Interaction method and device for multiple intelligent fusion terminals, equipment and medium

Abstract

The invention provides an interaction method and device for multiple intelligent fusion terminals, equipment and a medium, and belongs to the technical field of data interaction.The method comprises the steps that when it is detected that the network state of a current terminal is an offline state, agent capacity values of multiple terminals are predicted based on a trained agent capacity prediction model, and then an agent intelligent fusion terminal is determined; and sending a broadcast signal to the outside, so that the proxy intelligent fusion terminal sends the mark data of the current terminal to a master controller. When it is detected that the network state of the current terminal is recovered to an online state, sending a proxy data request signal to a general controller, receiving a proxy data list and a control instruction sent by the general controller, sending a collaborative parameter request signal to the multiple terminals, and receiving model key regulation and control parameters sent by the multiple associated terminals during the offline period of the current terminal; and sending unmarked data to a master controller based on the proxy data list and the control instruction, and updating model parameters based on the model key regulation and control parameters. The data interaction efficiency can be improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of data interaction, and more particularly relates to an interaction method and device of a multi-intelligent fusion terminal, equipment and a medium. BACKGROUND

[0002] In the existing intelligent fusion terminal application scenarios, multiple intelligent fusion terminals usually need to interact with the master control and work cooperatively to achieve various complex tasks and functions. However, in actual operation, the network state of the intelligent fusion terminal does not always remain stable online, and the online state is often switched to the offline state. When a certain intelligent fusion terminal is in the offline state, the normal data transmission and instruction interaction between it and the master control will be interrupted, which will cause the data accumulated by the terminal during the offline period to be unable to be uploaded to the master control in time, and the master control cannot issue new instructions to the terminal or obtain its related state information, which seriously affects the information interaction efficiency between multiple devices. SUMMARY

[0003] The purpose of the present application is to provide an interaction method and device of a multi-intelligent fusion terminal, equipment and a medium to improve the information interaction efficiency of the multi-intelligent fusion terminal when switching between online and offline states, and reduce the occurrence of fault problems.

[0004] The first aspect of the embodiment of the present application provides an interaction method of a multi-intelligent fusion terminal, comprising: In response to detecting that the network state of the current intelligent fusion terminal is switched from the online state to the offline state, predicting the agent capability value of the multiple target intelligent fusion terminals based on the trained agent capability prediction model, the trained agent capability prediction model being obtained by federated learning cooperative training based on the multiple target intelligent fusion terminals and the master control when the current intelligent fusion terminal is in the online state; Determining the agent intelligent fusion terminal from the multiple target intelligent fusion terminals based on the agent capability value; Sending a broadcast signal to the outside, the broadcast signal being used to make the agent intelligent fusion terminal send the mark data of the current intelligent fusion terminal to the master control; In response to detecting that the network state of the current intelligent fusion terminal is switched from the offline state to the online state, sending an agent data request signal to the master control, and receiving the agent data list and control instructions generated by the master control based on the agent data request signal; If the agent data list and control instructions sent by the master control are received, a cooperative parameter request signal is sent to the multiple target intelligent fusion terminals, and the model key control parameters during the offline period sent by the multiple target intelligent fusion terminals are received, the agent data list and control instructions being generated based on the agent data request signal; sending unmarked data to the master based on the agent data list and the control instruction, and updating model parameters of the trained agent capability prediction model based on the model key regulation parameters.

[0005] In a second aspect, the present application provides an interaction device of a multi-intelligent fusion terminal, which comprises: The first detection unit is configured to, in response to detecting that the network state of the current intelligent fusion terminal is switched from the online state to the offline state, predict agent capability values of the plurality of target intelligent fusion terminals based on the trained agent capability prediction model, which is obtained through federated learning collaborative training based on the plurality of target intelligent fusion terminals and the master when the current intelligent fusion terminal is in the online state. The first processing unit is configured to determine an agent intelligent fusion terminal from the plurality of target intelligent fusion terminals based on the agent capability values. The signal sending unit is configured to send a broadcast signal to the outside, the broadcast signal being used to cause the agent intelligent fusion terminal to send the marked data of the current intelligent fusion terminal to the master. The second detection unit is configured to, in response to detecting that the network state of the current intelligent fusion terminal is switched from the offline state to the online state, send an agent data request signal to the master, and receive an agent data list and a control instruction sent by the master based on the agent data request signal. The second processing unit is configured to, if the agent data list and the control instruction sent by the master are received, send a collaborative parameter request signal to the plurality of target intelligent fusion terminals, and receive model key regulation parameters during the offline period sent by the plurality of target intelligent fusion terminals, the agent data list and the control instruction being generated based on the agent data request signal. The model updating unit is configured to send unmarked data to the master based on the agent data list and the control instruction, and update model parameters of the trained agent capability prediction model based on the model key regulation parameters.

[0006] In a third aspect, the present application provides an electronic device, which comprises a memory, a processor, and a computer program stored in the memory and running on the processor, and the processor implements the steps of the interaction method of the multi-intelligent fusion terminal when executing the computer program.

[0007] In a fourth aspect, the present application provides a computer readable storage medium, which stores a computer program, and the processor implements the steps of the interaction method of the multi-intelligent fusion terminal when executing the computer program.

[0008] The interaction method and device of the multi-intelligent fusion terminal, the equipment and the medium provided by the embodiments of the present application have the following beneficial effects: The embodiments of the present application effectively improve the running stability and reliability of each terminal device under unstable network conditions. When the current intelligent fusion terminal is in an offline state, the proxy capability values of multiple target associated intelligent fusion terminals can be accurately predicted, and a proxy intelligent fusion terminal is determined therefrom. This process is based on a federated learning collaborative training model, fully utilizes the data and computing resources of each target associated intelligent fusion terminal, and ensures the accuracy and reliability of the prediction. When the current intelligent fusion terminal resumes the online state, a proxy data request signal is sent to the general control, and after receiving the proxy data list and control instructions, the model key control parameters during the offline period are requested from the multiple target intelligent fusion terminals, and finally the model parameters are updated based on this information and unmarked data is sent. The entire process forms a complete closed loop, so that the system can still maintain the continuity of the data and the stability of the model when the network state frequently switches, effectively avoiding the problems of data loss and model failure caused by offline, and ensuring that each terminal device can run stably and efficiently under different network environments. BRIEF DESCRIPTION OF DRAWINGS

[0009] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0010] Figure 1 A flowchart of an interaction method of a multi-intelligent fusion terminal provided by an embodiment of the present application; Figure 2 A structural block diagram of an interaction device of a multi-intelligent fusion terminal provided by an embodiment of the present application; Figure 3 A schematic block diagram of an electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0011] In the following description, specific details such as specific system structures, techniques, etc. are presented in order to thoroughly understand the embodiments of the present application. However, it should be clear to those skilled in the art that the present application can also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits and methods are omitted to avoid unnecessary details that hinder the description of the present application.

[0012] In order to make the purpose, technical solutions and advantages of the present application clearer, specific embodiments will be described below with reference to the drawings.

[0013] Reference Figure 1 , Figure 1This is a flowchart illustrating an embodiment of the interaction method for multiple intelligent fusion terminals provided in this application. The method is applied to a large-scale terminal interaction system, which includes multiple interconnected intelligent fusion terminals and a central control unit. Each intelligent fusion terminal interacts with the central control unit. The multiple interconnected intelligent fusion terminals include a current intelligent fusion terminal and multiple target intelligent fusion terminals. The current intelligent fusion terminal is any one of the multiple interconnected intelligent fusion terminals. The method is executed by the current intelligent fusion terminal and includes: S101: In response to detecting that the network status of the current intelligent fusion terminal has switched from online to offline, predict the agent capability values ​​of multiple target intelligent fusion terminals based on the trained agent capability prediction model.

[0014] Among them, the trained agent capability prediction model is obtained by federated learning collaborative training based on multiple target intelligent fusion terminals and the central control when the current intelligent fusion terminal is online; the multiple target intelligent fusion terminals are all devices that have established an association with the current intelligent fusion terminal, and the agent capability value is used to judge the reliability of the target intelligent fusion terminal in transmitting data on behalf of the current intelligent fusion terminal; In this embodiment, the large-scale terminal interaction system includes multiple interconnected intelligent fusion terminals and a central control unit. Each intelligent fusion terminal receives operating data from the electrical equipment in its assigned area and can make preliminary judgments based on this data, such as whether the electrical equipment has malfunctioned, for example, whether a certain electrical equipment is overloaded. If a malfunction is detected, the malfunction information is promptly uploaded to the central control unit. If no malfunction is detected, or only a trend towards a malfunction is detected, all received data is shared with other terminals associated with that intelligent fusion terminal, and all data is simultaneously sent to the central control unit for unified judgment.

[0015] Among them, the distribution area is the smallest power supply management unit in the power system, specifically referring to a specific geographical area covered by a distribution transformer or a distribution substation, such as a community, a village, or a small industrial and commercial area. It is the basic scope for a single intelligent fusion terminal to collect data and monitor equipment.

[0016] The central control unit is the core management and scheduling center of a large-scale terminal interactive system. It can receive equipment operation data uploaded by various intelligent converged terminals and perform unified analysis and judgment. When the central control unit detects that a terminal is offline, it can coordinate with other associated terminals to receive the data collected by the current terminal in the offline state and perform timely analysis and judgment to avoid delays in handling faults.

[0017] When all devices in the large terminal interaction system are in an online state, the current intelligent fusion terminal, multiple target intelligent fusion terminals and the master control cooperatively train the agent capability prediction model corresponding to each terminal through federated learning every certain time length, to obtain the trained agent capability prediction model corresponding to each terminal. When some intelligent fusion terminals are offline, these terminals can predict the agent capability value of other intelligent fusion terminals in an online state based on the trained agent capability prediction model of each terminal. The agent capability value is used to judge the reliability of the intelligent fusion terminal for data transmission. The higher the agent capability value, the greater the possibility of the intelligent fusion terminal in an offline state for data transmission.

[0018] In the embodiment, the training sample of the agent capability prediction model includes historical hardware performance data, historical agent task performance data, network environment and topology relationship data of the intelligent fusion terminal, and the corresponding agent capability value.

[0019] The historical hardware performance data includes the communication module performance data of the terminal, such as the signal strength, transmission rate and packet loss rate of wireless communication; the computing and storage capability data, such as the CPU processing speed and the remaining amount of memory / storage space; the endurance and stability data, such as the device power supply stability and the continuous operation fault-free time length, etc.

[0020] The historical agent task performance data includes the agent transmission success rate, transmission delay and data integrity. The agent transmission success rate is the proportion of successful and complete delivery when replacing other terminals to transmit data in the history; the transmission delay is the average time consumption and maximum delay from receiving to forwarding data to the master control; the data integrity is whether data loss, error or tampering occurs in the agent process.

[0021] The network environment and topology relationship data includes the physical distance from the offline intelligent fusion terminal, the network link quality from the master control and the network load in the transformer area.

[0022] The above training sample can be processed locally by each intelligent fusion terminal and the master control under the framework of federated learning, and only the model parameters are shared for collaborative training, so that the trained agent capability prediction model can accurately predict the agent capability value of each target intelligent fusion terminal when the current intelligent fusion terminal is offline.

[0023] S102: Determine the agent intelligent fusion terminal from the multiple target intelligent fusion terminals based on the agent capability value.

[0024] In the embodiment, when the network state of the current smart converged terminal is changed from online to offline, the current smart converged terminal can obtain the agent capability values of the target smart converged terminals based on the trained agent capability prediction model, and determine the agent smart converged terminal based on the agent capability values. For example, an agent capability threshold can be set, and when the agent capability value of a target smart converged terminal is greater than or equal to the agent capability threshold, the target smart converged terminal is determined as the agent smart converged terminal. There can be multiple agent smart converged terminals.

[0025] S103: externally sending a broadcast signal, the broadcast signal being used to make the agent smart converged terminal send the marker data of the current smart converged terminal to the master controller.

[0026] In the embodiment, after determining the agent smart converged terminal that can transmit the offline data, the current smart converged terminal sends information to the outside through a broadcast signal. The broadcast signal can carry the information that the current smart converged terminal has gone offline, the identifier (such as the device ID) of the selected agent smart converged terminal, and the indication that the marker data needs to be transmitted. The transmission range of the broadcast signal usually covers all target smart converged terminals associated with the current smart converged terminal, ensuring that the selected agent smart converged terminal can accurately receive and respond. In the embodiment, the indication that the marker data needs to be transmitted in the broadcast signal is used to make the agent smart converged terminal determine that the offline smart converged terminal needs help to transmit the offline data in a period of time.

[0027] In an embodiment, the current smart converged terminal filters the offline data in a preset time period based on a preset rule every preset time period to obtain the marker data. The preset time period and the preset rule can be set based on the historical offline data of the current smart converged terminal. The preset rule can include the type of historical data prone to failure, the data acquisition frequency lower than a preset data acquisition frequency, etc., and the preset time period can be set according to experience.

[0028] Suppose there is one agent smart converged terminal, the current smart converged terminal can send all the marked data to the agent smart converged terminal, and the agent smart converged terminal sends the marker data to the master controller. Suppose there are at least N agent smart converged terminals, N≥2, the marker data can be equally divided into N parts, or divided into N parts based on a specific classification rule. The current smart converged terminal can send each part of data to a different agent smart converged terminal, and each agent smart converged terminal sends the corresponding data to the master controller.

[0029] S104: In response to detecting that the network state of the current intelligent fusion terminal is switched from the offline state to the online state, sending an agent data request signal to the general control, receiving the agent data list and control instructions generated by the general control based on the agent data request signal.

[0030] In this embodiment, the network state of the current intelligent fusion terminal is switched from the online state to the offline state, and then from the offline state to the online state, constituting a cycle, that is, a complete network failure event. In response to detecting that the network state of the current intelligent fusion terminal is switched from the offline state to the online state, the current intelligent fusion terminal needs to send the unmarked data in the offline state to the general control, so it needs to send an agent data request signal to the general control first. The agent data request signal is a data reconciliation and demand application signal, and its core function is to prepare for subsequent offline unmarked data transmission. The agent data request signal includes the identity information, fault cycle information (offline duration) and confirmation instruction of the current intelligent fusion terminal.

[0031] After the current intelligent fusion terminal sends the agent data request signal to the general control, the general control searches for the marking information received by itself based on the instruction, identity information and fault cycle information, and generates an agent data list and a control instruction. The agent data list usually contains the type, quantity and timestamp of the marking information. The control instruction is an operation instruction generated by the general control and sent to the current intelligent fusion terminal, and the instruction content can be "please send the data not sent in the offline state based on the agent data list" or "please update the model parameters".

[0032] S105: If the agent data list and control instruction sent by the general control are received, a cooperative parameter request signal is sent to a plurality of target intelligent fusion terminals, and the model key control parameters during the offline period sent by the plurality of target intelligent fusion terminals are received.

[0033] The agent data list and control instruction are generated based on the agent data request signal.

[0034] In this embodiment, if the current intelligent fusion terminal receives the agent data list and control instruction sent by the general control, it can determine the unmarked data in the offline state and the unmarked data based on the agent data list, and send the above data to the general control for storage. In addition, the current intelligent fusion terminal also needs to send a cooperative parameter request signal to a plurality of target intelligent fusion terminals, which is used to obtain the model control data during the offline period. After receiving the signal, the plurality of target intelligent fusion terminals respectively send the model key control parameters in the local to the current intelligent fusion terminal, which include weight adjustment parameters, feature mapping parameters and loss function correction parameters, etc. After receiving the model key control parameters sent by each terminal, the current intelligent fusion terminal can adjust the local model parameters based on these data.

[0035] S106: sending unmarked data to the general control based on the agent data list and the control instruction, and updating model parameters of the trained agent capability prediction model based on the model key regulation parameters.

[0036] In the embodiment, the current intelligent fusion terminal determines unmarked data and un-sent marked data in the offline state based on the agent data list, and sends the data to the general control for storage. Through this operation, the "data gap" during the offline period of the current intelligent fusion terminal can be filled, and the integrity of the general control obtaining the data of the electrical equipment in the transformer area is ensured. In addition, the current intelligent fusion terminal integrates the respective weight adjustment parameters, feature mapping parameters and loss function correction parameters sent by the plurality of target intelligent fusion terminals, obtains effective regulation parameters, and updates the model parameters of the local agent capability prediction model based on the effective regulation parameters. The updated model parameters are fed back to the general control through federated learning, ensuring that the agent capability prediction model parameters of all terminals in the entire system remain consistent, and avoiding prediction result deviation caused by model parameter difference.

[0037] From the above, it can be concluded that the embodiment of the application is applied to a large terminal interaction system, effectively improving the running stability and reliability of each terminal device of the large terminal interaction system under unstable network conditions. When the current intelligent fusion terminal is in an offline state, the proxy capability values of a plurality of associated intelligent fusion terminals can be accurately predicted, and a proxy intelligent fusion terminal is determined therefrom. This process is based on federated learning to collaboratively train the model, fully utilizes the data and computing resources of each terminal, and ensures the accuracy and reliability of the prediction. When the terminal returns to an online state, the proxy data request signal is sent to the general control, and after receiving the agent data list and control instruction, the model key regulation parameters during the offline period are requested from the plurality of target intelligent fusion terminals, and finally the model parameters are updated based on these information and unmarked data is sent. The entire process forms a complete closed loop, so that the system can still maintain the continuity of the data and the stability of the model when the network state frequently switches, effectively avoiding the problems of data loss and model failure caused by offline, and ensuring that each terminal device in the large terminal interaction system can stably and efficiently operate under different network environments.

[0038] In an embodiment of the present application, determining a proxy intelligent fusion terminal from a plurality of target intelligent fusion terminals based on a proxy capability value comprises: Screening at least one terminal with a proxy capability value greater than or equal to a preset proxy capability value as an initial candidate set; Calculating the physical distance and the transformer area overlap degree between each terminal in the initial candidate set and the current intelligent fusion terminal; Based on the physical distance and the transformer area overlap degree, determine the proxy intelligent fusion terminal from the initial candidate set.

[0039] In the embodiment, the agent capability value is used to measure the comprehensive capability index of the intelligent fusion terminal when acting as a proxy to transmit data, including but not limited to computing capability, communication stability, load capacity, reliability and other dimensions. The higher the agent capability value, the stronger the agent capability. The preset agent capability value is a threshold value preset to filter terminals with basic agent qualifications. Terminals below this value are excluded due to insufficient capability. If the agent capability value of the intelligent fusion terminal is greater than or equal to the preset agent capability value, the intelligent fusion terminal is stored in the initial candidate set until all intelligent fusion terminals are judged, and the initial candidate set is obtained.

[0040] Because the terminals in the initial candidate set only meet the basic qualifications and do not consider adaptability, it is necessary to filter the terminals in the initial candidate set by physical distance and distribution transformer area overlap to obtain the proxy intelligent fusion terminal. The physical distance represents the physical distance between the intelligent fusion terminal and the previous intelligent fusion terminal. The closer the distance, the more matching it is. The distribution transformer area is a basic topology unit in the power network. High overlap means that the terminal has stronger relevance in the power network and the data interaction is more consistent with the coordination logic of the power system, which can reduce the complexity of cross-regional coordination.

[0041] Therefore, after normalizing and weighted summing the above two parameters, the weighted score is obtained, and the terminal with the highest weighted score is selected to improve the real-time performance, stability and coordination efficiency of the proxy.

[0042] In an embodiment of the present application, after determining the proxy intelligent fusion terminal from the plurality of target intelligent fusion terminals based on the agent capability value, the method further comprises: Classifying the local distribution transformer data obtained in the offline state, and calculating the comprehensive score corresponding to each type of data based on the urgency and timeliness of each type of data; Determining the data with a comprehensive score greater than a first threshold value as marked data; Sending the marked data to the proxy intelligent fusion terminal.

[0043] In the embodiment, the local distribution transformer data represents power-related data of the current intelligent fusion terminal in the station area (distribution transformer power supply area), which can include voltage, current, power, device state, fault information and power load, etc. Different data has different urgency and timeliness. For example, the distribution transformer data is classified into real-time operation data (including current and voltage, etc.), device state data (including fault state and running time, etc.), historical statistical data (including daily and monthly power consumption, etc.) and other categories.

[0044] For each class of data after classification, a comprehensive score is calculated based on its urgency and timeliness. The urgency reflects the size of the data's impact on the current running state of the current intelligent fusion terminal, and the timeliness reflects how quickly the data loses its value over time. In this embodiment, a certain algorithm, such as a weighted average algorithm, is used to set different weights for the urgency and timeliness, and the comprehensive score of each class of data is obtained by considering these two factors comprehensively.

[0045] A different first threshold is set for each class of data, or the first thresholds corresponding to each class of data are the same. The data with a comprehensive score greater than the first threshold is determined as the marked data in each class of data. These marked data are the data that are currently critical to the system operation and need to be processed in a timely manner.

[0046] In this embodiment, when it is detected that the current intelligent fusion terminal is in an offline state, the marked data is sent to the selected proxy intelligent fusion terminal every first time interval. In this way, important data can be gradually transmitted out in the offline state for subsequent processing and utilization. In addition, the marked data and unmarked data are also stored in the storage space, such as the memory, of the current intelligent fusion terminal, so that when the online state is restored, the data that have not been transmitted can be directly sent to the master control.

[0047] From the above, it can be seen that by calculating the comprehensive score based on the urgency and timeliness and determining the marked data, the embodiment can prioritize processing data that has a large impact on system operation and strong timeliness. This ensures that critical data is not missed or delayed in the offline state, improves the system's response speed to important data, and ensures the stable operation of the system.

[0048] In addition, in the case of the current intelligent fusion terminal being offline, important data can be transmitted to the proxy intelligent fusion terminal in a timely manner through the above method, so that some terminals can still maintain a certain data processing and transmission capability when the network is unstable or interrupted, enhancing the system's fault tolerance and adaptability to network failures and improving the stability and reliability of the entire large-scale terminal interaction system.

[0049] In an embodiment of the present application, the proxy intelligent fusion terminal includes at least two target intelligent fusion terminals; before sending the marked data to the proxy intelligent fusion terminal, the method further includes: receiving attribute information sent by each proxy intelligent fusion terminal respectively, the attribute information including at least one of network stability, remaining storage capacity, and historical data transmission success rate; wherein, sending the marked data to the proxy intelligent fusion terminal includes: The marked data is divided into at least two data subsets according to data volume or urgency, and each data subset is sent to a corresponding agent intelligent fusion terminal respectively; The priority order of each agent intelligent fusion terminal is determined based on the attribute information; Based on the priority order, the corresponding data subsets are sent to the agent intelligent fusion terminals in turn at every first time interval, wherein the time point of sending data to the agent intelligent fusion terminal with the highest priority is the first time interval from detecting the offline state, and the time points of sending data to the subsequent agent intelligent fusion terminals with lower priorities are sequentially delayed by a preset time difference.

[0050] For example, there are three data subsets A, B and C, three agent intelligent fusion terminals D1, D2 and D3, the first time interval is 10 minutes, and the preset time difference is 2 minutes. The priority order of the three intelligent fusion terminals is D1>D2>D3. Therefore, 10 minutes after the current intelligent fusion terminal switches to the offline state, the data subset A is sent to the terminal D1, 12 minutes after the current intelligent fusion terminal switches to the offline state, the data subset B is sent to the terminal D2, and 14 minutes after the current intelligent fusion terminal switches to the offline state, the data subset C is sent to the terminal D3.

[0051] In this embodiment, the attribute information of the agent intelligent fusion terminal includes at least one of network stability, remaining storage capacity and historical data transmission success rate. The network stability reflects the reliability of the network environment where the agent intelligent fusion terminal is located, and can be measured by indicators such as network packet loss rate and network delay fluctuation. High network stability means a lower probability of data interruption or error during transmission. The remaining storage capacity refers to the size of the free space available for storing data in the agent intelligent fusion terminal. Sufficient remaining storage capacity can ensure that the received marked data has a place to store, avoiding data loss due to insufficient storage space. The historical data transmission success rate refers to the ratio of the number of successful data transmissions to the total number of transmissions in a certain period of time. A high historical data transmission success rate indicates that the terminal has high reliability and stability in data transmission.

[0052] If the agent intelligent fusion terminal includes the above three attribute information, the three attribute information can be respectively valued and then weighted and summed to obtain a final evaluation score. The priority order of each agent intelligent fusion terminal is determined according to the order of the evaluation scores from large to small. The priority order is used to guide the corresponding transmission relationship of the data subsets in the following. The marked data is divided into at least two data subsets according to data volume or urgency, and the current intelligent fusion terminal sends different data subsets to the agent intelligent fusion terminals, for example, the data subset with the largest data volume is sent to the agent intelligent fusion terminal with the highest evaluation score / priority.

[0053] In the embodiment, every first time interval, the current intelligent fusion terminal needs to screen and mark the data obtained in the first time interval, and sequentially send the marked data to the proxy intelligent fusion terminal for data transmission. According to the time point, the data transmission is sequentially delayed, which can avoid the problem of channel congestion caused by the simultaneous competition of all proxy intelligent fusion terminals for the Bluetooth transmission channel. In the environment where the Bluetooth signal is easily disturbed, the time-sharing transmission reduces the possibility of signal conflict, reduces the risk of data transmission failure, and improves the success rate of overall data transmission.

[0054] In an embodiment of the present application, the model key regulation parameter is used to update the model parameter of the trained agent ability prediction model, comprising: Classify the received model key regulation parameters sent by the plurality of target intelligent fusion terminals to obtain weight adjustment parameters, feature mapping parameters and loss function correction parameters; For each type of parameter, based on the model contribution of each intelligent fusion terminal in the historical collaborative training, the differential weights of different intelligent fusion terminals sending the same type of parameter are determined; Based on the model key regulation parameters sent by different intelligent fusion terminals and the differential weights corresponding to each type of parameter, the same type of parameters are weighted and fused to obtain the fusion results corresponding to each type of parameter; Based on the fusion results corresponding to each type of parameter, the network layer corresponding to the trained agent ability prediction model is used to update the model parameter of the trained agent ability prediction model.

[0055] In the embodiment, the model key regulation parameter refers to the parameter that can have a key impact on the performance of the trained agent ability prediction model. These parameters are used to optimize the prediction effect of the model on the agent ability, involving the adjustment of different aspects of the model, and the model key regulation parameter includes weight adjustment parameters, feature mapping parameters and loss function correction parameters, etc. Among them, the weight adjustment parameter is a type of parameter used to adjust the weight of each network layer in the agent ability prediction model. The weight determines the importance of different input features or neurons in model calculation. By adjusting the weight, the model can change the way of processing data and the prediction result. The feature mapping parameter represents the parameter related to the feature mapping operation in the model. Feature mapping is the process of converting input data from the original feature space to another feature space. These parameters control the conversion method and rules, and affect the extraction and representation of data features by the model. The loss function correction parameter is a parameter used to correct the loss function of the agent ability prediction model. The loss function is used to measure the difference between the model prediction result and the true result. By correcting the loss function, the model can be guided to train and optimize in a more accurate direction.

[0056] In this embodiment, different intelligent fusion terminals have different contributions to the model in historical collaborative training. Some terminals have a large contribution to the improvement of model performance, while some terminals have a small contribution to the improvement of model performance. This embodiment evaluates the model contribution of each terminal by analyzing historical collaborative training data, and then assigns different weights to the same type of model key control parameters sent by each terminal according to the contribution.

[0057] In this embodiment, based on the model key control parameters sent by different intelligent fusion terminals and the different weights corresponding to each type of parameter, the same type of parameters is weighted and fused to obtain the fusion result corresponding to each type of parameter. For example, for the loss function correction parameter, the same type of parameters is weighted and fused to obtain the fusion result: .

[0058] Wherein, n represents the number of other intelligent fusion terminals except the current intelligent fusion terminal, j represents the jth intelligent fusion terminal, represents the loss function correction value of the jth intelligent fusion terminal, represents the different weight of the jth intelligent fusion terminal.

[0059] In one embodiment, based on the fusion result corresponding to each type of parameter, the model parameters of the trained agent capability prediction model are updated using the network layer corresponding to the trained agent capability prediction model, including: The hierarchical sensitivity of the fusion result corresponding to each type of parameter is checked, and based on the parameter sensitivity threshold of each network layer of the agent capability prediction model, abnormal parameter components exceeding the parameter sensitivity threshold are screened out. The parameter sensitivity threshold is determined based on the historical model adjustment influence degree, and the historical model adjustment influence degree is the influence degree of the parameter adjustment of each network layer in the historical training on the model prediction accuracy; For abnormal parameter components, based on the local interaction data features of the current intelligent fusion terminal in offline state, a parameter correction factor is generated. The distribution deviation degree of the parameter correction factor and the local interaction data features is positively correlated; The same type of abnormal parameter components corrected based on the parameter correction factor and the normal parameter components that do not exceed the parameter sensitivity threshold are spliced to obtain the parameter set to be updated of each network layer; Based on the parameter set to be updated of each network layer, the model parameters of the trained agent capability prediction model are updated using the network layer corresponding to the trained agent capability prediction model.

[0060] In this embodiment, the object of hierarchical sensitivity check is the fusion result corresponding to each type of parameter in the model key control parameter, that is, the parameter adjustment amount. "Hierarchical" means that the sensitivity of the model key control parameter is analyzed layer by layer according to the network layer structure of the agent capability prediction model.

[0061] The agent capability prediction model is usually composed of multiple network layers with different functions, such as an input layer, a convolution layer, a pooling layer, a fully connected layer, and the like. Each network layer undertakes a unique task in the prediction process of the model, and there are significant differences in the sensitivity to parameter changes. For example, the convolution layer is responsible for extracting features of an image or data, and a slight change in the parameters thereof can have a greater impact on the accuracy of feature extraction, and thus affect the prediction performance of the entire model; and the pooling layer is mainly used for dimension reduction and extraction of main features, and the sensitivity to parameter changes can be relatively low. Therefore, if a unified sensitivity standard is used to check the parameters of all network layers, some key network layer parameters can be ignored, or non-key network layer parameters can be overly sensitive, thereby affecting the optimization effect of the model. Through hierarchical sensitivity checking, more reasonable and accurate parameter checking standards can be developed for the characteristics of different network layers, and the effectiveness of model parameter updating can be improved.

[0062] After obtaining the fusion results corresponding to various parameters respectively, the fusion results are classified according to the network layers to which the parameters belong. Then, for each network layer, the fusion results of the parameters of the layer are compared with the pre-set parameter sensitivity threshold, and abnormal parameter components that exceed the parameter sensitivity threshold are screened out.

[0063] In this embodiment, for the abnormal parameter components, a parameter correction factor is generated based on the local interaction data features in the offline state of the intelligent fusion terminal, and the abnormal parameter components are corrected based on the parameter correction factor. The parameter correction factor is positively correlated with the distribution deviation degree of the local interaction data features. The distribution deviation degree of the local interaction data features represents the difference in distribution between the local interaction data features generated in the offline state of the intelligent fusion terminal and the "standard data features" used in the model training. The greater the difference, the greater the value of the parameter correction factor.

[0064] In this embodiment, the same type of abnormal parameter components corrected based on the parameter correction factor are spliced with normal parameter components that do not exceed the parameter sensitivity threshold, to obtain a set of to-be-updated parameters of each network layer; and based on the set of to-be-updated parameters of each network layer, the model parameters of the trained agent capability prediction model are updated by using the network layers corresponding to the trained agent capability prediction model.

[0065] From the above, the embodiment can precisely locate parameters that have a great impact on the prediction accuracy of the model by screening abnormal parameter components through hierarchical sensitivity verification, thereby improving the pertinence of parameter adjustment; the parameter correction factor is generated according to the characteristics of local interaction data, the abnormal parameters can be flexibly corrected in combination with the actual situation, and the adaptability of the model is enhanced; the updated model parameters after splicing the corrected abnormal parameters and normal parameters can not only retain effective information but also optimize the abnormal part, which is helpful to improve the prediction accuracy and stability of the prediction model of the agent capability, so that the prediction model can more accurately predict in different scenarios.

[0066] In an embodiment of the present application, the parameter adjustment of each network layer in historical training includes the number of parameter adjustments and the adjustment amplitude of the parameter of the lth layer each time; The parameter sensitivity threshold is determined based on the influence degree of historical model adjustment, comprising: The parameter sensitivity threshold is determined based on the following formula:

[0067] wherein, is the parameter sensitivity threshold of the lth network layer, is the weight coefficient, is the number of parameter adjustments of the lth network layer in historical training, is the change amount of the prediction accuracy of the model after the ith adjustment of the parameter of the lth layer, is the adjustment amplitude of the parameter of the lth layer the ith time, is the standard deviation of the parameter adjustment amplitude of the lth network layer in historical training.

[0068] The interaction method of the multi-intelligent fusion terminal corresponding to the above embodiment, Figure 2 is a structural block diagram of the interaction device of the multi-intelligent fusion terminal provided by an embodiment of the present application. The device is applied to a large terminal interaction system. In addition to the device, the large terminal interaction system also includes a plurality of interrelated intelligent fusion terminals and a general control. Each intelligent fusion terminal interacts with the general control. The plurality of interrelated intelligent fusion terminals include a current intelligent fusion terminal and a plurality of target intelligent fusion terminals. The current intelligent fusion terminal is any intelligent fusion terminal in the plurality of interrelated intelligent fusion terminals. For ease of illustration, only the parts related to the embodiments of the present application are shown. For details, refer to Figure 2 The interaction device of the multi-intelligent fusion terminal 20 includes a first detection unit 21, a first processing unit 22, a signal sending unit 23, a second detection unit 24, a second processing unit 25, and a model updating unit 26.

[0069] The first detection unit 21 is configured to, in response to detecting that the network state of the current smart fusion terminal is switched from online to offline, predict the agent capability values of the target smart fusion terminals based on the trained agent capability prediction model. The trained agent capability prediction model is obtained by federated learning collaborative training based on the target smart fusion terminals and the master control when the current smart fusion terminal is in the online state. The first processing unit 22 is configured to determine the agent smart fusion terminal from the target smart fusion terminals based on the agent capability values. The signal sending unit 23 is configured to send a broadcast signal to the outside, and the broadcast signal is used to make the agent smart fusion terminal send the marker data of the current smart fusion terminal to the master control. The second detection unit 24 is configured to, in response to detecting that the network state of the current smart fusion terminal is switched from offline to online, send an agent data request signal to the master control, and receive the agent data list and control instructions generated by the master control based on the agent data request signal. The second processing unit 25 is configured to, if the agent data list and control instructions sent by the master control are received, send a collaborative parameter request signal to the target smart fusion terminals, and receive the model key control parameters during offline of the target smart fusion terminals, and the agent data list and control instructions are generated based on the agent data request signal. The model updating unit 26 is configured to send unmarked data to the master control based on the agent data list and control instructions, and update the model parameters of the trained agent capability prediction model based on the model key control parameters.

[0070] In an embodiment of the present application, the first processing unit 22 is specifically configured to: Filter at least one terminal with an agent capability value greater than or equal to a preset agent capability value as an initial candidate set; Calculate the physical distance and the distribution area overlap between each terminal in the initial candidate set and the current smart fusion terminal; Determine the agent smart fusion terminal from the initial candidate set based on the physical distance and the distribution area overlap.

[0071] In an embodiment of the present application, after determining the agent smart fusion terminal from the target smart fusion terminals based on the agent capability values, the interaction device 20 of the multi-smart fusion terminal further comprises a third processing unit. The third processing unit is configured to classify the distribution data of the local area obtained in the offline state, and calculate the comprehensive scores corresponding to each type of data based on the emergency degree and timeliness of each type of data. Determine the data with a comprehensive score greater than a first threshold value as marker data. The marked data is sent to the proxy intelligent fusion terminal.

[0072] In an embodiment of the present application, the proxy intelligent fusion terminal comprises at least two target intelligent fusion terminals; before the marked data is sent to the proxy intelligent fusion terminal, the interaction device 20 of the multi-intelligent fusion terminal further comprises a receiving unit; The receiving unit is configured to receive attribute information sent by each proxy intelligent fusion terminal respectively, the attribute information comprising at least one of network stability, remaining storage capacity and historical data transmission success rate.

[0073] The third processing unit is specifically configured to: divide the marked data into at least two data subsets according to data volume or urgency, and send each data subset to a corresponding proxy intelligent fusion terminal respectively; determine a priority order of each proxy intelligent fusion terminal based on the attribute information; based on the priority order, send the corresponding data subset to each proxy intelligent fusion terminal in turn at every first time interval, wherein the time point of sending data to the proxy intelligent fusion terminal with the highest priority is the first time interval from detecting the offline state, and the time points of sending data to the subsequent proxy intelligent fusion terminals with lower priorities lag by a preset time difference value in turn.

[0074] In an embodiment of the present application, the model updating unit 26 is specifically configured to: classify the received model key control parameters sent by the plurality of target intelligent fusion terminals to obtain weight adjustment parameters, feature mapping parameters and loss function correction parameters; for each type of parameter, determine a differential weight of sending the parameter by different intelligent fusion terminals based on the model contribution of each intelligent fusion terminal in the historical collaborative training; based on the model key control parameters sent by different intelligent fusion terminals and the differential weight corresponding to each type of parameter, weight and fuse the same type of parameters to obtain a fusion result corresponding to each type of parameter respectively; based on the fusion result corresponding to each type of parameter respectively, update the model parameters of the trained agent capability prediction model using the network layer corresponding to the trained agent capability prediction model.

[0075] In an embodiment of the present application, the model updating unit 26 is specifically configured to: perform hierarchical sensitivity checking on the fusion result corresponding to each type of parameter, filter out abnormal parameter components that exceed a parameter sensitivity threshold based on the parameter sensitivity threshold of each network layer of the agent capability prediction model, the parameter sensitivity threshold being determined based on a historical model adjustment influence degree, the historical model adjustment influence degree being an influence degree of parameter adjustment of each network layer on model prediction accuracy in historical training. For abnormal parameter components, a parameter correction factor is generated based on the local interaction data characteristics of the current intelligent fusion terminal in offline state; the parameter correction factor is positively correlated with the distribution deviation of the local interaction data characteristics. By concatenating the abnormal parameter components of the same type after correction based on the parameter correction factor with the normal parameter components that do not exceed the parameter sensitivity threshold, the parameter set to be updated for each network layer is obtained. Based on the parameter set to be updated for each network layer, the model parameters of the trained agent capability prediction model are updated using the network layer corresponding to the trained agent capability prediction model.

[0076] In one embodiment of this application, the parameter adjustment of each network layer during historical training includes the number of parameter adjustments and the adjustment magnitude of the parameters of the l-th layer each time; Model update unit 26 is specifically used to: determine the parameter sensitivity threshold based on the following formula:

[0077] in, The parameter sensitivity threshold of the l-th network layer. These are the weighting coefficients. This refers to the number of parameter adjustments made to the l-th network layer during historical training. Let be the change in model prediction accuracy after the i-th adjustment of the parameters of the l-th layer. Let be the adjustment range of the parameters of the ith layer. denoted as the standard deviation of the parameter adjustment magnitude of the l-th layer network during historical training.

[0078] See Figure 3 , Figure 3 This is a schematic block diagram of an electronic device provided according to an embodiment of this application. Figure 3 The electronic device 300 in this embodiment may include one or more processors 301, one or more input devices 302, one or more output devices 303, and one or more memories 304. The processors 301, input devices 302, output devices 303, and memories 304 communicate with each other via a communication bus 305. The memories 304 store computer programs, including program instructions. The processors 301 execute the program instructions stored in the memories 304. Specifically, the processors 301 are configured to invoke the program instructions to perform the functions of the units in the above-described device embodiments, for example... Figure 2 The functions of the first detection unit 21, the first processing unit 22, the signal sending unit 23, the second detection unit 24, the second processing unit 25, and the model update unit 26 are shown.

[0079] It should be understood that, in the embodiments of the present application, the processor 301 can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor.

[0080] The input device 302 can include a touchpad, a fingerprint collection sensor (for collecting fingerprint information and direction information of a fingerprint of a user), a microphone, etc., and the output device 303 can include a display (LCD, etc.), a speaker, etc.

[0081] The memory 304 can include read-only memory and random access memory, and provide instructions and data for the processor 301. A portion of the memory 304 can also include non-volatile random access memory.

[0082] In specific implementations, the processor 301, the input device 302 and the output device 303 described in the embodiments of the present application can execute the implementation manners described in the interaction method of the multi-intelligent fusion terminal provided by the embodiments of the present application, and can also execute the implementation manners of the electronic device described in the embodiments of the present application, which will not be described here.

[0083] In another embodiment of the present application, a computer readable storage medium is provided, and the computer readable storage medium stores a computer program. The computer program includes program instructions, and the program instructions are executed by a processor to implement all or part of the processes of the above-mentioned embodiment methods. The computer program can also be used to instruct related hardware to complete, and the computer program can be stored in a computer readable storage medium. When the computer program is executed by the processor, the steps of the above-mentioned various method embodiments can be implemented. The computer program includes computer program code, which can be in the form of source code, object code, executable file or some intermediate form, etc. The computer readable medium can include any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal and software distribution medium, etc.

[0084] The computer readable storage medium can be an internal storage unit of the electronic device of any of the preceding embodiments, such as a hard disk or a memory of the electronic device. The computer readable storage medium can also be an external storage device of the electronic device, such as a plug-in hard disk, a Smart Media Card (SMC), a Secure Digital (SD) card, a Flash Card, or the like provided on the electronic device. Further, the computer readable storage medium can include both the internal storage unit and the external storage device of the electronic device. The computer readable storage medium is used to store a computer program and other programs and data required by the electronic device. The computer readable storage medium can also be used to temporarily store data that has been output or will be output.

[0085] Those skilled in the art can understand that the units and algorithm steps of the examples described in combination with the embodiments disclosed herein can be realized by electronic hardware, computer software or a combination of both. In order to clearly illustrate the interchangeability of hardware and software, the components and steps of the examples have been described in general terms in the foregoing description. Whether the functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. A person skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0086] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the electronic device and the units described above can refer to the corresponding processes in the foregoing method embodiments, which will not be described here.

[0087] In several embodiments provided in the present application, it should be understood that the disclosed electronic device and method can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of units is only a logical function division, and actual implementation can have another division manner. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the units shown or discussed can be indirect coupling or communication connection through some interfaces or units, and can also be electrical, mechanical or other forms of connection.

[0088] The units described as separate components can or can not be physically separated, and the components shown as units can or can not be physical units, i.e. can be located in one place or distributed on multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the embodiments of the present application.

[0089] In addition, each of the functional units in each embodiment of the present application can be integrated in one processing unit, or each unit can exist physically, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0090] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of various equivalent modifications or replacements within the technical range disclosed by the present application, and these modifications or replacements should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. An interaction method of a multi-intelligent fusion terminal, characterized in that, The method comprises the following steps: in response to detecting that the network state of the current intelligent fusion terminal is switched from online state to offline state, predicting the agent capability values of the plurality of target intelligent fusion terminals based on the trained agent capability prediction model, wherein the trained agent capability prediction model is obtained by federated learning of the plurality of target intelligent fusion terminals and the master control when the current intelligent fusion terminal is in online state; determining the agent intelligent fusion terminal from the plurality of target intelligent fusion terminals based on the agent capability values; sending a broadcast signal to the outside, wherein the broadcast signal is used to make the agent intelligent fusion terminal send the marker data of the current intelligent fusion terminal to the master control; in response to detecting that the network state of the current intelligent fusion terminal is switched from offline state to online state, sending an agent data request signal to the master control, and receiving the agent data list and control instructions generated by the master control based on the agent data request signal; if the agent data list and control instructions sent by the master control are received, sending a cooperative parameter request signal to the plurality of target intelligent fusion terminals, and receiving the model key control parameters during offline period sent by the plurality of target intelligent fusion terminals, wherein the agent data list and control instructions are generated based on the agent data request signal; sending the unmarked data to the master control based on the agent data list and control instructions, and updating the model parameters of the trained agent capability prediction model based on the model key control parameters.

2. The method of claim 1, wherein, The method further comprises the following steps after determining the agent intelligent fusion terminal from the plurality of target intelligent fusion terminals based on the agent capability values: screening at least one terminal with an agent capability value greater than or equal to a preset agent capability value as an initial candidate set; calculating the physical distance and the distribution area overlap degree between each terminal in the initial candidate set and the current intelligent fusion terminal; determining the agent intelligent fusion terminal from the initial candidate set based on the physical distance and the distribution area overlap degree.

3. The method of claim 1, wherein, The method further comprises the following steps after determining the agent intelligent fusion terminal from the plurality of target intelligent fusion terminals based on the agent capability values: classifying the distribution data of the local area obtained in offline state, and calculating the comprehensive scores corresponding to each type of data based on the emergency degree and timeliness of each type of data; determining the data with a comprehensive score greater than a first threshold value as marker data; sending the marker data to the agent intelligent fusion terminal.

4. The method of claim 3, wherein, The agent intelligent fusion terminal comprises at least two target intelligent fusion terminals; the method further comprises the following steps before sending the marker data to the agent intelligent fusion terminal: receiving attribute information sent by each agent intelligent fusion terminal respectively, wherein the attribute information comprises at least one of network stability, remaining storage capacity and historical data transmission success rate; wherein, the step of sending the marker data to the agent intelligent fusion terminal comprises: dividing the marker data into at least two data subsets according to data volume or emergency degree, and sending each data subset to the corresponding agent intelligent fusion terminal respectively; determining the priority order of each agent intelligent fusion terminal based on the attribute information. Based on the priority ranking, the corresponding data subsets are sequentially sent to each agent intelligent fusion terminal at every first time interval, wherein the time point for sending data to the agent intelligent fusion terminal with the highest priority is the first time interval from detecting the offline state, and the time points for sending data to subsequent priority agent intelligent fusion terminals are sequentially delayed by a preset time difference.

5. The method of claim 1, wherein, The model parameter of the trained agent capability prediction model is updated based on the model key regulation parameters, including: The received model key regulation parameters sent by the plurality of target intelligent fusion terminals are classified to obtain weight adjustment parameters, feature mapping parameters, and loss function correction parameters; For each type of parameter, the differential weights of different intelligent fusion terminals sending the parameters are determined based on the model contribution of each intelligent fusion terminal in the historical collaborative training; The same type of parameters are weighted and fused based on the model key regulation parameters sent by different intelligent fusion terminals and the differential weights corresponding to each type of parameter, to obtain a fusion result corresponding to each type of parameter; The model parameters of the trained agent capability prediction model are updated using the network layer corresponding to the trained agent capability prediction model based on the fusion result corresponding to each type of parameter.

6. The method of claim 5, wherein, The model parameter of the trained agent capability prediction model is updated based on the model key regulation parameters, including: The hierarchical sensitivity of the fusion result corresponding to each type of parameter is checked, and abnormal parameter components that exceed the parameter sensitivity threshold are filtered out based on the parameter sensitivity threshold of each network layer of the agent capability prediction model, the parameter sensitivity threshold is determined based on the historical model adjustment influence degree, and the historical model adjustment influence degree is the influence degree of the parameter adjustment of each network layer on the model prediction accuracy in the historical training; For the abnormal parameter components, a parameter correction factor is generated based on the local interaction data features under the offline state of the current intelligent fusion terminal; the parameter correction factor is positively correlated with the distribution deviation degree of the local interaction data features; The same type of abnormal parameter components corrected based on the parameter correction factor and the normal parameter components that do not exceed the parameter sensitivity threshold are spliced to obtain a set of parameters to be updated for each network layer; The model parameters of the trained agent capability prediction model are updated using the network layer corresponding to the trained agent capability prediction model based on the set of parameters to be updated for each network layer.

7. The method of claim 6, wherein, The parameter adjustment of each network layer in the historical training includes the number of parameter adjustments and the adjustment amplitude of the parameter of the lth layer each time; The parameter sensitivity threshold is determined based on the historical model adjustment influence degree, including: The parameter sensitivity threshold is determined based on the following formula: wherein, is a parameter sensitivity threshold of the l-th network layer, is a weight coefficient, is a number of parameter adjustments for the l-th network layer in the historical training, is a change in prediction accuracy of the model after the i-th adjustment of the l-th layer parameter, is an adjustment amplitude of the i-th adjustment of the l-th layer parameter, is a standard deviation of the parameter adjustment amplitude of the l-th network layer in the historical training.

8. An interactive device of a multi-intelligent fusion terminal, characterized in that, including: The first detection unit is configured to, in response to detecting that the network state of the current intelligent fusion terminal is switched from the online state to the offline state, predict the agent capability values of the plurality of target intelligent fusion terminals based on the trained agent capability prediction model, and the prediction model is obtained through federated learning collaborative training based on the plurality of target intelligent fusion terminals and the master control when the current intelligent fusion terminal is in the online state. The multiple target intelligent fusion terminals are devices associated with the current intelligent fusion terminal; The first processing unit is configured to determine a proxy intelligent fusion terminal from the multiple target intelligent fusion terminals based on proxy capability values of the multiple target intelligent fusion terminals; The signal sending unit is configured to send a broadcast signal to the outside, and the broadcast signal is used to make the proxy intelligent fusion terminal send marker data of the current intelligent fusion terminal to the general control; The second detection unit is configured to send a proxy data request signal to the general control in response to detecting that the network state of the current intelligent fusion terminal is switched from an offline state to an online state, and receive a proxy data list and a control instruction generated by the general control based on the proxy data request signal; The second processing unit is configured to send a cooperation parameter request signal to the multiple target intelligent fusion terminals if the proxy data list and the control instruction sent by the general control are received, and receive model key control parameters during the offline period sent by the multiple target intelligent fusion terminals, wherein the proxy data list and the control instruction are generated based on the proxy data request signal; The model updating unit is configured to send unmarked data to the general control based on the proxy data list and the control instruction, and update model parameters of the trained proxy capability prediction model based on the model key control parameters.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, The processor executes the computer program to implement the steps of the method of any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 9. The computer program is executed by the processor to implement the steps of the method of any one of claims 1 to 7.

Images