Dynamic optimization method and device for central Mesh network and medium

Abstract

The invention provides a dynamic optimization method and device of a central Mesh network and a medium, and relates to the technical field of dynamic optimization of the central Mesh network.The method comprises the steps that the current moment serves as a starting point, operation time sequence data and service request time sequence data of the Mesh network are collected, and the operation time sequence data and the service request time sequence data comprise node residual energy, link quality, bandwidth utilization rate and other information; classifying the service request data according to priorities, counting the request quantity and the data quantity of each type of service, and generating a service request feature vector; extracting features of network operation data by adopting a time sequence modeling method, dynamically distributing feature weights by utilizing an attention mechanism, generating a joint feature vector through dimension reduction processing, and providing input features for multi-objective optimization; and constructing a global optimization path planning model according to the generated joint feature vector, solving a multi-objective optimization function by applying a genetic algorithm, and searching a Pareto optimal solution, thereby outputting a path planning scheme with the highest fitness.

Description

Technical Field

[0001] The present invention relates to the technical field of dynamic optimization of a centralized Mesh network, and in particular to a method, device and medium for dynamic optimization of a centralized Mesh network. Background Art

[0002] The current path planning problem in complex network environments faces the challenge of multi-objective optimization. In particular, when considering multiple constraints such as latency, packet loss rate, and remaining energy of network nodes, traditional path planning algorithms often find it difficult to balance globality and diversity. Most existing methods use fixed weight models or simple heuristic algorithms. Although they are computationally efficient, they are prone to fall into local optimality in dynamic network environments, resulting in inflexible path selection and difficulty in meeting real-time and reliability requirements. In addition, the unbalanced use of network resources will cause some nodes to run out of energy prematurely, thereby shortening the network life and reducing overall performance. This limitation is particularly prominent in energy-constrained networks such as wireless sensor networks, and a more intelligent and adaptive optimization method is needed to balance various performance indicators.

[0003] In the prior art, the path planning problem of the centralized Mesh network is usually solved by some traditional algorithms with a single objective optimization as the core, mainly focusing on minimizing the path delay. Such methods have a fast calculation speed, but ignore factors such as packet loss rate and node energy, and have poor adaptability to resource-constrained network environments. Secondly, some improved heuristic algorithms can achieve multi-objective optimization to a certain extent, by adjusting parameter weights to balance performance indicators such as delay and energy consumption, but usually require fixed weight values ​​in advance, and it is difficult to adapt to dynamically changing network environments in real time. In addition, there are also some machine learning-based methods that predict the optimal path through training with historical data, but in practical applications, due to the high computational complexity and insufficient response to dynamics, it is difficult to promote to scenarios with high real-time requirements.

[0004] In actual centralized Mesh network applications, the above-mentioned existing technologies have obvious limitations, especially in dynamic network environments, showing the following deficiencies: First, traditional path planning methods are prone to fall into local optimal solutions and cannot find the comprehensive optimal path in multi-objective trade-offs, resulting in global performance degradation; second, heuristic algorithms with fixed weights lack the ability to respond to dynamic changes in network topology and business needs during path planning, and cannot achieve flexible path optimization; third, existing machine learning or other complex algorithms have large computational overhead and are not suitable for scenarios with limited energy or high real-time requirements. Therefore, existing technologies are difficult to meet actual needs in terms of how to balance the delay, packet loss rate and residual energy in the network and provide dynamic optimization capabilities.

[0005] The above information disclosed in this Background section is only for enhancement of understanding of the background of the present disclosure and therefore it may contain information that does not constitute the prior art that is already known to one of ordinary skill in the art. Summary of the invention

[0006] The purpose of the present invention is to provide a method, device and medium for dynamic optimization of a centralized Mesh network to solve the problems raised in the above background technology.

[0007] To achieve the above object, the present invention provides the following technical solutions: A dynamic optimization method for a centralized Mesh network, comprising the following specific steps: Step 1: Set a time window and collect the operation sequence data and service request sequence data of the Mesh network within the time window. The network operation sequence data includes the remaining energy of each Mesh network node, the link quality between nodes and the current bandwidth utilization rate of each link. The service request sequence data includes the data volume, priority and timeliness index data of all services. Step 2: Divide the collected service request time series data into high priority, medium priority and low priority services according to the service data volume, priority and timeliness indicators, and count the number of requests, total data volume and service proportion of each type of service in the time window k, and combine them into a service request statistical feature vector; Step 3: Use the time series modeling method to extract features from the collected network runtime time series data to obtain node energy trend features, link quality features, and bandwidth utilization features. Use the attention mechanism to assign feature weights, and generate network dimension reduction feature vectors through dimensionality reduction processing. Combined with the service request statistical feature vector, generate a joint feature vector for the current time window. Step 4: Based on the joint feature vector of the current time window, a global optimization path planning model is constructed, and a genetic algorithm is used to solve the multi-objective optimization function, find the Pareto optimal solution, and finally output the global optimization path planning solution with the highest fitness.

[0008] Furthermore, the logic for collecting the Mesh network's runtime data and service request data is as follows: Taking the current time t as the starting point of the collection, define the dynamic time window k. Collect network operation timing data and business request timing data within the time range; Data collection is started by "periodic triggering": periodic triggering, at fixed time intervals, the central node broadcasts collection instructions to the entire network, requiring all nodes to report the latest status information; The specific contents of the collection are as follows: The node residual energy is set to The collection method is that the central node collects the energy status of each node through polling instructions every 1 minute, and ,in The total energy of the node. The event triggering condition is that the energy decline rate is greater than 5% / minute, and the node needs to actively report the current remaining energy; Link quality data measurement method: Nodes exchange heartbeat packets every 10 seconds, and use ACK and RTT to count packet loss rate and delay. The central node monitors the RSSI and SNR of link communication, and calculates the weighted average of the heartbeat packet measurement value and the central node monitoring value in a weight ratio of 7:3 as the final link quality data; The link bandwidth usage collection protocol is that the central node obtains the traffic statistics of the switch port through SNMP. The calculation formula is as follows: ; in, is the bandwidth utilization rate of the i-th link, is the actual occupied bandwidth of the ith link, is the total bandwidth of the ith link; The business request dataset for each time window contains the following information: Business priorities are set as , define the priority field in the service request header: 0 for high, 1 for medium, 2 for low, and the initiating node digital signature is required; the service timeliness is set to , the maximum allowed end-to-end delay for each request, which is automatically filled in by the service type mapping table; the service data volume is set to ,When the service is initiated, the transmission data size is recorded through the request header.

[0009] Furthermore, the logic of dividing the collected service request time series data into high priority, medium priority and low priority services according to the service data volume, priority and timeliness indicators is as follows: Business requests are classified as high, medium, or low priority, and must meet any of the following conditions: For high priority business sets , if the business priority , that is, the request is clearly marked as high priority in the business request header, then the request needs to meet any of the following conditions: a. Business timeliness , that is, the maximum allowed end-to-end delay of the request is less than or equal to the delay threshold of the high priority; b. Business data volume and , that is, the amount of requested data is small and the timeliness requirement is high; among them, and are the set delay and quantity thresholds; For medium priority services , if the business request meets all of the following conditions: a. Business priority , that is, the request is clearly marked as medium priority in the business request header, b. Business timeliness , that is, the latency requirement of the request is between high priority and low priority; c. Business data volume , that is, the amount of requested data is moderate; among them, and is the set threshold value; For low priority services , if the business request meets any of the following conditions: a. If the business priority , that is, the request is clearly marked as low priority in the business request header, b. Business timeliness , that is, the maximum allowed end-to-end delay of the request is greater than the threshold value of the medium priority and the service data volume , that is, large data needs to be transmitted, and the latency requirement is not high; The logic of counting the number of requests, total data volume, and business proportion of each type of business within the statistical time window k and combining them into a business request statistical feature vector is as follows: For each type of business, namely high priority, medium priority and low priority, the following three key characteristics are counted respectively: Count the number of requests belonging to this type of business in the current time window k , the calculation formula is: ; in, is a collection of certain types of business, that is or , Represents an index, when When , it refers to high priority business, when When , it refers to the medium priority business, when When , it refers to low priority business; The transmission data volume of all services of this type in the current time window k is accumulated, and the calculation formula is: ; in, Indicates the data volume of the nth request. Statistics of total data volume for a certain type of business; Calculate the ratio of the total data volume of this type of business to the total data volume of all businesses in the current time window. The formula is as follows: ; in, is the total data volume of all services, The proportion of data volume for a certain type of business; like ,but and ;like , then all proportions are set to ; Combine the statistical results into a structured feature vector. The business request statistical feature vector is defined as follows: ; in, are the number of requests for high, medium, and low priority services, respectively. are the total data volume of high, medium and low priority services respectively, They are the total data volume proportions of high, medium and low priority services respectively; After the feature vector is generated, the statistical results need to be checked for consistency: request quantity consistency, ,in, is the total number of requests in the current time window k; data consistency, ; Total percentage check, .

[0010] Furthermore, the time series modeling method is used to extract features from the network operation time series data collected in step 1. The logic for extracting node energy trend features, link quality features, and bandwidth utilization features is as follows: In the time window The node and link timing data within, including: node remaining energy timing data , represents the residual energy change of the i-th node in the time window; Link quality characteristic timing data includes packet loss rate , represents the change of packet loss rate of link i over time in the time window, delay , represents the end-to-end delay and delay jitter of link i in the time window , represents the amplitude of delay fluctuation of link i in the time window; bandwidth utilization time series data , represents the bandwidth utilization rate of link i in the time window; Use LSTM to model various types of time series data, build independent time series models for each type of time series data, and set the input sequence of each type of time series data to ; For each input sequence, the output of LSTM is the hidden state in the time window or the final hidden state : ; in, is the input of the time series feature, are the hidden state and candidate state of the previous time step, respectively. It captures the dynamic representation of temporal features and the hidden state of the final output Represents a time window Trend characteristics of The final hidden state of the LSTM model , extract the following trend feature set: node energy trend feature, assuming that the total number of nodes in the network is indivual, ; Link quality trend characteristics, , where i represents the trend characteristic of the i-th link quality; bandwidth utilization trend characteristic, ,in is the total number of links in the network; The logic of using the attention mechanism to assign feature weights and generating network dimension reduction feature vectors through dimensionality reduction processing is: For each type of feature, the weight of each dimension of data is dynamically calculated through the attention mechanism to associate the importance of the feature with the target task. The formula is as follows: ; in, is the dynamic representation of feature o, which is the feature vector obtained after forward propagation in the latter layer of the neural network. Feature o represents the dimensional feature of the latter layer in the network. is a trainable parameter in the attention layer, used to dynamically represent the features Perform a linear transformation, yes The transposed matrix of is the importance weight of feature o, and ,satisfy ; Perform weighted summation on the features to obtain the aggregated representation after the attention mechanism : ; Among them, aggregation means By dynamically representing all features According to their importance weight The weighted sum is obtained; according to the aggregate representation formula, the aggregate representation of different features is calculated, and the weighted sum of the dynamic representation of each dimension of the node energy feature is calculated: ; in, is a dynamic representation of the node energy characteristics, is the corresponding importance weight; The link quality feature aggregation is expressed as follows: and bandwidth utilization feature aggregation is expressed as .

[0011] Furthermore, the principal component analysis method is used to reduce the dimension of the attention-weighted feature set, and the input feature set is: ; The PCA projection matrix is ​​obtained by eigenvalue decomposition of the characteristic covariance matrix , use the projection matrix to reduce the dimension of the original feature set, and calculate the feature vector after dimension reduction: ; in, is the original feature set, is the feature vector after dimensionality reduction; the output network state feature vector after dimensionality reduction is ; At the same time, combined with the business request statistical feature vector in step 2, the logic for generating the joint feature vector of the current time window is: The network status features after dimensionality reduction are concatenated with the service request statistical features in step 2 to generate the joint feature vector of the current time window. The formula is: ; in, Represents the trend characteristics of network operation status within the time window, Represents the statistical characteristics of business requests within the time window, and the joint feature vector It integrates network status characteristics and service request characteristics.

[0012] Furthermore, combined with the joint feature vector of the current time window generated in step 3, a global optimization path planning model is constructed, and the genetic algorithm is used to solve the multi-objective optimization function. The logic of finding the Pareto optimal solution is: Joint eigenvector The network status characteristics and business request characteristics are combined, and based on these characteristics, a global optimization path planning model is constructed. In path planning, the following core optimization goals are first defined: Minimum path delay: The total delay of the selected path to ensure the real-time performance of the service Does not exceed the maximum allowable end-to-end delay of the service , that is, target 1 is ,in is the total delay from the source node to the destination node, which is determined by the link delay characteristics; Minimize link congestion: Avoid overloading of links with high bandwidth utilization in the path to improve the overall stability of the network, that is, goal 2 is ,in is the maximum bandwidth utilization of all links in the path, The maximum safe bandwidth utilization threshold of the defined link; Prolong network life: Prioritize paths with higher residual energy to prevent premature node failure, that is, goal 3 is Maximize the remaining energy, where is the minimum residual energy of the nodes in the path, obtained through the node energy trend characteristics; Path planning must meet the following constraints: Business timeliness constraints: , ensuring that the selected path can meet the timeliness requirements of the business; link quality constraints: ,in, is the packet loss rate of the ith link, is the maximum allowed packet loss rate threshold, Represents a path; node energy constraint: ,in, is the residual energy of the ith node, is the minimum required threshold of node energy; Using the reduced-dimensional network state feature vector generated in step 3 and business request statistical feature vector , defining the dynamic weights of nodes and links in the network topology: Link Weight , comprehensively considering the link quality characteristics and bandwidth utilization: ; in, is the delay of link i, is the packet loss rate of link i, is the bandwidth utilization of link i, are weight parameters, and ; Node Weight , calculated based on the node residual energy: ; in, is the residual energy of node i, when When it is smaller, the node weight increases, thus avoiding the selection of nodes with high load and fast energy consumption.

[0013] Furthermore, a weighted shortest path algorithm based on feature weights is used to dynamically plan the optimal path by combining the feature vector in step 3 and the business characteristics in step 2. The specific logic is as follows: Modeling network topology as a weighted directed graph , define the node set represents all nodes in the network, Represents all links in the network, and the link weight is given by Definition, using the improved Dijkstra algorithm, planning the path based on the following path cost function: ; in, For path The total cost reflects the comprehensive weight of the selected path. is the set of all relay nodes on the path; Build a network topology diagram , the link weight and node weight Map it to the graph, input the starting point, end point and delay constraint of the business request, and for each candidate path , calculate its path cost , filter out paths that do not meet the following conditions: ; Select the total cost from the remaining candidate paths The smallest path is taken as the preliminary optimal path; After obtaining the initial optimal path, the genetic algorithm is used to further optimize the path selection to find the Pareto optimal solution: Use genetic algorithm to find the Pareto optimal solution of the optimization path and randomly generate the initial path population , where each path Contains a series of nodes and links between them. For each path Calculation: Path Delay , path packet loss rate , the minimum residual energy of the path ; The fitness function is defined as: ; Use the Pareto dominance relationship to select the Pareto optimal solution set. If the path Dominant Path , then Excluded; The tournament selection method is used to select paths with higher fitness from the current population as the parents of the next generation; the selected parents are crossovered to generate new path individuals; the generated new individuals are randomly mutated to increase population diversity; these steps are iterated until the termination condition is met, that is, the maximum number of iterations is reached, and the path population is updated after each iteration, and the path with the highest fitness is retained; Finally select a path from the Pareto optimal solution set , which has the best overall performance, as described below: ; Among them, here is the normalized comprehensive fitness value; Export Selected Path Detailed information, including: path node sequence, path link sequence, total path delay , the average packet loss rate of the path , the minimum residual energy of the path ; After outputting the path planning solution, analyze the optimization effect: compare the delay, packet loss rate and remaining energy changes between the initial path and the final path, and evaluate the utilization efficiency of the optimization solution on network resources.

[0014] The present invention also provides a dynamic optimization device for a centralized Mesh network, the device is used to implement the above-mentioned dynamic optimization method for a centralized Mesh network, comprising: The data collection module is used to set a time window and collect the operation sequence data and service request sequence data of the Mesh network within the time window. The network operation sequence data includes the remaining energy of each Mesh network node, the link quality between nodes and the current bandwidth usage rate of each link. The service request sequence data includes the data volume, priority and timeliness index data of all services. The data processing module is used to divide the collected service request time series data into high priority, medium priority and low priority services according to the service data volume, priority and timeliness indicators, and to count the number of requests, total data volume and service proportion of each type of service in the time window k, and combine them into a service request statistical feature vector; The feature extraction module is used to extract features from the collected network runtime data using a time series modeling method, obtain node energy trend features, link quality features, and bandwidth utilization features, use an attention mechanism to assign feature weights, and generate a network dimension reduction feature vector through dimensionality reduction processing. At the same time, it combines the service request statistical feature vector to generate a joint feature vector for the current time window; The optimization path module is used to build a global optimization path planning model based on the joint feature vector of the current time window, use genetic algorithms to solve multi-objective optimization functions, find the Pareto optimal solution, and finally output the global optimization path planning solution with the highest fitness.

[0015] The present invention further provides a storage medium, wherein the storage medium stores a computer program, and when the computer program is executed by a processor, each step in the above-mentioned method for dynamic optimization of a centralized Mesh network is implemented.

[0016] Compared with the prior art, the present invention has the following beneficial effects: The present invention collects the running sequence data and service request sequence data of the Mesh network. This data collection mechanism can not only reflect the running status of the network in real time and dynamically, but also provide key input information for path planning, thereby solving the problem of insufficient response capability to dynamic changes of network topology and service demand in the background technology; the collected service request sequence data is classified and counted, and a service request statistical feature vector is generated. By classifying the services by priority and combining the information such as the number of requests, total data volume and service proportion within the statistical time window, the overall distribution characteristics of the service demand can be more accurately evaluated.

[0017] The present invention adopts a time series modeling method to extract features from network runtime time series data, and assigns feature weights through an attention mechanism. This innovative technical feature can dynamically perceive changes in node energy trends, link quality, and bandwidth utilization, and assign different weights to these features, further highlighting the impact of key performance indicators on path planning. Compared with the prior art that simply processes time series data, the present invention significantly improves the ability to accurately model the dynamic environment of the network through the attention mechanism. In addition, the introduction of dimensionality reduction processing technology reduces the computational complexity by reducing the feature dimension, providing a more concise and efficient input feature for global optimization path planning.

[0018] The present invention combines the generated joint feature vector to construct a global optimization path planning model, and solves the multi-objective optimization function through a genetic algorithm to find the Pareto optimal solution. The introduction of the genetic algorithm enables the path planning model to have a global search capability, and can find the path with the highest fitness in the multi-objective trade-off, avoiding the defect that the traditional algorithm is prone to falling into the local optimum. In addition, the Pareto optimization technology enables this solution to simultaneously optimize multiple objectives such as the path delay, packet loss rate and residual energy, instead of focusing only on a single performance, thereby solving the limitation of insufficient multi-objective optimization capability in the prior art. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a schematic diagram of the overall method flow of the present invention; Figure 2 This is a flow chart of the overall system module of the present invention. DETAILED DESCRIPTION

[0020] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with specific embodiments.

[0021] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in the present invention should be understood by people with ordinary skills in the field to which the present invention belongs. The words "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprise" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0022] Example: See also Figure 1 , the present invention provides a technical solution: A dynamic optimization method for a centralized Mesh network, comprising the following specific steps: Step 1: Set a time window and collect the operation sequence data and service request sequence data of the Mesh network within the time window. The network operation sequence data includes the remaining energy of each Mesh network node, the link quality between nodes and the current bandwidth utilization rate of each link. The service request sequence data includes the data volume, priority and timeliness index data of all services. The logic for collecting the Mesh network's runtime data and service request data is as follows: Taking the current time t as the starting point of the collection, define the dynamic time window k. Collect network operation timing data and business request timing data within the time range; Data collection is started by "periodic triggering": periodic triggering, at fixed time intervals, the central node broadcasts collection instructions to the entire network, requiring all nodes to report the latest status information; The specific contents of the collection are as follows: The node residual energy is set to The collection method is that the central node collects the energy status of each node through polling instructions every 1 minute, and ,in The total energy of the node. The event triggering condition is that the energy decline rate is greater than 5% / minute, and the node needs to actively report the current remaining energy; Link Quality Data Measurement method: Nodes exchange heartbeat packets every 10 seconds, and use ACK and RTT to count packet loss rate and delay. The central node monitors the RSSI and SNR of link communication, and calculates the weighted average of the heartbeat packet measurement value and the central node monitoring value in a weight ratio of 7:3 as the final link quality data, that is, , and the collection of link quality data must meet the packet loss rate ; Indicates the overall quality of the link. A higher value indicates better link quality and more stable transmission performance. Reflects the signal strength. The higher the signal strength, the better the link quality. Indicates the ratio of signal to noise, the higher it is, the less signal interference; The link bandwidth usage collection protocol is that the central node obtains the traffic statistics of the switch port through SNMP. The calculation formula is as follows: ; in, is the bandwidth utilization rate of the i-th link, is the actual occupied bandwidth of the ith link, is the total bandwidth of the i-th link; when Increase, It also increases, indicating that the burden on the network increases. Vice versa, high link utilization leads to network congestion, affecting data transmission efficiency; The business request dataset for each time window contains the following information: Business priorities are set as , define the priority field in the service request header: 0 for high, 1 for medium, 2 for low, and the initiating node digital signature is required; the service timeliness is set to , the maximum allowed end-to-end delay for each request, which is automatically filled in by the service type mapping table; the service data volume is set to ,When the service is initiated, the size of the transmitted data is recorded through the request header; Affects resource allocation and scheduling. The higher the priority, the more timely it needs to be processed. Ensure that service requests meet latency requirements, affecting user experience. It directly affects the network bandwidth and resource usage. The larger the data volume, the more network resources are required. The relationship between business priority, timeliness and data volume: high-priority business requests usually require lower latency and higher bandwidth usage to ensure that the business can be completed in a timely manner. As the data volume increases, timeliness requirements are affected, and network resources need to be dynamically adjusted, causing changes in priority.

[0023] Step 2: Divide the collected service request time series data into high priority, medium priority and low priority services according to the service data volume, priority and timeliness indicators, and count the number of requests, total data volume and service proportion of each type of service in the time window k, and combine them into a service request statistical feature vector; The logic of dividing the collected business request time series data into high priority, medium priority and low priority businesses according to the business data volume, priority and timeliness indicators is as follows: Business requests are classified as high, medium, or low priority, and must meet any of the following conditions: For high priority business sets , if the business priority , that is, the request is clearly marked as high priority in the business request header, then the request needs to meet any of the following conditions: a. Business timeliness , that is, the maximum allowed end-to-end delay of the request is less than or equal to the delay threshold of the high priority; b. Business data volume and , that is, the amount of requested data is small and the timeliness requirement is high; among them, and It is the delay and quantity thresholds that are dynamically set according to the service type or application scenario; Threshold It is set dynamically, and defines high priority requirements according to the business type. The smaller the value, the higher the latency requirement of the service and the stronger the timeliness. , then the service meets the high-priority delay requirement; Indicates the latency requirement of the transmission request, reflecting the time sensitivity of the service; Indicates the transmission requirements of the service. The smaller the data volume, the easier it is to transmit it quickly, thus meeting the timeliness requirements. The two threshold values ​​jointly determine which services have high priority; The smaller and The smaller it is, the more likely it is that the business will meet high priority; For medium priority services , if the business request meets all of the following conditions: a. Business priority , that is, the request is clearly marked as medium priority in the business request header, b. Business timeliness , that is, the latency requirement of the request is between high priority and low priority; c. Business data volume , that is, the amount of requested data is moderate; among them, and It is the threshold value set for medium-delay services; The higher the value, the looser the latency requirement of the service. The latency requirement of medium priority is between high priority and low priority. Reflects the transmission needs of the business, with moderate data volume and meeting the requirements of medium priority; For low priority services , if the business request meets any of the following conditions: a. If the business priority , that is, the request is clearly marked as low priority in the business request header, b. Business timeliness , that is, the maximum allowed end-to-end delay of the request is greater than the threshold value of the medium priority and the service data volume , that is, large data needs to be transmitted, and the latency requirement is not high; Low-priority services have loose latency requirements and allow for larger latency. Low-priority services allow for larger data volumes and usually have lower requirements for transmission timeliness. The logic of counting the number of requests, total data volume, and business proportion of each type of business within the statistical time window k and combining them into a business request statistical feature vector is as follows: For each type of business, namely high priority, medium priority and low priority, the following three key characteristics are counted respectively: Count the number of requests belonging to this type of business in the current time window k , the calculation formula is: ; in, is a collection of certain types of business, that is or , Represents an index, when When , it refers to high priority business, when When , it refers to the medium priority business, when When , it refers to low priority business; The number of requests for a certain type of business is counted within the current time window k, and the total number of requests belonging to a certain type of business is reflected in the number distribution of the type of business in the total requests; The more requests there are, The bigger; The transmission data volume of all services of this type in the current time window k is accumulated, and the calculation formula is: ; in, Indicates the data volume of the nth request. Statistics of total data volume for a certain type of business; Statistics on the total amount of data transmitted by a certain type of service within the current time window k reflect the traffic distribution of this type of service. or The bigger, The bigger; Calculate the ratio of the total data volume of this type of business to the total data volume of all businesses in the current time window. The formula is as follows: ; in, is the total data volume of all services, The proportion of data volume for a certain type of business; Reflects the proportion of a certain type of business in the total traffic in the current time window k, and is used to measure the traffic distribution of services of different priorities. Increase, Increase, Increased reduce; like ,but and ;like , then all proportions are set to ; Combine the statistical results into a structured feature vector. The business request statistical feature vector is defined as follows: ; in, are the number of requests for high, medium, and low priority services, respectively. are the total data volume of high, medium and low priority services respectively, They are the total data volume proportions of high, medium and low priority services respectively; After the feature vector is generated, the statistical results need to be checked for consistency: request quantity consistency, ,in, is the total number of requests in the current time window k; data consistency, ; Total percentage check, .

[0024] Step 3: Use the time series modeling method to extract features from the collected network runtime time series data to obtain node energy trend features, link quality features, and bandwidth utilization features. Use the attention mechanism to assign feature weights, and generate network dimension reduction feature vectors through dimensionality reduction processing. Combined with the service request statistical feature vector, generate a joint feature vector for the current time window. The time series modeling method is used to extract features from the network runtime time series data collected in step 1. The logic for extracting node energy trend features, link quality features, and bandwidth utilization features is as follows: In the time window The node and link timing data within, including: node remaining energy timing data , represents the change of the remaining energy of the i-th node in the time window; the higher the energy value, the stronger the working ability and duration of the node, otherwise the node will be unable to continue to provide services; Link quality characteristic timing data includes packet loss rate , which represents the change of packet loss rate and delay of link i in the time window over time. , represents the end-to-end delay and delay jitter of link i in the time window , represents the amplitude of the delay fluctuation of link i within the time window. These characteristics describe the performance and stability of the link and affect the reliability and efficiency of data transmission; bandwidth utilization time series data , represents the bandwidth utilization rate of link i in the time window. A high bandwidth utilization rate indicates a heavy network load, which affects the service quality; Use LSTM to model various types of time series data, build independent time series models for each type of time series data, and set the input sequence of each type of time series data to ; For each input sequence, the output of LSTM is the hidden state in the time window or the final hidden state : ; in, is the input of the time series feature, are the hidden state and candidate state of the previous time step, respectively. It captures the dynamic representation of temporal features and the hidden state of the final output Represents a time window Trend characteristics of It is the hidden state of the current time step, capturing the dynamic representation of time series features. LSTM can output the hidden state by learning the dynamic changes of historical time series data. The higher the value, the better the model captures the current features; The final hidden state of the LSTM model , extract the following trend feature set: node energy trend feature, assuming that the total number of nodes in the network is indivual, ; Extract the energy trend characteristics of all nodes, including The final hidden state of the nodes; the higher the Indicates that the energy trend of node i is upward, enhancing the stability and service capabilities of the network; Link quality trend characteristics, , where i represents the i-th link quality trend feature; extract the trend features of the link packet loss rate, delay and delay jitter. The higher the characteristic value, the lower the link quality, such as high packet loss rate and high delay, which affects the reliability of the network; Bandwidth utilization trend characteristics, ,in is the total number of links in the network; extract the bandwidth usage trend characteristics of all links, including The final hidden state of the link is high A value of indicates that link i is heavily loaded, causing network congestion; The logic of using the attention mechanism to assign feature weights and generating network dimension reduction feature vectors through dimensionality reduction processing is: For each type of feature, the weight of each dimension of data is dynamically calculated through the attention mechanism to associate the importance of the feature with the target task. The formula is as follows: ; in, is the dynamic representation of feature o, which is the feature vector obtained after forward propagation in the latter layer of the neural network, reflecting the state of the feature in the current context. Feature o represents the dimensional feature of the latter layer in the network. is a trainable parameter in the attention layer, used to dynamically represent the features Perform a linear transformation, yes The transposed matrix of is the importance weight of feature o, reflecting the importance of the feature, and ,satisfy ; Weight The higher the value, the more important the feature is in the current task. The sum of the weights is 1, ensuring that the relative importance of the features can be normalized. In the attention mechanism, The main function of Converted into a query vector, so that the model can calculate the correlation between the feature and the target task. After the matrix transformation, the formed vector directly participates in the subsequent attention weight calculation, thereby determining the contribution of each feature in the aggregate representation; In the formula, As part of the query vector, it is used to calculate the dynamic representation of each feature The importance of , a scalar value can be obtained, indicating the similarity or relevance of feature o to the query, which is then indexed and normalized to generate the weight ,therefore, The transpose operation ensures the dimensional compatibility of matrix operations and plays a key role in the attention mechanism; Perform weighted summation on the features to obtain the aggregated representation after the attention mechanism : ; Among them, aggregation means By dynamically representing all features According to their importance weight The weighted sum is obtained; according to the aggregate representation formula, the aggregate representation of different features is calculated, and the weighted sum of the dynamic representation of each dimension of the node energy feature is calculated: ; in, is a dynamic representation of the node energy characteristics, is the corresponding importance weight; The link quality feature aggregation is expressed as follows: and bandwidth utilization feature aggregation is expressed as ; These aggregated representations not only retain the important information related to the target task in the original multi-dimensional features, but also reduce the data dimension and simplify the input complexity of the subsequent multi-objective optimization model; Use the principal component analysis method to reduce the dimension of the attention-weighted feature set and input the feature set: ; The PCA projection matrix is ​​obtained by eigenvalue decomposition of the characteristic covariance matrix , use the projection matrix to reduce the dimension of the original feature set, and calculate the feature vector after dimension reduction: ; in, is the original feature set, is the feature vector after dimensionality reduction; the output network state feature vector after dimensionality reduction is ; It is a projection matrix obtained by eigenvalue decomposition of the feature covariance matrix. After dimensionality reduction, the most important feature information is retained. Reducing the feature dimension helps the training and computational efficiency of subsequent models. At the same time, combined with the business request statistical feature vector in step 2, the logic for generating the joint feature vector of the current time window is: The network status features after dimensionality reduction are concatenated with the service request statistical features in step 2 to generate the joint feature vector of the current time window. The formula is: ; in, Represents the trend characteristics of network operation status within the time window, Represents the statistical characteristics of business requests within the time window, and the joint feature vector It combines network status characteristics and service request characteristics; Joint eigenvector The different parts of the network together provide a comprehensive view of network status and business requests, which helps in subsequent decision-making and optimization.

[0025] Step 4: Based on the joint feature vector of the current time window, a global optimization path planning model is constructed, and a genetic algorithm is used to solve the multi-objective optimization function, find the Pareto optimal solution, and finally output the global optimization path planning solution with the highest fitness; Combined with the joint feature vector of the current time window generated in step 3, a global optimization path planning model is constructed, and the genetic algorithm is used to solve the multi-objective optimization function. The logic of finding the Pareto optimal solution is: Joint eigenvector The network status characteristics and business request characteristics are combined, and based on these characteristics, a global optimization path planning model is constructed. In path planning, the following core optimization goals are first defined: Minimum path delay: The total delay of the selected path to ensure the real-time performance of the service Does not exceed the maximum allowable end-to-end delay of the service , that is, target 1 is ,in is the total delay from the source node to the destination node, which is determined by the link delay characteristics; That is, the total delay of the selected path, which reflects the total delay from the source node to the destination node, and describes the real-time performance and response speed of the network. The smaller it is, the faster the link transmission speed on the path is, which is conducive to meeting the timeliness requirements. It is the maximum allowable end-to-end delay of the service request, which is determined by the characteristics of the service request and constrains the upper limit of the path delay. like , the path does not meet the business timeliness requirements and will be filtered out. The larger it is, the looser the path delay limit allowed by the network; Minimize link congestion: Avoid overloading of links with high bandwidth utilization in the path to improve the overall stability of the network, that is, goal 2 is ,in is the maximum bandwidth utilization of all links in the path, Define the maximum safe bandwidth utilization threshold of the link and define the load upper limit of the link in the path to avoid network overload; , the maximum bandwidth utilization of all links in the path, reflects the load of the most congested link on the path. The larger its value, the closer the load of a link in the path is to its bandwidth limit, resulting in congestion; ,The path does not meet the link congestion requirement and will be filtered out. The larger it is, the higher the load on the link is allowed to be; Prolong network life: Prioritize paths with higher residual energy to prevent premature node failure, that is, goal 3 is Maximize the remaining energy, where It is the minimum residual energy of the nodes in the path, obtained through the node energy trend characteristics, reflecting the residual energy state of the weakest node in the path. The larger its value is, the more sufficient the energy of all nodes in the path is, which can extend the network life and node usage time; The optimization goal is to maximize , that is, select the path with higher residual energy to avoid premature node failure. If the residual energy of a node in the path is lower than the threshold, the path will be filtered; Path planning must meet the following constraints: Business timeliness constraints: , ensuring that the selected path can meet the timeliness requirements of the business; link quality constraints: ,in, is the packet loss rate of the ith link, is the maximum allowed packet loss rate threshold. The upper limit of the packet loss rate is determined by the network service requirements. represents a path; Reflects the reliability of link transmission. The larger the value, the more serious the packet loss of the link, resulting in data loss or retransmission. ,The path does not meet the link quality requirements and will be filtered out; Node energy constraints: ,in, is the residual energy of the ith node, is the minimum required threshold of node energy, which determines the lower limit of energy of all nodes in the path; Reflects the current energy state of the node. The smaller the value, the more energy the node consumes, which affects the service life of the path. , the path will be filtered out; Using the reduced-dimensional network state feature vector generated in step 3 and business request statistical feature vector , defining the dynamic weights of nodes and links in the network topology: Link Weight , comprehensively considering the link quality characteristics and bandwidth utilization: ; in, is the delay of link i, is the packet loss rate of link i, is the bandwidth utilization of link i, are weight parameters, and ; Reflects the combined impact of link delay, packet loss rate, and bandwidth utilization. The larger the delay of link i, the higher the weight. The packet loss rate of link i, the larger the value, the higher the weight. The bandwidth utilization of link i. The larger the value, the higher the weight. Link delay The weight of reflects the importance of delay to link weight. is the link packet loss rate The weight of reflects the importance of packet loss to the link weight. Bandwidth Utilization The weight of reflects the importance of bandwidth utilization to link weight; Set parameters in different network environments. If you are more sensitive to latency, such as applications with strong real-time requirements, set Larger, possible relationships are: If the packet loss rate has a significant impact on the application, such as voice calls or video streaming, you can set Larger, possible relationships are: ; If bandwidth utilization is the main consideration, such as large data transmission, you can set Larger, possible relationships are: ; Node Weight , calculated based on the node residual energy: ; in, is the residual energy of node i, when When it is small, the node weight increases, thus avoiding the selection of nodes with high load and fast energy consumption; Reflects the inverse ratio of the node's residual energy. The larger its value is, the lower the node's residual energy is and the higher its weight is, which reduces the possibility of selecting the node. The larger the node weight, The smaller it is, the more likely it is to choose nodes with sufficient remaining energy; The weighted shortest path algorithm based on feature weights is used to dynamically plan the optimal path by combining the feature vector in step 3 and the business characteristics in step 2. The specific logic is as follows: Modeling network topology as a weighted directed graph , define the node set represents all nodes in the network, Represents all links in the network, and the link weight is given by Definition, using the improved Dijkstra algorithm, planning the path based on the following path cost function: ; in, For path The total cost reflects the comprehensive weight of the selected path. is the set of all relay nodes on the path; Reflects the total weight of the links and nodes in the path. The smaller the value, the better the overall performance of the path. An increase in the weight of any link or node will result in a decrease in the total cost of the path. Increase, thereby reducing the priority of selecting this path; Build a network topology diagram , the link weight and node weight Map it to the graph, input the starting point, end point and delay constraint of the business request, and for each candidate path , calculate its path cost , filter out paths that do not meet the following conditions: ; Select the total cost from the remaining candidate paths The smallest path is taken as the preliminary optimal path; Choose the path with high fitness: In each generation, we first select the path with higher fitness from the current population. This is achieved through the tournament selection method. The specific process is as follows: The tournament selection method is used to select paths with higher fitness from the current population as the parents of the next generation; the selected parents are crossovered to generate new path individuals; the generated new individuals are randomly mutated to increase population diversity; these steps are iterated until the termination condition is met, that is, the maximum number of iterations is reached, and the path population is updated after each iteration, and the path with the highest fitness is retained; After obtaining the initial optimal path, the genetic algorithm is used to further optimize the path selection to find the Pareto optimal solution: Use genetic algorithm to find the Pareto optimal solution of the optimization path and randomly generate the initial path population , where each path Contains a series of nodes and links between them. For each path Calculation: Path Delay , path packet loss rate , the minimum residual energy of the path ; Fitness definition: The fitness of each path is determined by its latency, packet loss rate, and remaining energy. Usually, the fitness function takes these factors into consideration to ensure that the selected path performs well in multiple dimensions. The fitness function is a multi-objective optimization that reflects the performance of each path in terms of delay, packet loss rate, and residual energy. The fitness function is defined as: ; Use the Pareto dominance relationship to select the Pareto optimal solution set. If the path Dominant Path , then Excluded; Fitness optimization direction: Reflects the comprehensive performance of the path in terms of delay, packet loss rate, and remaining energy; and The smaller it is, the higher the fitness; The larger the The smaller it is, the higher the fitness; Crossover operation,The purpose of the crossover operation is to combine two parent paths into a new path individual,retaining the advantages of the parent paths and exploring new possibilities.,The crossover is usually done by exchanging some nodes or links in the parent paths.,The specific method can vary depending on the path encoding method; The mutation operation is to make random adjustments to the new path individuals after the crossover, which involves changing some nodes or links in the path, or randomly replacing some parts. The purpose of mutation is to prevent the algorithm from falling into the local optimal solution and explore new solution space; The Pareto optimal solution set means that no other path can perform better in all three dimensions: delay, packet loss rate, and residual energy. Mathematically, if the path Dominant Path , it means: ; That is, the path Either better than or equal to the path in all dimensions , and outperforms the path in at least one dimension ; Finally select a path from the Pareto optimal solution set , which performs best overall, that is, according to the normalized comprehensive fitness value To determine the optimal path, the specific expression is as follows: ; Among them, here It is the normalized comprehensive fitness value, which reflects the comprehensive performance of the path under multiple objectives, and finally selects the path with the best fitness as the optimal solution; Export Selected Path Detailed information, including: Path node sequence: list of nodes from the source node to the destination node and path link sequence: detailed information of each link on the path, which may include link delay, bandwidth utilization, etc., the total delay of the path That is, the sum of all link delays in the path, the average packet loss rate of the path That is, the average packet loss rate of all links on the path, the minimum residual energy of the path That is, the residual energy of the weakest node in the path; After the path planning solution is output, the optimization effect is analyzed: the delay, packet loss rate and residual energy changes of the initial path and the final path are compared to evaluate the utilization efficiency of the optimization solution on network resources; Delay change: By comparing the delay of the initial path and the optimized path, the optimization solution is evaluated to improve the timeliness of the service. Packet loss rate change: By comparing the change of packet loss rate, the reliability improvement of the optimized path is evaluated. Residual energy change: By comparing the minimum residual energy of the path, the utilization efficiency of the optimized path on network resources, especially node energy, is evaluated. By comparing the performance of the initial path and the optimized path, we analyze whether the optimization scheme has improved the utilization efficiency of network resources and whether it has allocated network resources more reasonably to improve the overall stability and life of the network.

[0026] See also Figure 2 The present invention also provides a dynamic optimization device for a centralized Mesh network, the device is used to implement the above-mentioned dynamic optimization method for a centralized Mesh network, including: The data collection module is used to set a time window and collect the operation sequence data and service request sequence data of the Mesh network within the time window. The network operation sequence data includes the remaining energy of each Mesh network node, the link quality between nodes and the current bandwidth usage rate of each link. The service request sequence data includes the data volume, priority and timeliness index data of all services. The data processing module is used to divide the collected service request time series data into high priority, medium priority and low priority services according to the service data volume, priority and timeliness indicators, and to count the number of requests, total data volume and service proportion of each type of service in the time window k, and combine them into a service request statistical feature vector; The feature extraction module is used to extract features from the collected network runtime data using a time series modeling method, obtain node energy trend features, link quality features, and bandwidth utilization features, use an attention mechanism to assign feature weights, and generate a network dimension reduction feature vector through dimensionality reduction processing. At the same time, it combines the service request statistical feature vector to generate a joint feature vector for the current time window; The optimization path module is used to build a global optimization path planning model based on the joint feature vector of the current time window, use genetic algorithms to solve multi-objective optimization functions, find the Pareto optimal solution, and finally output the global optimization path planning solution with the highest fitness.

[0027] The present invention further provides a storage medium, wherein the storage medium stores a computer program, and when the computer program is executed by a processor, each step in the above-mentioned method for dynamic optimization of a centralized Mesh network is implemented.

[0028] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0029] The above embodiments may be implemented in whole or in part by software, hardware, firmware or any other combination thereof. When implemented by software, the above embodiments may be implemented in whole or in part in the form of a computer program product. Those skilled in the art may appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein may be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed by hardware or software methods depends on the specific application and design constraints of the technical solution.

[0030] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, and may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0031] The above description is only a specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application.

Claims

1. A dynamic optimization method for a centralized Mesh network, characterized in that: The specific steps include: Step 1: Set a time window and collect the operation sequence data and service request sequence data of the Mesh network within the time window. The network operation sequence data includes the remaining energy of each Mesh network node, the link quality between nodes and the current bandwidth utilization rate of each link. The service request sequence data includes the data volume, priority and timeliness index data of all services. Step 2: Divide the collected service request time series data into high priority, medium priority and low priority services according to the service data volume, priority and timeliness indicators, and count the number of requests, total data volume and service proportion of each type of service in the time window k, and combine them into a service request statistical feature vector; Step 3: Use the time series modeling method to extract features from the collected network runtime time series data to obtain node energy trend features, link quality features, and bandwidth utilization features. Use the attention mechanism to assign feature weights, and generate network dimension reduction feature vectors through dimensionality reduction processing. Combined with the service request statistical feature vector, generate a joint feature vector for the current time window. Step 4: Based on the joint feature vector of the current time window, a global optimization path planning model is constructed, and a genetic algorithm is used to solve the multi-objective optimization function, find the Pareto optimal solution, and finally output the global optimization path planning solution with the highest fitness.

2. The method for dynamic optimization of a centralized Mesh network according to claim 1, characterized in that: The logic for collecting the Mesh network's runtime data and service request data is as follows: Taking the current time t as the starting point of the collection, define the time window k. Collect network operation timing data and business request timing data within the time range; Data collection is started through "periodic triggering": periodic triggering, at fixed time intervals, the central node broadcasts collection instructions to the entire network, requiring all nodes to report the latest status information; The specific contents of the collection are as follows: The node residual energy is set to The collection method is that the central node collects the energy status of each node through polling instructions every 1 minute, and ,in The total energy of the node. When the energy decrease rate is greater than 5% / minute, the node needs to actively report the current remaining energy; Link quality data measurement method: Nodes exchange heartbeat packets every 10 seconds, and use ACK and RTT to count packet loss rate and delay. The central node monitors the RSSI and SNR of link communication, and calculates the weighted average of the heartbeat packet measurement value and the central node monitoring value in a weight ratio of 7:3 as the final link quality data; The link bandwidth usage collection protocol is that the central node obtains the traffic statistics of the switch port through SNMP. The calculation formula is as follows: ； in, is the bandwidth utilization rate of the i-th link, is the actual occupied bandwidth of the ith link, is the total bandwidth of the ith link; The business request dataset for each time window contains the following information: Business priorities are set as , define the priority field in the service request header: 0 for high, 1 for medium, 2 for low, and the initiating node digital signature is required; the service timeliness is set to , the maximum allowed end-to-end delay for each request, which is automatically filled in by the service type mapping table; the service data volume is set to ,When the service is initiated, the transmission data size is recorded through the request header.

3. The method for dynamic optimization of a centralized Mesh network according to claim 2, characterized in that: The logic of dividing the collected business request time series data into high priority, medium priority and low priority businesses according to the business data volume, priority and timeliness indicators is as follows: Business requests are classified as high, medium, or low priority, and must meet any of the following conditions: For high priority business sets , if the business priority , that is, the request is clearly marked as high priority in the business request header, then the request needs to meet any of the following conditions: a. Business timeliness , that is, the maximum allowed end-to-end delay of the request is less than or equal to the delay threshold of the high priority; b. Business data volume and , that is, the amount of requested data is small and the timeliness requirement is high; among them, and are the set delay and quantity thresholds; For medium priority services , if the business request meets all of the following conditions: a. Business priority , that is, the request is clearly marked as medium priority in the business request header, b. Business timeliness , that is, the latency requirement of the request is between high priority and low priority; c. Business data volume , that is, the amount of requested data is moderate; among them, and is the set threshold value; For low priority services , if the business request meets any of the following conditions: a. If the business priority , that is, the request is clearly marked as low priority in the business request header, b. Business timeliness , that is, the maximum allowed end-to-end delay of the request is greater than the threshold value of the medium priority, and the service data volume , that is, large data needs to be transmitted, and the latency requirement is not high; The logic of counting the number of requests, total data volume, and business proportion of each type of business within the statistical time window k and combining them into a business request statistical feature vector is as follows: For each type of business, namely high priority, medium priority and low priority, the following three key characteristics are counted respectively: Count the number of requests belonging to this type of business in the current time window k , the calculation formula is: ； in, is a collection of certain types of business, that is or , Represents an index, when When , it refers to high priority business, when When , it refers to the medium priority business, when When , it refers to low priority business; The transmission data volume of all services of this type in the current time window k is accumulated, and the calculation formula is: ； in, Indicates the data volume of the nth request. Statistics of total data volume for a certain type of business; Calculate the ratio of the total data volume of this type of business to the total data volume of all businesses in the current time window. The formula is as follows: ； in, is the total data volume of all services, The proportion of data volume for a certain type of business; like ,but and ;like , then all proportions are set to ; Combine the statistical results into a structured feature vector. The business request statistical feature vector is defined as follows: ； in, are the number of requests for high, medium, and low priority services, respectively. are the total data volume of high, medium and low priority services respectively, They are the total data volume proportions of high, medium and low priority services respectively; After the feature vector is generated, the statistical results need to be checked for consistency: request quantity consistency, ,in, is the total number of requests in the current time window k; data consistency, ; Total percentage check, .

4. The method for dynamic optimization of a centralized Mesh network according to claim 2, characterized in that: The time series modeling method is used to extract features from the network runtime time series data collected in step 1. The logic for extracting node energy trend features, link quality features, and bandwidth utilization features is as follows: In the time window The node and link timing data within, including: node remaining energy timing data , represents the residual energy change of the i-th node in the time window; Link quality characteristic timing data includes packet loss rate , represents the change of packet loss rate of link i over time in the time window, delay , represents the end-to-end delay and delay jitter of link i in the time window , represents the amplitude of delay fluctuation of link i in the time window; bandwidth utilization time series data , represents the bandwidth utilization rate of link i in the time window; Use LSTM to model various types of time series data, build independent time series models for each type of time series data, and set the input sequence of each type of time series data to ; For each input sequence, the output of LSTM is the hidden state in the time window or the final hidden state : ； in, is the input of the time series feature, are the hidden state and candidate state of the previous time step, respectively. The dynamic representation of the temporal features is captured, and the hidden state of the final output is , indicating the time window Trend characteristics of The final hidden state of the LSTM model , extract the following trend feature set: node energy trend feature, assuming that the total number of nodes in the network is indivual, ; Link quality trend characteristics, , where i represents the trend characteristic of the i-th link quality; bandwidth utilization trend characteristic, ,in is the total number of links in the network; The logic of using the attention mechanism to assign feature weights and generating network dimension reduction feature vectors through dimensionality reduction processing is: For each type of feature, the weight of each dimension of data is dynamically calculated through the attention mechanism to associate the importance of the feature with the target task. The formula is as follows: ； in, is the dynamic representation of feature o, which is the feature vector obtained after forward propagation in the latter layer of the neural network. Feature o represents the dimensional feature of the latter layer in the network. is a trainable parameter in the attention layer, used to dynamically represent the features Perform a linear transformation, yes The transposed matrix of is the importance weight of feature o, and ,satisfy ; Perform weighted summation on the features to obtain the aggregated representation after the attention mechanism : ； Among them, aggregation means By dynamically representing all features According to their importance weight The weighted sum is obtained; according to the aggregate representation formula, the aggregate representation of different features is calculated, and the weighted sum of the dynamic representation of each dimension of the node energy feature is calculated: ； in, is a dynamic representation of the node energy characteristics, is the corresponding importance weight; The link quality feature aggregation is expressed as follows: and bandwidth utilization feature aggregation is expressed as .

5. The method for dynamic optimization of a centralized Mesh network according to claim 4, characterized in that: Use the principal component analysis method to reduce the dimension of the attention-weighted feature set and input the feature set: ; The PCA projection matrix is ​​obtained by eigenvalue decomposition of the characteristic covariance matrix , use the projection matrix to reduce the dimension of the original feature set, and calculate the feature vector after dimension reduction: ； in, is the original feature set, is the feature vector after dimensionality reduction; the output network state feature vector after dimensionality reduction is ; At the same time, combined with the business request statistical feature vector in step 2, the logic for generating the joint feature vector of the current time window is: The network status features after dimensionality reduction are concatenated with the service request statistical features in step 2 to generate the joint feature vector of the current time window. The formula is: ； in, Represents the trend characteristics of network operation status within the time window, Represents the statistical characteristics of business requests within the time window, and the joint feature vector It integrates network status characteristics and service request characteristics.

6. The method for dynamic optimization of a centralized Mesh network according to claim 5, characterized in that: Combined with the joint feature vector of the current time window generated in step 3, a global optimization path planning model is constructed, and the genetic algorithm is used to solve the multi-objective optimization function. The logic of finding the Pareto optimal solution is: Joint eigenvector The network status characteristics and business request characteristics are combined, and based on these characteristics, a global optimization path planning model is constructed. In path planning, the following core optimization goals are first defined: Minimum path delay: The total delay of the selected path to ensure the real-time performance of the service Does not exceed the maximum allowable end-to-end delay of the service , that is, target 1 is ,in is the total delay from the source node to the destination node, which is determined by the link delay characteristics; Minimize link congestion: Avoid overloading of links with high bandwidth utilization in the path to improve the overall stability of the network, that is, goal 2 is ,in is the maximum bandwidth utilization of all links in the path, The maximum safe bandwidth utilization threshold of the defined link; Prolong network life: Prioritize paths with higher residual energy to prevent premature node failure, that is, goal 3 is Maximize the remaining energy, where is the minimum residual energy of the nodes in the path, obtained through the node energy trend characteristics; Path planning must meet the following constraints: Business timeliness constraints: , ensuring that the selected path can meet the timeliness requirements of the business; link quality constraints: ,in, is the packet loss rate of the ith link, is the maximum allowed packet loss rate threshold, Represents a path; node energy constraint: ,in, is the residual energy of the ith node, is the minimum required threshold of node energy; Using the reduced-dimensional network state feature vector generated in step 3 and business request statistical feature vector , defining the dynamic weights of nodes and links in the network topology: Link Weight , comprehensively considering the link quality characteristics and bandwidth utilization: ； in, is the delay of link i, is the packet loss rate of link i, is the bandwidth utilization of link i, are weight parameters, and ; Node Weight , calculated based on the node residual energy: ； in, is the residual energy of node i, when When it is smaller, the node weight increases, thus avoiding the selection of nodes with high load and fast energy consumption.

7. The method for dynamic optimization of a centralized Mesh network according to claim 6, characterized in that: The weighted shortest path algorithm based on feature weights is used to dynamically plan the optimal path by combining the feature vector in step 3 and the business characteristics in step 2. The specific logic is as follows: Modeling network topology as a weighted directed graph , define the node set represents all nodes in the network, Represents all links in the network, and the link weight is given by Definition, using the improved Dijkstra algorithm, planning the path based on the following path cost function: ； in, For path The total cost reflects the comprehensive weight of the selected path. is the set of all relay nodes on the path; Build a network topology diagram , the link weight and node weight Map it to the graph, input the starting point, end point and delay constraint of the business request, and for each candidate path , calculate its path cost , filter out paths that do not meet the following conditions: ； Select the total cost from the remaining candidate paths The smallest path is taken as the preliminary optimal path; After obtaining the initial optimal path, the genetic algorithm is used to further optimize the path selection to find the Pareto optimal solution: Use genetic algorithm to find the Pareto optimal solution of the optimization path and randomly generate the initial path population , where each path Contains a series of nodes and links between them. For each path Calculation: Path Delay , path packet loss rate , the minimum residual energy of the path ; The fitness function is defined as: ； Use the Pareto dominance relationship to select the Pareto optimal solution set. If the path Dominant Path , then Excluded; The tournament selection method is used to select paths with higher fitness from the current population as the parents of the next generation; the selected parents are crossovered to generate new path individuals; the generated new individuals are randomly mutated to increase population diversity; these steps are iterated until the termination condition is met, that is, the maximum number of iterations is reached, and the path population is updated after each iteration, and the path with the highest fitness is retained; Finally select a path from the Pareto optimal solution set , which has the best overall performance, as described below: ； Among them, here is the normalized comprehensive fitness value; Export Selected Path Detailed information, including: path node sequence, path link sequence, total path delay , the average packet loss rate of the path , the minimum residual energy of the path ; After outputting the path planning solution, analyze the optimization effect: compare the delay, packet loss rate and remaining energy changes between the initial path and the final path, and evaluate the utilization efficiency of the optimization solution on network resources.

8. A dynamic optimization device for a centralized Mesh network, characterized in that: The device is used to execute a dynamic optimization method of a centralized Mesh network according to any one of claims 1 to 7, comprising: The data collection module is used to set a time window and collect the operation sequence data and service request sequence data of the Mesh network within the time window. The network operation sequence data includes the remaining energy of each Mesh network node, the link quality between nodes and the current bandwidth usage rate of each link. The service request sequence data includes the data volume, priority and timeliness index data of all services. The data processing module is used to divide the collected service request time series data into high priority, medium priority and low priority services according to the service data volume, priority and timeliness indicators, and to count the number of requests, total data volume and service proportion of each type of service in the time window k, and combine them into a service request statistical feature vector; The feature extraction module is used to extract features from the collected network runtime data using a time series modeling method, obtain node energy trend features, link quality features, and bandwidth utilization features, use an attention mechanism to assign feature weights, and generate a network dimension reduction feature vector through dimensionality reduction processing. At the same time, it combines the service request statistical feature vector to generate a joint feature vector for the current time window; The optimization path module is used to build a global optimization path planning model based on the joint feature vector of the current time window, use genetic algorithms to solve multi-objective optimization functions, find the Pareto optimal solution, and finally output the global optimization path planning solution with the highest fitness.

9. A storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by a processor, a dynamic optimization method for a centralized Mesh network as described in any one of claims 1 to 7 is implemented.

Images