Reputation mechanism based digital twin incentive method in industrial internet of things

Abstract

The present application belongs to the technical field of federated learning and edge computing, and discloses a digital twin incentive method based on a reputation mechanism in an industrial Internet of Things, which is completed by an incentive mechanism combined by a short-term incentive mechanism and a long-term incentive mechanism, namely a reputation management mechanism, and comprises the following steps: step 1, constructing a digital twin transmission model; step 2, designing the type of digital twin and classifying; step 3, using federated learning to aggregate the model of digital twin, compensating the edge digital twin, obtaining the utility of each digital twin device and the utility of the task publisher, and delaying the analysis of the federated learning of each digital twin; and step 4, designing a fair and stable federated learning model. The present application adds a hybrid long-term and short-term incentive mechanism into the federated learning of the digital twin in the edge server and the cloud, so as to avoid the information transmission error and inaccuracy between the digital twin and the cloud.

Description

Technical Field

[0001] This invention belongs to the field of federated learning and edge computing technology, specifically the design of a reputation-based digital twin incentive method for the Industrial Internet of Things. Background Technology

[0002] The concept of digital twins, proposed in recent years, is a conceptual system for the interaction between the physical world and digital space. It is both a new technology and a new paradigm. Since its formal inception in 2009, after a decade of development, digital twins have evolved into a new industry. Besides its deepening development in manufacturing, it has also found applications in smart cities, energy industries, healthcare, and defense industries. The combination of industrial IoT devices and artificial intelligence has recently attracted great interest, and the aggregation of digital twin-assisted co-sensing with fair incentives and robust federated learning has been researched. The application of digital twins and the advancement of artificial intelligence technology promise to realize digital twin-assisted co-sensing, opening up new horizons for various on-demand smart city applications such as environmental monitoring, public safety, and traffic control. With their flexible mobility and line-of-sight connectivity, digital twins can autonomously sense areas at any time and any location.

[0003] However, to date, no technology has been proposed that utilizes incentive mechanisms between edge digital twins and the cloud. Summary of the Invention

[0004] To address the aforementioned technical issues, this invention provides a reputation-optimized digital twin incentive method for the Industrial Internet of Things (IIoT). This method is applicable to fair and robust federated learning mechanisms for digital twin-assisted collective intelligence perception. Specifically, firstly, within the federated learning framework, a 5G heterogeneous network supported by edge computing is used to provide high-data-rate, low-latency near-end federated learning services to edge digital twins. Then, based on contract theory, an optimal incentive mechanism is designed to accurately and fairly incentivize edge digital twins to participate in cloud-based digital twin federated learning under conditions of information asymmetry.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solution:

[0006] This invention relates to a reputation-based digital twin incentive method for the Industrial Internet of Things (IIoT). The digital twin incentive method combines a short-term incentive mechanism (i.e., a contract-based incentive mechanism) with a long-term incentive mechanism (i.e., a reputation management mechanism), and specifically includes the following steps:

[0007] Step 1: Construct a digital twin transmission model for the Industrial Internet of Things;

[0008] Step 2: The types of digital twins were designed and classified.

[0009] Step 3: Use federated learning to aggregate models of digital twins, use contract theory to compensate edge digital twins accordingly, and obtain the utility of each digital twin device and the utility of the task issuer. Finally, perform a delay analysis on each federated learning of digital twins.

[0010] Step 4: Design a fair and stable federated learning model (FRFL) to address the threats that exist in the federated learning process.

[0011] A further improvement of the present invention is that: step 1, which constructs a digital twin transmission model for an industrial Internet of Things (IoT), includes several industrial IoT devices, edge digital twins corresponding to the industrial IoT devices, and cloud digital twins, specifically including two communication modes:

[0012] Edge Digital Twin to Central Server Communication Mode (D2C): Digital twins with high signal-to-noise ratio channels directly upload local parameters to the SBS via D2C communication;

[0013] Edge-to-edge digital twin (D2D) mode: When a digital twin encounters a low signal-to-noise ratio channel, D2D communication is used to transmit the data to a nearby edge server operating in D2C communication mode. The relay edge server then transmits the digital twin to the SBS.

[0014] A further improvement of this invention is that the modeling process for the edge digital twin to central server communication mode (D2C) in step 1 is as follows:

[0015] The path loss experienced by digital twin i in D2C transmission is:

[0016]

[0017] Among them, K δ It is an additional decay function, f c It is the carrier frequency, d i，h (t) is the Euclidean distance from the edge server where the digital twin is located to the terminal server, h∈{SBS,MBS};

[0018] set up Given the transmit power of digital twin i in time slot t, the average receive power of base station h reaches:

[0019]

[0020] In time slot t, the uplink data rate from digital twin i on the edge base station to base station h is...

[0021]

[0022] Among them, B D2C It is the sub-channel bandwidth. It is the noise power of additive white Gaussian noise;

[0023] A further improvement of this invention is that the modeling process of edge-to-edge digital twin (D2D) mode in step 1 is as follows: a free space path loss model is adopted, wherein the D2D path loss between digital twin i and digital twin j is distance-dependent, i.e.:

[0024]

[0025] Where d i，j (t) is the three-dimensional Euclidean distance between digital twins i and j, l represents the path loss exponent, and B D2D The bandwidth of D2D is given by the data transfer rate from digital twin i to j in time slot t.

[0026]

[0027] A further improvement of the present invention is that, in the federated learning process of step 3, each task publisher needs to periodically obtain a trained local model from a set of digital twins, and the task publisher of task k provides contracts for all independent digital twins.

[0028]

[0029] The maximum wait delay in each global iteration Ω represents the proportion of the allocated model profits of the digital twin. k For a set of J data reward contract projects Contract Item Ω j，k =(s j，k p j，k The required data size s for each type of digital twin is specified. j，k , service k and reward p j，k The relationship between them.

[0030] A further improvement of the present invention is that: the utility of each digital twin device obtained in step 3 is: in the nth global iteration, the contract term Ω is selected. j，k The utility of a J-type digital twin is represented as

[0031]

[0032] in: This refers to the total computation and transmission time of the digital twin in the nth global iteration of the federated learning task k. This is the reputation value of the digital twin. In the nth global iteration of task k, the number of participating digital twins is... I j The number of J-type digital twins

[0033]

[0034]

[0035] A further improvement of the present invention is that: the delayed analysis of each federated learning of the digital twin in step 3 is specifically as follows:

[0036] The delay of digital twin i in the nth global iteration of federated learning task k is determined by the model parameters. The local computing time of the device where the digital twin i resides and wireless transmission time Composition, including local computation time:

[0037]

[0038] Θ k It is the number of local iterations. f is the number of CPU cycles required by the device containing the digital twin i to train a model using a single data sample. i It is the CPU rotation speed frequency of the device where the digital twin i resides, s i，k It is the data size in task k of the digital twin i.

[0039] Transmission time for uploading model parameters to the cloud base station in both D2C and D2D modes of the transmission task. for:

[0040]

[0041] in yes The size of i′ is a nearby digital twin with a high signal-to-noise ratio. Data samples are transmitted to the cloud server through i′. h is the base station, and γ is the base station. i，h For the data rate from digital twin i to base station h, the transmission time via wired backhaul between mobile base station MBS and serving base station SBS is negligible. Total time:

[0042]

[0043] A further improvement of the present invention is that the utility of the task publisher in step 3 is specifically as follows:

[0044] For the task issuer of task k, the overall utility is the difference between their satisfaction and the total payoff for the digital twin during model training:

[0045]

[0046] A(s j，k ) represents a satisfaction function related to the number of data samples used in model training.

[0047] A further improvement of the present invention is that the Federated Learning Model (FRFL) in step 4 includes

[0048] Fair incentive mechanisms: including contract theory, feasibility studies of contract theory, and the design of the most ideal contract mechanism;

[0049] Robust model aggregation: includes contribution measures and fair profit distribution;

[0050] Dynamic reputation assessment: includes historical training records, behavioral impact, and fading-out effects over time.

[0051] A further improvement of this invention is that the reputation of digital twin i in dynamic reputation assessment is calculated as follows:

[0052]

[0053] Where t is the current time slot. These are the input parameters: v1 and v2 are the lower bound asymptotes, ψ1 is the growth rate, ψ2 determines the maximum growth rate, and λ0 depends on the initial value. v3 affects the point where the maximum asymptote growth occurs, therefore

[0054]

[0055] The digital twins are categorized based on their participation and training behaviors, and different fading weights and reputation updates are assigned to them. Behavioral and fading effects from historical learning records are utilized and aggregated to estimate the reputation update, i.e., the input parameters.

[0056]

[0057] Here, μ1 and μ2 are time elapsed weights, satisfying 0 < μ1 < μ2 < 1. A fading weight μ2 and a positive reputation decrease ΔR0 are used to penalize the non-participatory behavior of the digital twin. CI is the round contribution index of digital twin i in time slot t of task k. It is used to identify whether the contribution of digital twin i in the global iteration is higher than the average value. ΔR0 = {ΔR1, ΔR2} are reputation increments and decrements, satisfying 0 < ΔR1 < ΔR2. ΔR1 is the reputation increment of participating digital twins whose contribution index CI in a round is higher than the average value as a reward, and ΔR2 is the reputation decrement of participating digital twins whose contribution index CI in a round is lower than the average value as a penalty.

[0058] The beneficial effects of this invention are: by introducing a digital twin and an incentive mechanism, this invention incorporates a hybrid long-term and short-term incentive mechanism into the federated learning between the digital twin in the edge server and the cloud, thereby avoiding errors and inaccuracies in information transmission between the digital twin and the cloud.

[0059] This invention establishes a transmission model by introducing edge computing, enabling efficient transmission of digital twins with high and low signal-to-noise ratios.

[0060] This invention introduces federated learning to acquire information from the cloud and digital twins, uses contract theory to compensate for the costs of edge digital twins, and incentivizes them to participate in federated learning tasks, achieving low latency and high efficiency.

[0061] This invention introduces the FRFL mechanism and uses contract theory to design optimal contracts to maximize the utility of the task issuer, while ensuring the fairness, individual rationality, and incentive compatibility of the digital twin under asymmetric information conditions.

[0062] This invention collects relevant information about digital twins by applying robust aggregation rules and dynamic reputation estimation, filters out outliers with low-quality local updates, and encourages legitimate behavior of digital twins while suppressing their misbehavior, thereby enabling the recruitment of trustworthy digital twins in federated learning tasks. Attached Figure Description

[0063] Figure 1 This is a flowchart illustrating the excitation method of the present invention.

[0064] Figure 2 This is a system model diagram of the excitation method of the present invention.

[0065] Figure 3 This is a system model diagram of the federated learning model of this invention. Detailed Implementation

[0066] The embodiments of the present invention will be disclosed below with reference to the drawings. For clarity, many practical details will be described in the following description. However, it should be understood that these practical details are not intended to limit the invention. That is, in some embodiments of the invention, these practical details are not essential.

[0067] like Figure 1-3 As shown, this invention is a reputation-based digital twin incentive method for the Industrial Internet of Things (IIoT). This method combines a short-term incentive mechanism (contract-based) with a long-term incentive mechanism (reputation management). The invention provides a hybrid short- and long-term incentive mechanism incorporated into the digital twins on the edge server and a cloud-based federated learning method to solve the problem. Each edge digital twin possesses a certain computing power to meet the needs of performing simple local tasks. Conversely, the cloud-based digital twin possesses extremely strong computing power, capable of completing computational tasks, providing edge computing services to the edge digital twins, performing federated learning, and aggregating cloud-based digital twin models.

[0068] like Figure 1-2 As shown, the digital twin stimulation method of the present invention specifically includes the following steps:

[0069] Step 1: Construct a digital twin transmission model for the Industrial Internet of Things.

[0070] The digital twin transmission model for the Industrial Internet of Things (IIoT) constructed in this invention includes several IIoT devices, edge digital twins corresponding to the IIoT devices, and cloud digital twins. Specifically, it includes two communication modes: edge digital twin to central server communication mode (D2C) and edge digital twin to edge digital twin mode (D2D).

[0071] Edge Digital Twin to Central Server Communication Mode (D2C): Digital twins with high signal-to-noise ratio channels can directly upload local parameters to the SBS via D2C communication;

[0072] Edge-to-Edge Digital Twin (D2D) Mode: If the signal-to-noise ratio (SNR) of the direct D2C channel is low, it is difficult to provide a high data rate to support timely data transmission in D2D mode. Therefore, when encountering a digital twin with a low SNR, D2D communication can be used to transmit the data to a nearby edge server operating in D2C communication mode. The relay edge server can then pass the digital twin to the SBS.

[0073] The modeling process for the edge digital twin to central server communication mode (D2C) is as follows:

[0074] The path loss experienced by digital twin i in D2C transmission is:

[0075]

[0076] Among them, K δ It is an additional decay function, f c It is the carrier frequency, d i，h(t) is the Euclidean distance from the edge server where the digital twin is located to the terminal server, h∈{SBS,MBS};

[0077] set up Given the transmit power of digital twin i in time slot t, the average receive power of base station h reaches:

[0078]

[0079] In time slot t, the uplink data rate from digital twin i on the edge base station to base station h is...

[0080]

[0081] Among them, B D2C It is the sub-channel bandwidth. It is the noise power of additive white Gaussian noise;

[0082] The modeling process for edge-to-edge digital twin (D2D) mode is as follows: a free-space path loss model is adopted, where the D2D path loss between digital twin i and digital twin j is distance-dependent, i.e.:

[0083]

[0084] Where d i，j (t) is the three-dimensional Euclidean distance between digital twins i and j, l represents the path loss exponent, and B D2D Let the bandwidth of D2D be denoted as . Ignoring interference from other devices, then, in time slot t, the data transmission rate from digital twin i to j is:

[0085]

[0086] Step 2: The types of digital twins were designed and classified.

[0087] Because digital twins differ in computing power, the costs of collecting and training data also differ. The costs of collecting and training digital twins, such as time and energy costs, are proportional to the size of the training data used, i.e., c. i s i，k This invention employs marginal data usage cost (c). i This characterizes the heterogeneity of digital twins in federated learning. Generally, digital twins are categorized into J types based on the unit cost of data training. The marginal data usage cost belongs to the j-th cost level, c. j , The digital twin is called a J-type digital twin.

[0088] Assume 0 < c min =c1<…<c j =c max Here, c min and c max It is the lower bound of cj and c j The upper bound of the concept. This invention assumes that the type of each digital twin remains constant throughout the learning process of each task.

[0089] Step 3: Use federated learning to aggregate models of digital twins, use contract theory to compensate edge digital twins accordingly, and obtain the utility of each digital twin device and the utility of the task issuer. Finally, perform a delay analysis on each federated learning of digital twins.

[0090] In federated learning, each task issuer needs to periodically retrieve a trained local model from a set of digital twins. Because the types of digital twins are private, information asymmetry exists between the task issuer and the worker digital twins. In economics, contract theory studies contractual arrangements designed for different types of individuals under conditions of information asymmetry, and is applicable to the federated learning process between monopolistic task issuers and individual digital twins. This invention utilizes contract theory to compensate for the costs incurred by the digital twins and incentivize their participation in the federated learning task. The task issuer of task k provides contracts to all independent digital twins.

[0091]

[0092] For all types of individual digital twins, it includes the maximum wait latency in each global iteration, i.e. The proportion of distributed model profits involved in digital twins and a collection of J data reward contract projects Since artificial intelligence models are typically profitable. It can be viewed as a long-term bonus for participants, Λ k It represents the future profit from training the AI ​​model for task k. Contract item Ω j，k =(s j，k p j，k The required data size s for each type of digital twin is specified. j，k Task k and reward p j，k The relationship between them.

[0093] The utility of each digital twin device obtained in step 3 is: in the nth global iteration, the contract term Ω is selected. j，k The utility of a J-type digital twin is represented as

[0094]

[0095] in: This refers to the total computation and transmission time of the digital twin in the nth global iteration of the federated learning task k. It is the reputation value of the digital twin.

[0096] In the nth global iteration of task k, the number of participating digital twins is I j The number of J-type digital twins

[0097]

[0098]

[0099] Furthermore: The specific details of the delayed analysis for each federated learning iteration of the digital twin are as follows:

[0100] The delay of digital twin i in the nth global iteration of federated learning task k is determined by the model parameters. The local computing time of the device where the digital twin i resides and wireless transmission time Composition, including local computation time:

[0101]

[0102] Θ k It is the number of local iterations. f is the number of CPU cycles required by the device containing the digital twin i to train a model using a single data sample. i It is the CPU rotation speed frequency of the device where the digital twin i resides, s i，k It is the data size in task k of the digital twin i.

[0103] Considering the two modes of transmission tasks (i.e., D2C and D2D), the transmission time for uploading model parameters to the cloud's main base station. for:

[0104]

[0105] in yes The size of i′ is a nearby digital twin with a high signal-to-noise ratio. Data samples are transmitted to the cloud server through i′. h is the base station, and γ is the base station. i，h For the data rate from digital twin i to base station h, the transmission time via wired backhaul between mobile base station MBS and serving base station SBS is negligible. Total time:

[0106]

[0107] Furthermore, the specific utility of the task publisher in step 3 is as follows:

[0108] For the task issuer of task k, the overall utility is the difference between their satisfaction and the total payoff for the digital twin during model training:

[0109]

[0110] A(s j，k The function denoted by represents a satisfaction function related to the number of data samples used in model training. The satisfaction function is modeled...

[0111]

[0112] λ s It is the satisfaction coefficient, q j，k It represents the quality of service of the j-type digital twin in task k.

[0113] Step 4: Design a fair and stable federated learning model (FRFL) to address the threats in the federated learning process and optimize the problem to design a contract-based incentive and reputation management mechanism.

[0114] There are three types of threats that the system reconsiders, which can affect the security and efficiency of the federated learning process.

[0115] ① The selfishness of digital twins. Individual digital twins, as rational and selfish agents, are unwilling to participate in collaborative learning processes involving multiple senses. Therefore, there are insufficient participants in federated learning tasks, which deteriorates the accuracy of the learning model.

[0116] ② Free-rider attacks by digital twins. Self-interested digital twins may launch free attacks to unfairly gain benefits, such as stealing lucrative AI models and cheating on rewards in contracts without contributing to the federated learning process. Therefore, the enthusiasm of honest digital twins may be dampened.

[0117] ③ Byzantine digital twins. Byzantine digital twins may maliciously exit the learning process and merge and send arbitrary adversarial local updates to jeopardize the federated learning process, for example, by sharing meaningless, redundant, or even erroneous local model gradients.

[0118] To address the aforementioned threats in the federated learning process, this invention designs a fair and stable federated learning model (FRFL).

[0119] like Figure 3 As shown, the Federated Learning Model (FRFL) includes

[0120] Fair incentive mechanisms: including contract theory, feasibility studies of contract theory, and the design of the most ideal contract mechanism;

[0121] Robust model aggregation: includes contribution measures and fair profit distribution;

[0122] Dynamic reputation assessment: includes historical training records, behavioral impact, and fading-out effects over time.

[0123] This invention designs a short-term incentive mechanism, namely a fair incentive mechanism or a data reward contract, to compensate for the participation cost of the digital twin. This invention also designs a long-term incentive mechanism, namely robust model aggregation and dynamic reputation evaluation, namely reputation value and model profit distribution. The long-term incentive stimulates high-quality local model training of the digital twin.

[0124] First, this invention employs contract theory to design optimal contracts, maximizing the utility of task issuers while ensuring fairness, individual rationality (IR), and incentive compatibility (IC) of digital twins under asymmetric information conditions. Upon receiving local model updates from digital twins, to address the Byzantine digital twin problem and improve model performance, this invention utilizes a robust aggregation algorithm to filter out outliers of low-quality local updates. Based on the Contribution Index (CI) measurement, this invention also employs a fair allocation rule to equitably distribute future model profits among participants. Finally, based on historical learning records and user behavior, this invention dynamically estimates the long-term reputation of marginal digital twins to encourage legitimate behavior and suppress inappropriate behavior, aiming to recruit trustworthy digital twins in federated learning tasks.

[0125] Regarding short-term incentives, this invention first proposes an optimization problem, then conducts feasibility and optimization analyses. Finally, it demonstrates the contractual fairness of the incentive mechanism of this invention in terms of participation fairness and reward fairness.

[0126] Feasibility and optimality are the fundamental goals of contract mechanisms in practical deployment. Short-term incentives, i.e., contract-based incentive mechanisms, specifically include:

[0127] 1. Contractual Feasibility: A contract is considered feasible if each piece of equipment achieves the maximum non-negative utilization rate shown in the following formula by faithfully adopting the contractual items appropriate to its type:

[0128]

[0129] The contract is feasible if each device achieves its maximum non-negative utility.

[0130] 2. Contract Optimality: A contract is considered optimal if it maximizes the utility of the designer (i.e., the task issuer) among all feasible contracts.

[0131]

[0132] According to the principle of revelation, a viable contract is equivalent to saying that for all types of digital twins, both IR and IC constraints are satisfied simultaneously.

[0133] 3. Individual Rationality (IR): If each j-type digital twin is involved in a contract project Ω designed for its type... j，k =(s j，k p j，k A contract obtains non-negative utility, and mathematically speaking, a contract satisfies IR.

[0134]

[0135] 4. Individual Compatibility (IC): A contract satisfies IC, meaning that each j-type digital twin mathematically prefers contract items Ω prepared for its own type. j，k =(s j，k p j，k ), and not others.

[0136]

[0137] As the contract designer, the task publisher's goal is to design the optimal contract that maximizes its benefit during the federated learning process for each task k, denoted as...

[0138]

[0139] The optimization problem can be simulated as follows:

[0140] Question 1: max π k (Ω k )

[0141]

[0142] In the network, it was observed that the task publisher in each global iteration (i.e. The utility functions in the equation are independent of each other. Therefore, in the following equation... The maximum value is equal to each of the global iterations of task k. Based on this observation, the optimal contract design problem, i.e., Problem 1, can be simplified as follows.

[0143] Question 2: max

[0144]

[0145] let Let represent the optimal data size strategy. To solve problem 2, we first represent a relaxed problem without monotonicity constraint c1, and then check whether the obtained solution satisfies c1.

[0146] Finally, the fairness of the designed optimal contract is analyzed.

[0147] 5. Fairness of Participation: Fairness of participation is guaranteed if any rational and self-interested digital twin is incentivized to honestly follow the contractual agreements, meaning they have no incentive to withdraw from the federated learning process and to report the cost of dishonesty in exchange for compensation.

[0148] 6. Reward Fairness: If, during model training, the participating digital twins with higher local model quality receive higher rewards while the non-participating digital twins receive no rewards, then reward fairness is guaranteed.

[0149] 7. Contract fairness: If both participation fairness and reward fairness are guaranteed, then the contract is fair.

[0150] Regarding long-term incentives, namely robust model aggregation and dynamic reputation assessment, after the learning task is completed, the corresponding reputation value is updated according to the dynamic reputation assessment component. This component utilizes historical training records to evaluate the credibility of the digital twin from a long-term perspective. Generally, a user's reputation tends to gradually increase when performing a series of legitimate actions, while dishonest behavior can quickly damage it. Intuitively, after observing a digital twin's high learning contribution to the task, its reputation will increase slightly. Conversely, if the digital twin provides Byzantine updates—for example, meaningless, redundant, and erroneous ones—its reputation will decrease considerably. The widely adopted logistic function, known as the Richard curve, is used for modeling the behavior of digital twins because it grows fastest in the middle and slowest on the left and right sides. The reputation of digital twin i is calculated as follows:

[0151]

[0152] Where t is the current time slot. These are the input parameters: v1 and v2 are the lower bound asymptotes, ψ1 is the growth rate, ψ2 determines the maximum growth rate, and λ0 depends on the initial value. v3 affects the point where the maximum increase of the asymptote occurs. Clearly,

[0153] Generally speaking, recent learning records are more important than previous records.

[0154] This invention assigns a higher weight to the most recent records than to previous ones. Furthermore, the rates of increase and decrease in reputation values ​​should be different. Therefore, this invention classifies digital twins based on their engagement and training behaviors and assigns them different fading weights and reputation updates. Based on the above observations, we utilize behavioral and fading effects from historical learning records and summarize them to estimate the reputation update, i.e., the input parameters.

[0155]

[0156] Here, μ1 and μ2 are time-elapse weights, satisfying 0 < μ1 < μ2 < 1. A higher fading weight μ2 and a positive reputation decrease ΔR0 are used to penalize the non-participatory behavior of the digital twin. CI is the round number of digital twin i in time slot t of task k.

[0157] It is used to identify whether the contribution of digital twin i in the global iteration is higher than the average. ΔR0 = {ΔR1, ΔR2} are reputation increments and decrements, satisfying 0 < ΔR1 < ΔR2. That is, ΔR1 is the reputation increment of participating digital twins with a CI above the average level in a round as a reward, while ΔR2 is the reputation decrement of participating digital twins with a CI below the average level in a round as a penalty.

[0158] This invention introduces a hybrid long-term and short-term incentive mechanism into the federated learning process between digital twins on edge servers and the cloud, avoiding errors and inaccuracies in information transmission between the digital twins and the cloud. It establishes a transmission model using edge computing to efficiently transmit digital twins with both high and low signal-to-noise ratios. Furthermore, by incorporating federated learning, it acquires information from both the cloud and the digital twins, using contract theory to compensate for the costs of edge digital twins and incentivize their participation in federated learning tasks. This achieves low latency and high efficiency. The invention employs contract theory to design optimal contracts, maximizing the utility of task publishers while ensuring fairness, individual rationality, and incentive compatibility of digital twins under asymmetric information conditions. Finally, by applying robust aggregation rules and dynamic reputation estimation, it collects relevant information from digital twins, filtering out outliers with low-quality local updates while encouraging legitimate behavior and suppressing inappropriate behavior, thus recruiting trustworthy digital twins in federated learning tasks.

[0159] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A digital twin incentive method based on a reputation mechanism in an industrial Internet of Things, characterized in that: The digital twin incentive method is completed by a short-term incentive mechanism, i.e., a contract-based incentive mechanism, and a long-term incentive mechanism, i.e., a reputation management mechanism, and specifically includes the following steps: Step 1: Constructing an industrial Internet of Things digital twin transmission model; Step 2: Designing the type of digital twin and classifying the digital twin; Step 3: Using federated learning to aggregate the model of the digital twin, using contract theory to compensate the corresponding edge digital twin, and at the same time deriving the utility of each digital twin device and the utility of the task publisher, and finally analyzing the delay of each federated learning of the digital twin; Step 4: Designing a fair and stable federated learning model FRFL to deal with threats in the federated learning process and optimize the problem to design a contract-based incentive mechanism and a reputation management mechanism, wherein In the federated learning process of step 3, each task publisher needs to regularly obtain a trained local model from a group of digital twins, and the task publisher of task k provides a contract to all independent digital twins a maximum waiting delay for each global iteration, a proportion of the allocated model profit for the digital twin, a set of J data reward contract items a contract item specifies the data size required for each kind of digital twin a relationship between task k and reward ; The utility of each digital twin device resulting from step 3 is: in the nth global iteration, the utility of the j-type digital twin of the contract item is expressed as wherein: denotes the total computation and transmission time of a digital twin in the n-th global iteration of the federated learning task k, is the reputation value of a digital twin, in the n-th global iteration of the task k, with the number of participating digital twins being = , is the number of j-type digital twins = = ； The delay analysis of the federated learning of each digital twin in step 3 is specifically: the delay of the digital twin i in the n-th global iteration of the federated learning task k consists of the local computation time of the device where the digital twin i resides in the wireless transmission time and the wireless transmission time where the local computation time: / is the number of local iterations, is the number of CPU cycles of the device where the digital twin i is located when using one data sample for model training, is the CPU rotation frequency of the device where the digital twin i is located, is the data size in the task k of the digital twin i, Transmission time of D2C and D2D modes of transmission tasks to upload model parameters to the cloud total base station For: wherein is the size of is a nearby digital twin with a high signal-to-noise ratio, by transmitting the data sample to a cloud server, h is a base station, is the data rate from digital twin i to base station h, the transmission time between mobile base station MBS and serving base station SBS via wired backhaul is negligible, total time: = + ； The utility of the task publisher in step 3 is specifically: For the task publisher of task k, the overall utility is the difference between the satisfaction and the total payment to the digital twin in model training: )+(1- represents a satisfaction function related to the number of data samples for model training, the satisfaction function modeling A (1+ ) is a satisfaction coefficient, is the quality of service of the j-type digital twin in task k; The federated learning model FRFL in step 4 includes Fair incentive mechanism: including contract theory, contract theory feasibility study, and the most ideal contract mechanism design; Robust model aggregation: including contribution measures and fair profit distribution; Dynamic reputation assessment: including historical training records, behavior impact, and fading effect over time, The fair incentive mechanism is a short-term incentive that compensates the digital twin for its participation cost, The robust model aggregation and dynamic reputation assessment are long-term incentives that stimulate high-quality local model training of the digital twin; The reputation calculation of the digital twin i in the dynamic reputation assessment is: where t is the current time slot, is an input parameter, and is the lower asymptote, is the growth rate, determines the maximum growth rate, depends on the initial value ), influences the point at which the asymptote maximum growth occurs, thus ∈（0,1） According to the participation behavior and training behavior of the digital twin, it is classified and different extinction weights and reputation updates are assigned, the behavior effect and extinction effect in the historical learning record are utilized and summarized to estimate the reputation update, i.e. input parameters ）= wherein, and are time elapse weights, satisfying 0 < < 1, <1, adopting a decay weight and positive reputation decrement to punish non-participating digital twins, is the round contribution index CI of digital twin i in time slot t of task k, is used to identify whether the contribution of digital twin i in the global iteration is higher than the average, is the reputation increment / decrement value, satisfying 0 < < 1, , is the reputation increment of participating digital twins whose round contribution index CI is higher than the average as a reward, is the reputation decrement of participating digital twins whose round contribution index CI is lower than the average as a punishment.​​ 2. The method as claimed in claim 1, wherein the method is based on a reputation mechanism in an industrial internet of things (IIoT) for incentivizing digital twins. The step 1 of constructing an industrial Internet of Things digital twin transmission model includes a plurality of industrial Internet of Things devices, edge digital twins corresponding to the industrial Internet of Things devices, and cloud digital twins, and specifically includes two communication modes: Edge digital twin to center server communication mode D2C: digital twins with high signal-to-noise ratio channels directly upload local parameters to SBS through D2C communication; Edge digital twin to edge digital twin mode D2D: when encountering a digital twin with a low signal-to-noise ratio channel, D2D communication is used, and the data is transmitted to a nearby edge server running in D2C communication mode, and the digital twin is delivered to SBS by the relay edge server.

3. The method of claim 2, wherein: The modeling process of the edge digital twin to center server communication mode D2C in step 1 is: The path loss experienced by digital twin i in D2C transmission is: (t)= +20 log(4π (t) / c, wherein, is an additional attenuation function, is a carrier frequency, is the Euclidean distance of the edge server to the terminal server where the digital twin resides, h e {SBS, MBS}; Let (t) is the transmit power of digital twin i at time slot t, the average received power of base station h reaches: (t)= (t) / At time slot t, the uplink data rate from digital twin i on the edge base station to base station h is wherein is a subchannel bandwidth, is a noise power of additive white Gaussian noise.

4. The method as claimed in claim 2, wherein the method further comprises: The modeling process of edge digital twin to edge digital twin mode D2D in step 1 is: the free space path loss model is adopted, in which the D2D path loss between digital twin i and digital twin j is related to the distance, that is: ， wherein ) is the three-dimensional Euclidean distance between digital twins i and j, denotes the path loss exponent, denotes the bandwidth for D2D, the data transmission rate from digital twin i to j at time slot t is gamma i,j (t) = B D2D log2(1 + P i Tr(t) ≡ i,j D2D(t) / phi2).

Images