Time-sensitive perception federated learning participation device incentive method in industrial internet of things

Abstract

The application provides a time-sensitive perception federated learning participating device incentive method in an industrial Internet of Things, and the steps are as follows: an edge server selects candidate devices according to resource condition requirements and contract item of a federated learning task design; the data freshness of the candidate devices is measured by using an average AoI, the comprehensive reputation value of the candidate devices is calculated, the candidate devices are selected as participating devices by using the comprehensive reputation value, and the participating devices perform a federated learning task to train a local model; the benefit function of the edge server is established as a target function according to the contract item of the participating devices, a constraint condition is established, the optimal contract design problem under the conditions of information asymmetry and information symmetry is deduced, and the optimal contract item of the participating devices is obtained by using a convex optimization tool; and the edge server pays the participating devices according to the rewards of the optimal contract item by using a blockchain. The application can effectively guarantee the benefits of the edge server and the devices, and guarantee the fairness of the incentive and the enthusiasm of the participating devices.

Description

Technical Field

[0001] This invention relates to the technical field of federated learning, and more particularly to a method for incentivizing participating devices in federated learning with timeliness awareness in the Industrial Internet of Things. Background Technology

[0002] The Industrial Internet of Things (IIoT) is considered a key technology for upgrading traditional industries to a new intelligent stage. IIoT is characterized by ubiquitous connectivity, covering almost the entire industrial chain. Widely deployed IIoT devices generate unprecedented amounts of big data every day. With the rapid development of artificial intelligence, the sensory data from IIoT devices can be fully explored and utilized. However, this data can be privacy-sensitive. Improper handling may lead to privacy and security issues, thus hindering the pace of industrial intelligence development. This patent application primarily explores an incentive mechanism approach for federated learning.

[0003] Federated learning is an emerging distributed machine learning paradigm where devices can train machine learning models locally without sending raw data to others. Federated learning has recently been widely adopted to protect data privacy. Despite these significant benefits, the practical application of federated learning still faces considerable challenges. A major challenge is that devices participating in federated learning tasks need to contribute their resources and data while also facing the risk of privacy breaches. Therefore, without sufficient compensation, most devices are unwilling to participate in federated learning training. This lack of participants will significantly impact the performance of federated learning. Therefore, designing an effective incentive mechanism for federated learning to encourage more devices to participate in federated training is of great importance.

[0004] Incentive mechanisms have been extensively studied in various contexts (such as P2P networks and mobile crowd sensing networks). However, designing incentive mechanisms for federated learning remains challenging due to several obstacles. The first is the difficulty in fairly quantifying the contribution of each participant in federated learning; the second is how to leverage the nature of federated learning to design incentive mechanisms for participants, thereby improving the performance of federated learning models. To address these challenges, much research has been based on different technologies (such as game theory, auction theory, and blockchain) and metrics (such as data quality, data volume, and reputation value). However, these studies have not considered the impact of data freshness. In some industrial IoT applications, such as autonomous vehicles and drones, data freshness is a crucial factor, significantly impacting accuracy and application efficiency.

[0005] Furthermore, due to the mobility of devices and the dynamic nature of edge networks, edge servers cannot obtain specific information about the devices' behavior during federated learning tasks, and the behavior of Industrial IoT devices depends entirely on their own ethical constraints. This leads to information heterogeneity between edge servers and devices. Because of information asymmetry, edge servers may not know the exact information of each device, such as reputation value, local data quality, and available resources. This information asymmetry between edge servers and devices can impose additional costs on the entire federated learning task. Therefore, designing an effective mechanism that meets the requirements of different devices and mitigates the impact of information asymmetry is meaningful. Summary of the Invention

[0006] To address the technical problem that information asymmetry between edge servers and devices in existing federated learning systems can lead to additional costs for the entire federated learning task, this invention proposes a timeliness-aware federated learning incentive method for participating devices in the Industrial Internet of Things (IIoT). This method can effectively guarantee the efficiency of edge servers and devices, and ensure the fairness of incentives and the enthusiasm of participating devices.

[0007] To achieve the above objectives, the technical solution of the present invention is implemented as follows: a federated learning method for incentivizing devices with timeliness awareness in the Industrial Internet of Things, comprising the following steps:

[0008] Step 1. The edge server designs resource requirements and contract terms based on the federated learning task. After the resource requirements are met, the industrial IoT device is selected as a candidate device.

[0009] Step 2. Measure the data freshness of candidate devices using the average AoI, calculate the comprehensive reputation score of candidate devices, select candidate devices as participating devices based on the comprehensive reputation score, and have participating devices perform federated learning tasks to train the local model;

[0010] Step 3. Based on the contract terms of the participating devices, establish the benefit function of the edge server as the objective function. Based on the individual rationality constraints and incentive compatibility constraints of the participating devices, establish the constraints, derive the optimal contract design problem under the conditions of information asymmetry and information symmetry, and use convex optimization tools to obtain the optimal contract terms of the participating devices.

[0011] Step 4. The edge server uses the blockchain to pay the participating devices based on the reward of the best contract item. The edge server cannot refuse to pay.

[0012] Preferably, the resource requirements for the federated learning task design include data size, data type, CPU frequency, required training accuracy, and training time threshold.

[0013] The contract terms designed based on the federated learning task are (R) n (f n ), fn ), f n R is the computing resources required for an industrial IoT device n to perform a federated learning task. n This refers to the corresponding reward received by industrial IoT device n.

[0014] Preferably, the method for calculating the comprehensive reputation value of candidate devices is as follows: comprehensive reputation value for:

[0015]

[0016] in, For direct reputational opinion, As an indirect reputation opinion, Weight

[0017] Preferably, indirect reputation opinions It consists of three elements: trust value Distrust value and uncertainty value and For a time window {t1,…,t…} a ,…,t X}, where X represents the total number of time windows.

[0018] Indirect credit opinion for:

[0019]

[0020] Where k is a given constant;

[0021] Using λ1 and λ2 to represent the reputation weights for positive and negative interactions respectively, and λ1 + λ2 = 1, λ1, λ2 ∈ (0, 1), λ1 ≤ λ2, we obtain:

[0022]

[0023] in, and These represent the number of positive and negative interactions, respectively. If the performance of this local model update is lower than that of the previous local model update, it is defined as a negative interaction. This represents a variable that influences the uncertainty of reputational opinions.

[0024] Preferably, the direct reputation value of the candidate device for:

[0025]

[0026] Where a is the weight value, a∈(0,1), the weight value a is greater than the weights y1 and y2 of the most recent and past experiences, y1, y2∈(0,1), y1+y2=1; This represents the average AoI of candidate device n;

[0027] The ratio of the number of times candidate device n performs federated learning tasks to the total number of times it participates in competition within a time slot T is defined as: E n =p w / p c p c p represents the total number of times a candidate device participates in the competition. w p w ≤p c The timescale for recent and past experiences is T. gap The interaction experience of candidate device n satisfies T≤T within the most recent time slot T. gap The most recent ratio T > T gap Ratio of time in the past

[0028] Preferably, the method for calculating the average AoI is as follows:

[0029] Assume the data are at times t1,...,t z ,...,t Z Generate, and subsequently in u1,...,u z ,...,u Z The instantaneous information age A of candidate device n is received at any time. n (t z ) is represented as A n (t z )=t z -u z The freshness of data is defined by the average AoI of candidate devices n over a time slot T:

[0030] Preferably, if the overall reputation value of a candidate device is greater than a predefined reputation threshold, then the candidate device is selected as a participating device to perform the federated learning task. Then, the participating device selects the optimal contract according to its own type to train the local model, and the result makes the accuracy of the global model reach a predefined value.

[0031] The process of performing a federated learning task includes: participating devices iteratively training a shared global model using their local data and generating local models; the participating devices updating the local models to the edge servers and performing global model aggregation; repeating the above process until the global model accuracy reaches a predefined value;

[0032] After the federated learning task is completed, the edge server updates its direct reputation opinion on the participating devices based on the interaction history, and adds the direct reputation opinion to the reputation blockchain after block verification and consensus scheme.

[0033] Preferably, the method for deriving the optimal contract design problem under conditions of information asymmetry in step three is as follows:

[0034] Optimize the expected benefits of edge servers using the statistical distribution of device types, defining the parameter θ. j To assess data quality, It is a coefficient for the number of iterations of the local model, depending on the data quality θ. j The participating devices are divided into J types and sorted in ascending order: θ1≤…≤θ j ≤…≤θ J ;

[0035] According to the contract terms (R) j (f j ), f j The benefit function for designing type j equipment is:

[0036]

[0037] In the formula, ε is a predefined weighting parameter for energy consumption; This represents the energy consumption for transferring local model updates during global iteration, and P j It is the transmission power, σ is the data size for local model updates, and e j The packet error rate is caused by an uncertain transmission environment, h j It is the channel gain of the point-to-point link between type j device and edge server, N j B is background noise, and B is the transmission bandwidth. This represents the energy consumption of a single local iteration, and C j D j These represent the number of CPU cycles required for a type j device to execute one data sample during local model training and the size of the local data sample used, respectively. ζ represents the effective capacitance parameter of the computing chipset of the type j device.

[0038] According to the contract terms (R) j (f j ), f j The benefit function for designing the edge server is:

[0039]

[0040] In the formula, u represents the edge server's satisfaction parameter with the participating devices, and l represents the unit reward cost. T represents the total time of one global iteration. max This represents the maximum time that the edge server can tolerate for federated learning tasks, and This represents the computation time for local model iteration. The transmission time represents the local model update;

[0041] Based on the benefit function of device type j, an optimal contract is designed using contract theory. The optimal contract design problem is to maximize the benefit function of the edge server.

[0042]

[0043] In the formula, J represents the total number of participating devices, and q j This represents the probability that the participating device belongs to type j.

[0044] The constraints that different types of participating devices must meet, namely, individual rationality constraints and incentive compatibility constraints, are as follows:

[0045]

[0046]

[0047] 0≤f1<…<f j , j∈[2,...,J];

[0048] T j cmp ≤T max ,j∈[1,2,...,J];

[0049]

[0050] Among them, R max This represents the total reward budget for the edge servers;

[0051] Based on individual rationality constraints and incentive compatibility constraints, the problem is simplified to a relaxed optimal contract design problem. The solution to the relaxed optimal contract problem is:

[0052] make get:

[0053]

[0054] In the formula, π1 = 0, and type k represents the values ​​1, 2, ..., j;

[0055] but:

[0056]

[0057] In the formula,

[0058] available:

[0059]

[0060] The constraints are:

[0061] 0≤f1<…<f j ,j∈[2,...,J];

[0062] T j cmp ≤t max ,j∈[1,...,J];

[0063]

[0064] By using the standard convex optimization tool CVX to solve the problem, the optimal computational resource f can be obtained. j and corresponding rewards R j .

[0065] Preferably, under conditions of information symmetry, the edge server accurately knows the types of all participating devices, and any contract item (R) j (f j ), f j All of them should be satisfied.

[0066] By eliminating the benefit of each Industrial IoT device to zero, the optimization problem becomes:

[0067]

[0068] The constraints are:

[0069]

[0070] T j cmp ≤T max ,j∈[1,...,J];

[0071]

[0072] The optimal computational resource f is obtained using the convex optimization tool CVX. j and corresponding rewards R j .

[0073] Preferably, when the edge server uses the blockchain for payment, the reward is automatically executed through a smart contract on the blockchain, that is, the contract automatically executes the transaction according to the predetermined rules; if the participating device fails to complete the federated learning task, the edge server will be protected by the smart contract and refuse to make payment.

[0074] Compared with existing technologies, the beneficial effects of this invention are as follows: It proposes a reputation evaluation scheme that combines Average Information Age (AoI) to assess the freshness of data from industrial IoT devices. Edge servers can evaluate the data quality of each participant and guide the design of the optimal contract. Furthermore, it proposes a blockchain-based reputation management system to ensure the authenticity of data provided by industrial IoT devices. Then, it designs a timeliness-aware incentive mechanism based on contract theory to supervise the active participation of industrial IoT devices in federated tasks and to provide high-quality data, thereby maximizing the benefits of edge servers to obtain the optimal contract. Simulation results show that as the reputation value increases, this invention effectively improves the efficiency of edge servers. Attached Figure Description

[0075] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0076] Figure 1 This is a schematic diagram of the process of the present invention.

[0077] Figure 2 This is a schematic diagram illustrating the continuous change of AoI over time according to the present invention.

[0078] Figure 3 This is a schematic diagram illustrating the direct reputation values ​​of devices under different average AoI values ​​according to the present invention.

[0079] Figure 4 This is a schematic diagram illustrating the benefits of the edge server of the present invention across different reputation ranges.

[0080] Figure 5 This diagram illustrates the benefits of the edge server of the present invention in different scenarios. Detailed Implementation

[0081] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0082] like Figure 1 As shown, based on the Industrial Internet of Things (IIoT), this invention proposes a federated learning method for incentivizing participating devices in the IIoT with timeliness awareness, specifically including the following steps:

[0083] Step 1. The edge server designs resource requirements and contract terms based on the federated learning task. After the resource requirements are met, the industrial IoT device is selected as a candidate device and is eligible to perform the federated learning task.

[0084] The resource requirements for designing a federated learning task specifically include data size, data type, CPU frequency, required training accuracy, and training time threshold. Setting these resource requirements ensures the quality of federated learning.

[0085] The contract terms designed based on the federated learning task are (R) n (f n ), f n ), f n R refers to the computing resources required for an industrial IoT device n to perform a federated learning task. n This refers to the corresponding reward received by industrial IoT device n, with reward R. n and computing resources f n It can be obtained by solving the objective function.

[0086] Step 2. Selecting high-quality candidate devices for the federated learning task based on a reputation-based evaluation scheme: The data freshness of candidate devices is evaluated using the average AoI metric. A comprehensive reputation score for each candidate device is calculated, and these devices are selected as participating devices based on their comprehensive reputation scores. These participating devices then perform the federated learning task to train the local model. After the federated learning task is completed, the edge server updates its direct reputation opinion on the reputation blockchain based on the interaction history.

[0087] Calculate the overall reputation score of the candidate devices. Direct credit opinion and indirect reputation opinions The composition and calculation formula are as follows:

[0088]

[0089] in, Weight For a time window {t1,...,t...} x ,...,t X}, where X represents the total number of time windows. Indirect reputation opinion It consists of three elements: trust value Distrust value and uncertainty value and Indirect credit opinion The calculation formula is:

[0090]

[0091] Where k is a given constant representing the level of influence on reputation uncertainty. A behavioral effect is introduced to calculate indirect reputation opinions: positive interactions increase the reputation value of candidate devices, and vice versa. To prevent negative behavioral events, a heavier penalty is imposed in cases of negative interactions. λ1 and λ2 represent the reputation calculation weights for positive and negative interactions, respectively. λ1 + λ2 = 1, λ1, λ2 ∈ (0, 1), λ1 ≤ λ2, yielding the following formula:

[0092]

[0093] in, and These represent the number of positive and negative interactions, respectively. If the performance of this local model update is lower than that of the previous local model update, it is defined as a negative interaction. This represents a variable that influences the uncertainty of reputational opinions.

[0094] Direct reputation value of candidate devices The calculation formula is:

[0095]

[0096] Where 'a' is the weight value, a∈(0,1), and the weight value 'a' is greater than y1 and y2. y1 and y2 represent the weights of recent and past experiences, respectively, y1, y2∈(0,1), y1+y2=1. The ratio of the number of times candidate device n performs federated learning tasks to the total number of times it participates in competition within a time slot T is defined as: E n =p w / p c p c p represents the total number of times a candidate device participates in the competition. w p w ≤p c The timescales of recent and past experiences are determined by T. gapDefined as follows: the interaction experience of candidate device n within the most recent time slot T satisfies T≤T gap The most recent ratio T > T gap Ratio of time in the past This represents the average information age (AoI) of candidate device n.

[0097] We introduce the Average Information Age (AoI) to evaluate the data freshness of candidate devices. Assume the data is available at times t1,...,t... z ,...,t Z Generate, and subsequently in u1,...,u z ,...,u Z The instantaneous information age A of candidate device n is received at any time. n (t z ) is represented as A n (t z )=t z -u z The freshness of data is defined by the average information age of candidate devices n over a time slot T, and its calculation method is as follows:

[0098]

[0099] The lower the average AoI, the fresher the data and the higher the quality of local model updates.

[0100] If a candidate device's overall reputation score is greater than a predefined reputation threshold (set to 0.5), then that candidate device is selected as a participating device for the federated learning task. The participating device then selects the optimal contract based on its own type to train the model. Upon completion of the task, it receives a corresponding reward. The device will choose a suitable contract based on its own conditions; essentially, it performs federated learning for local model training, and the resulting global accuracy must reach a predefined value.

[0101] The process of performing a federated learning task specifically includes: the industrial IoT devices participating in the federated learning task iteratively train a shared global model using their local data (the simulation dataset used to train the model is the MNIST dataset) and generate a local model; the industrial IoT devices update the local model to the edge server and perform global model aggregation; the above process is repeated until the accuracy of the global model reaches a predefined value.

[0102] After the federated learning task is completed, the edge server updates its direct reputation opinion on the participating devices based on the interaction history. After block verification and consensus, the direct reputation opinion is added to the reputation blockchain. The blockchain guarantees the authenticity of the data provided by industrial IoT devices.

[0103] Step 3. Develop a time-sensitive federated learning incentive mechanism in local model training: Based on the contract terms in Step 1, establish the benefit function of the edge server as the objective function, establish constraints based on the individual rationality constraints and incentive compatibility constraints of the participating devices, derive the optimal contract design problem under information asymmetry and information symmetry conditions, and use convex optimization tools to obtain the optimal contract terms of the participating devices.

[0104] Under conditions of information asymmetry, a time-aware federated learning incentive mechanism is used to solve a contract problem. Although the edge server does not have exact information about the devices, it can classify the participating devices into different types and use the statistical distribution of device types to optimize the expected benefits of the edge server. A parameter θ is defined. j To assess data quality, It is a coefficient for the number of iterations of the local model, depending on the data quality θ. j The participating devices are divided into J types and sorted in ascending order: θ1≤…≤θ j ≤…≤θ J .

[0105] Comprehensive credit score Normalization is performed to obtain the normalized comprehensive reputation value r. j ∈(0,1), using This represents the number of iterations required to update the local model when the global accuracy is fixed. During simulation, a target global accuracy value can be preset; once reached, local model training stops. The device's overall reputation score. The higher the value, the higher the quality of the local data, and thus the fewer iterations are needed to update the local model.

[0106] According to the contract terms (R) j (f j ), f j To design the benefit function for device type j, the specific steps are as follows:

[0107]

[0108] In the formula, ε is a predefined weighting parameter for energy consumption, and ε = 1 during simulation. This represents the energy consumption for transferring local model updates during global iteration, and P j It is the transmission power, σ is the data size for a local model update, and eh The packet error rate is caused by an uncertain transmission environment, h h It is the channel gain of the point-to-point link between type j device and edge server, N h B is the background noise, and B is the transmission bandwidth. Without loss of generality, since the wireless communication environment (such as channel bandwidth and channel capacity) remains almost constant between consecutive time slots, it is assumed that the network state of all participating devices is the same, i.e. This represents the energy consumption of a single local iteration, and C j D j These represent the number of CPU cycles required for a type j device to execute one data sample during local model training and the size of the local data sample used, respectively. ζ represents the effective capacitance parameter of the computing chipset of the type j device.

[0109] According to the contract terms (R) j (f j ), f j Design the benefit function for the edge server, specifically:

[0110] U s (R j ) = uln(T max -T j t )-lR j

[0111] In the formula, u represents the edge server's satisfaction parameter with the participating devices, and l represents the unit reward cost. T represents the total time of one global iteration. max This represents the maximum time that the edge server can tolerate for federated learning tasks, and T represents the computation time for local model iteration. j com =σ / (1-e j )Bln(1+P j h j / N j ) represents the transmission time of a local model update.

[0112] Based on the benefit function of device type j, an optimal contract is designed using contract theory. The optimal contract design problem is to maximize the benefit function of the edge server.

[0113]

[0114] In the formula, J represents the total number of participating devices, and q j This represents the probability that the participating device belongs to type j.

[0115] Different types of participating devices need to meet individual rationality constraints and incentive compatibility constraints, with the following constraints:

[0116]

[0117]

[0118] 0≤f1<···<f j , j∈[2,...,J];

[0119] T j cmp ≤T max ,j∈[1,2,...,J];

[0120]

[0121] Among them, R max This represents the total reward budget for the edge servers.

[0122] To derive the optimal contract, based on individual rationality constraints and incentive compatibility constraints, we can simplify it to obtain a relaxed optimal contract design problem. The solution to the relaxed optimal contract is:

[0123] make get:

[0124]

[0125] In the formula, π1 = 0, and type k represents taking the values ​​1, 2, ..., j.

[0126] Substitute formula (6) into In the middle, we get

[0127]

[0128] In the formula,

[0129] Substitute formula (7) into formula (5), and then remove all rewards R. j We can obtain:

[0130]

[0131] The constraints are:

[0132] 0≤f1<…<f j ,j∈[2,...,J];

[0133] T jcmp ≤T nax ,j∈[1,...,J];

[0134]

[0135] Note that the problem described above is a concave function, and the constraint set is a convex set. Using standard convex optimization tools, CVX solves equation (8) to obtain the optimal computational resource f. j and corresponding rewards R j .

[0136] Under conditions of information symmetry, a time-aware incentive mechanism based on federated learning is used to solve a contract problem. Under information symmetry, the edge server can accurately know the types of all participating devices, and any contract item (R...)... j (f j ), f j All of them should be satisfied. That is, the benefit of any participating equipment is zero.

[0137] By making the benefit of each industrial IoT device zero, the optimization problem can be formulated as:

[0138]

[0139] The constraints are:

[0140]

[0141] T j cmp ≤T max ,j∈[1,...,J];

[0142]

[0143] Under conditions of information symmetry, the benefit of any participating device is zero. Assume there exists an optimal contract term (R). j (f j ), f j ), In other words, assuming Edge servers can further increase their effectiveness by increasing computing resources f j The quantity until This contradicts the assumption, which is invalid. Therefore, under conditions of information symmetry, the benefit of any device is zero.

[0144] This problem can be solved by using the convex optimization tool CVX to determine the optimal computational resource f. j and corresponding rewards R j .

[0145] Step 4. Based on the reward of the optimal contract, the edge server uses the blockchain to pay the participating devices. The edge server cannot refuse to pay; the smart contract can help supervise the transfer of rewards from the edge server to the participating devices.

[0146] When edge servers use blockchain for payments, rewards can be automatically executed through smart contracts on the blockchain, meaning the contract automatically executes the transaction according to pre-defined rules. If a participating device fails to complete the federated learning task, the edge server will be protected by the smart contract conditions and refuse payment.

[0147] like Figure 2 As shown, in most Industrial Internet of Things (IIoT) scenarios (e.g., autonomous driving, telemedicine), devices need to monitor the surrounding physical environment and system status in real time to provide timely and effective information for intelligent decision-making and control. Therefore, Average AoI is crucial for real-time monitoring systems and status updates in IIoT, and has become an important indicator for evaluating information freshness. Assume data is generated at times t1, t2, ... t... Z Generate, and in time u1, u2, ... u Z Accept them sequentially. The lower the average AoI, the fresher the data.

[0148] like Figure 3 As shown, when calculating the direct reputation value, all parameters except the average AoI remain unchanged. The smaller the average AoI, the fresher the data, and the larger the direct reputation value.

[0149] When calculating the benefits of edge servers, all parameters remain unchanged except for the overall reputation score. Figure 4 As shown, the benefits of edge servers increase with the increase of the overall reputation range of devices. An increase in reputation range implies an increase in the number of reputable and high-quality devices. High device reputation has a positive impact on the benefits of edge servers. The introduction of average AoI enables more accurate reputation calculation and encourages well-known devices to contribute accurate and reliable data to the learning task, thus achieving a more reliable federated learning task.

[0150] Information-symmetric solution of equations, such as Figure 5 As shown, the edge server's benefit is higher under symmetric information conditions than under asymmetric information conditions. While the proposed asymmetric information scheme encourages devices to find contracts that match their type, the edge server cannot obtain the exact device type. Under symmetric information conditions, each device's benefit is zero, resulting in the minimum total reward, thus maximizing the edge server's benefit.

[0151] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A federated learning method for incentivizing devices with timeliness perception in the Industrial Internet of Things, characterized in that, The steps are as follows: Step 1. The edge server designs resource requirements and contract terms based on the federated learning task. After the resource requirements are met, the industrial IoT device is selected as a candidate device. Step 2. Measure the data freshness of candidate devices using the average AoI, calculate the comprehensive reputation score of candidate devices, select candidate devices as participating devices based on the comprehensive reputation score, and have participating devices perform federated learning tasks to train the local model; Step 3. Based on the contract terms of the participating devices, establish the benefit function of the edge server as the objective function. Based on the individual rationality constraints and incentive compatibility constraints of the participating devices, establish the constraints, derive the optimal contract design problem under the conditions of information asymmetry and information symmetry, and use convex optimization tools to obtain the optimal contract terms of the participating devices. Step 4. The edge server uses the blockchain to pay the participating devices based on the reward of the optimal contract item. The edge server cannot refuse to pay. The method for deriving the optimal contract design problem under the condition of information asymmetry in step three is as follows: Optimize the expected benefits of edge servers using the statistical distribution of device types, defining the parameter θ. j To assess data quality, It is a coefficient for the number of iterations of the local model, depending on the data quality θ. j The participating devices are divided into J types and sorted in ascending order: θ1≤···≤θ j ≤···≤θ J ; According to the contract terms (R) j (f j ), f j The benefit function for designing type j equipment is: In the formula, ε is a predefined weighting parameter for energy consumption; This represents the energy consumption for transferring local model updates during global iteration, and P j It is the transmission power, σ is the data size for local model updates, and e j The packet error rate is caused by an uncertain transmission environment, h j It is the channel gain of the point-to-point link between type j device and edge server, N j B is background noise, and B is the transmission bandwidth. This represents the energy consumption of a single local iteration, and C j D j These represent the number of CPU cycles required for a type j device to execute one data sample during local model training and the size of the local data sample used, respectively. ζ represents the effective capacitance parameter of the computing chipset of the type j device. According to the contract terms (R) j (f j ), f j The benefit function for designing the edge server is: In the formula, u represents the edge server's satisfaction parameter with the participating devices, and l represents the unit reward cost. T represents the total time of one global iteration. max This represents the maximum time that the edge server can tolerate for federated learning tasks, and This represents the computation time for local model iterations. The transmission time represents the local model update; Based on the benefit function of device type j, an optimal contract is designed using contract theory. The optimal contract design problem is to maximize the benefit function of the edge server. In the formula, J represents the total number of participating devices, and q j This represents the probability that the participating device belongs to type j. The constraints that different types of participating devices must meet, namely, individual rationality constraints and incentive compatibility constraints, are as follows: 0≤f1<···<f j ,j∈[2,...,J]; Among them, R max This represents the total reward budget for the edge servers; Based on individual rationality constraints and incentive compatibility constraints, the problem is simplified to a relaxed optimal contract design problem. The solution to the relaxed optimal contract problem is: make get: In the formula, π1 = 0, and type k represents the values ​​1, 2, ..., j; but: In the formula, available: The constraints are: 0≤f1<···<f j ,j∈[2,...,J]; By using the standard convex optimization tool CVX to solve the problem, the optimal computational resource f can be obtained. j and corresponding rewards R j ; Under conditions of information symmetry, the edge server accurately knows the types of all participating devices, and any contract item (R) j (f j ), f j All of them should be satisfied. By eliminating the benefit of each Industrial IoT device to zero, the optimization problem becomes: The constraints are: The optimal computational resource f is obtained using the convex optimization tool CVX. j and corresponding rewards R j .

2. The method for incentivizing participating devices in federated learning for timeliness perception in the Industrial Internet of Things according to claim 1, characterized in that, The resource requirements for the design of the federated learning task include data size, data type, CPU frequency, required training accuracy, and training time threshold. The contract terms designed based on the federated learning task are (R) n (f n ), f n ), f n R is the computing resources required for an industrial IoT device n to perform a federated learning task. n This refers to the corresponding reward received by industrial IoT device n.

3. The method for incentivizing participating devices in federated learning for timeliness perception in the Industrial Internet of Things according to claim 1 or 2, characterized in that, The method for calculating the comprehensive reputation value of candidate devices is as follows: Comprehensive reputation value for: in, For direct reputational opinion, As an indirect reputation opinion, Weight 4. The method for incentivizing participating devices in federated learning for timeliness perception in the Industrial Internet of Things according to claim 3, characterized in that, Indirect credit opinion It consists of three elements: trust value Distrust value and uncertainty value and For a time window {t1,…,t…} x ,…,t x }, where X represents the total number of time windows. Indirect credit opinion for: Where k is a given constant; Using λ1 and λ2 to represent the reputation weights for positive and negative interactions respectively, and λ1 + λ2 = 1, λ1, λ2 ∈ (0, 1), λ1 ≤ λ2, we obtain: in, and These represent the number of positive and negative interactions, respectively. If the performance of this local model update is lower than that of the previous local model update, it is defined as a negative interaction. This represents a variable that influences the uncertainty of reputational opinions.

5. The method for incentivizing participating devices in federated learning for timeliness perception in the Industrial Internet of Things according to claim 4, characterized in that, The direct reputation value of the candidate device for: Where a is the weight value, a∈(0,1), the weight value a is greater than the weights y1 and y2 of the most recent and past experiences, y1, y2∈(0,1), y1+y2=1; This represents the average AoI of candidate device n; The ratio of the number of times candidate device n performs federated learning tasks to the total number of times it participates in competition within a time slot T is defined as: E n =p w / p c p c p represents the total number of times a candidate device participates in a competitive federated learning task. w p is the number of times a federated learning task is performed. w ≤p c The timescale for recent and past experiences is T. gap The interaction experience of candidate device n satisfies T≤T within the most recent time slot T. gap The most recent ratio T > T gap Ratio of time in the past 6. The method for incentivizing participating devices in federated learning for timeliness perception in the Industrial Internet of Things according to claim 5, characterized in that, The method for calculating the average AoI is as follows: Data at times t1,...,t z ,...,t Z Generate, and subsequently in u1,...,u z ,...,u z The instantaneous information age A of candidate device n is received at any time. n (t z ) is represented as A n (t z )=t z -u z The freshness of data is defined by the average AoI of candidate devices n over a time slot T:

7. The method for incentivizing participating devices in federated learning for timeliness perception in the Industrial Internet of Things according to any one of claims 4-6, characterized in that, If the overall reputation value of a candidate device is greater than the predefined reputation threshold, then the candidate device is selected as a participating device to perform the federated learning task. Then, the participating device selects the optimal contract according to its own type to train the local model, and the result makes the accuracy of the global model reach the predefined value. The process of performing a federated learning task includes: participating devices iteratively training a shared global model using their local data and generating local models; the participating devices updating the local models to the edge servers and performing global model aggregation; repeating the above process until the global model accuracy reaches a predefined value; After the federated learning task is completed, the edge server updates its direct reputation opinion on the participating devices based on the interaction history, and adds the direct reputation opinion to the reputation blockchain after block verification and consensus scheme.

8. The method for incentivizing participating devices in federated learning for timeliness perception in the Industrial Internet of Things according to claim 1, characterized in that, When edge servers use blockchain for payments, rewards are automatically executed through smart contracts on the blockchain, meaning the contract automatically executes the transaction according to pre-determined rules; if a participating device fails to complete the federated learning task, the edge server will be protected by the smart contract and refuse to make payment.

Images