Event triggering type distributed neural network working method and system

Abstract

The invention belongs to the technical field of data processing, and discloses an event trigger type distributed neural network working method and system, and the method comprises the steps: selecting a main control processor through a competition voting mechanism; dynamically adjusting the main control processor based on the main control time; dynamically dividing calculation tasks; on the basis of a multi-level redundant backup mechanism, the calculation subtasks are distributed to the corresponding terminal nodes through a task distribution algorithm; the terminal nodes receive the corresponding calculation subtasks, execute the calculation subtasks and reserve intermediate calculation data and calculation results; and screening the calculation results of the terminal nodes according to preset credibility to obtain a final calculation result. The invention provides an event triggering type distributed neural network working mechanism, a main control processor is selected through competitive voting, computing resources are integrated, tasks are divided, equipment with different computing capacities is adapted, and in order to deal with terminal disconnection, multi-level redundant backup and credible level screening are adopted, and it is ensured that a computing result is reliable.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to an event-triggered distributed neural network working method and system. Background Art

[0002] Digital logistics technology refers to the application of information technology and logistics management concepts to optimize and manage the entire logistics process, and realize the digitalization, intelligence and integration of the logistics process. Digital logistics technology realizes the informatization of the logistics process through information technology, including logistics information platform, logistics information collection, logistics information processing, logistics information sharing, etc. Through big data, Internet of Things, cloud computing, neural network and other technical means, it realizes the automation and intelligence of the logistics process, including intelligent loading, intelligent scheduling, intelligent warehousing, intelligent distribution, etc., to improve logistics efficiency, reduce logistics costs, and improve the quality and level of logistics services.

[0003] Digital logistics technology has a wide range of applications, covering all aspects of the logistics process, including procurement, production, warehousing, distribution, etc. It can be applied to various logistics models, including traditional logistics, e-commerce logistics, supply chain logistics, etc. Digital logistics technology can monitor and control the entire logistics process, improve the visibility and controllability of logistics, provide data support and decision-making basis for logistics companies, and promote the transformation, upgrading and intelligent development of the logistics industry.

[0004] However, in traditional logistics solutions, due to technical and cost reasons, it is usually impossible to achieve intelligent data operations on the entire process of transporting goods. That is, existing technologies can often only store key information about goods at certain nodes. With the evolution of digital technology and neural network technology, the computing power of each device can be further utilized to fully explore the value of goods.

[0005] Therefore, how to provide an event-triggered distributed neural network working method and system is a problem that needs to be solved urgently. Summary of the invention

[0006] The embodiments of the present invention provide an event-triggered distributed neural network working method and system to solve the above-mentioned technical problems in the prior art.

[0007] In order to have a basic understanding of some aspects of the disclosed embodiments, a brief summary is given below. This summary is not a general review, nor is it intended to identify key / important components or to describe the scope of protection of these embodiments. Its only purpose is to present some concepts in a simple form as a prelude to the detailed description that follows.

[0008] According to a first aspect of an embodiment of the present invention, an event-triggered distributed neural network working method is provided.

[0009] In one embodiment, the event-triggered distributed neural network working method includes: Based on the neural network training and reasoning task attributes, a competitive voting mechanism is used to select the master control processor within the preset computing network range; the master control time is calculated according to the computing power and task complexity of the master control processor, and the master control processor is dynamically adjusted based on the master control time; The main control processor integrates the current computing resources, dynamically divides the computing tasks according to the computing power requirements and task characteristics, and obtains several computing subtasks; based on the multi-level redundant backup mechanism, the computing subtasks are assigned to the corresponding terminal nodes using the task allocation algorithm; The terminal node receives the corresponding computing subtask, executes the computing subtask and retains the intermediate computing data and computing results; the computing results of the terminal node are screened according to the preset credibility to obtain the final computing result.

[0010] In one embodiment, a competitive voting mechanism is used to select a master processor within a preset computing network range, including: Qualify the terminal nodes within the preset computing network range, and use the terminal nodes that pass the qualification as candidate terminals, and monitor the performance data of the candidate terminals in real time, where the performance data includes computing power, memory usage and network delay; According to the preset evaluation criteria, the score of each candidate terminal is calculated; based on the trust level of the candidate terminal, a corresponding weight is assigned to each candidate terminal, and the candidate terminal participates in the voting according to the score result and the corresponding weight to determine the main control processor.

[0011] In one embodiment, the calculating the master control time according to the computing power and task complexity of the master control processor, and dynamically adjusting the master control processor based on the master control time includes: The master control time is calculated according to the computing power and task complexity of the current master control processor, and the duty time of the current master control processor is monitored in real time. When the duty time exceeds the master control time, the master control processor selection mechanism is re-initiated and a new master control processor is confirmed; The calculation formula of the master control time is:

[0012] In the formula, T Indicates the master time. K represents the safety factor, C represents the complexity of the task, P Indicates the computing power of the main control processor. L Indicates the current load of the main control processor. H represents the historical performance coefficient of the main control processor, D Indicates network latency.

[0013] In one embodiment, the main control processor integrates the current computing resources and dynamically divides the computing tasks according to the computing power requirements and task characteristics of the computing tasks to obtain a number of computing subtasks; based on the multi-level redundant backup mechanism, the computing subtasks are allocated to the corresponding terminal nodes using the task allocation algorithm, including: The main control processor dynamically divides the computing task according to the changes in the current environment and the status information of each terminal node, combined with the computing power requirements and task characteristics of the computing task, and obtains several computing subtasks; according to the priority of the computing subtask and the resource status of the terminal node, the scheduling order of the computing subtask is adjusted; Obtain the status information and resource availability information of each terminal node, and assign corresponding computing subtasks to each terminal node based on the multi-level redundant backup mechanism and task allocation algorithm, combined with the status information and resource storage information of the terminal node.

[0014] In one embodiment, the terminal node receives the corresponding computing subtask, executes the computing subtask and retains the intermediate computing data and computing results; the computing results of the terminal node are screened according to the preset credibility, and the final computing results include: Each terminal node receives the corresponding computing subtask and executes the computing subtask according to the adjusted scheduling order; the main control processor monitors the computing progress and results of each terminal node in real time and retains the intermediate data and results; The main control processor collects and summarizes the calculation results of all terminal nodes. If there is only one calculation result, it will be directly used as the final calculation result. If there are multiple calculation results, the calculation results of different terminal nodes corresponding to the same calculation subtask will be obtained, and the calculation results of different terminal nodes will be analyzed to see whether they are the same. If so, one of the calculation results will be randomly selected as the calculation result of the calculation subtask. If not, the calculation result selected by majority voting will be used as the calculation result of the calculation subtask. Based on the computing power of each terminal node and the distance from the main control processor, a corresponding trust level is set for each terminal node, and the calculation results of all computing subtasks are screened and merged according to the feasible level of each terminal node to obtain the final calculation result.

[0015] According to a second aspect of an embodiment of the present invention, an event-triggered distributed neural network operating system is provided.

[0016] In one embodiment, the event-triggered distributed neural network operating system includes: The master control processor selection module is used to select the master control processor within the preset computing network range based on the neural network training and reasoning task attributes using a competitive voting mechanism; calculate the master control time according to the computing power and task complexity of the master control processor, and dynamically adjust the master control processor based on the master control time; The computing task division and allocation module is used for the main control processor to integrate the current computing resources, dynamically divide the computing tasks according to the computing power requirements and task characteristics of the computing tasks, and obtain several computing subtasks; based on the multi-level redundant backup mechanism, the computing subtasks are allocated to the corresponding terminal nodes using the task allocation algorithm; The calculation result screening and merging module is used for the terminal node to receive the corresponding calculation subtask, execute the calculation subtask and retain the intermediate calculation data and calculation results; the calculation results of the terminal node are screened according to the preset credibility to obtain the final calculation result.

[0017] In one embodiment, when the master processor selection module selects the master processor within the preset computing network range using a competitive voting mechanism, it performs qualification authentication on the terminal nodes within the preset computing network range, and uses the terminal nodes that have passed the authentication as candidate terminals, and monitors the performance data of the candidate terminals in real time, wherein the performance data includes computing power, memory usage, and network delay status; calculates the score of each candidate terminal according to preset evaluation criteria; and assigns a corresponding weight to each candidate terminal based on the trust level of the candidate terminal, and the candidate terminals participate in voting according to the scoring results and the corresponding weights to determine the master processor.

[0018] In one embodiment, the master processor selection module calculates the master control time according to the computing power and task complexity of the master processor and dynamically adjusts the master processor based on the master control time. It calculates the master control time according to the computing power and task complexity of the current master processor and monitors the current master processor's duty time in real time. When the duty time exceeds the master control time, the master processor selection mechanism is re-initiated and a new master processor is confirmed. The calculation formula of the master control time is:

[0019] In the formula, T Indicates the master time. K represents the safety factor, C represents the complexity of the task, P Indicates the computing power of the main control processor. L Indicates the current load of the main control processor. H represents the historical performance coefficient of the main control processor, D Indicates network latency.

[0020] In one embodiment, the computing task partitioning and allocating module includes a computing task partitioning module and a computing task allocating module; The computing task division module is used for the main control processor to dynamically divide the computing task according to the changes in the current environment and the status information of each terminal node, combined with the computing power requirements and task characteristics of the computing task, to obtain a number of computing subtasks; according to the priority of the computing subtask and the resource status of the terminal node, the scheduling order of the computing subtask is adjusted; The computing task allocation module is used to obtain the status information and resource availability information of each terminal node, and allocate corresponding computing subtasks to each terminal node based on the multi-level redundant backup mechanism and task allocation algorithm in combination with the status information and resource storage information of the terminal node.

[0021] In one embodiment, the calculation result screening and merging module includes a calculation subtask execution module, a calculation result summary and analysis module, and a calculation result screening and output module; The computing subtask execution module is used for each terminal node to receive the corresponding computing subtask and execute the computing subtask according to the adjusted scheduling order; the main control processor monitors the computing progress and computing results of each terminal node in real time and retains the intermediate data and results; The calculation result summary and analysis module is used for the main control processor to collect and summarize the calculation results of all terminal nodes. If there is only one calculation result, the calculation result is directly used as the final calculation result. If there are multiple calculation results, the calculation results of different terminal nodes corresponding to the same calculation subtask are obtained, and the calculation results of different terminal nodes are analyzed to see whether they are the same. If so, one of the calculation results is randomly selected as the calculation result of the calculation subtask. If not, the calculation result selected by majority voting is used as the calculation result of the calculation subtask. The calculation result screening output module sets a corresponding trust level for each terminal node based on the computing power of each terminal node and the distance from the main control processor, and screens and merges the calculation results of all calculation subtasks according to the feasible level of each terminal node to obtain the final calculation result.

[0022] According to a third aspect of an embodiment of the present invention, a computer device is provided.

[0023] In some embodiments, the computer device includes a memory and a processor, the memory stores a computer program, and the processor implements the steps of the above method when executing the computer program.

[0024] According to a fourth aspect of embodiments of the present invention, a computer-readable storage medium is provided.

[0025] In one embodiment, the computer-readable storage medium stores a computer program, and the computer program implements the steps of the above method when executed by a processor.

[0026] The technical solution provided by the embodiment of the present invention may have the following beneficial effects: The present invention proposes an event-triggered distributed neural network working mechanism. Based on the attributes of neural network training and reasoning tasks, in an environment with digital interconnection functions, a master processor is selected through an innovative competitive voting mechanism, and the master processor integrates the current computing resources and divides the computing tasks to adapt to different computing power facilities, and retains the intermediate data and results of the calculation. Furthermore, since small neural network computing terminals may be offline, a multi-level redundant backup mechanism is used when assigning tasks to them, that is, multiple terminals calculate the same neural network task and screen the final results by setting a trust level, so as to more completely ensure the availability of the final calculation results. The present invention can be deployed in parks, roads, vehicles, personnel and other devices that support digital logistics.

[0027] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0029] Figure 1 is a flow chart of an event-triggered distributed neural network working method according to an exemplary embodiment; Figure 2 is a schematic diagram of a distributed neural network working mechanism in an event-triggered distributed neural network working method according to an exemplary embodiment; Figure 3 is a structural block diagram of an event-triggered distributed neural network working system according to an exemplary embodiment; Figure 4 The figure is a schematic diagram showing the structure of a computer device according to an exemplary embodiment. DETAILED DESCRIPTION

[0030] The following description and accompanying drawings fully illustrate the specific embodiments of this article so that those skilled in the art can practice them. Parts and features of some embodiments may be included in or replace parts and features of other embodiments. The scope of the embodiments of this article includes the entire scope of the claims, as well as all available equivalents of the claims. Herein, the terms "first", "second", etc. are only used to distinguish one element from another, without requiring or implying any actual relationship or order between these elements. In fact, the first element can also be called the second element, and vice versa. Moreover, the terms "include", "comprise" or any other variant thereof are intended to cover non-exclusive inclusion, so that the structure, device or equipment including a series of elements includes not only those elements, but also other elements that are not explicitly listed, or also include elements inherent to such structure, device or equipment. In the absence of more restrictions, the elements defined by the sentence "including one..." do not exclude the existence of other identical elements in the structure, device or equipment including the elements. Each embodiment is described in a progressive manner herein, and each embodiment focuses on the differences from other embodiments, and the same and similar parts between the embodiments can be referred to each other.

[0031] The terms "longitudinal", "lateral", "upper", "lower", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc. in this document indicate the orientation or position relationship based on the orientation or position relationship shown in the drawings, and are only for the convenience of describing this document and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operate in a specific orientation, and therefore cannot be understood as a limitation on the present invention. In the description of this document, unless otherwise specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a mechanical connection or an electrical connection, it can also be the internal communication of two elements, it can be a direct connection, or it can be an indirect connection through an intermediate medium. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0032] As used herein, the term "plurality" means two or more than two, unless otherwise specified.

[0033] In this document, the character " / " indicates that the preceding and following objects are in an "or" relationship. For example, A / B means: A or B.

[0034] In this article, the term "and / or" is a description of the association relationship between objects, indicating that three relationships may exist. For example, A and / or B means: A or B, or, A and B.

[0035] It should be understood that, although the various steps in the flow chart are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear description in this article, the execution of these steps is not strictly limited in order, and these steps can be executed in other orders. Moreover, at least a portion of the steps in the figure may include multiple sub-steps or multiple stages, and these sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these sub-steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a portion of the sub-steps or stages of other steps.

[0036] Each module in the device or system of the present application can be implemented in whole or in part by software, hardware, or a combination thereof. The above modules can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute operations corresponding to the above modules.

[0037] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0038] Figure 1 An embodiment of an event-triggered distributed neural network working method of the present invention is shown.

[0039] In this optional embodiment, the event-triggered distributed neural network working method includes: Step S101: Based on the neural network training and reasoning task attributes, a competitive voting mechanism is used to select a master control processor within a preset computing network range; the master control time is calculated according to the computing power and task complexity of the master control processor, and the master control processor is dynamically adjusted based on the master control time; Step S103: The main control processor integrates the current computing resources, dynamically divides the computing tasks according to the computing power requirements and task characteristics of the computing tasks, and obtains a number of computing subtasks; based on the multi-level redundant backup mechanism, the computing subtasks are assigned to the corresponding terminal nodes using the task allocation algorithm; Step S105: The terminal node receives the corresponding computing subtask, executes the computing subtask and retains the intermediate computing data and computing results; the computing results of the terminal node are screened according to the preset credibility to obtain the final computing results.

[0040] In this optional embodiment, a competitive voting mechanism is used to select a master processor within a preset computing network range, including: Qualify the terminal nodes within the preset computing network range, and use the terminal nodes that pass the qualification as candidate terminals, and monitor the performance data of the candidate terminals in real time, where the performance data includes computing power, memory usage and network delay; According to the preset evaluation criteria, the score of each candidate terminal is calculated; based on the trust level of the candidate terminal, a corresponding weight is assigned to each candidate terminal, and the candidate terminal participates in the voting according to the score result and the corresponding weight to determine the main control processor.

[0041] In this optional embodiment, the calculating the master control time according to the computing power of the master control processor and the task complexity, and dynamically adjusting the master control processor based on the master control time includes: The master control time is calculated according to the computing power and task complexity of the current master control processor, and the duty time of the current master control processor is monitored in real time. When the duty time exceeds the master control time, the master control processor selection mechanism is re-initiated and a new master control processor is confirmed; In this optional embodiment, the main control processor integrates the current computing resources, and dynamically divides the computing tasks according to the computing power requirements and task characteristics of the computing tasks to obtain a number of computing subtasks; based on the multi-level redundant backup mechanism, the computing subtasks are assigned to the corresponding terminal nodes using the task allocation algorithm, including: The main control processor dynamically divides the computing task according to the changes in the current environment and the status information of each terminal node, combined with the computing power requirements and task characteristics of the computing task, and obtains several computing subtasks; according to the priority of the computing subtask and the resource status of the terminal node, the scheduling order of the computing subtask is adjusted; Obtain the status information and resource availability information of each terminal node, and assign corresponding computing subtasks to each terminal node based on the multi-level redundant backup mechanism and task allocation algorithm, combined with the status information and resource storage information of the terminal node.

[0042] In this optional embodiment, the terminal node receives the corresponding computing subtask, executes the computing subtask and retains the intermediate computing data and computing results; the computing results of the terminal node are screened according to the preset credibility, and the final computing results include: Each terminal node receives the corresponding computing subtask and executes the computing subtask according to the adjusted scheduling order; the main control processor monitors the computing progress and results of each terminal node in real time and retains the intermediate data and results; The main control processor collects and summarizes the calculation results of all terminal nodes. If there is only one calculation result, it will be directly used as the final calculation result. If there are multiple calculation results, the calculation results of different terminal nodes corresponding to the same calculation subtask will be obtained, and the calculation results of different terminal nodes will be analyzed to see whether they are the same. If so, one of the calculation results will be randomly selected as the calculation result of the calculation subtask. If not, the calculation result selected by majority voting will be used as the calculation result of the calculation subtask. Based on the computing power of each terminal node and the distance from the main control processor, a corresponding trust level is set for each terminal node, and the calculation results of all computing subtasks are screened and merged according to the feasible level of each terminal node to obtain the final calculation result.

[0043] Figure 3 An embodiment of an event-triggered distributed neural network operating system of the present invention is shown.

[0044] In this optional embodiment, the event-triggered distributed neural network operating system includes: The master processor selection module 201 is used to select the master processor within the preset computing network range by using a competitive voting mechanism based on the neural network training and reasoning task attributes; calculate the master control time according to the computing power and task complexity of the master control processor, and dynamically adjust the master control processor based on the master control time; The computing task division and allocation module 203 is used for the main control processor to integrate the current computing resources, dynamically divide the computing tasks according to the computing power requirements and task characteristics of the computing tasks, and obtain a number of computing subtasks; based on the multi-level redundant backup mechanism, the computing subtasks are allocated to the corresponding terminal nodes using the task allocation algorithm; The calculation result screening and merging module 205 is used for the terminal node to receive the corresponding calculation subtask, execute the calculation subtask and retain the intermediate calculation data and calculation results; screen the calculation results of the terminal node according to the preset credibility to obtain the final calculation result.

[0045] In order to facilitate understanding of the above technical solutions of the present invention, the above technical solutions of the present invention are further described below from the perspective of architecture and principle, as follows: The meaning of the event-triggered distributed neural network working method is as follows: Event-triggered: "Event-triggered" means that the working mechanism starts or executes related computing tasks when certain specific events occur. Under this mechanism, the system does not need to run continuously, but triggers computing based on external or internal events (such as data arrival, state changes, resource availability, etc.). This design can effectively improve resource utilization efficiency, reduce unnecessary computing, and reduce energy consumption. For example, when a sensor detects specific conditions (such as temperature changes, traffic flow changes, etc.), the system will immediately start the corresponding neural network task for analysis and reasoning. This event-driven approach makes the system more flexible, responsive, and able to adapt to dynamically changing environments.

[0046] Distributed neural network working mechanism: "Distributed neural network working mechanism" means that the computing tasks of this solution are carried out in parallel on multiple different computing nodes. These nodes can be terminal devices with different computing capabilities (such as edge computing devices, cloud servers, etc.), which work together through the network, such as Figure 2 As shown in the figure. Under this mechanism, the master processor is responsible for integrating and managing the computing resources of each node, and reasonably divides the tasks to different nodes to adapt to their computing capabilities. For example, nodes with stronger computing capabilities can undertake more complex tasks, while nodes with weaker computing capabilities handle simple calculations. This distributed approach can effectively improve computing efficiency and the overall performance of the system.

[0047] The present invention proposes an event-triggered distributed neural network working mechanism. Based on the attributes of neural network training and reasoning tasks, in an environment with digital interconnection functions, a master processor is selected through an innovative competitive voting mechanism, and the master processor integrates the current computing resources and divides the computing tasks to adapt to different computing power facilities, and retains the intermediate data and results of the calculation. Furthermore, since small neural network computing terminals may be offline, a multi-level redundant backup mechanism is used when assigning tasks to them, that is, multiple terminals calculate the same neural network task and screen the final results by setting a trust level, so as to more completely ensure the availability of the final calculation results. It can be deployed in parks, roads, vehicles, personnel and other devices that support digital logistics.

[0048] In the present invention, the triggering conditions of the "event-triggered distributed neural network working mechanism" refer to events or changes that occur in a specific environment. These events can be: environmental changes (such as changes in sensor data, changes in user behavior, etc.), task requirements (such as the arrival of new computing tasks, changes in the status of existing tasks, etc.) and computing resource status (such as changes in terminal computing power, equipment online or offline, etc.); These events activate the training and reasoning tasks of the neural network, prompting the system to respond accordingly.

[0049] "Neural network training and reasoning task attributes" specifically include the following aspects: Task Type: Training tasks: involve adjusting the parameters of a neural network to improve the model's predictive power for a specific dataset.

[0050] Reasoning task: Use the trained neural network to predict or classify new data.

[0051] Data characteristics: Dimension and nature of input data: Different tasks may require different forms of data input.

[0052] Computational complexity: Training tasks are usually more complex than inference tasks and require higher computing resources.

[0053] Real-time requirements: Different tasks have different real-time requirements. Some tasks may require a quick response, while other tasks can accept a longer processing time.

[0054] Resource allocation strategy: Depending on the attributes of the task, the system may choose different resource allocation strategies. For example, training tasks may use high-performance computing resources intensively, while inference tasks may be distributed to multiple terminals.

[0055] 1. Competitive voting mechanism The competitive voting mechanism means that the various devices covered by the present invention can report and broadcast their own spare computing power and device types under network coverage. When reporting and broadcasting information, they must clearly mark this capability for competitive needs, and select the strongest computing power device within a certain range of the network as the main control processor. Low computing power terminals (i.e., terminal nodes) automatically obey the computing task allocation of high computing power terminals. After completing the selection of the main control processor, the main control processor integrates the current computing resources and divides the computing tasks to adapt to different computing power facilities, and retains the intermediate computing data and results. Typical computing tasks include logistics cargo identification, path planning, autonomous driving decision-making and other tasks. The specific operation is to execute the relevant artificial neural network on the corresponding processor to obtain the calculation results and compare them with the preset data.

[0056] 1) Overview of competitive voting mechanism: The competitive voting mechanism is a dynamic decision-making method for selecting the master processor. In a distributed neural network environment, since there are multiple terminals with computing capabilities, these terminals can participate in the selection process of the master processor. Through this mechanism, the system can evaluate each candidate terminal according to certain standards and rules to ensure that the selected master processor has the best performance in terms of computing power, resource management capabilities, and network environment adaptability.

[0057] a. Real-time data monitoring: The system collects performance data of each terminal in real time through monitoring tools and sensors, including computing power, memory usage, network latency, etc. These data provide the basis for evaluation and selection.

[0058] b. Multi-dimensional evaluation model: Based on the collected data, the system uses a multi-dimensional evaluation model to score each terminal. These scores not only reflect the current performance status, but also combine historical performance to ensure the comprehensiveness and accuracy of the selection process.

[0059] c. Trigger and decision-making mechanism: The system sets a series of trigger conditions, such as excessive load, slow node response, etc. Once these conditions are met, the system will start the re-selection process. This mechanism can ensure that the system can respond quickly in emergencies and ensure the continuity and efficiency of tasks.

[0060] 2) Contents of the competitive voting mechanism: a. Dynamic selection and adaptability: Traditional master processor selection is often static and difficult to cope with environmental changes. The competitive voting mechanism has the ability to dynamically select and respond to changes in system status in real time. For example, when the load of a terminal decreases or its computing power increases, the system can re-vote and select according to the current computing needs to adapt to the new environment.

[0061] b. Multi-dimensional evaluation criteria: This mechanism not only relies on computing power, but also takes into account a variety of factors, such as network latency, energy consumption, load balancing, historical performance, etc., and scores each candidate terminal through multi-dimensional evaluation. This diversified evaluation method can ensure that the selected main control processor has advantages in comprehensive performance.

[0062] c. Decentralized decision-making: The competitive voting mechanism makes decisions in a decentralized manner, avoiding the bottleneck of a single control point. All participating terminals can participate in voting according to their own status, which not only improves the fault tolerance of the system, but also enhances the overall reliability.

[0063] d. Trust level and weight allocation: During the voting process, the system can assign different weights to each terminal according to its trust level. Terminals with higher historical performance and resource management capabilities have a greater say in the voting. This strategy can not only improve the selection quality of the main control processor, but also motivate each terminal to continuously optimize its own performance.

[0064] 3) Voting process: a. Qualification authentication: Before the selection, the system first authenticates the qualifications of all candidate terminals to ensure that they are capable of participating in the voting, so as to avoid the voting being affected by low-performance or unstable terminals.

[0065] b. Grading evaluation: By monitoring the status of each terminal in real time, the system conducts a grading evaluation for each candidate terminal and assigns scores based on multiple evaluation criteria.

[0066] The calculation formula of terminal score is as follows:

[0067] In the formula, S i Indicates candidate terminal i Total score of W f Indicates the weight of the candidate terminal in terms of computing power, that is, the influence of computing power on the total score; F i Indicates candidate terminal i The computing power score is usually evaluated based on the terminal's hardware configuration, processing speed, and available resources; W r Indicates the weight of the candidate terminal in terms of reliability, that is, the influence of reliability on the total score; R i Indicates candidate terminal i The reliability score is usually evaluated based on historical mission completion, failure rate and other data; W t represents the weight of the candidate terminal in terms of response time, that is, the impact of response time on the total score, and W f + W r + W t =1; T i Indicates candidate terminal i The response time score is usually evaluated based on indicators such as task processing delay and network delay.

[0068] c. Voting execution: All eligible terminals participate in the voting according to the evaluation results to select the most qualified master processor. The voting results can be calculated through a weighted algorithm to ensure that the terminal finally selected has the best comprehensive capabilities.

[0069] In addition, after determining the score of each candidate terminal, a corresponding weight may be assigned to each candidate terminal based on its trust level. For example, the trust level of the candidate terminal may be set to A , adjust the final voting weight according to this level V i = S i · A i ,Finally, the master processor is determined through weighted voting.

[0070] d. Result confirmation and execution: After the voting results are confirmed, the system will immediately activate the selected master processor and start to be responsible for resource integration and task scheduling. During the execution process, the system will continue to monitor the performance of the master processor so that it can vote again when necessary.

[0071] The present invention further proposes a rotating master processor mechanism, that is, setting a master control time T for the master processor at the current moment, and the master processor only serves as the master processor within the master control time T, and re-initiates the selection mechanism beyond the master control time T. The advantage of such a setting is that it avoids the increase in computing power redundancy and data loss risk caused by a single master processor. At the same time, it is further proposed that the master control time T is calculated by the computing power of the master processor.

[0072] The master time T is dynamically calculated based on the computing power and task complexity of the master processor in the rotating master processor mechanism. This process involves the evaluation of multiple factors, such as processor performance, task load, network latency, etc. The following is a detailed expansion of the calculation process of the master time 𝑇 and the corresponding calculation formula.

[0073] Master time T The calculation formula is:

[0074] In the formula, K Represents the safety factor, which can be set based on experience to cope with emergencies, usually between 1.2 and 2.0; C Indicates the task complexity, that is, the amount of computation required to complete the task, which can be estimated by the amount of input data of the task, the complexity of the algorithm, etc., and is usually expressed in the number of computational cycles required to run; P Indicates the computing power of the main control processor, usually expressed in the number of calculations or floating-point operations per second (FLOPS); L Indicates the current load of the main control processor, that is, the current utilization rate of the main control processor or the load of other processors. The value range is 0 (no load) to 1 (full load). This is a key factor. When the load is high, the response speed of the processor may decrease. H Indicates the historical performance coefficient of the main control processor, which can be obtained by analyzing the past execution performance of the system. It is usually greater than 0 to reflect the past execution efficiency; D Represents network latency, taking into account the time for communication and data transfer.

[0075] The calculation process example is as follows: Assume the following parameters: Processor power P = 5 × 10 8 FLOPS; task complexity C = 1×10 8 Calculation cycle; current load L = 0.3 (30% load); historical performance coefficient H = 1.2; network delay D = 0.05 seconds; safety factor K = 1.5; Substitute the parameters into the formula to calculate that T is approximately equal to 0.5643 seconds.

[0076] In addition, in practical applications, the master control time T can also be dynamically adjusted according to the real-time load and task characteristics of the system. This can be achieved through the following mechanisms: Real-time monitoring: During the operation of the main control processor, its computing power and task execution status are monitored in real time, especially in the early stage of task execution, and the values ​​of C and L are adjusted according to real-time feedback.

[0077] Feedback loop: After each master control time ends, the actual execution time and effect of the task are evaluated to adjust the parameters for the next calculation of T.

[0078] 2. Task division and resource integration The main control processor dynamically divides the computing tasks according to the changes in the current environment and the status information of each node to ensure the load balancing of each node and the rational use of resources. In a distributed computing environment, the computing power, storage resources and network bandwidth of different nodes may vary. Through effective task division and resource integration, it is possible to avoid the situation where some nodes are overloaded while other nodes are idle. This not only improves the resource utilization of the system, but also significantly improves the speed and efficiency of task processing, ensuring that the system can still run stably under high load conditions.

[0079] 1) Implementation mechanism of task division a. Real-time status monitoring: The main control processor needs to monitor the status of each node in real time, including computing power, load conditions, network delay, etc. By collecting this data, the system can evaluate the current capabilities of each node.

[0080] b. Dynamic task partitioning strategy: Based on the status monitoring data, the master processor formulates a dynamic task partitioning strategy. Tasks can be appropriately allocated based on their characteristics (such as computationally intensive or I / O intensive) and the capabilities of the nodes. For example, if a node has strong computing power, more complex computing tasks can be assigned to that node, while simple tasks can be assigned to other nodes.

[0081] c. Task allocation algorithm: The specific method of implementing task division can be achieved through a variety of algorithms, such as load balancing algorithm, shortest task first, minimum cost first, etc. These algorithms can be adjusted according to the real-time node status to ensure that the load of each node is relatively balanced.

[0082] 2) Implementation mechanism of resource integration a. Resource availability monitoring: During the dynamic task partitioning process, the master processor also needs to monitor the resource availability of each node, including CPU, memory, and storage, etc. This helps the system understand which nodes have sufficient resources at a specific moment and can take on additional tasks.

[0083] b. Resource scheduling strategy: The effectiveness of resource integration depends on a reasonable scheduling strategy. The master processor can adjust the scheduling order of tasks according to the priority of the tasks and the resource status of the nodes to ensure that high-priority tasks are processed first. At the same time, in the case of tight resources, non-critical tasks can be delayed or rescheduled.

[0084] c. Redundancy and fault tolerance mechanism: Redundancy and fault tolerance mechanism are equally important in task division and resource integration. The master processor can replicate tasks or data between multiple nodes to prevent a node from failing. This redundant design enhances the reliability of the system and ensures that other nodes can take over tasks in time when a node fails.

[0085] 3) Implementation of load balancing a. Load evaluation: The master processor continuously tracks the performance indicators of each node and evaluates the load in real time. Load evaluation can be based on multiple dimensions such as task completion time and resource utilization rate for comprehensive analysis.

[0086] b. Dynamic adjustment: Once it is found that the load of a certain node is significantly higher than that of other nodes, the main control processor can dynamically adjust the task allocation and even re-divide the assigned tasks. This flexibility ensures the maximum utilization of resources and avoids the generation of bottlenecks.

[0087] 3. Multi-level redundant backup mechanism In order to deal with the risk of disconnection of small neural network computing terminals, the system adopts a multi-level redundant backup mechanism when allocating tasks. Multiple terminals perform the same task at the same time, and screen the final results according to the set trust level, thereby ensuring the reliability of the final calculation results.

[0088] In a distributed system, the disconnection of a computing terminal may cause the interruption of computing tasks, thus affecting the overall performance and reliability of the system. By implementing a multi-level redundant backup mechanism, the system can maintain the continuity of tasks and the effectiveness of results when some nodes fail. This is especially important for application scenarios that require high reliability and real-time performance (such as financial analysis, medical diagnosis, etc.).

[0089] 1) Implementation of multi-level redundant backup mechanism a. Parallel distribution of tasks: In the multi-level redundant backup mechanism, the master processor will distribute each task to multiple terminals. The specific degree of parallelism can be determined according to the complexity of the task and the resource status of the system. For example, for a computationally intensive task, it can be distributed to 3 to 5 terminals for parallel execution to ensure sufficient redundancy.

[0090] b. Trust level assessment: When each terminal performs a task, it will be assigned a trust level based on its performance indicators (such as computing power, historical success rate, etc.). This level will affect the screening of the final results. For example, if some terminals have historically shown higher stability and accuracy, their results will be given a higher weight in the final screening.

[0091] c. Summarization and screening of results: When all terminals have completed their tasks, the system will summarize the results returned by each terminal. Through the set trust level, the system uses a variety of algorithms (such as voting mechanism, weighted average, etc.) to screen the results. The final results will be selected based on this evaluation mechanism to ensure their correctness and reliability.

[0092] 2) Fault tolerance and recovery mechanism a. Node detection and failure processing: The system needs to monitor the status of each terminal in real time. Once a node is found to be offline or its returned results are beyond the reasonable range, the system will automatically exclude the results of the node and rely on the results of other valid nodes again. In this way, even if a node is offline, the system can still calculate through the results of other normal nodes.

[0093] b. Dynamically adjust the redundancy level: During task execution, the system can dynamically adjust the redundancy level according to changes in the current environment (such as node performance fluctuations, network conditions, etc.). For example, if the performance of a node drops significantly, the system can temporarily assign more tasks to other nodes with high credibility to ensure that the overall computing power is not affected.

[0094] Application examples: Consider a real-time image processing system that needs to analyze video streams from multiple cameras. In this case, in order to ensure the reliability of the analysis results, the system can adopt a multi-level redundant backup mechanism: Task allocation: The data analysis task of each camera is assigned to 3 different image processing nodes.

[0095] Trust Level: Based on historical analysis success rate and processing speed, these nodes are assigned different trust levels.

[0096] Result summary: After all nodes are processed, the system will vote or weight the analysis results according to the trust level of the nodes and select the final analysis result.

[0097] This mechanism ensures that even if a processing node fails to complete a task due to a failure, the system can still make decisions based on the results of other nodes, thereby improving the reliability of the overall system.

[0098] The multi-level redundant backup mechanism means that when the main control processor is performing calculations, there may be a small neural network computing power terminal that is offline, resulting in the unavailability of the calculation results. Therefore, when assigning computing tasks, the same computing tasks can be configured for multiple small computing power terminals to ensure that the final calculation results are obtained. At the same time, when multiple terminals calculate and process tasks, if they all get the processing results, it is necessary to screen the final results based on the set trust level, that is, set a higher trust level for the device with strong computing power and closer to the main control processor, and vice versa. When multiple terminals calculate and process tasks, only one gets the processing result. At this time, this processing result can be directly used as the final processing result.

[0099] The specific workflow of the present invention is as follows: 1. After the digital logistics network is deployed, it usually includes multiple available networks in a park, which are integrated at the top level, and all networks are connected to the Internet for unified data management and computing power allocation.

[0100] 2. Check and broadcast the computing power of all terminals in the computing network. All types of devices can report and broadcast their own spare computing power and device types under network coverage. When reporting and broadcasting information, they must clearly mark this capacity for competitive needs and select the strongest computing power device as the main control processor within a certain range of the network. Low computing power terminals will automatically obey the computing task allocation of high computing power terminals. Specifically: This solution is a computing power coordination solution based on a computing network. Its workflow includes the following steps: Perform a computing power inventory on all terminals in the computing network: By performing computing power detection on all terminals in the computing network, information such as the available computing power and device type of each terminal can be obtained.

[0101] Broadcast computing power information: All types of devices can report and broadcast their own free computing power and device type information under network coverage, so that other devices can choose the optimal computing power resources.

[0102] Competition demand identification: When reporting and broadcasting information, it is necessary to clearly identify this capability as being used for competition demand so that other devices are aware of its availability.

[0103] Select the device with the strongest computing power: Within a certain range of the network, by comparing and screening the computing power information of all devices, the device with the strongest computing power is selected as the main control processor.

[0104] Task allocation for low-computing-power terminals: Under the command of the main control processor, low-computing-power terminals automatically obey the computing task allocation of high-computing-power terminals, thereby achieving computing power coordination.

[0105] Through this method of computing power coordination, the computing power resources in the computing network can be utilized to the maximum extent, improving computing efficiency and reducing computing costs.

[0106] This solution can be used to build a distributed computing network. All terminals of the computing network will be included in it. Various types of devices can report their computing capabilities and types. The main control processor will select the device with the strongest computing power as the main control processor by broadcasting information. Low-computing power terminals will obey the computing task allocation of high-computing power terminals, so that when the computing tasks are executed, the computing power resources of each terminal in the network can be fully utilized, thereby improving computing efficiency. At the same time, in the process of broadcasting and reporting information, each terminal needs to clearly identify this capability for competitive needs to ensure the fairness and transparency of computing tasks.

[0107] Benefits of the program include: Maximize the utilization of resources in the computing network: Through computing power inventory and broadcasting, all terminal resources in the computing network can be better utilized to achieve optimal computing task allocation and utilization.

[0108] Improve computing efficiency: By selecting the most powerful computing device as the main control processor, the efficiency of computing task allocation can be guaranteed, thereby improving computing efficiency.

[0109] Cost savings: In computing networks, resource allocation can be more balanced and reasonable, avoiding some devices from being idle or overused, thus saving costs.

[0110] Realize computing power coordination: Through computing power inventory and broadcasting, various terminals can work together to achieve more efficient computing power coordination, thereby realizing more efficient computing tasks.

[0111] Adapt to changing environments: In computing networks, various terminal devices and network environments are unstable and changeable. Through computing power inventory and broadcasting, the allocation strategy can be adjusted in time to adapt to the ever-changing environment.

[0112] In addition, the present invention also includes the end and rotation of the master control time T: End of master control time T: When the master control time T reaches the preset end point, the master control processor will stop its master control function and start the final finishing and verification of the task.

[0113] Result output: The main control processor integrates the results of each processor and outputs the final result to ensure the integrity and accuracy of the data.

[0114] Initiate a new selection mechanism: When the master control time T ends, the system automatically initiates a new master control processor selection and enters the next cycle.

[0115] Fault detection: If any processor is detected to be faulty within the master control time T of the master processor, the master processor will immediately start the redundant backup mechanism and call other processors on standby.

[0116] Reassign tasks: The affected tasks will be reassigned to other available processors, ensuring that the tasks can continue without being affected.

[0117] Data backup and recovery: In any case, the master processor will keep a backup of the intermediate results to prevent the loss of important data in the event of a switch or failure.

[0118] 3. After the selection of the master processor is completed, the master processor integrates the current computing resources and divides the computing tasks to adapt to different computing power facilities, and retains the intermediate computing data and results. Typical computing tasks include logistics cargo identification, path planning, and autonomous driving decision-making. The specific operation is to execute the relevant artificial neural network on the corresponding processor to obtain the calculation results and compare them with the preset data.

[0119] After completing the selection of the master processor, the master processor will summarize the terminal computing power information received and evaluate and sort the available computing power.

[0120] Then, the main control processor will divide the tasks according to the computing power requirements and task characteristics, and assign the tasks to the terminal devices. The low computing power terminals will receive the task assignment and start to execute their own tasks.

[0121] The main control processor monitors the calculation progress and results of each terminal device and retains the intermediate data and results.

[0122] Once a terminal device completes its computing task, the main control processor will collect the computing results of the device and aggregate the results to generate the final computing result.

[0123] During the entire process, the main control processor will reallocate tasks as needed to fully utilize computing resources and improve computing efficiency.

[0124] 4. There are small neural network computing terminals that are disconnected, resulting in unavailable calculation results. Therefore, when allocating computing tasks, the same computing tasks can be configured for multiple small computing terminals to ensure that the final calculation results are obtained.

[0125] This solution can indeed improve the fault tolerance and reliability of computing tasks and avoid the unavailability of computing results due to the disconnection of a single small computing power terminal.

[0126] When allocating the same computing task to multiple small computing power terminals, the computing task can be split into multiple subtasks, and the subtasks can be assigned to different small computing power terminals for calculation. Finally, the calculation results of the subtasks can be summarized and integrated to obtain the final calculation result.

[0127] In this way, even if a small computing power terminal is disconnected or fails, other terminals can still continue to complete the computing tasks, thus ensuring the reliability and stability of the entire computing process.

[0128] It should be noted that when allocating computing tasks, the performance and computing power of small computing terminals need to be taken into account to avoid the situation where computing tasks are overly concentrated on a certain terminal, resulting in insufficient utilization of resources of other terminals.

[0129] Configuring the same computing task for multiple small computing power terminals is also called task redundancy. Its purpose is to improve the reliability and availability of the computing results. When assigning tasks, the same task can be assigned to multiple small computing power terminals, and their calculation results can be compared. If the calculation results of multiple terminals are the same, it means that the calculation results are reliable and their results can be used as the final results. However, if there are different calculation results, further processing is required to obtain the final result.

[0130] A common method is to use majority voting, which selects the result with the most occurrences as the final result. For example, if there are three small computing power terminals calculating the same task, two of which get the same result, while the other terminal gets a different result, then the same result can be selected as the final result because it appears the most times. If the number of occurrences is the same, a result can be randomly selected as the final result.

[0131] In addition, in order to ensure the reliability of the calculation results, backup and synchronization can be performed between multiple small computing power terminals. For example, the calculation results can be stored on multiple terminals and synchronized regularly to ensure that even if one of the terminals fails, it will not affect the availability of the calculation results. This method of backup and synchronization is also called data redundancy.

[0132] 5. When multiple terminals calculate small processing tasks, if they all get processing results, it is necessary to screen the final results based on the set trust level, that is, set a higher trust level for devices with strong computing power and closer to the main control processor, and vice versa.

[0133] After multiple terminals complete small processing tasks, if each terminal obtains the processing results, these results need to be screened and merged to obtain the final calculation results. In order to ensure the reliability and accuracy of the calculation results, each terminal needs to be evaluated and assigned a credibility.

[0134] Among them, the reliability of the device closer to the main control processor is higher, because the communication delay between it and the main control processor is lower, and the data transmission is more stable and reliable. Therefore, a higher credibility level can be set for these terminals. Conversely, the credibility of the terminal farther away from the main control processor is lower, because the communication delay between it and the main control processor is higher, and the data transmission is more susceptible to factors such as network fluctuations. Therefore, a lower credibility level can be set for these terminals.

[0135] According to the credibility level of the terminal, different merging strategies can be used to obtain the final calculation results. For example, for terminals with higher credibility, their calculation results can be simply weighted averaged to retain their greater contribution; while for terminals with lower credibility, a more conservative strategy can be adopted, such as removing their results or assigning a smaller weight to their results. In this way, while ensuring the reliability of the calculation results, the computing power resources of each terminal can be fully utilized to improve the computing efficiency.

[0136] For multiple terminals calculating small processing tasks, if all terminals obtain the processing results, some technical means can be used to ensure the reliability of the final results. For example, encryption technology can be used to ensure the security of the calculation process and results, and multiplexing technology can be used to increase the reliability of communication and bandwidth utilization.

[0137] Multiplexing is an effective communication technology that combines multiple signals or data streams into one signal and transmits it on a shared transmission medium, thereby improving communication reliability and bandwidth utilization. Specifically, multiplexing can be applied in the following ways to ensure the reliability of the final results of multiple terminal computing and processing tasks.

[0138] 1) Types of multiplexing technology a. Time Division Multiplexing (TDM): In time division multiplexing, multiple signals are divided into different time slices in time, and each signal is transmitted within its allocated time slice. This method ensures that each terminal occupies the transmission medium within a specific time, thereby reducing the risk of data collision and improving the reliability of transmission.

[0139] b. Frequency Division Multiplexing (FDM): Frequency Division Multiplexing divides the available bandwidth into multiple frequency channels, and each signal is sent on a different frequency. This method effectively improves bandwidth utilization, and the signals of each terminal are transmitted simultaneously on different frequencies without interfering with each other.

[0140] c. Code Division Multiplexing (CDM): Code Division Multiplexing uses different coding techniques to transmit multiple signals simultaneously on the same frequency. Each signal uses a unique code, allowing the receiver to correctly distinguish different signals. This can effectively improve data transmission efficiency and reliability.

[0141] 2) Increase communication reliability a. Data redundancy and retransmission mechanism: On the basis of multiplexing, data redundancy and retransmission mechanism can be added. For example, when performing time division multiplexing, redundant data (such as checksum) can be added to the end of each signal to ensure that even if some data is lost during transmission, the receiver can still obtain complete information through retransmission.

[0142] b. Communication monitoring and error detection: By monitoring the status of the communication link in real time, the system can evaluate the integrity of the signal. Once an error is detected, it can quickly switch to other signal paths or signal sources to ensure reliable data transmission.

[0143] 3) Improve bandwidth utilization a. Dynamic bandwidth allocation: Multiplexing technology allows the system to dynamically adjust bandwidth allocation according to the current network load. For example, if a terminal has a heavy processing task, its transmission bandwidth can be temporarily increased, while the bandwidth of other terminals with lighter transmission loads can be reduced to optimize overall resource utilization.

[0144] b. Data packet aggregation: By aggregating multiple data packets from different terminals for batch transmission, the packet header overhead in the network can be reduced and the overall network performance can be improved. This measure is particularly effective in the high-frequency transmission of small data packets.

[0145] 4) Application Examples In a distributed computing system, multiple terminals simultaneously process the same task and return the results. The system can take advantage of multiplexing technology in the following ways: Task allocation: Each terminal sends the calculation results back to the main control processor through a shared channel within a different time slice (TDM).

[0146] Encryption and coding: When sending data, each terminal uses a different coding method (CDM) to encrypt the data to ensure data security during transmission.

[0147] Monitoring and feedback: The system monitors the status of each channel in real time to ensure that any signal loss can be remedied through redundant data or retransmission mechanisms.

[0148] In addition, in practical applications, it is also necessary to consider the impact of factors such as network topology, bandwidth, and latency on the allocation of computing power and the merging of results. For example, in the case of a complex topology or high network latency, it may be necessary to design some distributed algorithms to implement task allocation and result merging in order to improve the efficiency and reliability of the algorithm.

[0149] In a complex network topology and high network latency environment, it is particularly important to design an effective distributed algorithm to achieve task allocation and result merging. The following are some key design considerations and specific algorithm solutions.

[0150] 1) Consideration of network topology The complexity of network topology will directly affect the strategies of task allocation and result merging. Common topologies include star, ring, mesh, etc. Different topologies have different communication delays, bandwidths, and fault tolerance capabilities.

[0151] Star topology: The central node is directly connected to all terminals, which is suitable for centralized task scheduling.

[0152] Ring topology: Each node directly communicates with only two adjacent nodes, which is suitable for frequent inter-node communication.

[0153] Mesh topology: Any nodes can communicate directly with each other, which is suitable for highly parallel task distribution.

[0154] 2) Distributed algorithm for task allocation a. Task allocation based on load balancing In a distributed environment, load balancing algorithms (such as polling, minimum number of connections, minimum response time, etc.) can be used to dynamically allocate tasks. The specific steps are as follows: State perception: Each node regularly reports its own load status (CPU, memory, network bandwidth, etc.) to the master node.

[0155] Task scheduling strategy: The master node dynamically adjusts task allocation based on the load information of each node. For example, new tasks are preferentially allocated to the node with the smallest load.

[0156] Feedback mechanism: After the node is executed, the execution time and results are fed back to the master node for the next step of scheduling optimization.

[0157] b. Hierarchical-based task allocation In complex network topologies, a hierarchical task allocation mechanism can be used to divide the network into multiple levels, with each level responsible for different task types.

[0158] Top-level control: The top-level node is responsible for the division of global tasks and monitoring the global status.

[0159] Middle-layer coordination: The middle-layer nodes are responsible for allocating tasks to the actual execution terminals within this layer.

[0160] Low-level execution: The terminal node is responsible for executing tasks and returning results.

[0161] This approach can reduce the communication burden on the network and improve the flexibility of task allocation.

[0162] 3) Distributed algorithm for result merging a. Distributed computing and result aggregation During the result merging process, distributed computing frameworks such as MapReduce can be used to process the aggregation of results.

[0163] Map phase: Each terminal first processes its calculation results locally to generate intermediate results.

[0164] Shuffle phase: Group the intermediate results according to specific keys and send the data to the appropriate nodes.

[0165] Reduce phase: perform the merge operation to obtain the final result.

[0166] This method can effectively improve the merging efficiency and reduce network traffic.

[0167] b. Merge results based on voting mechanism When multiple terminals calculate the same task, a voting mechanism can be used to merge the results: Result collection: The master node collects results from all terminals.

[0168] Voting algorithm: Merge the results according to preset voting rules (such as majority voting, weighted voting, etc.) to ensure the accuracy and reliability of the final results.

[0169] Anomaly detection: If a result deviates significantly from other results, it can be marked as an anomaly and other results can be selected for merging.

[0170] 4) Strategies to improve efficiency and reliability Dynamic adjustment: Dynamically adjust the algorithm's parameters and strategies based on network latency and bandwidth conditions to improve the adaptability of task allocation and result merging.

[0171] Fault-tolerant mechanism: Design a fault-tolerant mechanism to ensure that in the event of node failure or network instability, the system can automatically retry task allocation and result merging to ensure the reliability of the final result.

[0172] Data compression: During the result transmission process, data compression technology can be used to reduce the amount of transmitted data and improve the utilization of network bandwidth.

[0173] In addition, we also need to consider the balance of resource utilization, that is, how to reasonably allocate computing resources to meet the needs of different computing tasks. We can determine the appropriate allocation strategy based on the characteristics of different tasks and the demand for resources, such as using optimization methods such as dynamic programming or greedy algorithms to achieve balanced resource utilization.

[0174] In resource allocation, it is a complex problem to reasonably allocate computing resources to meet the needs of different computing tasks. Dynamic programming and greedy algorithms are two commonly used optimization methods that can effectively achieve balanced resource utilization. The following is a detailed description of these two algorithms and their application in resource allocation.

[0175] Dynamic programming: Dynamic programming is an algorithm that breaks down complex problems into simpler subproblems. It is applicable to problems with overlapping subproblems and optimal substructure properties. In resource allocation, dynamic programming can be used to solve multi-stage decision problems where the resource requirements and available resources for each task may change at different stages.

[0176] Dynamic programming is suitable for situations where the problem scale is small and has overlapping sub-problems, such as multi-task resource allocation, knapsack problem, etc. In the problem of resource allocation, if the resource requirements between tasks have complex dependencies, dynamic programming is more appropriate.

[0177] Example: Suppose there is a series of computing tasks, each of which requires different amounts of resources and has different benefits. Through dynamic programming, we can effectively determine how to allocate limited resources to each task to maximize the overall benefit.

[0178] Greedy algorithm: A greedy algorithm is an algorithm based on local optimal selection, which selects the option that currently appears to be the best at each step. Greedy algorithms are usually suitable for solving optimization problems, but it should be noted that they do not always obtain the global optimal solution.

[0179] Greedy algorithms are suitable for situations where the problem structure is relatively simple and the local optimal solution can lead to the global optimal solution. In simple resource allocation, if the relationship between resource demand and benefit between tasks is clear, the greedy algorithm can quickly find a feasible solution.

[0180] Example: When resources are limited, the benefits and resource requirements of tasks vary. Through the greedy algorithm, we can prioritize tasks with the highest unit resource benefits, thereby effectively utilizing limited resources.

[0181] This step can improve the reliability and accuracy of the calculation results. Since the disconnection of a small neural network computing terminal may cause the calculation results to be unavailable, configuring the same computing task for multiple terminals can increase the availability of the calculation results. At the same time, setting the trust level according to the distance from the main control processor can further filter the final calculation results and improve the accuracy and reliability of the calculation results. This ensures that the correct calculation results can still be obtained under the distribution and state changes of various devices in the computing network, thereby improving the computing efficiency and accuracy, and improving the availability and reliability of the entire computing network.

[0182] 6. When multiple terminals calculate and process tasks, only one gets the processing result. In this case, this processing result can be directly used as the final processing result.

[0183] 7. Through the aforementioned process, the various types of devices covered by the present invention can implement an event-triggered distributed neural network working mechanism in the computing network to achieve the various computing tasks required for digital logistics and further enhance the intrinsic computing power value of digital logistics.

[0184] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as follows: Figure 4 As shown. The computer device includes a processor, a memory and a network interface connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store static information and dynamic information data. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the steps in the above method embodiment are implemented.

[0185] Those skilled in the art will understand that Figure 4 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present invention, and does not constitute a limitation on the computer device to which the solution of the present invention is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0186] In addition, the present invention also provides a computer device, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above method embodiment when executing the computer program.

[0187] In addition, the present invention further provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps in the above method embodiment are implemented.

[0188] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided by the present invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory or optical memory, etc. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0189] The present invention is not limited to the structures which have been described above and shown in the drawings, and various modifications and changes may be made without departing from the scope thereof. The scope of the present invention is limited only by the appended claims.

Claims

1. An event-triggered distributed neural network working method, characterized in that: include: Based on the neural network training and reasoning task attributes, a competitive voting mechanism is used to select the master control processor within the preset computing network range; Calculate the master control time according to the computing power of the master control processor and the task complexity, and dynamically adjust the master control processor based on the master control time; The main control processor integrates the current computing resources, dynamically divides the computing tasks according to the computing power requirements and task characteristics, and obtains several computing subtasks; Based on the multi-level redundant backup mechanism, the task allocation algorithm is used to allocate computing subtasks to corresponding terminal nodes; The terminal node receives the corresponding computing subtask, executes the computing subtask and retains the intermediate computing data and computing results; The calculation results of the terminal nodes are screened according to the preset credibility to obtain the final calculation results.

2. The event-triggered distributed neural network working method according to claim 1, characterized in that: A competitive voting mechanism is used to select the master processor within the preset computing network, including: Qualify the terminal nodes within the preset computing network range, and use the terminal nodes that pass the qualification as candidate terminals, and monitor the performance data of the candidate terminals in real time, where the performance data includes computing power, memory usage and network delay; According to the preset evaluation criteria, the score of each candidate terminal is calculated; based on the trust level of the candidate terminal, a corresponding weight is assigned to each candidate terminal, and the candidate terminal participates in the voting according to the score result and the corresponding weight to determine the main control processor.

3. The event-triggered distributed neural network working method according to claim 1, characterized in that: The calculating the master control time according to the computing power and task complexity of the master control processor and dynamically adjusting the master control processor based on the master control time includes: The master control time is calculated according to the computing power and task complexity of the current master control processor, and the duty time of the current master control processor is monitored in real time. When the duty time exceeds the master control time, the master control processor selection mechanism is re-initiated and a new master control processor is confirmed; The calculation formula of the master control time is: ； In the formula, T Indicates the master time. K represents the safety factor, C represents the complexity of the task, P Indicates the computing power of the main control processor. L Indicates the current load of the main control processor. H represents the historical performance coefficient of the main control processor, D Indicates network latency.

4. The event-triggered distributed neural network working method according to claim 1, characterized in that: The main control processor integrates the current computing resources and dynamically divides the computing tasks according to the computing power requirements and task characteristics of the computing tasks to obtain a number of computing subtasks; Based on the multi-level redundant backup mechanism, the task allocation algorithm is used to allocate computing subtasks to corresponding terminal nodes, including: The main control processor dynamically divides the computing task according to the changes in the current environment and the status information of each terminal node, combined with the computing power requirements and task characteristics of the computing task, and obtains several computing subtasks; according to the priority of the computing subtask and the resource status of the terminal node, the scheduling order of the computing subtask is adjusted; Obtain the status information and resource availability information of each terminal node, and assign corresponding computing subtasks to each terminal node based on the multi-level redundant backup mechanism and task allocation algorithm, combined with the status information and resource storage information of the terminal node.

5. The event-triggered distributed neural network working method according to claim 1, characterized in that: The terminal node receives the corresponding computing subtask, executes the computing subtask and retains the intermediate computing data and computing results; The calculation results of the terminal nodes are screened according to the preset credibility, and the final calculation results include: Each terminal node receives the corresponding computing subtask and executes the computing subtask according to the adjusted scheduling order; the main control processor monitors the computing progress and results of each terminal node in real time and retains the intermediate data and results; The main control processor collects and summarizes the calculation results of all terminal nodes. If there is only one calculation result, it will be directly used as the final calculation result. If there are multiple calculation results, the calculation results of different terminal nodes corresponding to the same calculation subtask will be obtained, and the calculation results of different terminal nodes will be analyzed to see whether they are the same. If so, one of the calculation results will be randomly selected as the calculation result of the calculation subtask. If not, the calculation result selected by majority voting will be used as the calculation result of the calculation subtask. Based on the computing power of each terminal node and the distance from the main control processor, a corresponding trust level is set for each terminal node, and the calculation results of all computing subtasks are screened and merged according to the feasible level of each terminal node to obtain the final calculation result.

6. An event-triggered distributed neural network operating system, characterized in that: include: The master control processor selection module is used to select the master control processor within the preset computing network range based on the neural network training and reasoning task attributes using a competitive voting mechanism; calculate the master control time according to the computing power and task complexity of the master control processor, and dynamically adjust the master control processor based on the master control time; The computing task division and allocation module is used for the main control processor to integrate the current computing resources, dynamically divide the computing tasks according to the computing power requirements and task characteristics of the computing tasks, and obtain several computing subtasks; Based on the multi-level redundant backup mechanism, the task allocation algorithm is used to allocate computing subtasks to corresponding terminal nodes; The calculation result screening and merging module is used for the terminal node to receive the corresponding calculation subtasks, execute the calculation subtasks and retain the intermediate calculation data and calculation results; The calculation results of the terminal nodes are screened according to the preset credibility to obtain the final calculation results.

7. The event-triggered distributed neural network operating system according to claim 6, characterized in that: When the master processor selection module selects the master processor within the preset computing network range by using the competitive voting mechanism, it performs qualification authentication on the terminal nodes within the preset computing network range, and uses the terminal nodes that have passed the authentication as candidate terminals, and monitors the performance data of the candidate terminals in real time, wherein the performance data includes computing power, memory usage and network delay status; calculates the score of each candidate terminal according to the preset evaluation criteria; and assigns a corresponding weight to each candidate terminal based on the trust level of the candidate terminal, and the candidate terminals participate in voting according to the scoring results and the corresponding weights to determine the master processor.

8. The event-triggered distributed neural network operating system according to claim 6, characterized in that: The master processor selection module calculates the master control time according to the computing power and task complexity of the master processor, and dynamically adjusts the master processor based on the master control time. It calculates the master control time according to the computing power and task complexity of the current master processor, and monitors the duty time of the current master processor in real time. When the duty time exceeds the master control time, the master processor selection mechanism is re-initiated and a new master processor is confirmed. The calculation formula of the master control time is: ； In the formula, T Indicates the master time. K represents the safety factor, C represents the complexity of the task, P Indicates the computing power of the main control processor. L Indicates the current load of the main control processor. H represents the historical performance coefficient of the main control processor, D Indicates network latency.

9. The event-triggered distributed neural network operating system according to claim 6, characterized in that: The computing task division and allocation module includes a computing task division module and a computing task allocation module; The computing task division module is used for the main control processor to dynamically divide the computing task according to the changes in the current environment and the status information of each terminal node, combined with the computing power requirements and task characteristics of the computing task, to obtain a number of computing subtasks; according to the priority of the computing subtask and the resource status of the terminal node, the scheduling order of the computing subtask is adjusted; The computing task allocation module is used to obtain the status information and resource availability information of each terminal node, and allocate corresponding computing subtasks to each terminal node based on the multi-level redundant backup mechanism and task allocation algorithm in combination with the status information and resource storage information of the terminal node.

10. The event-triggered distributed neural network operating system according to claim 6, characterized in that: The calculation result screening and merging module includes a calculation subtask execution module, a calculation result summary and analysis module and a calculation result screening and output module; The computing subtask execution module is used for each terminal node to receive the corresponding computing subtask and execute the computing subtask according to the adjusted scheduling order; the main control processor monitors the computing progress and computing results of each terminal node in real time and retains the intermediate data and results; The calculation result summary and analysis module is used for the main control processor to collect and summarize the calculation results of all terminal nodes. If there is only one calculation result, the calculation result is directly used as the final calculation result. If there are multiple calculation results, the calculation results of different terminal nodes corresponding to the same calculation subtask are obtained, and the calculation results of different terminal nodes are analyzed to see whether they are the same. If so, one of the calculation results is randomly selected as the calculation result of the calculation subtask. If not, the calculation result selected by majority voting is used as the calculation result of the calculation subtask. The calculation result screening output module sets a corresponding trust level for each terminal node based on the computing power of each terminal node and the distance from the main control processor, and screens and merges the calculation results of all calculation subtasks according to the feasible level of each terminal node to obtain the final calculation result.

Images