Vehicle pile network cross-platform learning fairness and energy consumption optimization method and system based on lossy communication
By optimizing the resource allocation of equipment in the zero-trust federated learning framework, the problem of unfair selection of vehicle equipment under lossy communication conditions is solved, the system energy consumption is minimized and the learning process is stable, and the variety of learning models is improved Sex and generalization ability.
Patent Information
- Application Number
- CN202410928255.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-11
- Publication Date
- 2024-11-08
AI Technical Summary
Under lossy communication conditions, in the zero-trust federated learning framework, unfair device selection of vehicle devices leads to a decrease in accuracy and stability of the global model, and limited communication resources and battery reserves make participating in the ZTFL framework challenging.
By solving the pre-built joint optimization model in each learning round, the communication resources and computing resources allocation of the equipment are optimized to minimize system energy consumption while taking into account long-term fair constraints to ensure that each device is selected on average the number of times selected Meet the set expectations.
Realize zero-trust federated learning under lossy communication conditions, improves the fairness and efficiency of the learning process, reduces system energy consumption, extends device battery life, and promotes wider device participation, thereby enhancing the learning model diversity and generalization ability.
Smart Images

Figure CN118921702A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of vehicle networking, and specifically relates to a fairness and energy consumption optimization method and system for cross-platform learning of a vehicle-pile network based on lossy communication. Background Art
[0002] With the recent development of the "car-pile-network" cross-platform technology, data transmission is growing rapidly and constantly challenging the limits of network transmission. The work of uploading all prediction data to cloud computing servers is becoming increasingly tight. Artificial intelligence (AI) at the edge and AI algorithms running on edge devices have become a promising key enabler for realizing the grand vision of ubiquitous intelligence.
[0003] Federated Learning (FL), an edge AI training model proposed by Google, has become a model for edge AI training. The core feature of FL is that user devices in FL train machine learning models without revealing original data. However, this method of training models does not involve data exchange and requires a lot of computing and communication resources in each round of training. In this case, it is expected that only a subset of models to be trained on behalf of all models is selected. Integrating the zero-trust security model into the FL learning process enables continuous system verification and authentication, which helps to formulate effective network access policies. Zero Trust FL (ZTFL) promotes the application of "car-pile-network" cross-platform technology in real-world fields.
[0004] Compared with traditional federated learning, the devices in the ZTFL framework are usually vehicles. Not only do they work under more unfavorable communication conditions, but they also have much smaller battery reserves than ordinary mobile devices. In the context of CIoT, this brings several challenges to real-world ZTFL: 1) Limited bandwidth prevents some edge servers from successfully uploading their trained models despite their participation in the ZTFL process; 2) Due to limited communication resources and the dynamic state of edge servers, some servers must be selected to perform training on behalf of all servers. Unreasonable vehicle device selection schemes lead to degraded accuracy and stability of the global model. Absurd vehicle device selection schemes are often referred to as unfair selection schemes; 3) Due to the large scale of current models, it is very challenging for devices with low battery and limited computing resources to participate in the ZTFL framework. Summary of the invention
[0005] In view of this, the present invention aims to provide a fairness and energy consumption optimization method and system for cross-platform learning of vehicle-pile network based on lossy communication, which is used to solve the above-mentioned problems existing in the ZTFL framework under lossy communication conditions.
[0006] In order to achieve the above object, the technical solution provided by the present invention is as follows:
[0007] In the first aspect, the present invention provides a fairness and energy consumption optimization method for cross-platform learning of a vehicle-pile network based on lossy communication. In the zero-trust federated learning framework ZTFL, in each learning round, the pre-built optimization model of energy consumption and fairness of the joint optimization system is solved for the participating devices in the framework, and the selected device in the current learning round, the communication resource allocation under lossy communication conditions, and the computing resource allocation are obtained, thereby performing zero-trust federated learning;
[0008] The optimization model aims to minimize the system energy consumption while considering long-term fairness constraints;
[0009] The system energy consumption includes the energy consumption related to uploading communication and the energy consumption related to training models, which are determined by the allocated communication resources and computing resources respectively;
[0010] The long-term fairness constraint is that under long-term conditions, the average number of times each device is selected satisfies the condition that it is at least not less than the set expected value.
[0011] Furthermore, the expression of the optimization model is as follows:
[0012]
[0013] Where n is the set of participating devices The index of the selected device in , τ is the set of global iteration rounds The indices of different iteration rounds in , is the communication energy consumption, For training energy consumption, represents the average number of times a device is selected at time T-1, β is the setting value to ensure fairness, and x τ,n =1 indicates that the nth vehicle device participates in federated learning in the τth round.
[0014] Furthermore, when solving the optimization model, the long-term fairness constraint is converted into an objective function through the Lyapunov optimization theory for solution. The conversion process includes:
[0015] A virtual queue is constructed for each device, and the virtual queue is used to represent the current backlog of the device due to time constraints;
[0016] A quadratic Lyapunov function is established based on the virtual queue, and the quadratic Lyapunov function is used to constrain the increase of the cardinality of the virtual queue set;
[0017] Based on Lyapunov drift theorem, a drift function is constructed to represent the increment of the virtual queue caused by a single training device;
[0018] The drift function is combined with the original objective function of the optimization model to obtain the drift plus cost function;
[0019] Minimizing the drift plus the cost function is used as the objective function of the transformed optimization model.
[0020] Furthermore, the expression of the drift plus cost function is as follows:
[0021]
[0022] Among them, Δ(Θ(τ)) is the drift function, and V is the penalty factor;
[0023] The expression of the drift function is as follows:
[0024]
[0025] In the formula, is a quadratic Lyapunov function, Θ(τ) is the set length of the virtual queue;
[0026] The expression of the quadratic Lyapunov function is as follows:
[0027]
[0028] In the formula, Z τ,n is the virtual queue of the nth vehicle device in the τth round;
[0029] The expression of the virtual queue is as follows:
[0030] Z τ+1,n =max[Z τ,n +β-x τ,n ,0].
[0031] Furthermore, the objective function of the transformed optimization model is as follows:
[0032] minf(x τ,n ,β,V)
[0033] In the formula, f(x τ,n ,β,V) is the upper bound of the drift plus cost function, and the specific relationship is as follows:
[0034]
[0035] Furthermore, the constraints of the optimization model include at least:
[0036] Channel reachability constraints, channel condition constraints, bandwidth allocation ratio constraints, device selection constraints, bandwidth constraints, upload power constraints and computing resource constraints.
[0037] Furthermore, the channel reachability constraint is as follows:
[0038]
[0039] In the formula, a n is the probability of channel damage, m is the waterfall threshold, B is the total bandwidth of the central server, and b n is the bandwidth allocation ratio of each vehicle equipment channel, N0 represents the spectral power density of Gaussian noise, p n is the upload power, h n is the gain of the channel, and ε is the set threshold.
[0040] Furthermore, when solving the objective function of the transformed optimization model, the convex optimization theory is used to further transform the objective function as follows:
[0041] minf τ,n ,p * g(f,p * )
[0042]
[0043] f min <f τ,n <f max
[0044] In the formula, f τ,n represents the CPU frequency selected by vehicle device n during the τth global iteration, p * represents the optimal value of the upload power p, g(·) is the converted objective function; C τ,n represents the number of CPU cycles of device n in the τth global iteration, D τ,n represents the data set size of device n in the τth round of global iteration, t e represents the number of local iterations, ξ represents the CPU capacitor parameter, T Limit is the upper limit of system time consumption, is the optimal value of the upload power of device n in the τth round, f max and f min They are the upper and lower limits of the CPU frequency respectively.
[0045] Furthermore, in ZTFL, whether a device has been selected is determined according to the following formula:
[0046]
[0047] In the formula, is an indicator used to evaluate whether the device completes training within the time constraint. If yes, it is set to 1, otherwise it is 0; An indicator used to evaluate whether the device successfully uploads related parameters. If yes, it is set to 1, otherwise it is 0; x′ τ,n Indicates whether the nth vehicle device is selected in the τth round. If so, it is set to 1, otherwise it is 0.
[0048] In a second aspect, the present invention provides a fairness and energy consumption optimization system for cross-platform learning of a vehicle-pile network based on lossy communication, including a central server and vehicle equipment, wherein the central server and vehicle equipment constitute a zero-trust federated learning framework ZTFL;
[0049] In the zero-trust federated learning framework ZTFL, in each learning round, the pre-built joint optimization system energy consumption and fairness optimization model is solved for the participating devices in the framework to obtain the selected devices in the current learning round, the communication resource allocation under lossy communication conditions, and the computing resource allocation, so as to perform zero-trust federated learning.
[0050] The optimization model aims to minimize the system energy consumption while considering long-term fairness constraints;
[0051] The system energy consumption includes the energy consumption related to uploading communication and the energy consumption related to training models, which are determined by the allocated communication resources and computing resources respectively;
[0052] The long-term fairness constraint is that under long-term conditions, the average number of times each device is selected satisfies the condition that it is at least not less than the set expected value.
[0053] In summary, the present invention provides a fairness and energy consumption optimization method and device for cross-platform learning of a vehicle-pile network based on lossy communication. In the zero-trust federated learning framework ZTFL, for each learning round, the pre-built joint optimization system energy consumption and fairness optimization model is solved for the participating devices in the framework, and the selected device in the current learning round, the communication resource allocation under lossy communication conditions, and the computing resource allocation are obtained, so as to perform zero-trust federated learning; the optimization model aims to minimize the system energy consumption, while considering long-term fairness constraints; the system energy consumption includes the energy consumption related to the upload communication and the energy consumption related to the training model, which are determined by the allocated communication resources and computing resources respectively; the long-term fairness constraint is that under long-term conditions, the average number of times each device is selected satisfies the condition of at least not less than the set expected value. By comprehensively considering energy consumption, fairness and communication condition restrictions, the present invention provides an efficient and fair solution for cross-platform machine learning applications in complex environments such as vehicle-pile networks, and also solves the problems existing in the existing ZTFL framework under lossy communication conditions. BRIEF DESCRIPTION OF THE DRAWINGS
[0054] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0055] Figure 1 A schematic diagram of a federated learning framework provided for an embodiment of the present invention;
[0056] Figure 2 The numerical result diagram of the accuracy of the simulation of the model proposed in the present invention and other models under the condition of parameter γ=10 and set data set provided in the embodiment of the present invention;
[0057] Figure 3 A numerical result diagram of the accuracy of simulation of the model proposed in the present invention and other models under the condition of parameter γ=1 and set data set provided in an embodiment of the present invention;
[0058] Figure 4 A numerical result diagram of the accuracy of simulation of the model proposed in the present invention and other models under the condition of parameter γ=10, set data set and real mildly lossy communication provided in an embodiment of the present invention;
[0059] Figure 5 A numerical result diagram of the accuracy of simulation of the model proposed in the present invention and other models under the condition of parameter γ=1, set data set and real mildly lossy communication provided in an embodiment of the present invention;
[0060] Figure 6 The numerical result diagram of the simulation accuracy of the model proposed in the present invention and other models under the condition of parameter γ=1, set data set and actual medium-loss communication provided in the embodiment of the present invention. DETAILED DESCRIPTION
[0061] In order to make the purpose, features and advantages of the present invention more obvious and easy to understand, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described below are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0062] This embodiment provides a fairness and energy consumption optimization method for cross-platform learning of a vehicle-pile network based on lossy communication. In the zero-trust federated learning framework ZTFL, in each learning round, the pre-built optimization model of energy consumption and fairness of the joint optimization system is solved for the participating devices in the framework, and the selected device in the current learning round, the communication resource allocation under lossy communication conditions, and the computing resource allocation are obtained, so as to perform zero-trust federated learning.
[0063] The optimization model aims to minimize the system energy consumption while considering long-term fairness constraints;
[0064] The system energy consumption includes the energy consumption related to uploading communication and the energy consumption related to training models, which are determined by the allocated communication resources and computing resources respectively;
[0065] The long-term fairness constraint is that under long-term conditions, the average number of times each device is selected satisfies the condition that it is at least not less than the set expected value.
[0066] It should be noted that this embodiment is a learning method designed for a cross-platform environment of vehicle-charging network (vehicles, charging stations, networks), aiming to optimize fairness and energy consumption issues under lossy communication conditions. Its core lies in the use of Zero-Trust Federated Learning (ZTFL) framework.
[0067] Among them, cross-platform learning of vehicle-charging station networks refers to the realization of data sharing and model learning in a multi-platform environment involving vehicles, charging stations and a wider network infrastructure to improve service, management or operational efficiency. Lossy communication recognizes that there may be data loss, delays or errors in network communications, requiring solutions to adapt to these unreliable factors. Fairness and energy consumption optimization aims to solve two key issues of resource allocation, ensuring that all devices participating in learning can contribute and benefit fairly while reducing the energy consumption of the overall system.
[0068] In this embodiment, the zero-trust federated learning framework (ZTFL) adopts a federated learning approach, allowing data to remain on local devices and only model parameters to be shared between devices, thereby enhancing data privacy protection. Adding the "zero trust" principle on this basis means that no participant is automatically trusted before verification, further strengthening security. At the beginning of each round of learning, the algorithm solves a preset optimization model that aims to simultaneously minimize the total energy consumption of the system and ensure long-term fairness. This means that it not only considers how to reduce energy use, but also ensures that each participating device has a fair chance of being selected to participate in learning.
[0069] The optimization model also performs resource allocation, including communication resource allocation and computing resource allocation. In particular, for lossy communication environments, the optimization model determines how to allocate communication resources to minimize the energy consumption caused by data transmission while ensuring that data can be effectively communicated. At the same time, tasks are reasonably allocated according to the computing power of the device to reduce the energy consumption caused by computing-intensive operations.
[0070] In addition, the optimization model also takes into account long-term fairness constraints to ensure that after multiple learning cycles, the number of times each participating device is selected to participate in learning is not less than a preset expected value, avoiding certain devices from being frequently excluded due to resource constraints and achieving fairness in participation.
[0071] The method provided in this embodiment can improve the efficiency and fairness of cross-platform learning while ensuring data privacy and security, especially under the conditions of limited resources and unstable network environment, it can still ensure the stability and efficiency of the learning process. The optimized resource allocation strategy helps to extend the battery life of the device, reduce operating costs, and promote more extensive device participation, thereby enhancing the diversity and generalization ability of the learning model.
[0072] See also Figure 1 , Figure 1 This is a schematic diagram of the federated learning framework. Figure 1 And an application example introduces some embodiments of the fairness and energy consumption optimization method of the vehicle-pile network cross-platform learning based on lossy communication of the present invention. In order to implement the method, the application example is introduced through four chapters: problem definition, channel conditions and reachability, long-term fairness constraints, and system energy consumption.
[0073] 1. Problem Definition
[0074] First, in order to explain the framework of the present invention, it is necessary to explain some important mathematical symbols. The set indexed by τ Represents different global iteration rounds in ZTFL. represents each vehicle device in the system (indexed by n). In addition, the present invention assumes that in each round of ZTFL, the maximum number of vehicle device training is n l , the vehicle device is the device that participates in federated learning in the context of the cross-platform vehicle-pile network of the present invention. It is understandable that the device can be other types of devices in other frameworks. Another important symbol is x′ τ,n , indicating whether the nth vehicle device is selected in the τth round. It is the main object of optimization under the fairness constraint. In some embodiments, if the vehicle device is selected, the strict mathematical expression is as follows:
[0075]
[0076] In realistic dynamic application scenarios, each vehicle faces the problem of unavailability of training and communication. However, once training is provided, the central server only knows the specific challenges of the vehicle. Therefore, the central server faces a dynamic selection problem to solve this problem. In practice, due to channel and vehicle parameter constraints, the vehicle may still need to complete training or upload parameters within a specified time after training. These situations cause the central server to select these vehicles to participate in model aggregation. However, they must participate in model aggregation, so that the fairness constraint is not met. This situation requires the central server to dynamically update the vehicle selection based on a fairness algorithm to ensure that long-term time constraints are met during the learning process. To this end, in some embodiments, a mathematical formula is given to update whether a vehicle has been selected.
[0077]
[0078] here, is used as an evaluation metric to measure whether the vehicle device completes training within the time constraint. If it meets the training within the time constraint, will be set to 1, otherwise, it will be 0. In addition, It is evaluated as an evaluation indicator of whether the vehicle equipment successfully uploads the relevant parameters. If the model parameters are successfully uploaded within the constraint time, then will be set to 1, otherwise, it will be 0. In addition, It is often related to the channel environment. In the following section, the present invention will introduce channel parameters and transmission packet rates so that the simulation matches the real world.
[0079] 2. Channel conditions and reachability
[0080] Each selected vehicle device must upload its model to the central server after completing the training. The central server then aggregates these models to form a global model that is distributed to other vehicle devices. In this process, Shannon's theorem is often used to calculate the time required to upload data. The maximum data upload rate for the nth vehicle device is:
[0081]
[0082] Therefore, the maximum upload time for the nth device is:
[0083]
[0084] Here, B represents the total bandwidth of the central server, b n is the bandwidth allocation ratio of each vehicle equipment channel, p n is the upload power, N0 represents the spectral power density of Gaussian noise, S n is the amount of data uploaded for the model, h n is the gain of the channel. It should be noted that h n is a random variable with a typical random distribution, which becomes an important consideration for subsequent optimization.
[0085] An unstable channel will corrupt the uploaded data, and as the communication power continues to increase, the probability of the transmitted data being corrupted is even higher. Since the stability of the central server channel is much better than that of a single vehicle device, this invention assumes that data corruption only occurs during the upload process. The probability of corruption is:
[0086]
[0087] 3. Long-term fairness constraints
[0088] However, in ZTFL with fairness constraints, treating fairness as a hard constraint is often different from the approach adopted. If fairness is imposed as a hard constraint, then during the optimization process, it will take precedence over other optimization conditions to be satisfied. This concept of fairness greatly limits the search range of feasible solutions. It can lead to extreme cases of vehicle equipment allocation, limiting the impact of fairness on model generalization to a local scope. This leads to constraints that can more effectively enhance the generalization ability of the model, but will significantly increase energy consumption. Therefore, in subsequent optimizations, the present invention needs to modify the constraint and treat it as a less strict restriction when solving the problem equation. In order to meet the requirements of soft fairness constraints in simulation experiments, the present invention improves the short-term fairness constraint into a long-term time constraint. Ensure that local minor violations of constraints are not reduced without reducing the overall model generalization ability, as shown below:
[0089]
[0090] In the above formula, the use of mathematical expectation as a way to express fairness constraints is particularly noteworthy, where β represents the average number of times a client is selected. The present invention expects to keep β at a stable value, so as to ensure that each client can be selected by fixed parameters. Represents the average number of times the client is selected at time t-1. Since the selection of the client is a random event, an expectation must be put on the outside to be correctly analyzed and calculated using analytical language. This method regards network problems as random systems and recognizes the existence of random events, time changes and uncertainties in real-world communication scenarios. The present invention uses mathematical expectations to convert random vectors in random systems into computable vectors, thereby achieving easier-to-handle analysis and optimization. This constraint ensures that vehicle equipment meets fairness standards during long-term experiments. In the following sections, it replaces the short-term fairness constraint metric used in other ZTFL scenarios.
[0091] 4. System energy consumption
[0092] To describe the model training process of ZTFL, let t e represents the number of local iterations; f τ,n represents the CPU frequency selected by vehicle device n during the τth round of global iteration; C τ,n represents the number of CPU cycles of vehicle device n in the τth round of global iteration, D τ,n represents the dataset size of vehicle device n in the τth round of global iteration. The time used for global iteration in the ZTFL process is as follows:
[0093]
[0094] The energy consumption during ZTFL training is:
[0095]
[0096] Where ξ represents the CPU capacitance parameter.
[0097] In the above equation, the present invention derives the following communication energy consumption:
[0098]
[0099] In combination with the introduction of the above application examples, in some embodiments, the expression of the optimization model of the present invention is as follows, which is recorded as the first optimization problem P1:
[0100]
[0101] Where n is the set of participating devices The index of the selected device in , τ is the set of global iteration rounds The indices of different iteration rounds in , is the communication energy consumption, For training energy consumption, represents the average number of times a device is selected at time T-1, β is the setting value to ensure fairness, and x τ,n =1 indicates that the nth vehicle device participates in federated learning in the τth round.
[0102] In this embodiment, the energy consumption of the system is divided into two parts: the first part is the energy related to the upload communication, and the second part is the energy used to train the model to obtain parameters. Based on the above application examples, the model can also consider the following constraints:
[0103]
[0104] Where ε represents the limiting factor of the packet error rate, T Limit is the upper limit of system time consumption. The parameters that need to be optimized are: τ,n ,b τ,n ,p τ,n ,f τ,n , i.e., the vehicle equipment selected in the τth round of iteration, the corresponding bandwidth allocation, upload power, and CPU frequency.
[0105] However, the expression of the fairness index in the above optimization model cannot be directly incorporated into the objective function value as a penalty function. Therefore, in some embodiments, Lyapunov optimization theory is used to convert offline and online constraints. The present invention transforms the random problem into a Markov chain problem. Here, the present invention introduces a virtual queue for each vehicle device, representing the current backlog of vehicle devices due to time constraints.
[0106] The recursive function of backlog is expressed as:
[0107] Z τ+1,n =max[Z τ,n +β-x τ,n ,0] (11)
[0108] The present invention specifically adopts a positive definite method to process this virtual queue. Compared with other methods, the particularity of this method is that it ensures that the backlog is not negative. If the backlog is negative, it may cause the negative backlog of the k1th vehicle device to offset the positive backlog of the k2th vehicle device, making it impossible to accurately measure the backlog in each vehicle device. Simply put, the larger the backlog value, the worse the fairness of vehicle device selection, that is, when Z τ,n =∞, the selection is completely unfair; on the contrary, the smaller the hoarding value, the better the fairness of vehicle equipment selection. τ,n = 0, the selection is fair. If all virtual queues (whose backlog is determined by Z τ,n If the average rate of the ZTFL is kept stable during the whole ZTFL process, the long-term time average constraint equation (Equation 6) will be established.
[0109] If fairness is not maintained during the ZTFL learning process, then Z τ,n The value of diverges to βτ over time max , β is fairness, τ max is the maximum number of training rounds.
[0110] Now the virtual sequence assumed in the previous text of the present invention is that if this client is not selected, the value of the virtual sequence increases by β, so if it is unfair, from the perspective of analysis, the value of the virtual sequence will reach βτ in the most extreme case max . ) Under long-term time constraints, this value will soar to infinity. To ensure fairness, the growth of the sequence value should stabilize after each iteration. In addition, note that in the set In , elements are selected from the set as queue lengths to ensure fairness constraints. Obviously, if no vehicle equipment is specified in the first round, its queue length will increase significantly. Using Theorem 1, the present invention transforms the complex time coupling constraint terms into a set of cardinality constraints. It is necessary to ensure that the growth of the cardinality of the set is smooth and eventually converges to a certain integer. A common method to constrain the size of the cardinality of a set is to use Lyapunov optimization constraints to constrain the increase in the cardinality and minimize the objective function required by the present invention. The present invention establishes a quadratic Lyapunov function below:
[0111]
[0112] here, represents the length of all virtual queue sets. In order to be able to measure the The present invention uses the Lyapunov drift theorem. Mathematically, it is expressed as:
[0113]
[0114] In ZTFL, the central server allows the calculation of different queue lengths in real time through online scheduling. Therefore, this value is used as a condition in the Lyapunov drift. The above optimization scheme successfully transforms the time-coupled backlog into the size of the Lyapunov drift, thereby transforming the original constraints into the objective function. Therefore, the penalty function integrates these constraints into the objective function. Formally, the present invention calls this combination a drift plus cost function. Its mathematical form is as follows:
[0115]
[0116] V ≥ 0 is a penalty factor that aims to strike a balance between minimizing the objective and satisfying the fairness constraint.
[0117] In a further embodiment of the present invention, the penalty factor parameter of the aforementioned embodiment directly affects the performance of the optimization algorithm. The present invention will further analyze the range of parameter V in the experiment. By some simple scaling, the upper bound of the modified objective function can be obtained as follows:
[0118]
[0119] in this case, It is easy to see that the right side of equation (15) is the upper bound of the objective function. If the present invention minimizes this part, it can not only minimize the objective function but also ensure the stability of the queue.
[0120] Therefore, in a further embodiment of the present invention, equation (16) transforms the optimization problem P1 into P2, and introduces the distributed online scheduling problem through P2. Specifically,
[0121] minf(x τ,n ,β,V)(16)
[0122] The constraints considered by P2 are consistent with those of P1, namely:
[0123]
[0124] The parameters that need to be optimized are: x τ,n ,b τ,n ,p τ,n ,f τ,n .
[0125] In the above embodiments, the Lyapunov optimization fairness problem in the FDMA-based ZTFL system is studied, which requires considering how the dynamic queuing quantifies the level of vehicle equipment participation in the context of packet error rate and providing a fair solution algorithm for the allocation of vehicle equipment.
[0126] However, due to the objective function in the equation, the variables to be optimized appear in nonlinear logarithmic form, and there are complex 0-1 mixed variables and nonlinear constraints in the equation. Existing solvers cannot directly solve this complex situation, making it very easy to fall into local optimality. To solve this problem, in the following part of the text, convex theory will be applied to optimization and function scaling.
[0127] In equation (15), The central server is required to perform calculations each time a vehicle device is selected during actual operation. Since many uncontrollable factors in the ZTFL process make it unusable for vehicle devices, offline calculations must be performed in advance. In the following, the present invention proposes a new mathematical optimization scheme to estimate x τ,n The optimal value of is used as an offline parameter to guide the subsequent ZTFL process.
[0128] First, the recursive function is constructed as G(τ). The present invention has:
[0129]
[0130] So, about It can be expressed as:
[0131]
[0132] The vehicle equipment meets the backlog arrangement, and the present invention can construct an arrangement to assume the relationship between the selection of vehicle equipment and the backlog quantity. n =β, indicating that the vehicle equipment has not been selected; a n =β-1 indicates that the vehicle equipment has been selected. The constructed virtual sequence is expressed as:
[0133]
[0134] In combination with the application of the above convex optimization theory method in the ZTFL process, in one embodiment of the present invention, the optimization problem P2 is further converted into P3 for solution.
[0135] First, consider the value of β. Through classification discussion, we can know that the scenario when β∈[3 / 7,1 / 2] is difficult to solve and needs further processing. In this part, formulas (17) and (18) use Lyapunov optimization theory to solve the problem. The control method of the virtual queue controls the backlog of each client to ensure that the backlog of each client is minimized. This is for x τ,n Optimize and transform the four-parameter optimization problem of problem P2 into a three-parameter optimization problem. Then further use optimization theory, use inequality optimization scheme, and use scaling scheme to optimize b parameters respectively. Finally, use segmented optimization to optimize parameter p and transform it into p*, as follows:
[0136] minf τ,n ,p * g(f,p * ) (19)
[0137]
[0138] f min <f τ,n <f max
[0139] Furthermore, the optimization problem P2 in the previous text is transformed into the optimization problem P3. τ,n ,b τ,n ,p τ,n ,f τ,n These four variables are optimized mathematically. τ,n is the number of clients selected. In some studies, this value is fitted by random sampling, but in fact, when the number of training rounds is small, it will have a large error. The training process of the Internet of Vehicles system is often small, which brings problems:
[0140] It cannot be guaranteed that each client is selected evenly. Some clients are selected too many times, and their local data is repeatedly learned as important data in the ZTFL process, resulting in a decrease in the generalization of the model. The entire model is overfitted to this client that has been selected many times. However, using 0-1 planning solver instead of random sampling is too large to solve directly because each selection is a separate 0-1 dimension.
[0141] To solve this problem, the present invention uses a dynamic system method to convert the situation where the client is selected into a virtual number series. Through mathematical methods, it is ensured that the increase in each round of this virtual number series is stable (that is, the stability of the dynamic system). This is the purpose of using Lyapunov, which is an optimization method that is often used to consider complex system problems such as neural networks.
[0142] Through this optimization method, the present invention can solve the x that cannot be solved by 0-1 programming. τ,n
[0143] Now, the present invention successfully transforms the four-variable problem into a three-variable problem. τ,n , thus the present invention analyzes that the situation where it is linearly proportional to the local data set is the optimal situation.
[0144] Now the three-variable problem is transformed into a two-variable problem, and the transmission power constraint can be transformed into the transmission time and packet loss rate constraint. The transmission time constraint is related to the training time. This requires considering p τ,n ,f τ,n For further optimization, the present invention considers using the packet loss rate to directly obtain the information about p τ,n The multi-value optimization result is analyzed: in the most likely case p τ,n There are only three possible values, which are then jointly optimized. This greatly reduces the feasible domain of the search, and finally the optimization problem P3 is obtained.
[0145] As a result, the present invention simultaneously ensures the constraints of fairness and energy consumption, and ensures that they are combined to obtain the best. In the above embodiment, by analyzing the objective function, the present invention determines the proportional relationship of converting a small group of non-convex functions into convex functions. Then, the present invention solves this part separately to obtain analytical solutions of some variables to be optimized, and then obtains analytical solutions of all parameters to be optimized.
[0146] The performance of the proposed method is verified in combination with the data sets MNIST, CIFAR-10, and CIAFAR-100.
[0147] In the ZTFL process, the present invention first verifies the accuracy of the proposed fairness constraint based on lossless communication.
[0148] The present invention establishes a fairness comparison experiment, using the FedAvg model of randomly selected vehicle equipment, the FedProx model of randomly selected vehicle equipment, and the FedAvg model with increased fairness for comparison. Figure 2 and Figure 3 As shown in the figure, the simulation experiment shows that the fairness constraint proposed in this patent effectively alleviates the problems related to non-iid data. In the subsequent simulation experiment, real-world communication conditions will be introduced. The vehicle equipment uses offline to verify the fairness constraint to alleviate the impact caused by offline problems.
[0149] The experiments in this section further explore the role of fairness as a constraint by introducing realistic metrics for vehicle devices that are offline due to network issues. The current discussion focuses on the learning efficiency under two different communication conditions: low-interference channels and medium-interference channels. Simulated experiments with low interference in natural environments such as Figure 4 and 5 As shown in the figure, the simulation experiment of moderate disturbance in natural environment is as follows Figure 6 As shown. The simulation of the above experiments shows that the fairness constraint outperforms other baselines in offline conditions. Theoretically, it is generated by reducing the randomness of vehicle device selection. When facing the problem of network packet loss, different algorithms for randomly selecting vehicle devices further reduce the expected value of vehicle device selection. Fairness constraints are introduced to recalculate the vehicle device selection scheme after the vehicle device is offline, and a re-estimated vehicle device scheme is obtained to ensure that no vehicle device cannot participate in the model aggregation process for a long time due to random vehicle device selection during the learning process.
[0150] Imposing fairness constraints requires fewer rounds of model training, thereby reducing the energy consumption of the vehicle during the entire training process. Imposing fairness constraints requires fewer rounds of model training, thereby reducing the energy consumption of vehicle equipment during the entire training process. The threshold for vehicle equipment to participate in federated learning is significantly reduced, and Table 1 shows the average energy consumption of vehicle equipment.
[0151] Table 1 Average energy consumption of vehicles in actual communication environment
[0152]
[0153] This study explores the fairness problem of ZTFL through Lyapunov optimization in frequency division multiple access (FDMA) systems. It proposes a dynamic queue quantization method for vehicle equipment participation considering packet loss rate. Through the analysis of the objective function, the present invention transforms some non-convex functions into convex functions and gives analytical solutions. In addition, the present invention uses MNIST, Cifar10 and Cifar100 datasets to experimentally evaluate the model of the present invention. The results show that in the case of lossy communication, the model proposed by the present invention can significantly improve the model accuracy while maintaining an average reduction of 26% in the communication completion rate.
[0154] The above is a detailed introduction to an embodiment of a fairness and energy consumption optimization method for cross-platform learning of a vehicle-pile network based on lossy communication of the present invention. The following is a detailed introduction to an embodiment of a fairness and energy consumption optimization system for cross-platform learning of a vehicle-pile network based on lossy communication of the present invention.
[0155] This embodiment provides a fairness and energy consumption optimization system for cross-platform learning of a vehicle-pile network based on lossy communication, including a central server and vehicle equipment, wherein the central server and vehicle equipment constitute a zero-trust federated learning framework ZTFL;
[0156] In the zero-trust federated learning framework ZTFL, in each learning round, the pre-built joint optimization system energy consumption and fairness optimization model is solved for the participating devices in the framework to obtain the selected devices in the current learning round, the communication resource allocation under lossy communication conditions, and the computing resource allocation, so as to perform zero-trust federated learning.
[0157] The optimization model aims to minimize the system energy consumption while considering long-term fairness constraints;
[0158] The system energy consumption includes the energy consumption related to uploading communication and the energy consumption related to training models, which are determined by the allocated communication resources and computing resources respectively;
[0159] The long-term fairness constraint is that under long-term conditions, the average number of times each device is selected satisfies the condition that it is at least not less than the set expected value.
[0160] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A fairness and energy consumption optimization method for cross-platform learning of vehicle-pile networks based on lossy communication, characterized in that: In the zero-trust federated learning framework ZTFL, in each learning round, the pre-built joint optimization system energy consumption and fairness optimization model is solved for the participating devices in the framework to obtain the selected devices in the current learning round, the communication resource allocation under lossy communication conditions, and the computing resource allocation, so as to perform zero-trust federated learning. The optimization model aims to minimize the system energy consumption while considering long-term fairness constraints; The system energy consumption includes energy consumption related to upload communication and energy consumption related to training model, which are determined by allocated communication resources and computing resources respectively; The long-term fairness constraint is that under long-term conditions, the average number of times each device is selected satisfies a condition that is at least not less than a set expected value.
2. The fairness and energy consumption optimization method for cross-platform learning of vehicle-pile network based on lossy communication according to claim 1 is characterized in that: The expression of the optimization model is as follows: Where n is the set of participating devices The index of the selected device in , τ is the set of global iteration rounds The indices of different iteration rounds in , is the communication energy consumption, For training energy consumption, represents the average number of times a device is selected at time T-1, β is the setting value to ensure fairness, and x τ,n =1 means that the nth device participates in federated learning in the τth round.
3. The fairness and energy consumption optimization method for cross-platform learning of vehicle-pile network based on lossy communication according to claim 2 is characterized in that: When solving the optimization model, the long-term fairness constraint is converted into an objective function through Lyapunov optimization theory for solving, and the conversion process includes: Building a virtual queue for each device, the virtual queue being used to represent a current backlog of the device due to time constraints; Establishing a quadratic Lyapunov function based on the virtual queue, wherein the quadratic Lyapunov function is used to constrain the increase of the cardinality of the virtual queue set; Based on Lyapunov drift theorem, constructing a drift function for representing the increment of the virtual queue caused by a single training device; Combining the drift function with the original objective function of the optimization model to obtain a drift plus cost function; Minimizing the drift plus cost function is used as the objective function of the transformed optimization model.
4. The fairness and energy consumption optimization method for cross-platform learning of vehicle-pile network based on lossy communication according to claim 3 is characterized in that: The expression of the drift plus cost function is as follows: Wherein, Δ(Θ(τ)) is the drift function, and V is the penalty factor; The expression of the drift function is as follows: In the formula, is the quadratic Lyapunov function, Θ(τ) is the set length of the virtual queue; The expression of the quadratic Lyapunov function is as follows: In the formula, Z τ,n is the virtual queue of the nth vehicle device in the τth round; The expression of the virtual queue is as follows: WITH τ+1,n =max[Z τ,n +β-x τ,n ,0]。 5. The fairness and energy consumption optimization method for cross-platform learning of vehicle-pile network based on lossy communication according to claim 4 is characterized in that: The objective function of the transformed optimization model is as follows: minf(x τ,n ,β,V) In the formula, f(x τ,n ,β,V) is the upper bound of the drift plus cost function, and the specific relationship is as follows:
6. The fairness and energy consumption optimization method for cross-platform learning of vehicle-pile network based on lossy communication according to claim 2 or 5 is characterized in that: The constraints of the optimization model include at least: Channel reachability constraints, channel condition constraints, bandwidth allocation ratio constraints, device selection constraints, bandwidth constraints, upload power constraints and computing resource constraints.
7. The fairness and energy consumption optimization method for cross-platform learning of vehicle-pile network based on lossy communication according to claim 6 is characterized in that: The channel reachability constraint is as follows: In the formula, a n is the probability of channel damage, m is the waterfall threshold, B is the total bandwidth of the central server, and b n is the bandwidth allocation ratio of each vehicle equipment channel, N0 represents the spectral power density of Gaussian noise, p n is the upload power, h n is the gain of the channel, and ε is the set threshold.
8. The fairness and energy consumption optimization method for cross-platform learning of vehicle-pile network based on lossy communication according to claim 5 is characterized in that: When solving the objective function of the transformed optimization model, the objective function is further transformed using convex optimization theory as follows: min f τ,n ,p * g(f,p * ) f min <f τ,n <f max In the formula, f τ,n represents the CPU frequency selected by vehicle device n during the τth global iteration, p * represents the optimal value of the upload power p, g(·) is the converted objective function; C τ,n represents the number of CPU cycles of device n in the τth global iteration, D τ,n represents the data set size of device n in the τth round of global iteration, t e represents the number of local iterations, ξ represents the CPU capacitor parameter, T Limit is the upper limit of system time consumption, is the optimal value of the upload power of device n in the τth round, f max and f min They are the upper and lower limits of the CPU frequency respectively.
9. The fairness and energy consumption optimization method for cross-platform learning of vehicle-pile network based on lossy communication according to claim 2 is characterized in that: In ZTFL, whether a device has been selected is determined according to the following formula: In the formula, is an indicator used to evaluate whether the device completes training within the time constraint. If yes, it is set to 1, otherwise it is 0; An indicator used to evaluate whether the device successfully uploaded related parameters. If yes, it is set to 1, otherwise it is 0; x' τ,n Indicates whether the nth vehicle device is selected in the τth round. If so, it is set to 1, otherwise it is 0.
10. A fairness and energy consumption optimization system for cross-platform learning of vehicle-pile networks based on lossy communication, characterized in that: It includes a central server and a vehicle device, wherein the central server and the vehicle device constitute a zero-trust federated learning framework ZTFL; In the zero-trust federated learning framework ZTFL, in each learning round, the pre-built joint optimization system energy consumption and fairness optimization model is solved for the participating devices in the framework to obtain the selected devices in the current learning round, the communication resource allocation under lossy communication conditions, and the computing resource allocation, thereby performing zero-trust federated learning; The optimization model aims to minimize the system energy consumption while considering long-term fairness constraints; The system energy consumption includes energy consumption related to upload communication and energy consumption related to training model, which are determined by allocated communication resources and computing resources respectively; The long-term fairness constraint is that under long-term conditions, the average number of times each device is selected satisfies a condition that is at least not less than a set expected value.
