CN117575291A - Federal learning data collaborative management method based on edge parameter entropy - Google Patents

Abstract

The application discloses a federal learning data collaborative management method based on edge parameter entropy, and belongs to the field of data management. The method comprises the following steps: acquiring initial parameters of a central model, performing local model training, and generating a local update model; calculating the edge parameter entropy of the local update model, and if the updated edge parameter entropy is larger than a preset iteration threshold, transmitting the local update model to a central end; the iteration is repeated until the edge parameter entropy of the local model is less than the threshold. And receiving and utilizing the global center model to determine a power dispatching plan of the electromechanical system, and sending the power dispatching plan to the corresponding electromechanical system. The scheme is transmitted to the central terminal only when the local model update change is obvious, so that the communication cost and the calculation burden of the central terminal are reduced; an update plan is set, so that a local terminal can be helped to update a model according to requirements, and model training efficiency is improved; by setting the power dispatching plan, the power resource waste can be reduced, and the utilization rate and the operation efficiency of the electromechanical system to the power are improved.

Description

Federal learning data collaborative management method based on edge parameter entropy

Technical Field

The application relates to the field of data management, in particular to a federal learning data collaborative management method based on edge parameter entropy.

Background

Along with the continuous development of industrial informatization, the electromechanical system is developed, the quantity of the constituent equipment of the electromechanical system is large, the structure is complex, faults are easy to occur in the running process, one fault usually corresponds to a plurality of possible reasons, and a plurality of diagnosis operation actions are needed to confirm the types of the possible reasons. The operation and maintenance work of the electromechanical system faces a plurality of challenges, the mode mainly comprising manual inspection is low in efficiency, high in cost and difficult to quantify, the maintenance of the electromechanical system depends on the experience of electromechanical maintenance personnel, and the experience of the electromechanical maintenance personnel is limited by factors such as knowledge level. Because the occurrence probabilities of various possible reasons are different, the cost (comprehensive consideration of time, personnel, material cost and the like) for diagnosing the operation actions of the possible reasons is also different, the fault decision cannot be efficiently made, the faults of the electromechanical system are fundamentally eliminated, the management cost of the operation data is high, the efficiency is low, meanwhile, the electromechanical system has the power resource waste, the utilization rate of the electromechanical system to the power is low, and the operation efficiency is low.

In addition, federal learning is an advanced machine learning method that has been developed due to the need for privacy protection and security. Federal learning solves the problem of privacy disclosure caused by traditional centralized machine learning by performing model training on local equipment, sharing only model updates but not original data. In the existing federal learning technology, the bottom-layer device trains a local model through local data and transmits model update to the top-layer device. The top level device gathers and integrates updates from the bottom level device, performing aggregation of the global model. The new global model updates are again transmitted to the underlying devices, allowing the model of each device to be updated. The process is iterated, the local information of all the devices is continuously integrated, and a global model is gradually formed. However, in the prior art, in the iterative process, if the model update generated by the bottom device is not changed greatly, but is still transmitted to the top device, the communication cost and the calculation pressure of the top device are increased. And training is only carried out according to global model parameters during each time of iterative training of the local end, and other references are not available, so that the number of iterative training times is excessive, and the model training efficiency is low.

Disclosure of Invention

In order to overcome the defects, the embodiment of the application provides a federal learning data collaborative management method based on edge parameter entropy, which solves the problems of high management cost and lower efficiency of the operation data of an electromechanical system in the prior art, and meanwhile, the electromechanical system has power resource waste, and has low utilization rate of power and lower operation efficiency; and the problem of communication cost and calculation pressure increase of the central terminal caused by the fact that the change condition of the local update model generated during each iteration is not evaluated and is transmitted to the central terminal during each local update model generation. And training is only carried out according to global model parameters during each time of iterative training of the local end, and other references are not available, so that the number of iterative training times is excessive, and the model training efficiency is low. According to the technical scheme, the communication cost between the local end and the central end can be reduced, and the calculation burden of the central end can be reduced; meanwhile, the power scheduling plan is set, so that the power resource waste can be reduced, and the utilization rate and the operation efficiency of the electromechanical system to the power are improved.

In a first aspect, an embodiment of the present application provides a federal learning data collaborative management method based on edge parameter entropy, where the method is executed by a local end, and the method includes:

Acquiring initial model parameters, performing local model training according to the initial model parameters and the pre-stored local data to obtain a local update model, and determining update model parameters according to the local update model;

calculating the edge parameter entropy of the local update model according to the initial model parameters, the update model parameters and an edge parameter entropy determination formula;

the edge parameter entropy determination formula is as follows:

；

wherein,is the edge parameter entropy; m is the total number of parameters; j is a parameter index; i is the iteration training round of the local model; />The method comprises the steps of locally updating initial model parameters before model training in an ith iteration; />Updating model parameters after the model training for the local updating in the ith iteration;

calculating the minimum value of the edge parameter entropy according to the edge parameter entropy and an edge parameter entropy minimization formula, and comparing the minimum value of the edge parameter entropy with a preset edge parameter entropy threshold;

if the minimum value of the edge parameter entropy is larger than a preset edge parameter entropy threshold, transmitting the local updating model to a central terminal for the central terminal to update a global model according to the local updating model;

And receiving a global model transmitted by a central terminal, acquiring the running state and the demand information of the electromechanical system, determining a power dispatching plan of the electromechanical system according to the global model, the running state and the demand information of the electromechanical system, and transmitting the power dispatching plan to the electromechanical system for the electromechanical system to perform power distribution adjustment according to the power dispatching plan.

Further, after comparing the minimum value of the edge parameter entropy with a preset edge parameter entropy threshold, the method further includes:

if the minimum value of the edge parameter entropy is smaller than a preset edge parameter entropy threshold, receiving a global model parameter transmitted by a central terminal, and determining an update plan for a local update model according to the global model parameter and a prestored data collaborative management strategy;

and carrying out local model training according to the update plan and the pre-stored local data, and updating the local update model.

Further, after updating the local update model, the method further includes:

calculating the edge parameter entropy of the updated local update model and the minimum value of the edge parameter entropy, if the minimum value of the edge parameter entropy is larger than a preset edge parameter entropy threshold, transmitting the updated local update model to a central terminal for the central terminal to update a global model according to the local update model, otherwise, receiving global model parameters transmitted by the central terminal, determining an update plan for the local update model, performing local model training according to the update plan and the prestored local data, and updating the local update model until a training stopping instruction transmitted by the central terminal is received.

Further, after calculating the edge parameter entropy of the locally updated model, the method further comprises:

and optimizing the local updating model according to the edge parameter entropy, a preset local model training optimization formula and the local updating model to obtain an optimized local updating model.

Further, after transmitting the local update model to the central terminal, the method further includes:

acquiring the transmission time of a local update model, determining the transmission delay of the local update model according to the transmission time, and determining whether the transmission delay reaches a preset transmission delay standard;

if the transmission delay does not reach the preset transmission delay standard, transmitting an instruction for acquiring the performance requirement of the global model to the central terminal;

and receiving the performance requirement of the global model transmitted by the central terminal, and re-determining the edge parameter entropy threshold according to the performance requirement of the global model, the transmission delay and a preset edge parameter entropy threshold determination standard.

Further, after transmitting the local update model to the central terminal, the method further includes:

acquiring network traffic used for local update model transmission, and determining whether the network traffic exceeds a preset network traffic use threshold;

If the network flow exceeds a preset network flow use threshold, transmitting an instruction for acquiring the performance requirement of the global model to a central terminal;

and receiving the performance requirement of the global model transmitted by the central terminal, and re-determining the edge parameter entropy threshold according to the performance requirement of the global model, the network flow and a preset edge parameter entropy threshold determination standard.

Further, the edge parameter entropy minimization formula is as follows:

；

b is a training sample set of the local update model; θ is the parameter set of the locally updated model;representing the j-th parameter after the local update model is trained.

Further, the local model training optimization formula is as follows:

；

wherein k is a local end number; t is the iteration number;is the i-th parameter; />Training an optimized objective function for the local model; />Is a cross entropy function; />Is an edge entropy function; />Is the coefficient multiplied by the edge entropy function.

Further, after sending the power dispatch plan to an electromechanical system, the method further comprises:

and monitoring the electricity consumption condition of the electromechanical system in real time according to the global model, if abnormal electricity consumption information is identified, updating the power dispatching plan according to the abnormal electricity consumption information, and sending the updated power dispatching plan to the electromechanical system for the electromechanical system to carry out power distribution adjustment again according to the power dispatching plan.

In the embodiment of the application, initial model parameters are obtained, local model training is carried out according to the initial model parameters and the pre-stored local data, a local update model is obtained, and update model parameters are determined according to the local update model; calculating the edge parameter entropy of the local update model according to the initial model parameters, the update model parameters and an edge parameter entropy determination formula; calculating the minimum value of the edge parameter entropy according to the edge parameter entropy and an edge parameter entropy minimization formula, and comparing the minimum value of the edge parameter entropy with a preset edge parameter entropy threshold; if the minimum value of the edge parameter entropy is larger than a preset edge parameter entropy threshold, transmitting the local updating model to a central terminal for the central terminal to update a global model according to the local updating model; receiving global model parameters transmitted by a central terminal, and determining an update plan for a local update model according to the global model parameters and a prestored data collaborative management strategy; and carrying out local model training according to the update plan and the pre-stored local data, and updating the local update model. By the federal learning data collaborative management method based on the edge parameter entropy, the data collaborative management method is transmitted to the central terminal only when the local update model is obviously changed, so that the communication cost between the local terminal and the central terminal and the calculation burden of the central terminal can be reduced. Meanwhile, an updating plan is set, so that the local terminal can update the model according to the requirements, the iterative training times are reduced, and the model training efficiency is improved. According to the technical scheme, the communication cost between the local end and the central end can be reduced, and the calculation burden of the central end can be reduced; meanwhile, the power scheduling plan is set, so that the power resource waste can be reduced, and the utilization rate and the operation efficiency of the electromechanical system to the power are improved.

Drawings

Fig. 1 is a flow chart of a data collaborative management method based on federal learning of edge parameter entropy according to an embodiment of the present application;

FIG. 2 is a schematic diagram of a federal learning data collaborative management policy based on edge parameter entropy according to an embodiment of the present application;

fig. 3 is a flow chart of a data collaborative management method based on federal learning of edge parameter entropy according to a second embodiment of the present application;

fig. 4 is a flow chart of a data collaborative management method based on federal learning of edge parameter entropy according to a third embodiment of the present application;

fig. 5 is a flow chart of a federal learning data collaborative management method based on edge parameter entropy according to a fourth embodiment of the present application.

Detailed Description

For the purpose of making the objects, technical solutions and advantages of the present application more apparent, the following detailed description of specific embodiments thereof is given with reference to the accompanying drawings. It is to be understood that the specific embodiments described herein are merely illustrative of the application and not limiting thereof. It should be further noted that, for convenience of description, only some, but not all of the matters related to the present application are shown in the accompanying drawings. Before discussing exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although a flowchart depicts operations (or steps) as a sequential process, many of the operations can be performed in parallel, concurrently, or at the same time. Furthermore, the order of the operations may be rearranged. The process may be terminated when its operations are completed, but may have additional steps not included in the figures. The processes may correspond to methods, functions, procedures, subroutines, and the like.

Technical solutions in the embodiments of the present application will be clearly described below with reference to the drawings in the embodiments of the present application, and it is apparent that the described embodiments are some embodiments of the present application, but not all embodiments. All other embodiments obtained by a person of ordinary skill in the art based on the embodiments in the present application are within the scope of the protection of the present application.

The terms first, second and the like in the description and in the claims, are used for distinguishing between similar objects and not necessarily for describing a particular sequential or chronological order. It is to be understood that the data so used may be interchanged, as appropriate, such that embodiments of the present application may be implemented in sequences other than those illustrated or described herein, and that the objects identified by "first," "second," etc. are generally of a type and not limited to the number of objects, e.g., the first object may be one or more. Furthermore, in the description and claims, "and/or" means at least one of the connected objects, and the character "/", generally means that the associated object is an "or" relationship.

The data collaborative management method based on the federation learning of the edge parameter entropy provided by the embodiment of the application is described in detail below through specific embodiments and application scenes thereof with reference to the accompanying drawings.

Embodiment one: fig. 1 is a flow chart of a federal learning data collaborative management method based on edge parameter entropy according to an embodiment of the present application. As shown in fig. 1, the method specifically comprises the following steps:

s101, obtaining initial model parameters, performing local model training according to the initial model parameters and the pre-stored local data to obtain a local update model, and determining update model parameters according to the local update model.

Firstly, the use scene of the scheme can be a scene in which a local end performs local model training, after a local update model is obtained, the edge parameter entropy of the local update model and the minimum value of the edge parameter entropy are calculated, if the minimum value of the edge parameter entropy is larger than a preset edge parameter entropy threshold, the local update model is transmitted to a central end, and otherwise, the next local model training is performed.

Based on the above usage scenario, it can be appreciated that the execution subject of the present application may be a local end, which is not limited herein.

In this solution, the method is performed by the local side.

The local end may be a separate device or entity that participates in training, such as a mobile device, sensor, or node in other distributed systems. Each local terminal holds local data and performs local model updating.

The initial model parameters may be initial parameters at the beginning of each iteration at the local end. The first iteration may be the model parameters initialized by the central terminal and distributed to the local terminals. At the beginning of a subsequent iteration, the initial model parameters may be parameters of the previous round of global model.

The pre-stored local data refers to a data set for training stored locally by the local end, and specifically may be local data on the user equipment. For example, if the local side is a smart phone, the pre-stored local data may include application usage records of the user, location information, and device sensor data. In industrial production, the pre-stored local data may include industrial parameters such as temperature, humidity, and pressure measured by the sensor.

The local update model may be model training by the data acquisition device using local data, and the obtained model parameter update. The locally updated model is a partially trained model. For example, in an automotive sensor, if the task is to analyze driving behavior, the locally updated model may contain model parameter updates regarding vehicle speed, acceleration, braking, single trip mileage, and steering, etc., to reflect the behavior of the vehicle and the driver.

Updating the model parameters can be to adjust the weight, bias and other parameters of the model through an optimization algorithm after one or more rounds of training, so that the model performs better on training data, and the updated parameters are parameters used by the trained model.

After the central terminal completes global model training each time, global model parameters are distributed to the local terminal, the local terminal can firstly store the global model parameters when receiving the global model parameters, and when local model training is required, the pre-stored global model parameters are obtained to serve as an initial model of the local model training. The local end then uses the obtained initial model parameters and the local pre-stored data to perform local model training, which may include calculating a loss function using the local data set, and then adjusting the model parameters by an optimization algorithm such as back propagation and gradient descent to reduce the loss. After the local model training is completed, the local end obtains a local update model, and the model reflects the adjustment of local data to the global model. The local update model only contains information learned on local data, and does not contain data of other local ends. When the local update model training is completed, updated model parameters may be saved, which may include, in particular, weights, biases, and other parameters learned during the training process of the model, and then stored in a file for later use.

S102, calculating the edge parameter entropy of the local update model according to the initial model parameter, the update model parameter and an edge parameter entropy determination formula.

Edge parameter entropy may be an indicator of uncertainty or complexity of model parameters. In this scheme, the edge parameter entropy may refer to uncertainty or information entropy of a model parameter given a locally updated model. The larger the value of the edge parameter entropy is, the more uncertain the model is to the parameter, the stronger the adaptability of the model to training data is, and the larger the contribution of the model to the global model generated by the center is correspondingly. When the value of the edge parameter entropy is small, the change of the local update model can be regarded as small, the contribution to the generation of the global model by the central terminal is small, and the model parameters at the moment can not be uploaded to the central terminal. And the lower the value of the edge parameter entropy is, the higher the similarity between the initial model parameter and the updated model parameter is, and the smaller the model weight distribution difference is; whereas the larger the value of the edge parameter entropy, the more divergent the parameter distribution. The divergence of the model results in a lower global model diagnostic rate, and therefore the entropy of the edge parameters as part of the loss function can be used to control the degree of divergence of the model.

The edge parameter entropy determination formula can be stored in the local end in advance, and after the initial model parameters are obtained and the model parameters are updated, the stored edge parameter entropy determination formula can be called, and the initial model parameters and the local updated model are used as formula parameters to calculate the edge parameter entropy. Specifically, when the edge parameter entropy needs to be calculated, the local end can call the pre-stored initial model parameters and the updated model parameters stored in the file, take the initial model parameters and the updated model parameters as formula parameters, and substitute the formula parameters into an edge parameter entropy determining formula to calculate so as to obtain the edge parameter entropy.

On the basis of the above technical solution, optionally, the edge parameter entropy determining formula is as follows:

；

wherein,is the edge parameter entropy; m is the total number of parameters; j is a parameter index; i is the iteration training round of the local model; />The method comprises the steps of locally updating initial model parameters before model training in an ith iteration; />And (5) updating model parameters after the model training for the local updating in the ith iteration.

In this scheme, in the calculation of the edge parameter entropy, the parameter is usually expressed in the form of a vector or matrix, and M is the total number of elements in the vector or matrix of the parameter. When the parameters are represented in the form of vectors, j may represent an element index in the vector; when the parameters are represented in the form of a matrix, j may represent a rank index in the matrix. Because the local end and the central end perform iterative training, when i is 1, the local end can be represented to perform local model training for the first time, when i is 2, the local end can be represented to perform local model training for the second time, and so on.

S103, calculating the minimum value of the edge parameter entropy according to the edge parameter entropy and an edge parameter entropy minimization formula, and comparing the minimum value of the edge parameter entropy with a preset edge parameter entropy threshold.

The edge parameter entropy minimization formula may be a formula that minimizes uncertainty of model parameters by adjusting the model parameters.

The edge parameter entropy minimization formula and the edge parameter entropy threshold value can be stored in a local end, and after the edge parameter entropy is obtained, the formula is automatically called to calculate the minimum value of the edge parameter entropy, and the minimum value of the edge parameter entropy is compared with the edge parameter entropy threshold value.

On the basis of the above technical solution, optionally, the edge parameter entropy minimization formula is as follows:

；

b is a training sample set of the local update model; θ is the parameter set of the locally updated model;representing the j-th parameter after the local update model is trained.

In this scenario, the training sample set of the locally updated model may refer to a data set used at the local end for training the model, which may be collected and stored by the local device or system. The training sample set is an input data sample used by the model when training on the local side.

Locally updating the parameter set of the model may refer to the set of parameters such as the weight and bias of the model during the local training process. These parameters can be adjusted during the training process by an optimization algorithm to make the model perform better on the local data. Once the local model training is complete, the updated parameter set may be parameters of the local update model.

And S104, if the minimum value of the edge parameter entropy is larger than a preset edge parameter entropy threshold, transmitting the local updating model to a central terminal for the central terminal to update the global model according to the local updating model.

If the minimum value of the edge parameter entropy is larger than the preset edge parameter entropy threshold, the local update model is larger in change, and the contribution to the generation of the global model at the central end is larger. Correspondingly, the local end transmits the local update model to the central end through the secure channel. The central end can integrate the local update models transmitted by the local ends, and in particular, can be completed by different integration algorithms, such as a federal average algorithm and the like. The central terminal can update the global model by using the integrated local update model, specifically, the central terminal can be realized by applying the integrated gradient information on the parameters of the global model, and the update modes can include gradient descent, random gradient descent and other modes.

On the basis of the above technical solution, optionally, after comparing the minimum value of the edge parameter entropy with a preset edge parameter entropy threshold, the method further includes:

if the minimum value of the edge parameter entropy is smaller than a preset edge parameter entropy threshold, receiving a global model parameter transmitted by a central terminal, and determining an update plan for a local update model according to the global model parameter and a prestored data collaborative management strategy;

and carrying out local model training according to the update plan and the pre-stored local data, and updating the local update model.

In the scheme, if the minimum value of the edge parameter entropy is smaller than the preset edge parameter entropy threshold, the change of the local update model is not large, and the contribution to the central end training global model is not large, so that the local end cannot transmit the local update model to the central end, and the obtained parameters of the local update model can be frozen. However, there may be a case that the minimum value of the edge parameter entropy corresponding to the local update model obtained after the local model training of other local ends is greater than the preset edge parameter entropy threshold, at this time, the other local ends meeting the standard transmit the local update model to the central end, and the central end uses the local update models to update the global model, and transmits the obtained global model parameters to all the local ends through the secure channel. The local side, which now freezes the locally updated model parameters, receives these global model parameters and analyzes them. And comparing the difference between the global model parameters and the current local update model parameters to know which parts need to be updated. And determining a processing scheme for the local data according to the pre-stored data collaborative management strategy. For example, if the policy is to prioritize certain data categories, the update plan may need to focus on model parameters associated with those categories. It is also necessary to consider the characteristics and distribution of the local data, to know which data has a greater impact on the model, and how to adjust the model to better accommodate the local data. And finally, by combining the information, designing a specific updating plan, specifically, determining which model parameters need to be updated, updating modes, updating sequences and the like.

In the scheme, the local updating model parameters are frozen when the local updating model obtained after the local end performs local model training is not obvious, and the next iteration training is directly performed, so that unnecessary communication cost between the local end and the central end can be reduced, meanwhile, the calculation burden of the central end is reduced, and the stability of the whole system is improved. Meanwhile, an updating plan is set, so that the local terminal can update the model according to the requirements, the iterative training times are reduced, and the model training efficiency is improved.

S105, receiving a global model transmitted by a central terminal, acquiring the running state and the demand information of the electromechanical system, determining a power dispatching plan of the electromechanical system according to the global model, the running state and the demand information of the electromechanical system, and transmitting the power dispatching plan to the electromechanical system for the electromechanical system to perform power distribution adjustment according to the power dispatching plan.

The electromechanical system may be a comprehensive system, and in particular may include components of a power plant, a power transmission network, a substation, etc. for generating, transmitting, and distributing power.

The running state may refer to the working condition of each component in the electromechanical system, specifically, may include the output power of the generator, the running state of the transformer, the load condition of the power grid, and the like, and these information may be collected in real time through the sensor and the monitoring device.

The demand information may refer to load demands in the power system, that is, real-time demand conditions of power by users, and specifically, may include load curves and predicted load demands in different time periods.

The local end can receive the global model through a wireless communication technology, and can acquire real-time running state and load demand information of the electromechanical system through a sensor and monitoring equipment connected to the electromechanical system. And then, utilizing the received global model, combining the running state and the demand information of the electromechanical system, determining an optimal power scheduling plan, wherein the optimal power scheduling plan can comprise power adjustment of a power station, coordinated running of equipment and the like, and transmitting the determined power scheduling plan to the electromechanical system. After the electromechanical system receives the power scheduling plan, each power station and related equipment can be informed to carry out corresponding power adjustment according to the plan. For example, the output power of the generator may be increased or decreased, parameters of the transformer may be adjusted, etc. By building the federal learning model, each electromechanical system performs model training locally, determines whether the updated local model needs to be transmitted to the central end according to comparison between the edge parameter entropy and the threshold value, and uploads the locally updated model parameters to the central end for aggregation when the edge parameter entropy is greater than a preset iteration threshold value. The central terminal updates the global model by integrating the local data of each system to realize the sharing and fusion of model parameters, and then issues the global model to the local terminal, and the local terminal uses the global model to carry out the power dispatching plan of the electromechanical system. Therefore, the communication cost and the calculation burden of the central end can be reduced, the power dispatching plan can be predicted more accurately, the collaborative management of data is realized through federal learning, and the overall performance of the model is improved.

According to the technical scheme provided by the embodiment, initial model parameters are obtained, local model training is carried out according to the initial model parameters and the pre-stored local data, a local update model is obtained, and update model parameters are determined according to the local update model; calculating the edge parameter entropy of the local update model according to the initial model parameters, the update model parameters and an edge parameter entropy determination formula; calculating the minimum value of the edge parameter entropy according to the edge parameter entropy and an edge parameter entropy minimization formula, and comparing the minimum value of the edge parameter entropy with a preset edge parameter entropy threshold; if the minimum value of the edge parameter entropy is larger than a preset edge parameter entropy threshold, transmitting the local updating model to a central terminal for the central terminal to update a global model according to the local updating model; and receiving a global model transmitted by a central terminal, acquiring the running state and the demand information of the electromechanical system, determining a power dispatching plan of the electromechanical system according to the global model, the running state and the demand information of the electromechanical system, and transmitting the power dispatching plan to the electromechanical system for the electromechanical system to perform power distribution adjustment according to the power dispatching plan. By the federal learning data collaborative management method based on the edge parameter entropy, the data collaborative management method is transmitted to the central end only when the local update model is obviously changed, so that the communication cost between the local end and the central end and the calculation burden of the central end can be reduced; meanwhile, the power scheduling plan is set, so that the power resource waste can be reduced, and the utilization rate and the operation efficiency of the electromechanical system to the power are improved.

On the basis of the above technical solution, optionally, after updating the local update model, the method further includes:

calculating the edge parameter entropy of the updated local update model and the minimum value of the edge parameter entropy, if the minimum value of the edge parameter entropy is larger than a preset edge parameter entropy threshold, transmitting the updated local update model to a central terminal for the central terminal to update a global model according to the local update model, otherwise, receiving global model parameters transmitted by the central terminal, determining an update plan for the local update model, performing local model training according to the update plan and the prestored local data, and updating the local update model until a training stopping instruction transmitted by the central terminal is received.

In this scheme, fig. 2 is a schematic diagram of federal learning data collaborative management policy based on edge parameter entropy according to the first embodiment of the present application, and as shown in fig. 2, the instruction for stopping training may be an instruction sent by the central terminal to the local terminal to stop the local terminal from performing local model training. After the central terminal updates the global model each time, whether the global model reaches a preset evaluation standard can be inspected, and if the global model does not reach the standard, the central terminal can send the latest global model parameters to the local terminal for the local terminal to perform the next local model training according to the global model parameters. If the standard is reached, a training stopping instruction can be sent to the local end to inform the local end that the global model reaches the standard, and iterative training is not needed.

After the local end updates the local update model, the edge parameter entropy of the local update model and the minimum value of the edge parameter entropy can be calculated again, whether the minimum value of the edge parameter entropy is larger than a preset edge parameter entropy threshold value is estimated again, if so, the local update model updated at the time is considered to be obvious in change, the contribution to global model training of the central end is large, and the local update model can be transmitted to the central end. If not, the local update model updated at the time is smaller in change, the contribution to global model training of the central terminal is smaller, transmission of the local update model is not needed, and next iteration training can be directly carried out according to an update plan. And after each training, the edge parameter entropy and the minimum value of the edge parameter entropy are calculated, and the minimum value of the edge parameter entropy is compared with a preset edge parameter entropy threshold value, so that the local model training is not stopped until a training stopping instruction transmitted by the central end is received.

In the scheme, the local end can be ensured to capture data variation in time by automatically and continuously carrying out iterative training of the local model, and the local update model with obvious variation is transmitted to the central end, so that the accuracy of the global model is improved.

On the basis of the above technical solution, optionally, after the power scheduling plan is sent to the electromechanical system, the method further includes:

and monitoring the electricity consumption condition of the electromechanical system in real time according to the global model, if abnormal electricity consumption information is identified, updating the power dispatching plan according to the abnormal electricity consumption information, and sending the updated power dispatching plan to the electromechanical system for the electromechanical system to carry out power distribution adjustment again according to the power dispatching plan.

In this scheme, the abnormal electricity consumption information may refer to an electricity consumption condition of the electromechanical system, which is inconsistent with a normal electricity consumption mode, and specifically may include a sudden electricity fluctuation, an abnormal increase or decrease of an electrical load, and the like.

The power usage data of each device and area in the electromechanical system can be collected in real time through sensors or monitoring devices, and specifically, the power usage data can comprise information such as power load, power and voltage. And analyzing the power use data acquired in real time by using a global model. A pattern or incident that does not correspond to the normal electricity usage pattern is identified and identified as abnormal electricity usage information. Upon detection of abnormal power usage information, the system may generate a new power schedule according to predefined coping strategies or optimization algorithms to accommodate the abnormal situation, and send the updated power schedule to the electromechanical system via wireless communication techniques to guide the actual power distribution and operation. For example, an abnormal electricity use condition suddenly occurs in a certain area in the electromechanical system, and the electric load sharply increases. This anomaly is detected by the global model and a new power schedule is generated using a predefined optimization algorithm. The new schedule may include adjusting the output of other devices, starting a backup generator, or balancing the power load by adjusting the use of the energy storage system, and then sending the updated power schedule to the electromechanical system for the electromechanical system to re-distribute power.

In the scheme, the power consumption condition is monitored in real time, the emergency can be responded in time, the power distribution adjustment is carried out again, and the stability of the electromechanical system is improved.

Embodiment two: fig. 3 is a flow chart of a federal learning data collaborative management method based on edge parameter entropy according to a second embodiment of the present application, and as shown in fig. 3, the specific method includes the following steps:

s301, obtaining initial model parameters, performing local model training according to the initial model parameters and the pre-stored local data to obtain a local update model, and determining update model parameters according to the local update model.

S302, calculating the edge parameter entropy of the local update model according to the initial model parameter, the update model parameter and an edge parameter entropy determination formula.

S303, optimizing the local update model according to the edge parameter entropy, a preset local model training optimization formula and the local update model, and obtaining an optimized local update model.

In the scheme, a preset local model training optimization formula can be a formula which ensures that a local update model of each local end can have higher classification precision, so that parameters of each local update model can be in the same order of magnitude, and the edge parameter entropy can be one of parameters of the formula.

After the edge parameter entropy is obtained, the obtained edge parameter entropy can be input into the formula to obtain a specific local model training optimization function, and the model parameters are directly adjusted on the basis of the original local update model by utilizing the function to obtain the optimized local update model.

In the scheme, the local updating model parameters of each local end can be balanced by optimizing the local updating model, so that the local updating model parameters are in the same order of magnitude, and the influence of the model parameters of certain edge devices on the global model is prevented from being excessive.

Based on the above technical solution, optionally, the local model training optimization formula is as follows:

；

wherein k is a local end number; t is the iteration number;is the i-th parameter; />Training an optimized objective function for the local model; />Is a cross entropy function; />Is an edge entropy function; />Is the coefficient multiplied by the edge entropy function.

In this scheme, the cross entropy function may be a commonly used loss function for measuring the performance of the model in the distributed learning environment. If in federal learning there are multiple local ends, each with its own local data set. The global model is updated on the central side, and the cross entropy loss function can be used to measure the difference between the predictions of the global model and the real labels.

The objective function of each local side contains cross entropy and edge parameter entropy. In calculating these two losses, which are very far apart in order to optimize the two functions with the same degree of importance during model training, a coefficient may be multiplied in front of the parameter entropyIt will have different values depending on the dataset and model. The improvement can ensure the accuracy of local data and enable the trained model parameters to be close to the global model parameters when the local end trains the local model. The lower the value of the edge parameter entropy is, the higher the qi and pi similarity is, and the smaller the model weight distribution difference is; whereas the larger the value of the edge parameter entropy, the more divergent the parameter distribution. The divergence of the model results in a lower global model diagnostic rate, and therefore edge parameter entropy is used to control the degree of divergence of the model as part of the objective function of the local model training optimization.

S304, calculating the minimum value of the edge parameter entropy according to the edge parameter entropy and an edge parameter entropy minimization formula, and comparing the minimum value of the edge parameter entropy with a preset edge parameter entropy threshold.

S305, if the minimum value of the edge parameter entropy is larger than a preset edge parameter entropy threshold, the local updating model is transmitted to a central terminal, and the central terminal updates a global model according to the local updating model.

And S306, receiving a global model transmitted by a central terminal, acquiring the running state and the demand information of the electromechanical system, determining a power dispatching plan of the electromechanical system according to the global model, the running state and the demand information of the electromechanical system, and transmitting the power dispatching plan to the electromechanical system for the electromechanical system to perform power distribution adjustment according to the power dispatching plan.

Embodiment III: fig. 4 is a flow chart of a federal learning data collaborative management method based on edge parameter entropy according to a third embodiment of the present application, as shown in fig. 4, and the specific method includes the following steps:

s401, obtaining initial model parameters, performing local model training according to the initial model parameters and the pre-stored local data to obtain a local update model, and determining update model parameters according to the local update model.

S402, calculating the edge parameter entropy of the local update model according to the initial model parameter, the update model parameter and an edge parameter entropy determination formula.

S403, calculating the minimum value of the edge parameter entropy according to the edge parameter entropy and an edge parameter entropy minimization formula, and comparing the minimum value of the edge parameter entropy with a preset edge parameter entropy threshold.

S404, if the minimum value of the edge parameter entropy is larger than a preset edge parameter entropy threshold, the local updating model is transmitted to a central terminal, and the central terminal updates a global model according to the local updating model.

S405, acquiring the transmission time of the local update model, determining the transmission delay of the local update model according to the transmission time, and determining whether the transmission delay reaches a preset transmission delay standard.

The transmission time of the local update model may refer to the time it takes for the local end to transmit the local update model to the central end.

The preset transmission delay standard may be a measurement transmission delay standard set according to specific application scenarios and requirements, and in particular, the standard may include requirements on instantaneity, response speed, and the like. For example, the local update model may be transmitted within a certain time range, so as to ensure that the global model can obtain the latest local update in time.

The transmission delay may refer to a time delay from the local end to the central end receiving the model during the model transmission.

When the local end starts to transmit the local update model, the sent time stamp is recorded. This represents the time at which the transmission starts. The locally updated model is then transmitted to the central end over the network. When the central terminal receives the complete local update model, the received time stamp is recorded. This indicates the time at which the transmission is completed. By calculating the difference between the receive timestamp and the transmit timestamp, the transmission delay of the local update model can be obtained. And then the local end compares the calculated transmission delay with a preset transmission delay standard. If the transmission delay is less than or equal to the preset standard, the transmission delay is within the acceptable range; otherwise, the transmission delay exceeds the standard.

S406, if the transmission delay does not reach the preset transmission delay standard, transmitting an instruction for acquiring the performance requirement of the global model to the central terminal.

The local side may generate a performance requirement request according to the local performance requirement and the application scenario, and specifically may include a desire for the performance of the global model, such as an accuracy and a delay requirement that the model is expected to achieve on a specific task. The generated performance requirement request is then transmitted to the central terminal over the network. After receiving the request for performance requirement of the local end, the central end can provide corresponding response according to the content of the request, specifically, the method can include informing the local end of the performance condition of the current global model, or providing advice, adjustment scheme and the like.

S407, receiving the performance requirement of the global model transmitted by the central terminal, and re-determining the edge parameter entropy threshold according to the performance requirement of the global model, the transmission delay and a preset edge parameter entropy threshold determination standard.

The performance requirement of the global model may be generally the expected performance or requirement of the central end on the locally updated model uploaded by each local end, and specifically may include performance requirements in terms of accuracy, convergence speed, robustness and the like of the model.

The preset edge parameter entropy threshold determination criterion may be a reference criterion for adjusting the edge parameter entropy threshold. For example, if the locally updated model transmitted over a period of time has less improvement in global model performance, the edge parameter entropy threshold may be lowered to reduce the frequency of transmission. Otherwise, if the task progress is obvious, the performance is improved, the threshold value can be moderately improved, and the communication overhead is reduced; if the communication overhead is too high, namely the transmission delay or the network traffic occupation is large, the improvement of the edge parameter entropy threshold can be considered to reduce unnecessary model transmission, so that the communication cost can be reduced while the task performance requirement is met; for more time sensitive tasks, more frequent model updates may be required so that the edge parameter entropy threshold may be lowered to increase the frequency of transmission.

When the transmission delay does not reach the preset transmission delay standard, the delay is higher, and the entropy threshold of the edge parameter can be properly increased to reduce the transmission frequency of the local update model. The local end can evaluate the time cost of model transmission according to the actual transmission delay condition, and can formulate a set of rules according to the performance requirement, the transmission delay, the edge parameter entropy threshold and other information of the global model, specifically, the indexes about the aspects of model performance, communication efficiency, resource utilization and the like, and redetermine the edge parameter entropy threshold so as to ensure that the edge parameter entropy threshold is in a proper range. For example, a federal learning system is used to train a speech recognition model. The local end comprises a smart phone and edge equipment, and the central end is a cloud server. The global model performance requirements for the central side transmission include requirements for speech recognition accuracy, e.g., up to 90% accuracy. The local end establishes standards according to the performance requirement and the transmission delay of the global model, for example, if the required accuracy is higher and the transmission delay is higher, the local end needs to reduce the transmission frequency of the local update model and appropriately improve the threshold value of the edge parameter entropy.

S408, receiving a global model transmitted by a central terminal, acquiring the running state and the demand information of the electromechanical system, determining a power dispatching plan of the electromechanical system according to the global model, the running state and the demand information of the electromechanical system, and transmitting the power dispatching plan to the electromechanical system for the electromechanical system to perform power distribution adjustment according to the power dispatching plan.

In this embodiment, by measuring the transmission delay, the edge parameter entropy threshold may be adjusted in real time when the network changes, so as to adapt to different communication environments, and reduce the communication cost while maintaining the task performance.

Embodiment four: fig. 5 is a flow chart of a federal learning data collaborative management method based on edge parameter entropy according to a fourth embodiment of the present application, and as shown in fig. 5, the specific method includes the following steps:

s501, obtaining initial model parameters, performing local model training according to the initial model parameters and the pre-stored local data to obtain a local update model, and determining update model parameters according to the local update model.

S502, calculating the edge parameter entropy of the local update model according to the initial model parameter, the update model parameter and an edge parameter entropy determination formula.

S503, calculating the minimum value of the edge parameter entropy according to the edge parameter entropy and an edge parameter entropy minimization formula, and comparing the minimum value of the edge parameter entropy with a preset edge parameter entropy threshold.

And S504, if the minimum value of the edge parameter entropy is larger than a preset edge parameter entropy threshold, transmitting the local updating model to a central terminal for the central terminal to update the global model according to the local updating model.

S505, obtaining network traffic used by local update model transmission, and determining whether the network traffic exceeds a preset network traffic use threshold.

The network traffic used may be the amount of data transmitted over the network during a local update model transmission, and in particular may be expressed in units of bytes, kilobytes, megabytes, and the like.

The preset network traffic usage threshold may be a preset threshold, which represents a maximum network traffic usage allowed by the system, and in particular, the threshold may be determined according to factors such as a requirement of the system, a network condition, and a device performance.

The network traffic of the local device may be monitored during the model transmission, in particular by recording the network data volume at the beginning and end of the transmission, or by using special network monitoring tools. And then accumulating the network traffic to obtain the total network traffic transmitted by the local update model, and comparing the accumulated network traffic with a preset network traffic use threshold.

S506, if the network flow exceeds a preset network flow use threshold, transmitting an instruction for acquiring the performance requirement of the global model to the central terminal.

S507, receiving the performance requirement of the global model transmitted by the central terminal, and redetermining the edge parameter entropy threshold according to the performance requirement of the global model, the network flow and a preset edge parameter entropy threshold determining standard.

The local end can receive the performance requirement of the global model transmitted by the central end through the secure channel, and the performance requirement of the global model transmitted by the central end can comprise indexes in terms of accuracy, delay, throughput and the like of the model. Considering the use condition of network traffic, the edge parameter entropy threshold needs to be adjusted under the condition of network traffic limitation so as to balance between the performance requirement and the use of the network traffic, and specifically, the performance requirement of the global model and the use of the network traffic can be comprehensively considered. If the performance requirements are higher, a larger edge parameter entropy threshold is needed in order to transmit the locally updated model more frequently. If the network traffic is limited, the edge parameter entropy threshold needs to be correspondingly reduced, and the transmission frequency is reduced. Based on the preset edge parameter entropy threshold determination standard, the edge parameter entropy threshold is reset in combination with the global model performance and the network traffic, and the threshold should enable the transmission of the local update model to control the use of the network traffic while meeting the performance requirement.

And S508, receiving the global model transmitted by the central terminal, acquiring the running state and the demand information of the electromechanical system, determining the power dispatching plan of the electromechanical system according to the global model, the running state and the demand information of the electromechanical system, and transmitting the power dispatching plan to the electromechanical system for the electromechanical system to perform power distribution adjustment according to the power dispatching plan.

In this embodiment, the edge parameter entropy threshold is adjusted in real time according to the network traffic, so that unnecessary model transmission can be reduced, and the network traffic is reduced, thereby more effectively utilizing limited bandwidth, enabling the system to better adapt to the change of the network condition, and improving the response speed of the system.

The foregoing description is only of the preferred embodiments of the present application and the technical principles employed. The present application is not limited to the specific embodiments described herein, but is capable of numerous obvious changes, rearrangements and substitutions as will now become apparent to those skilled in the art without departing from the scope of the present application. Therefore, while the present application has been described in connection with the above embodiments, the present application is not limited to the above embodiments, but may include many other equivalent embodiments without departing from the spirit of the present application, and the scope of the present application is determined by the scope of the claims.

Claims (9)

1. A federal learning data collaborative management method based on edge parameter entropy, wherein the method is performed by a local side, the method comprising:

acquiring initial model parameters, performing local model training according to the initial model parameters and the pre-stored local data to obtain a local update model, and determining update model parameters according to the local update model;

calculating the edge parameter entropy of the local update model according to the initial model parameters, the update model parameters and an edge parameter entropy determination formula;

the edge parameter entropy determination formula is as follows:

；

wherein,is the edge parameter entropy; m is the total number of parameters; j is a parameter index; i is the iteration training round of the local model; />The method comprises the steps of locally updating initial model parameters before model training in an ith iteration; />Updating model parameters after the model training for the local updating in the ith iteration;

calculating the minimum value of the edge parameter entropy according to the edge parameter entropy and an edge parameter entropy minimization formula, and comparing the minimum value of the edge parameter entropy with a preset edge parameter entropy threshold;

if the minimum value of the edge parameter entropy is larger than a preset edge parameter entropy threshold, transmitting the local updating model to a central terminal for the central terminal to update a global model according to the local updating model;

And receiving a global model transmitted by a central terminal, acquiring the running state and the demand information of the electromechanical system, determining a power dispatching plan of the electromechanical system according to the global model, the running state and the demand information of the electromechanical system, and transmitting the power dispatching plan to the electromechanical system for the electromechanical system to perform power distribution adjustment according to the power dispatching plan.

2. The federal learning data collaborative management method based on edge parameter entropy according to claim 1, wherein after comparing the minimum value of the edge parameter entropy with a preset edge parameter entropy threshold, the method further comprises:

if the minimum value of the edge parameter entropy is smaller than a preset edge parameter entropy threshold, receiving a global model parameter transmitted by a central terminal, and determining an update plan for a local update model according to the global model parameter and a prestored data collaborative management strategy;

and carrying out local model training according to the update plan and the pre-stored local data, and updating the local update model.

3. The federal learning data collaborative management method based on edge parameter entropy according to claim 1 or 2, wherein after updating a local update model, the method further comprises:

Calculating the edge parameter entropy of the updated local model and the minimum value of the edge parameter entropy, and if the minimum value of the edge parameter entropy is larger than a preset edge parameter entropy threshold, transmitting the updated local model parameter to a central end for the central end model to update a global model; otherwise, the local model training iteration is finished, global model parameters transmitted by the central terminal are received, and the local model is determined.

4. The federal learning data collaborative management method based on edge parameter entropy according to claim 1, wherein after computing edge parameter entropy of the locally updated model, the method further comprises:

and optimizing the local updating model according to the edge parameter entropy, a preset local model training optimization formula and the local updating model to obtain an optimized local updating model.

5. The federal learning data collaborative management method based on edge parameter entropy according to claim 1, further comprising, after transmitting the local update model to a central side:

acquiring the transmission time of a local update model, determining the transmission delay of the local update model according to the transmission time, and determining whether the transmission delay reaches a preset transmission delay standard;

If the transmission delay does not reach the preset transmission delay standard, transmitting an instruction for acquiring the performance requirement of the global model to the central terminal;

and receiving the performance requirement of the global model transmitted by the central terminal, and re-determining the edge parameter entropy threshold according to the performance requirement of the global model, the transmission delay and a preset edge parameter entropy threshold determination standard.

6. The federal learning data collaborative management method based on edge parameter entropy according to claim 1, further comprising, after transmitting the local update model to a central side:

acquiring network traffic used for local update model transmission, and determining whether the network traffic exceeds a preset network traffic use threshold;

if the network flow exceeds a preset network flow use threshold, transmitting an instruction for acquiring the performance requirement of the global model to a central terminal;

and receiving the performance requirement of the global model transmitted by the central terminal, and re-determining the edge parameter entropy threshold according to the performance requirement of the global model, the network flow and a preset edge parameter entropy threshold determination standard.

7. The federal learning data collaborative management method based on edge parameter entropy according to claim 1, wherein the edge parameter entropy minimization formula is as follows:

；

b is a training sample set of the local update model; θ is the parameter set of the locally updated model;representing the j-th parameter after the local update model is trained.

8. The federal learning data collaborative management method based on edge parameter entropy according to claim 6, wherein the local model training optimization formula is as follows:

；

wherein,knumbering the local end;tthe iteration times;is the i-th parameter; />Training an optimized objective function for the local model; />Is a cross entropy function; />Is an edge entropy function; />Is the coefficient multiplied by the edge entropy function.

9. The federal learning data collaborative management method based on edge parameter entropy according to claim 1, further comprising, after sending the power dispatch plan to an electromechanical system:

and monitoring the electricity consumption condition of the electromechanical system in real time according to the global model, if abnormal electricity consumption information is identified, updating the power dispatching plan according to the abnormal electricity consumption information, and sending the updated power dispatching plan to the electromechanical system for the electromechanical system to carry out power distribution adjustment again according to the power dispatching plan.