculates its hash value, generates a server signature using its private key, and distributes the initial packet, its hash value, and the server signature to several approved clients. Including the initial global model Training hyperparameters Privacy and security configuration At the same time, the hash value and server signature of the initial package are sent to the approved clients for consistency verification, ensuring the integrity of the data and the trustworthiness of its source;

[0063] Step 2: Continuously collect the first status indicators of the approved clients in each global training round, select the selected clients based on the first status indicators and generate the candidate set for the current global training round;

[0064] Step 2.1: In the first Within each global training round, the first-state indicators of each approved client are continuously collected, including online rate, latency jitter, failure rate, energy consumption, and bandwidth usage. These first-state indicators are then aggregated according to preset weights to obtain a stability score for each approved client. ;

[0065] The first status indicator of each approved client is compared with a preset security threshold. If any one of the first status indicators of the approved client exceeds the preset security threshold, the approved client is determined to be in an abnormal state, and a risk label is generated for the approved client. ,in, Indicates high risk. Indicates low risk. Indicates the first One global training round;

[0066] Step 2.2: Based on the preset stability threshold and the risk labels of approved clients Selected clients are chosen based on their stability score, which is greater than or equal to a preset stability threshold. Furthermore, the risk label of the approved client is low risk, meaning... At that time, clients that pass the review will be added to the candidate set for the current global training round. ;

[0067] If the current global training round candidate set The number of clients approved in China exceeded the concurrency limit. All approved clients were sorted in descending order of stability score, and the top-performing clients were selected. One approved client is selected as the selected client;

[0068] This embodiment generates a stability score and risk label for each approved client, selects eligible clients based on preset stability thresholds and risk labels, and generates a candidate set for the current global training round, eliminating unstable or high-risk nodes to ensure training safety and availability.

[0069] Step 3: Within the current global training round, perform several rounds of local training on each selected client based on the initial global model and the local data of the selected clients, generate the gradient of each selected client in the current global training round, and upload it to the server;

[0070] Step 3.1: Analyze the candidate set for the current global training round. Selected clients In the In the first global training round, the first round... The final global model obtained after one global training round As the initial global model trained locally, it is selected by the client. Local data Up Secondary local training yields selected clients. No. Local model in each global training round ;

[0071] In the In each global training epoch, the initial global model is used. As the initial global model for local training; in the... In the first global training round, the first... The final global model obtained after one global training round As the initial global model for local training;

[0072] Step 3.2: Based on the selected clients No. Local model in each global training round and the The initial global model in each global training round Calculate the selected clients No. Gradient of each global training epoch And upload it to the server;

[0073] Selected Clients No. Gradient of each global training epoch As shown in formula (1):

[0074] ;

[0075] Step 4: Based on the gradients uploaded by the selected clients, preprocess and cluster the selected clients to obtain a set of strategic equivalence groups;

[0076] Step 4.1: For the gradient of each selected client, retain the gradient with the largest absolute value. One component is set to zero in each position, and the rest are set to zero to obtain the gradient sparse vector of each selected client.

[0077] Selected Clients gradient sparse vector As shown in formula (2) below:

[0078] ;

[0079] in, This indicates selecting and retaining the first component based on its absolute value. One component is set to zero, and the remaining components are set to zero.

[0080] To eliminate scale differences, the gradient sparse vector of each selected client is L2 normalized to obtain a normalized sparse vector for each selected client; a set of normalized sparse vectors is generated based on the normalized sparse vector of each selected client.

[0081] Selected Clients Normalized sparse vectors As shown in formula (3) below:

[0082] ;

[0083] in, For selected clients gradient sparse vector Dimensions For selected clients gradient sparse vector The One component;

[0084] Step 4.2: Use the top-down K-means method to cluster the normalized sparse vector set to generate a set of strategic equivalence groups, including several strategic equivalence groups;

[0085] Clustering is performed using a normalized sparse vector set as the initial cluster. The proportion of common non-zero dimensions in the initial cluster is calculated. If the proportion of common non-zero dimensions in the initial cluster is not lower than a preset threshold, then the cluster is considered a cluster. If the initial cluster is strategically equivalent, the selected clients in the initial cluster are determined to be strategically equivalent, clustering is stopped, and the initial cluster is treated as a strategically equivalent group; otherwise, the initial cluster is split into two sub-clusters, the proportion of common non-zero dimensions of the two sub-clusters is calculated, and the proportion of common non-zero dimensions of the two sub-clusters is compared with the threshold. The size of the subcluster is used to determine whether to split it into subclusters, until the proportion of the common non-zero dimension of all subclusters is not less than the threshold. Alternatively, the total number of all clusters may reach the target number of clusters; after clustering, each cluster is considered a strategic equivalence group, and a set of strategic equivalence groups is generated based on all strategic equivalence groups. , For the first A strategic equivalent set;

[0086] This embodiment reduces the computational complexity and time cost of subsequent contribution evaluation by constructing a group-level candidate participant set, thereby compressing the combination size from the source.

[0087] Step 5: Perform group-level contribution evaluation on the strategic equivalence group set to obtain the individual contribution weight of each selected client;

[0088] This embodiment uses validation set loss reduction as the characteristic function. By comparing the loss changes caused by a certain group of clients joining other client groups with different distributions, the marginal contribution of each group is obtained. Then, the contribution value is obtained by using the cooperative Shapley value method, which effectively avoids the bias caused by Non-IID data to contribution evaluation and significantly reduces complexity.

[0089] Step 5.1: The server calculates the gradients uploaded by all selected clients. Calculate all client groups Group model, client group Candidate set for the current global training round A subset for any client group Client Group Group model As shown in the formula below:

[0090] ;

[0091] Step 5.2: Based on the first The initial global model in each global training round and client group Group model Establish characteristic functions As shown in the formula below:

[0092] ;

[0093] in, To verify the loss function;

[0094] Step 5.3: Calculate the value of each selected client in the set of strategic equivalence groups. The value of cooperation with Shapley ;

[0095] Since each selected client within the same strategic equivalence group makes the same marginal contribution to the global model, all selected clients within the same strategic equivalence group have the same cooperative Shapley value. For clients including... The first selected client Strategic equivalent groups The first selected client The cooperative Shapley value is defined as shown in formula (6):

[0096] ;

[0097] in, , and They represent the first Strategic equivalent groups , No. Strategic equivalent groups and the Strategic equivalent groups The number of clients selected , , The loop variable for the summation symbol is divided into two parts, representing the current combination calculation, assuming it originates from the strategic equivalent group. , , The number of selected clients. This represents the total number of all selected clients. It is the index variable of the strategic equivalence set. Indicates from the first The number of clients selected from each strategic equivalence group This indicates that the minimum value of all expressions within the parentheses is selected. Indicates rounding down. Represents the set of indices for all strategic equivalence sets. To remove Indexes of other strategic equivalence sets, Indicates from the first The number of clients selected from each strategic equivalence group Indicates coprime. In order to be selected clients Add one of the selected clients from the previously formed subset of selected clients. and These represent the selected clients. and selected clients The type of strategy. Selected clients. The value of cooperation with Shapley It satisfies the axioms of empty player, symmetry, and additivity.

[0098] Step 5.4: For each selected client in the set of strategic equivalence groups The collaboration Shapley value is normalized to obtain the individual contribution weight of each selected client;

[0099] Reward selected clients who contribute to model improvement; that is, only select clients with positive cooperative Shapley values ​​will be rewarded. All selected clients with positive cooperative Shapley values ​​will be added to the set. The relative contribution of each selected client is measured by calculating the proportion of the cooperative Shapley value of each selected client to the total, and the normalized cooperative Shapley value is used as the individual contribution weight of each selected client.

[0100] Selected Clients Individual contribution weight As shown in the formula below:

[0101] ;

[0102] Individual contribution weight It can be directly used for subsequent strategies such as fair and personalized reward allocation and aggregated weight adjustment.

[0103] Step 6: Based on the gradient and individual contribution weight of each selected client, perform contribution-weighted aggregation on the gradient of the selected clients, and use the aggregation result to update the initial global model of the current global training round to generate the final global model of the current global training round; based on the individual contribution weight of each selected client, generate the personalized local model of the selected client and distribute it to the corresponding selected client.

[0104] Step 6.1: The server performs weighted aggregation of the gradients received from each selected client based on the individual contribution weight of each selected client, and uses the aggregation result to perform weighted aggregation on the gradients received from each selected client. The initial global model in each global training round Update the model to obtain the final global model for the current global training round. :

[0105] ;

[0106] Among them, the individual contribution weight of each selected client satisfy ;

[0107] Step 6.2: Based on the individual contribution weight of each selected client, generate a personalized local model for the selected client and distribute it to the corresponding selected client;

[0108] Clients that contribute more should receive better local models, while clients that contribute less should not, yet expect to benefit from the global model. To penalize these low-contribution participants, the server will differentially affect the quality of their local models to reduce their benefits from the global model. By generating personalized models for each selected client based on their individual contribution weight and their local models, the server can improve global performance while balancing individual benefits and fairness.

[0109] Selected Clients Personalized local models As shown in formula (9):

[0110] ;

[0111] Step 7: Based on the individual contribution weight and stability score of the selected clients, execute the reward and penalty settlement for the selected clients, store the evidence on the blockchain, update the digital twin of the selected clients, complete the current global training round, and enter the next global training round.

[0112] Step 8: Repeat steps 2 to 7 until the final global model obtained after several global training rounds meets the convergence condition and the performance of the final global model reaches a stable state. Deploy the final global model with stable performance on the server side for centralized fault reasoning and prediction based on the collected device operation data.

[0113] Example 2:

[0114] This embodiment proposes an electronic device, including: one or more processors, and a memory for storing instructions, which, when executed by the one or more processors, cause the one or more processors to perform the digital twin-driven federated learning fair contribution assessment and personalized reward method.

[0115] The electronic device may be a mobile phone, computer, or tablet computer, etc., and includes a memory and a processor. The memory stores a computer program that, when executed by the processor, implements the digital twin-driven federated learning fair contribution assessment and personalized reward method as described in the embodiments. It is understood that the electronic device may also include input / output (I / O) interfaces and communication components.

[0116] The processor is used to execute all or part of the steps in the digital twin-driven federated learning fair contribution assessment and personalized reward method as described in the above embodiments. The memory is used to store various types of data, which may include, for example, instructions for any application or method in the electronic device, as well as application-related data.

[0117] The processor may be implemented as an Application Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), Programmable Logic Device (PLD), Field Programmable Gate Array (FPGA), controller, microcontroller, microprocessor, or other electronic components, and is used to execute the digital twin-driven federated learning fair contribution assessment and personalized reward method described in the above embodiments.

[0118] Example 3:

[0119] This embodiment proposes a computer-readable storage medium that stores executable instructions. When these instructions are executed, if they are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium.

[0120] The computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the digital twin-driven federated learning fair contribution assessment and personalized reward method described in the various embodiments of this application.

[0121] The aforementioned storage media include: flash memory, hard disks, multimedia cards, card-type memory (e.g., SD (Secure Digital Memory Card) or DX (Memory Data Register, MDR) memory), random access memory (RAM), static random-access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic storage, disks, optical discs, servers, APP (Application) app stores, and other media capable of storing program verification codes. These media store computer programs, which, when executed by a processor, can implement the various steps of the aforementioned digital twin-driven federated learning fair contribution assessment and personalized reward method.

[0122] Example 4:

[0123] This embodiment proposes a computer program product, including a computer program or instructions, which, when executed by a processor, implements the aforementioned digital twin-driven federated learning fair contribution assessment and personalized reward method.

[0124] Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a computer program product.

[0125] The various embodiments in this application are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0126] The scope of protection of this application is not limited to the embodiments described above. Obviously, those skilled in the art can make various modifications and variations to this disclosure without departing from the scope and spirit of this disclosure. If such modifications and variations fall within the scope of this disclosure and its equivalents, then the intent of this disclosure also includes these modifications and variations.

Claims

1. A digital twin-driven method for fair contribution assessment and personalized rewards in federated learning, characterized in that, Includes the following steps: The physical entities of industrial equipment participating in the joint training of the equipment failure prediction model are used as clients, and client registration and identity verification are completed on the blockchain. The system reviews whether clients are allowed to join the global training, and builds a digital twin for each approved client. In each global training round, the first status index of the approved clients is continuously collected. Based on the first status index, the selected clients are selected and the candidate set for the current global training round is generated. Within the current global training round, several rounds of local training are performed on each selected client based on the initial global model and the local data of the selected clients, generating the gradient of each selected client in the current global training round and uploading it to the server; Based on the gradients uploaded by the selected clients, the selected clients are preprocessed and clustered to obtain a set of strategic equivalence groups; The group-level contribution assessment is performed on the strategic equivalence group set to obtain the individual contribution weight of each selected client; Based on the gradient and individual contribution weight of each selected client, the final global model for the current global training round is generated; Based on the individual contribution weight of each selected client, a personalized local model for the selected client is generated and distributed to the corresponding selected client. Based on the individual contribution weight and stability score of the selected clients, the reward and penalty settlement of the selected clients is executed and notarized on the blockchain. The digital twins of the selected clients are updated, the current global training round is completed and the next global training round is entered, until the final global model obtained after several global training rounds satisfies the convergence condition and the final global model with stable performance is obtained. The final global model, whose performance has reached a stable state, is deployed on the server side for centralized fault reasoning and prediction based on the collected device operation data.

2. The method for fair contribution assessment and personalized rewards in digital twin-driven federated learning according to claim 1, characterized in that, The specific method for completing client registration and identity verification on the blockchain is as follows: The client generates a key pair, including a public key and a private key, using an encryption algorithm; the client submits registration information to the identity management contract and generates a digital signature of the registration information using the private key to prove the client's identity; The identity management contract uses the client's public key to verify the identity of the digital signature of the registration information; After successful identity verification, the identity management contract sets the client's registration status to available and enables the client's access permissions and settlement rules.

3. The method for fair contribution assessment and personalized rewards in digital twin-driven federated learning according to claim 2, characterized in that, The specific method for reviewing whether a client is allowed to join the global training is as follows: When the client requests to join the During each global training round, the client sends a join request based on the client's private key. Generate a digital signature for the join request and submit it to the identity management contract for review; The identity management contract verifies the client's registration status and access permissions based on the digital signature of the join request. If the client has completed registration and has valid access permissions, the verification is successful, and the identity management contract allows the client to join the [new client's] ... Each global training round; hashing the client's request to join the global training and writing it to the blockchain to obtain an auditable record; The server creates an initial packet, calculates its hash value, generates a server signature using its private key, and distributes the initial packet, its hash value, and the server signature to several approved clients. Including the initial global model Training hyperparameters Privacy and security configuration The system distributes the hash of the initial package and the server signature to approved clients.

4. The method for fair contribution assessment and personalized rewards in digital twin-driven federated learning according to claim 1, characterized in that, The first status indicators include online rate, latency jitter, failure rate, energy consumption, and bandwidth usage; The specific method for selecting clients and generating the candidate set for the current global training round based on the first state index is as follows: In the Within each global training round, the first state indicators of each approved client are continuously collected, and the first state indicators are summarized according to preset weights to obtain the stability score of each approved client. The first status indicator of each approved client is compared with a preset security threshold. If any of the first status indicators of the approved client exceeds the preset security threshold, the approved client is determined to be in an abnormal state, and a risk label is generated for the approved client. The risk label is used to indicate the risk status of the approved client, including high risk and low risk. Clients are selected based on a preset stability threshold and the risk label of approved clients. When the stability score of an approved client is greater than or equal to the preset stability threshold and the risk label of the approved client is low risk, the approved client is added to the candidate set of the current global training round. If the number of clients that have passed the review in the current global training round candidate set exceeds the concurrency limit. All approved clients were sorted in descending order of stability score, and the top-performing clients were selected. One client that passes the review will be selected as the selected client.

5. The method for fair contribution assessment and personalized rewards in digital twin-driven federated learning according to claim 1, characterized in that, The specific method for performing several rounds of local training on each selected client based on the initial global model and the local data of the selected clients is as follows: Selected clients from the current global training round candidate set In the In the first global training round, the first round... The final global model obtained after one global training round As the initial global model trained locally, it is selected by the client. Perform several local training iterations on local data to obtain the selected clients. No. Local model in each global training round ; Among them, in the first In each global training epoch, the initial global model is used. As the initial global model for local training; in the... In the first global training round, the first... The final global model obtained after one global training round As the initial global model for local training; The specific method for generating the gradient for each selected client in the current global training round is as follows: Based on selected clients No. Local model in each global training round and the The initial global model in each global training round Calculate the selected clients No. Gradient of each global training epoch .

6. The method for fair contribution assessment and personalized rewards in digital twin-driven federated learning according to claim 1, characterized in that, The specific methods for preprocessing and clustering the selected clients are as follows: For each selected client's gradient, retain the gradient with the largest absolute value. One component is set to zero in each position, and the rest are set to zero to obtain the gradient sparse vector of each selected client. The gradient sparse vector of each selected client is L2 normalized to obtain the normalized sparse vector of each selected client. A set of normalized sparse vectors is generated based on the normalized sparse vector of each selected client; A top-down bisecting K-means method is used to cluster the normalized sparse vector set to generate a set of strategic equivalence groups.

7. The method for fair contribution assessment and personalized rewards in digital twin-driven federated learning according to claim 6, characterized in that, The specific method for clustering a normalized sparse vector set using the top-down bisecting K-means method is as follows: Clustering is performed using a normalized sparse vector set as the initial cluster. The proportion of common non-zero dimensions in the initial cluster is calculated. If the proportion of common non-zero dimensions in the initial cluster is not lower than a preset threshold, then the cluster is considered a cluster. If the initial cluster is strategically equivalent, the selected clients in the initial cluster are determined to be strategically equivalent, clustering is stopped, and the initial cluster is treated as a strategically equivalent group; otherwise, the initial cluster is split into two sub-clusters, the proportion of common non-zero dimensions of the two sub-clusters is calculated, and the proportion of common non-zero dimensions of the two sub-clusters is compared with the threshold. The size of the subcluster is used to determine whether to split it into subclusters, until the proportion of the common non-zero dimension of all subclusters is not less than the threshold. Alternatively, the total number of all clusters may reach the target number of clusters. After clustering is completed, each cluster is treated as a strategic equivalence group, and a set of strategic equivalence groups is generated based on all strategic equivalence groups, including several strategic equivalence groups.

8. The method for fair contribution assessment and personalized reward of digital twin-driven federated learning according to claim 1, characterized in that, The specific method for evaluating the group-level contribution of the strategic equivalence group set is as follows: Based on the gradients uploaded by all selected clients, calculate the gradients for all client groups. Group model, client group It is a subset of the current global training round candidate set; Based on the Initial global model and client group for each global training round The group model establishes feature functions; Calculate the cooperation Shapley value for each selected client in the set of strategic equivalence groups; The cooperation Shapley value of each selected client in the strategic equivalence set is normalized to obtain the individual contribution weight of each selected client.

9. The method for fair contribution assessment and personalized rewards in digital twin-driven federated learning according to claim 1, characterized in that, The specific method for generating the final global model for the current global training round is as follows: The gradients received from each selected client are weighted and aggregated based on the individual contribution weight of each selected client, and the aggregation results are used to analyze the gradients of the selected clients. The initial global model of each global training round is updated to obtain the final global model of the current global training round; The specific method for generating personalized local models for selected clients is as follows: Based on the individual contribution weight of each selected client, a personalized model for each selected client is generated by combining the local model of each selected client.

Images