Cluster weak network micro cooperation method and system based on adaptive fusion of guidance strategy and local experience

Abstract

The invention discloses a cluster weak network micro cooperation method and system based on adaptive fusion of a guidance strategy and local experience. The method comprises the following steps: generating a lightweight guidance plan on line based on an improved genetic algorithm, and dynamically updating in combination with task requirements and network states; sensing cluster nodes and environment states through dynamic weight fuzzy comprehensive evaluation, and dynamically adjusting an execution strategy; on the basis of communication quality constraint Q-learning, local intuitive micro-collaborative execution of nodes is realized in a weak network environment; according to the guiding confidence and the local credibility, dynamically distributing weights, and fusing and outputting an optimal decision through a D-S evidence theory; network states (bandwidth, packet loss rate and time delay) are monitored in real time, and network state feedback is provided for each module. Through four core technologies of lightweight guidance plan generation, context awareness dynamic adjustment, local intuitive micro-collaboration and adaptive weight fusion, the reliability and adaptability problems of cluster collaboration in a weak network environment are effectively solved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of cluster cooperative control, and in particular to a cluster weak network micro-cooperative method and system based on adaptive fusion of guidance strategies and local experience. BACKGROUND

[0002] With the wide application of unmanned cluster technology in fields such as logistics distribution, emergency rescue, and environmental monitoring, the reliability and adaptability of cluster cooperative control have become a core technical bottleneck. Especially in a weak network environment (such as a communication bandwidth ≤ 1 Mbps, a packet loss rate ≥ 30%, and a time delay ≥ 500 ms), traditional cooperative methods face significant challenges:

[0003] 1. Centralized guidance strategies rely on stable communication links to transmit global instructions, and instruction loss or delay in a weak network environment can lead to cluster cooperative collapse, and heavy pre-plan files consume a large amount of communication resources, further exacerbating network congestion.

[0004] 2. Distributed execution schemes mostly use fixed rules for adjustment, lack the ability to perceive dynamic situations (defined as the comprehensive situation of “cluster node state and external environment state”, such as sudden appearance of obstacles and node failure), and cannot adapt to environmental changes in real time.

[0005] 3. Existing local experience-driven methods are mostly based on a single reinforcement learning model, without considering the limitations of experience interaction in a weak network, and are prone to local optimal traps, leading to a decline in the overall cooperative performance of the cluster. Existing “reinforcement learning based on communication quality” technology only corrects the reward function through fixed coefficients, lacks risk adaptation and experience screening mechanisms, and has insufficient weak network adaptability.

[0006] 4. The fusion of guidance strategies and local experience mostly uses fixed weights, which cannot be dynamically adjusted according to network quality and situation complexity, and has insufficient adaptability. Existing situation perception methods mostly use “fixed window weight + independent index evaluation”, which is difficult to balance the sensitivity of sudden situations and the stability of evaluation, and ignores the coupling interference between indexes.

[0007] Therefore, there is an urgent need for a cluster cooperative method that can achieve lightweight guidance, dynamic situation adaptation, local intelligent cooperation, and adaptive fusion in a weak network environment. SUMMARY

[0008] The present application aims to address the deficiencies in the prior art by providing a cluster weak network micro-cooperative method based on adaptive fusion of guidance strategies and local experience. Through four core technologies of lightweight guidance pre-plan generation, situation perception dynamic adjustment, local intuition micro-cooperation, and adaptive weight fusion, the method solves the problems of reliability and adaptability of cluster cooperative control in a weak network environment.

[0009] The technical solution for achieving the object of the application is: on the one hand, a cluster weak network micro-collaboration method based on adaptive fusion of guidance strategy and local experience is provided, and the method comprises the following steps:

[0010] Step 1, online lightweight guidance plan generation based on improved genetic algorithm: combining cluster task demand and network initial state, a lightweight guidance plan considering global optimization and transmission efficiency is generated;

[0011] Step 2, situation awareness adjustment based on dynamic weight fuzzy comprehensive evaluation: during distributed execution, the cluster node state and environment situation are sensed in real time, and the execution strategy is dynamically adjusted;

[0012] Step 3, local intuition micro-collaboration based on communication quality constraint Q-learning: under weak network environment, nodes autonomously realize micro-collaboration execution of local cluster based on historical experience and real-time communication quality;

[0013] Step 4, weight adaptive fusion based on confidence degree dynamic allocation: according to the execution confidence degree of the guidance plan and the decision confidence degree of the local experience, the fusion weight is dynamically allocated, and the optimal collaborative decision is output.

[0014] Further, the specific process of step 1 comprises:

[0015] Step 1-1, define the plan coding rule: adopt binary-real hybrid coding, wherein the binary bits represent the task execution sequence, and the real values represent the node resource allocation ratio;

[0016] Step 1-2, construct the fitness function: comprehensively consider the task completion rate, network transmission overhead and resource utilization rate, introduce dynamic weight coefficients to adapt to different task scenarios, and the fitness function F is as follows: ; In the formula, is the task completion rate, the value range is [0, 1], is the current completed task amount / total task amount; B is the plan transmission bandwidth occupation, unit: Kbps, B=plan file size / transmission time length; is the network maximum available bandwidth, unit: Kbps; U is the resource utilization rate, the value range is [0, 1]; , are all dynamic weight coefficients, which are determined according to the task type through the analytic hierarchy process;

[0017] Step 1-3, introduce plan simplification operator: perform redundant information pruning on the optimal individuals generated by genetic algorithm iteration, and the process comprises: ① delete repeated instructions; ② eliminate low-weight constraint items with resource allocation ratio≤preset resource allocation ratio threshold; ③ merge adjacent node same type task sequence;

[0018] Step 1-4, online dynamic update: every interval T time step, according to the network state feedback, repeat the implementation of step 1-1 to step 1-3 to update the guide plan, realize the real-time matching of the plan and the network state.

[0019] Further, the encoding length in step 1-1 is controlled within 1 / 3 of the traditional encoding.

[0020] Further, step 2 specifically includes:

[0021] Step 2-1, define the situation evaluation index system: including four core indexes of node communication quality C, resource remaining amount R, environmental interference degree I and task progress Q, each index is divided into four evaluation levels of "excellent, good, medium and poor";

[0022] Step 2-2, build a double-window dynamic weight calculation model: adopt parallel calculation of short-term sensitive window and long-term stable window, and combine with the situation fluctuation coefficient for adaptive fusion; specifically including:

[0023] ①Calculate the single-window variation coefficient:

[0024] For the short-term sensitive window, ;

[0025] For the long-term stable window, ;

[0026] In the formula, is the short-term variation coefficient of the i-th core index of the short-term sensitive window, is the long-term variation coefficient of the i-th core index of the long-term stable window, are the standard deviation of the i-th core index of the short-term sensitive window and the standard deviation of the i-th core index of the long-term stable window, are the mean of the i-th core index of the short-term sensitive window and the mean of the i-th core index of the long-term stable window;

[0027] ②Calculate the situation fluctuation coefficient :

[0028]

[0029] ③The fusion weight calculation is shown in the following formula:

[0030]

[0031] Wherein, is the weight of the i-th core index, forming a dynamic weight matrix ;

[0032] Step 2-3, index correlation correction: calculate the correlation between indexes by mutual information , if the correlation between the i-th index and the j-th index is greater than a preset threshold, use principal component analysis to decouple the weights, and the correction formula is: ; ; represents the correlation between the i-th core index and the j-th core index.

[0033] Step 2-4, fuzzy comprehensive evaluation calculation: determine the evaluation level membership degree of each core index by the membership function.

[0034] Construct a fuzzy evaluation matrix , combined with the corrected dynamic weight matrix , calculate the comprehensive evaluation result by weighted average method , as shown in the following formula:

[0035]

[0036] In the formula, is a fuzzy synthesis operator, is the membership degree of the i-th core index to the j-th evaluation level, =1,2,3,4.

[0037] Step 2-5, execute strategy adjustment: according to the comprehensive evaluation result , if the evaluation level is "poor", trigger strategy adjustment to adjust the amplitude and evaluation result negatively.

[0038] Further, the membership function expression in step 2-4 is as follows:

[0039]

[0040]

[0041]

[0042]

[0043] For different core indexes, adjust the parameters according to the quantitative standard , , , ; , , and represent the membership degrees of "excellent, good, medium, and poor" four evaluation levels of each core index.

[0044] Further, step 3 specifically includes: ​

[0045] Step 3-1, define state space and action space: state space S includes local node state, neighbor node state and environment state; action space A includes cooperative execution action and communication adjustment action;

[0046] Step 3-2, build communication quality constraint reward function: the reward function considers the cooperative effect and communication cost comprehensively, as shown in the following formula:

[0047]

[0048] wherein, represents the reward function, is a communication quality factor, with a value range of [0, 1], is a cooperative effect reward, is a communication cost penalty, is a decision risk value, with a value range of [0, 1], predicted by a random forest model trained by historical failure data, with the input being the current state and action , and the output being the action execution failure probability, is a risk weight, dynamically adjusted according to the task importance level;

[0049] Step 3-3, improve value update rule: introduce experience decay coefficient , build value update formula as follows:

[0050]

[0051] wherein, is a learning rate, is a discount factor, is the next state, is the optimal action in the next state, is an experience decay coefficient; represents the Q value that can be obtained after taking action a in state s, i.e. the expected cumulative return, represents the Q value that can be obtained after taking action in state , represents the maximum expected cumulative return, i.e. the maximum Q value, that can be obtained among all possible executed actions in the next state ;

[0052] Step 3-4, cross-node experience distillation mechanism: ① Experience screening: only receive experience samples of nodes with signal quality ≥ threshold value, filter low credibility experience; ② Distillation fusion: adopt knowledge distillation technology, compress the filtered multi-node experience into a unified experience model, i.e. teacher model, and the local node learns experience through the loss function optimization of student model and teacher model;

[0053] Step 3-5, local intuitive decision: when the communication delay > preset delay threshold or the packet loss rate > preset packet loss rate, the node executes local micro-collaboration based on the action with the maximum current value, and caches the decision experience for subsequent weight fusion.

[0054] Further, the communication quality factor is negatively correlated with the packet loss rate P.

[0055] Further, in step 4, the confidence-based weight adaptive fusion guides the decision and the local decision through D-S evidence theory fusion, and dynamically allocates the fusion weight, specifically including:

[0056] Step 4-1, calculate the decision confidence, including the guide strategy confidence and the local experience credibility ;

[0057] Among them, the guide strategy confidence is calculated based on the success rate of pre-plan execution and the network matching degree:

[0058]

[0059] The local experience credibility is calculated based on the historical decision accuracy rate and the current situation matching degree:

[0060]

[0061] In the formula, is the historical execution success rate of the guide pre-plan, is the current network bandwidth, is the historical accuracy rate of the local decision, is the similarity between the current situation and the historical optimal situation, is the pre-plan transmission bandwidth occupation, is the maximum available bandwidth of the network;

[0062] Step 4-2, dynamically allocate fusion weight: the weight is positively correlated with the confidence, and satisfies + =1, as shown in the following formula:

[0063]

[0064] wherein, is the guidance policy weight, is the local experience weight;

[0065] Step 4-3, D-S evidence theory fusion: taking the guidance decision and the local decision as two independent evidence sources, calculating the basic probability assignment function, and fusing to obtain the final decision through the Dempster combination rule, as shown in the following formula:

[0066]

[0067] wherein, is the basic probability assignment of the decision A after fusion, , respectively, are the basic probability assignments of the guidance decision and the local decision, the basic probability assignment of the guidance decision is calculated based on the guidance policy confidence , and the basic probability assignment of the local decision is calculated based on the local experience reliability . Further, the system also comprises: On the other hand, a cluster weak network micro-collaboration system is provided, and the system comprises:

[0068] A preplan generation module is configured to generate a lightweight guidance preplan online based on an improved genetic algorithm, and dynamically update the preplan in combination with a task demand and a network state;

[0069] A situation adjustment module is configured to perceive a cluster node and an environment state through dynamic weight fuzzy comprehensive evaluation, and dynamically adjust an execution strategy;

[0070] A local collaboration module is configured to realize micro-collaboration execution of a node local intuition in a weak network environment based on a communication quality constraint Q-learning;

[0071] A weight fusion module is configured to dynamically allocate weights according to a guidance confidence and a local reliability, and output an optimal decision through D-S evidence theory fusion.

[0072] Further, the system also comprises:

[0073] A communication adaptation module is configured to monitor a network state in real time, and provide network state feedback for each module.

[0074] Compared with the prior art, the present application has the following advantages:

[0075] The present application has the following advantages:

[0076] ​(1) Through the improved genetic algorithm, a lightweight guide plan is generated, the plan volume can be reduced from 128KB to 72KB, the reduction is 44%, the transmission cost is reduced by more than 40%, and the plan update success rate in a weak network environment is improved by 50%.

[0077] (2) Through dynamic weight fuzzy comprehensive evaluation, the real-time adjustment of situation perception is realized, and through a large number of simulation verification, the strategy adaptation accuracy is more than 92%.

[0078] (3) Through the communication quality constraint Q-learning, the reliability of local decision-making in a weak network is enhanced, and the micro-collaboration execution success rate is improved by 35%.

[0079] (4) Through the confidence adaptive fusion weight, the conflict problem of guide decision and local experience is solved, and finally the collaborative task completion rate is improved by 28%-45%, which is suitable for weak network scenes such as emergency rescue and field exploration.

[0080] The application will be further described in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF DRAWINGS

[0081] Figure 1 It is a flowchart of a cluster weak network micro-collaboration method based on adaptive fusion of guide strategy and local experience of the application in an embodiment.

[0082] Figure 2 It is a lightweight plan generation flowchart of the improved genetic algorithm in an embodiment.

[0083] Figure 3 It is a situation adjustment logic diagram of dynamic weight fuzzy comprehensive evaluation in an embodiment.

[0084] Figure 4 It is a local collaborative decision-making model diagram of communication quality constraint Q-learning in an embodiment.

[0085] Figure 5 It is a decision output result diagram of weight adaptive fusion in an embodiment.

[0086] Figure 6 It is a structure diagram of a cluster weak network micro-collaboration system based on adaptive fusion of guide strategy and local experience in an embodiment. DETAILED DESCRIPTION

[0087] In order to make the purpose, technical scheme and advantages of the application clearer, the application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the application and not to limit the application.

[0088] In an embodiment, the application is combined with Figure 1The application provides a cluster weak network micro-collaboration method based on adaptive fusion of guidance strategy and local experience, and the method comprises the following steps:

[0089] Step 1: online lightweight guidance plan generation based on an improved genetic algorithm: combining cluster task requirements and initial network state, a lightweight guidance plan is generated, which takes into account global optimization and transmission efficiency;

[0090] Step 2: situation awareness adjustment based on dynamic weight fuzzy comprehensive evaluation: during distributed execution, the cluster node state and environmental situation are sensed in real time, and the execution strategy is dynamically adjusted;

[0091] Step 3: local intuition micro-collaboration based on communication quality constraint Q-learning: under a weak network environment, nodes autonomously realize micro-collaboration execution of a local cluster based on historical experience and real-time communication quality;

[0092] Step 4: weight adaptive fusion based on confidence degree dynamic allocation: according to the execution confidence degree of the guidance plan and the decision confidence degree of the local experience, the fusion weight is dynamically allocated, and an optimal collaborative decision is output.

[0093] Further, in one of the embodiments, the method is combined with Figure 2 The specific process of step 1 comprises the following steps.

[0094] Step 1-1: defining a plan coding rule: a binary-real hybrid coding is adopted, wherein a binary bit represents a task execution sequence, and a real value represents a node resource allocation ratio;

[0095] Preferably, in some embodiments, the coding length in step 1-1 is controlled to be within 1 / 3 of the traditional coding.

[0096] Step 1-2: constructing a fitness function: considering the task completion rate, network transmission overhead and resource utilization rate, a dynamic weight coefficient is introduced to adapt to different task scenarios, and the fitness function F is as follows:

[0097]

[0098] In the formula, P is the task completion rate, the value range is [0, 1], P = current completed task amount / total task amount; B is the transmission bandwidth occupation of the plan, the unit is Kbps, B = plan file size / transmission time length; is the maximum available bandwidth of the network (monitored in real time by a communication adaptation module), the unit is Kbps; U is the resource utilization rate, the value range is [0, 1]; , are dynamic weight coefficients, which are determined according to the task type through an analytic hierarchy process;

[0099] Preferably, in some embodiments, The specific values are shown in Table 1.

[0100] Table 1 Specific values

[0101]

[0102] Step 1-3, introduce a pre-plan reduction operator: perform redundancy information pruning on the optimal individual generated by the genetic algorithm iteration, the process includes: ① delete duplicate instructions; ② eliminate low-weight constraint items with resource allocation ratio ≤ preset resource allocation ratio threshold; ③ merge adjacent node task sequences of the same type;

[0103] Here, the volume of the reduced pre-plan can be reduced by more than 40%.

[0104] Preferably, in some embodiments, the preset resource allocation ratio threshold is 5%.

[0105] Step 1-4, online dynamic update: every interval T time step, according to the network state feedback, repeat steps 1-1 to 1-3 to update the guide pre-plan, realize the real-time matching of the pre-plan and the network state.

[0106] Preferably, in some embodiments, T ∈ [5, 30] s, when the packet loss rate ≥ 40%, T = 5 s, and when the packet loss rate < 30%, T = 30 s.

[0107] Further, in combination with Figure 3 , step 2 specifically includes:

[0108] Step 2-1, define the context evaluation index system: including four core indexes of node communication quality C, resource remaining amount R, environmental interference degree I and task progress Q, each index is divided into four evaluation levels of “excellent, good, medium and poor”;

[0109] Preferably, in some embodiments, the quantification standard table is shown in Table 2.

[0110] Table 2 Quantification standard table

[0111]

[0112] Step 2-2, build a double-window dynamic weight calculation model: adopt parallel calculation of a short-term sensitive window (N1 = 5, capturing sudden context) and a long-term stable window (N2 = 20, suppressing short-term fluctuations), combined with context fluctuation coefficient adaptive fusion; specifically includes:

[0113] ① Calculate the single-window coefficient of variation:

[0114] For the short-term sensitive window, ;

[0115] For long-term stable window, ;

[0116] In the formula, is the short-term coefficient of variation of the i-th core index in the short-term sensitive window, is the long-term coefficient of variation of the i-th core index in the long-term stable window, is the standard deviation of the i-th core index in the short-term sensitive window, respectively, the standard deviation of the i-th core index in the long-term stable window, is the mean of the i-th core index in the short-term sensitive window, respectively, the mean of the i-th core index in the long-term stable window;

[0117] ②Calculate the context fluctuation coefficient :

[0118]

[0119] Here, the greater the fluctuation , the greater the value range [0.3, 0.7].

[0120] ③The fusion weight is calculated as follows:

[0121]

[0122] Where, is the weight of the i-th core index; thus obtaining the weight matrix W;

[0123] Step 2-3, index association correction: calculate the association degree between indexes by mutual information (value range [0, 1]), for highly associated index pairs ≥ preset threshold, use principal component analysis to decouple the weight, and the correction formula is: ; represents the association degree between the i-th core index and the j-th core index.

[0124] Preferably, in some embodiments, the preset threshold is 0.6.

[0125] Step 2-4, fuzzy comprehensive evaluation calculation: determine the evaluation grade membership of each core index by membership function;

[0126] Construct a fuzzy evaluation matrix , combined with the corrected dynamic weight matrix , the weighted average method is used to calculate the comprehensive evaluation result , as shown in the following formula:

[0127]

[0128] In the formula, is the fuzzy synthesis operator, The membership degree of the ith core index to the jth evaluation level, =1,2,3,4;

[0129] Preferably, in some embodiments, the membership function expression in step 2-4 is as follows:

[0130]

[0131]

[0132]

[0133]

[0134] According to the quantitative standard, adjust the parameters for different core indexes and represent the membership degrees of the four evaluation levels of “excellent, good, medium, and poor” for each core index, respectively.

[0135] Exemplarily, taking the communication quality C (packet loss rate) as an example, the membership function expression is as follows:

[0136]

[0137]

[0138]

[0139]

[0140] Step 2-5, execute strategy adjustment: according to the comprehensive evaluation result , if the evaluation level is “poor”, trigger the strategy adjustment (such as increasing resource allocation, adjusting the communication frequency) to adjust the amplitude and the negative correlation with the evaluation result.

[0141] Preferably, in some embodiments, the adjustment rules are shown in Table 3.

[0142] Table 3 Adjustment rules

[0143]

[0144] ​​​​​​​Further, in one embodiment, in step 3, by "communication quality-decision risk" double coupling Q-learning, by fusing communication quality constraints and decision risk evaluation, combined with cross-node experience distillation, realize weak network reliable decision-making, combined with Figure 4 , specifically including:

[0145] Step 3-1, define state space and action space: state space S includes local node state (such as resource, power), neighbor node state (such as communication quality, task state) and environment state (such as interference, obstacle); action space A includes cooperative execution action (such as data forwarding, task division) and communication adjustment action (such as frequency switching, power adjustment);

[0146] Step 3-2, build communication quality constraint reward function: the reward function comprehensively considers the cooperative effect and communication cost, as shown in the following formula:

[0147]

[0148] Among them, represents the reward function, is the communication quality factor, the value range is [0, 1], is the cooperative effect reward (based on task progress improvement amount calculation), is the communication cost penalty (based on bandwidth occupation amount calculation), is the decision risk value, the value range is [0, 1], which is predicted by a random forest model trained by historical failure data, the input is the current state and action , the output is the action execution failure probability, is the risk weight (preferably, the value range is [0.1, 0.3]), which is dynamically adjusted according to the task importance level (emergency rescue task =0.3, environmental monitoring =0.1);

[0149] Step 3-3, improve value update rule: introduce experience decay coefficient (in weak network increase, enhance the weight of historical experience), build value update formula as follows:

[0150]

[0151] Among them, is the learning rate (preferably, the value range is [0.1, 0.3]), is the discount factor (preferably, the value range is [0.7, 0.9]), is the next state, the optimal action for the next state, an experience decay coefficient (increased in weak network environment); represents the Q value that can be obtained after taking action a in state s, i.e., the expected cumulative return, represents the Q value that can be obtained after taking action a in state s, represents the Q value that can be obtained after taking action a in state s, represents the maximum expected cumulative return, i.e., the maximum Q value, that can be obtained among all possible actions a in the next state s,

[0152] Step 3-4, cross-node experience distillation mechanism: ① Experience screening: only receive experience samples from nodes with communication quality ≥ threshold (preferably, packet loss rate ≤ 40%), filter low-confidence experience; ② Distillation fusion: use knowledge distillation technology to compress the filtered multi-node experience into a unified experience model, i.e., a teacher model, and the local node learns experience through the loss function optimization of the student model and the teacher model;

[0153] Step 3-5, local intuitive decision: when the communication delay > preset delay threshold or the packet loss rate > preset packet loss rate, the node executes local micro-collaboration based on the action with the maximum current value, and caches the decision experience for subsequent weight fusion.

[0154] Preferably, in some embodiments, the preset delay threshold is 200 ms and the preset packet loss rate is 40%.

[0155] Preferably, in some embodiments, the communication quality factor is negatively correlated with the packet loss rate P, .

[0156] Further, in one of the embodiments, in step 4, the confidence-based weight adaptive fusion guides the decision and the local decision through D-S evidence theory fusion, dynamically allocates the fusion weight, and combines Figure 5 , specifically including:

[0157] Step 4-1, calculate the decision confidence, including the guide strategy confidence and the local experience confidence ;

[0158] The guide strategy confidence is calculated based on the pre-plan execution success rate and the network matching degree:

[0159]

[0160] The local experience confidence ​​​Calculated based on the matching degree between historical decision accuracy and current context:

[0161]

[0162] In the formula, To guide the historical success rate of the contingency plan, Given the current network bandwidth, For the historical accuracy of local decision-making, The similarity between the current situation and the historical best situation. To contingency plan for transmission bandwidth usage, The maximum available bandwidth of the network;

[0163] Step 4-2, Dynamically allocate fusion weights: The weights are positively correlated with the confidence level and satisfy the following conditions: + =1, as shown in the following formula:

[0164]

[0165] In the formula, To guide the strategy weights, Weighted by local experience;

[0166] Step 4-3, Dempster Evidence Theory Fusion: Treating the guided decision and the local decision as two independent sources of evidence, calculate the basic probability assignment function, and fuse them using the Dempster synthesis rule to obtain the final decision, as shown in the following equation:

[0167]

[0168] In the formula, To assign the basic probability to decision A after fusion. , These are the basic probability assignments for guided decision-making and local decision-making, respectively. Basic probability allocation Based on the confidence level of the guidance strategy Computation, Local Decision Basic probability allocation Reliability based on local experience calculate.

[0169] In one embodiment, a clustered weak network micro-cooperative system is provided, the system comprising:

[0170] The contingency plan generation module is used to generate lightweight guidance contingency plans online based on an improved genetic algorithm, and to dynamically update them in combination with task requirements and network status.

[0171] The situation adjustment module is configured to adjust the execution strategy dynamically by fuzzy comprehensive evaluation of the perception cluster node and the environment state through dynamic weights.

[0172] The local coordination module is configured to realize the micro-coordination execution of the node local intuition in a weak network environment based on the communication quality constraint Q-learning.

[0173] The weight fusion module is configured to dynamically allocate weights according to the guidance confidence and the local confidence, and output an optimal decision through D-S evidence theory fusion.

[0174] Further, the system further comprises:

[0175] The communication adaptation module is configured to monitor the network state in real time and provide network state feedback for each module.

[0176] Further, in one of the embodiments, the preplan generation module comprises:

[0177] The encoding unit is configured to encode the preplan by using a binary-real hybrid encoding rule;

[0178] The fitness calculation unit is configured to construct a fitness function and calculate a fitness value;

[0179] The simple element is configured to trim the redundant information of the optimal individual through a preplan simplification operator;

[0180] The updating unit is configured to update the guidance preplan dynamically online according to the network state feedback.

[0181] Further, in one of the embodiments, the situation adjustment module comprises:

[0182] The index collection unit is configured to collect the node communication quality, resource remaining amount, environment interference degree and task progress index;

[0183] The weight calculation unit is configured to calculate the index weight through a double-window dynamic weight calculation model and perform index correlation correction;

[0184] The fuzzy evaluation unit is configured to obtain a comprehensive evaluation result through a membership function and fuzzy comprehensive evaluation calculation;

[0185] The strategy adjustment unit is configured to perform corresponding strategy adjustment measures according to the comprehensive evaluation result.

[0186] Further, in one of the embodiments, the local coordination module comprises:

[0187] The space definition unit is configured to define the state space and the action space;

[0188] The reward calculation unit is configured to construct a communication quality constraint reward function and calculate a reward value;

[0189] a Q value updating unit configured to update the Q value based on an improved Q value updating rule;

[0190] an experience fusion unit configured to filter and fuse multi-node experiences through a cross-node experience distillation mechanism;

[0191] a local decision unit configured to execute a local intuitive decision and cache experiences when a trigger condition is met.

[0192] Further, in one of the embodiments, the weight fusion module comprises:

[0193] a confidence calculation unit configured to calculate a guidance strategy confidence and a local experience credibility, respectively;

[0194] a weight allocation unit configured to dynamically allocate a fusion weight according to the confidence;

[0195] a decision fusion unit configured to fuse the guidance decision and the local decision through a D-S evidence theory to output an optimal decision.

[0196] Further, in one of the embodiments, the communication adaptation module monitors network states using an SNMP protocol, a sampling frequency of 10 Hz, a state feedback delay of ≤200 ms, and real-time pushing of network state data such as bandwidth, packet loss rate, and time delay to other modules.

[0197] For specific definitions of the cluster weak network micro-collaboration system based on adaptive fusion of guidance strategies and local experiences, reference can be made to the definitions of the cluster weak network micro-collaboration method based on adaptive fusion of guidance strategies and local experiences in the foregoing, which will not be repeated here. Each module in the cluster weak network micro-collaboration system based on adaptive fusion of guidance strategies and local experiences can be realized in whole or in part by software, hardware, or a combination thereof. Each module can be embedded in or independent of a processor in a computer device in hardware form, or can be stored in a memory in a computer device in software form, so as to be called and executed by a processor to perform operations corresponding to each module.

[0198] In one embodiment, a computer device is provided, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor implements the following when executing the computer program:

[0199] Step 1: Online lightweight guidance plan generation based on an improved genetic algorithm: combining cluster task requirements and initial network states, a lightweight guidance plan is generated that takes into account global optimization and transmission efficiency;

[0200] Step 2: Situation awareness adjustment based on dynamic weight fuzzy comprehensive evaluation: during distributed execution, the cluster node states and environmental situations are sensed in real time, and the execution strategy is dynamically adjusted.

[0201] Step 3, local intuitive micro-collaboration based on communication quality constraint Q-learning: in a weak network environment, nodes autonomously implement micro-collaboration execution of a local cluster based on historical experience and real-time communication quality;

[0202] Step 4, adaptive fusion of weights based on dynamic allocation of confidence: according to the execution confidence of the guidance plan and the decision confidence of the local experience, dynamically allocate the fusion weights, and output the optimal collaborative decision.

[0203] For specific definitions of each step, refer to the definitions of the cluster weak network micro-collaboration method based on adaptive fusion of guidance strategies and local experience in the foregoing, which will not be repeated here.

[0204] In one embodiment, a computer-readable storage medium is provided, which stores a computer program that is executed by a processor to implement:

[0205] Step 1, online lightweight guidance plan generation based on improved genetic algorithm: combining cluster task requirements and network initial state, generate a lightweight guidance plan that takes into account global optimization and transmission efficiency;

[0206] Step 2, context-aware adjustment based on dynamic weight fuzzy comprehensive evaluation: during distributed execution, real-time perception of cluster node state and environment context, dynamic adjustment of execution strategy;

[0207] Step 3, local intuitive micro-collaboration based on communication quality constraint Q-learning: in a weak network environment, nodes autonomously implement micro-collaboration execution of a local cluster based on historical experience and real-time communication quality;

[0208] Step 4, adaptive fusion of weights based on dynamic allocation of confidence: according to the execution confidence of the guidance plan and the decision confidence of the local experience, dynamically allocate the fusion weights, and output the optimal collaborative decision.

[0209] For specific definitions of each step, refer to the definitions of the cluster weak network micro-collaboration method based on adaptive fusion of guidance strategies and local experience in the foregoing, which will not be repeated here.

[0210] As a specific example, in some embodiments, the present application is further verified and described.

[0211] Embodiment 1

[0212] Embodiment 1 provides a cluster weak network micro-collaboration method based on adaptive fusion of guidance strategies and local experience, applied to a field exploration cluster system composed of 10 unmanned aerial vehicle nodes, in a weak network scenario with a packet loss rate of 30%-50% and a bandwidth of 512 Kbps-1 Mbps, specifically including:

[0213] S1, online lightweight guidance plan generation based on improved genetic algorithm;

[0214] S101, binary-real hybrid coding is adopted, binary bits (8 bits) represent exploration path sequence (for example, "0101" represents that the 2nd and 4th node groups perform exploration tasks), and real values (2 decimal places) represent power distribution ratio (for example, "0.65" represents that 65% power is used for task execution);

[0215] S102, set fitness function weight =0.5 (task completion rate), =0.3 (transmission overhead), =0.2 (resource utilization rate), under initial network state =0.6, =300 Kbps, =1 Mbps, =0.7, substitute into fitness function formula calculation =0.5x0.6+0.3x(1-300 / 1000)+0.2x0.7=0.65;

[0216] S103, plan simplification operator deletes repeated path instructions (for example, "0101" coding is retained once for two consecutive times), and after simplification, the plan volume is reduced from 128 KB to 72 KB, with a reduction of 44%;

[0217] S104, set update period T=10 s, update the plan every 10 s according to the network bandwidth change feedback by the node, and when the bandwidth decreases to 512 Kbps, recalculate the fitness function to adjust the resource allocation ratio.

[0218] S2, dynamic weight fuzzy comprehensive evaluation based on double window correlation correction;

[0219] S201, communication quality (C), power remaining amount (R), environmental interference degree (I), and exploration progress (Q) are evaluation indexes, and the evaluation grades of each index are shown in Table 4 as follows:

[0220] Table 4 Evaluation grade division of each index

[0221]

[0222] S202, short-term window N1=5, long-term window N2=20, and the short-term variation coefficients of each index are calculated =[0.32, 0.18, 0.25, 0.25], and the long-term variation coefficients are =[0.21, 0.15, 0.19, 0.20], context fluctuation coefficient a = 0.45, substitute into the fusion weight calculation formula to calculate the initial weight =[0.30, 0.16, 0.23, 0.31];

[0223] S203, calculate the index mutual information: = 0.68 (high correlation), weight after principal component analysis correction =[0.21, 0.16, 0.16, 0.31];

[0224] S204, calculate the membership matrix by the triangular membership function , combined with the weight matrix , get the comprehensive evaluation result =[0.28, 0.35, 0.27, 0.10], the evaluation level is "good";

[0225] S205, because the evaluation level is "good", only fine-tune the communication frequency (switch from 2.4GHz to 5.8GHz), no need to adjust the resource allocation greatly.

[0226] S3, local intuitive micro-collaboration based on communication quality constraint Q-learning.

[0227] S301, the state space S includes power (such as "60%"), neighbor node packet loss rate (such as "35%"), environmental interference (such as "30dB"), exploration progress (such as "50%"), risk level (medium); the action space A includes 6 types of actions such as "data forwarding to node 3", "increasing exploration range by 10%", "switching communication frequency";

[0228] S302, the current P = 35%, calculate = 1 - 1.2 * 0.35 = 0.58; the reward of collaboration effect = 0.15, the punishment of communication cost = 0.08; the risk model prediction = 0.2, = 0.2 (exploration task), get = 0.58 * 0.15 - 0.42 * 0.08 - 0.2 * 0.2 = 0.0354;

[0229] S303, set = 0.2, = 0.8, = 1.0 + 0.01 * (50% - 35%) = 1.15, the current = 0.5, the next state optimal value = 0.6, substitute into formula (5) to get = 0.5 + 0.2 x (0.0354 + 0.8 x 0.6 - 0.5) x 1.15 = 0.524;

[0230] S304, receiving 3 node experiences satisfying the communication quality threshold, updating the local model after distillation fusion;

[0231] S305, when the communication delay reaches 250ms, selecting the action "data forwarding to node 3 and switching to 5.8GHz communication" with the largest value, executing local micro-collaboration, and buffering the decision experience.

[0232] S4, adaptive fusion based on dynamically allocated weights according to confidence.

[0233] S401, guiding the historical execution success rate of the plan = 0.75, the current bandwidth = 600Kbps, and calculating = 0.6 x 0.75 + 0.4 x (1 - |300 - 600| / 1000) = 0.54; the local decision historical accuracy = 0.8, the context similarity = 0.7, and calculating = 0.7 x 0.8 + 0.3 x 0.7 = 0.77;

[0234] S402, calculating the fusion weight = 0.54 / (0.54 + 0.77) = 0.41, = 0.59;

[0235] S403, the basic probability of the guiding decision "continue exploring along the original path" = 0.5, the basic probability of the local decision "adjust the path to avoid the interference area" = 0.7, and the fusion result = 0.82, finally executing the decision of "adjusting the path to avoid the interference area".

[0236] Embodiment 2

[0237] On the basis of embodiment 1, embodiment 2 provides a more specific weak network adaptability verification scheme, which optimizes the communication quality constraint Q-learning of step S3 for the extreme weak network scenario of packet loss rate 50% and bandwidth 512Kbps:

[0238] Adjusting the experience decay coefficient = 1.5 to further enhance the weight of historical experience in the weak network; at the same time, to compensate for the decay of the collaboration effect reward under high packet loss rate, the communication quality factor in the reward function is optimized, and its relationship is adjusted to ​=1.2x(1 - packet loss rate), when the packet loss rate is 50%, the coefficient 1.2 is determined by pre-experiment, for balancing between moderate incentive cooperative behavior and control communication cost, =1.2x(1 - 0.5)=0.6;

[0239] The experimental data show that the success rate of the optimized local micro-collaboration is increased from 82% in embodiment 1 to 91%, and the data transmission success rate is increased from 65% to 83%; the decision accuracy after weight fusion reaches 94%, which is significantly improved compared with single guide strategy (accuracy 68%) or single local strategy (accuracy 85%).

[0240] It should be noted that the same or similar parts in this embodiment as in embodiment 1 can be mutually referred to, and will not be described in detail in this application.

[0241] Embodiment 3

[0242] Based on embodiments 1 and 2, this embodiment 3 provides a cluster weak network micro-collaboration system based on adaptive fusion of guide strategy and local experience, as shown in Figure 6 , comprising:

[0243] The preplan generation module adopts a lightweight preplan generation model improved by genetic algorithm, integrates an encoding module, a fitness calculation module and a simplified update module, and the preplan generation time is ≤2s, and the update period can be configured to 5-30s;

[0244] The situation adjustment module includes an index acquisition unit, a dynamic weight calculation unit and a fuzzy evaluation unit, and adopts a sliding window to update the weight in real time, and the evaluation delay is ≤500ms;

[0245] The local collaboration module has a built-in communication quality monitoring unit and a Q-learning decision engine, and supports local micro-collaboration of 10-50 nodes, and the decision response time is ≤100ms;

[0246] The weight fusion module integrates a confidence calculation unit and a D-S fusion unit, and the fusion delay is ≤300ms, and the decision output accuracy is ≥90%;

[0247] The communication adaptation module adopts SNMP protocol to monitor network state, and the sampling frequency is 10Hz, and the state feedback delay is ≤200ms.

[0248] The above shows and describes the basic principles, main features and advantages of the present application. It should be understood by those skilled in the art that the present application is not limited to the above embodiments, and the above embodiments and descriptions in the specification are only to illustrate the principles of the present application, and any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A clustered weak network micro-cooperative method based on the adaptive fusion of guidance strategy and local experience, characterized in that, The method includes the following steps: Step 1, online lightweight guidance plan generation based on improved genetic algorithm: combine cluster task requirements and initial network state to generate a lightweight guidance plan that takes into account both global optimization and transmission efficiency. Step 2, Context-aware adjustment based on dynamic weighted fuzzy comprehensive evaluation: During distributed execution, the status of cluster nodes and environmental context are perceived in real time, and the execution strategy is dynamically adjusted. Step 3, Local Intuitive Micro-Coordination Based on Communication Quality Constraint Q-learning: In a weak network environment, nodes autonomously implement micro-coordination execution of local clusters based on historical experience and real-time communication quality. Step 4, Adaptive fusion of weights based on dynamic weight allocation of confidence: Based on the execution confidence of the guidance plan and the decision credibility of local experience, the fusion weights are dynamically allocated to output the optimal collaborative decision.

2. The cluster weak network micro-cooperative method based on adaptive fusion of guidance strategy and local experience as described in claim 1, characterized in that, Step 1 includes the following specific steps: Step 1-1, Define the contingency plan coding rules: Use binary-real number hybrid coding, where binary bits represent the task execution sequence and real numbers represent the node resource allocation ratio; Step 1-2, Construct the fitness function: Taking into account task completion rate, network transmission overhead, and resource utilization, dynamic weight coefficients are introduced to adapt to different task scenarios. The fitness function F is shown in the following formula: ； In the formula, The task completion rate is defined as [0, 1]. =Current completed task volume / Total task volume; B is the planned transmission bandwidth usage, unit: Kbps, B = Planned file size / Transmission time; The maximum available bandwidth of the network is in Kbps; U is the resource utilization rate, with a value range of [0,1]. , All are dynamic weighting coefficients, determined by the analytic hierarchy process based on the task type. Steps 1-3 introduce a pre-plan simplification operator: redundant information is pruned from the optimal individual generated by the genetic algorithm iteration. The process includes: ① deleting duplicate instructions; ② removing low-weight constraint terms whose resource allocation ratio is ≤ preset resource allocation ratio threshold; ③ merging adjacent nodes with the same type of task sequences. Steps 1-4, online dynamic update: Every time interval T, based on network status feedback, repeat steps 1-1 to 1-3 to update the guidance plan, so as to achieve real-time matching between the plan and the network status.

3. The cluster weak network micro-cooperative method based on adaptive fusion of guidance strategy and local experience according to claim 2, characterized in that, In step 1-1, the encoding length is controlled to be within 1 / 3 of that of traditional encoding.

4. The cluster weak network micro-cooperative method based on adaptive fusion of guidance strategy and local experience according to claim 1, characterized in that, Step 2 specifically includes: Step 2-1, define the situation evaluation index system: including four core indicators: node communication quality C, remaining resources R, environmental interference I, and task progress Q. Each indicator is divided into four evaluation levels: "excellent, good, medium, and poor". Step 2-2, Construct a dual-window dynamic weight calculation model: This involves parallel calculation using a short-term sensitive window and a long-term stable window, combined with adaptive fusion of the scenario fluctuation coefficient; specifically including: ① Calculate the single-window coefficient of variation: For short-term sensitive windows ; For long-term stable windows ; In the formula, Let be the short-term coefficient of variation of the i-th core indicator within the short-term sensitive window. Let be the long-term coefficient of variation of the i-th core indicator in the long-term stable window. These represent the standard deviations of the i-th core indicator in the short-term sensitive window and the i-th core indicator in the long-term stable window, respectively. These are the mean values ​​of the i-th core indicator in the short-term sensitive window and the mean values ​​of the i-th core indicator in the long-term stable window, respectively. ② Calculate the situational fluctuation coefficient : ； ③ The fusion weight is calculated as follows: ； in, The weights of the i-th core indicator are used to form a dynamic weight matrix. ; Steps 2-3, Index Correlation Correction: Calculate the correlation between indicators using mutual information. ,right For highly correlated index pairs that are ≥ a preset threshold, principal component analysis is used for weight decoupling, and the modified formula is as follows: ; This indicates the correlation between the i-th core indicator and the j-th core indicator; Steps 2-4, Fuzzy comprehensive evaluation calculation: Determine the evaluation level membership degree of each core indicator through membership function; Constructing a fuzzy evaluation matrix Combined with the corrected dynamic weight matrix The weighted average method was used to calculate the comprehensive evaluation results. As shown in the following formula: ； In the formula, For fuzzy synthesis operators, For the i-th core indicator, the i-th core indicator is the i-th core indicator. Membership degree of each evaluation level =1,2,3,4; Steps 2-5: Adjust strategy based on comprehensive evaluation results. If the evaluation level is "poor", a strategy adjustment will be triggered, with the adjustment magnitude being negatively correlated with the evaluation result.

5. The cluster weak network micro-cooperative method based on adaptive fusion of guidance strategy and local experience according to claim 4, characterized in that, The membership function expression in steps 2-4 is as follows: ； ； ； ； Adjust parameters according to quantitative standards for different core indicators. , , , ; , , and These represent the membership degree of each core indicator in the four evaluation levels of "excellent, good, average, and poor".

6. The cluster weak network micro-cooperative method based on adaptive fusion of guidance strategy and local experience according to claim 1, characterized in that, Step 3 specifically includes: Step 3-1, define the state space and action space: the state space S includes the local node state, the neighbor node state, and the environment state; the action space A includes collaborative execution actions and communication adjustment actions; Step 3-2, construct the communication quality constraint reward function: This reward function comprehensively considers the collaborative effect and communication cost, as shown in the following formula: ； in, Represents the reward function, This is the communication quality factor, with a value range of [0,1]. Rewards for synergistic effects, Penalty for communication costs, The decision risk value, ranging from [0,1], is predicted by a random forest model trained on historical failure data, with the current state as the input. With action The output is the probability of the action failing. Risk weights are dynamically adjusted based on the level of task importance. Step 3-3, Improvement Value update rule: Introduce an empirical decay coefficient , build The value update formula is as follows: ； in, For learning rate, As a discount factor, For the next state, The optimal action for the next state. This is the empirical attenuation coefficient; This represents the Q-value, or expected cumulative reward, that can be obtained after taking action a in state s. Representing state Take action below The Q value that can be obtained afterward Indicates the next state Below, all possible actions to be performed. The maximum expected cumulative return that can be obtained is the maximum Q value; Steps 3-4, cross-node experience distillation mechanism: ① Experience screening: Only experience samples from nodes with communication quality ≥ threshold are accepted, and low-confidence experiences are filtered out; ② Distillation fusion: Knowledge distillation technology is used to compress the screened multi-node experience into a unified experience model, i.e., the teacher model. Local nodes learn experience by optimizing the loss function of the student model and the teacher model. Steps 3-5, Local Intuitive Decision Making: When communication latency > preset latency threshold or packet loss rate > preset packet loss rate, the node makes a decision based on the current... The action with the highest value is executed locally through micro-coordination, while decision-making experience is cached for subsequent weight fusion.

7. The cluster weak network micro-cooperative method based on adaptive fusion of guidance strategy and local experience according to claim 6, characterized in that, The communication quality factor It is negatively correlated with the packet loss rate P.

8. The cluster weak network micro-cooperative method based on adaptive fusion of guidance strategy and local experience according to claim 1, characterized in that, In step 4, the confidence-based weighted adaptive fusion integrates guided decision-making and local decision-making through DS evidence theory, dynamically allocating fusion weights, specifically including: Step 4-1: Calculate the decision confidence level, including the confidence level of the guidance strategy. Credibility of local experience ; Among them, the confidence level of the guidance strategy Calculated based on the success rate of contingency plan execution and network matching degree: ； Local experience credibility Calculated based on the matching degree between historical decision accuracy and current context: ； In the formula, To guide the historical success rate of the contingency plan, Given the current network bandwidth, For the historical accuracy of local decision-making, The similarity between the current situation and the historical best situation. To contingency plan for bandwidth usage, The maximum available bandwidth of the network; Step 4-2, Dynamically allocate fusion weights: The weights are positively correlated with the confidence level and satisfy the following conditions: + =1, as shown in the following formula: ； In the formula, To guide the strategy weights, Weighted by local experience; Step 4-3, Dempster Evidence Theory Fusion: Treating the guided decision and the local decision as two independent sources of evidence, calculate the basic probability assignment function, and fuse them using the Dempster synthesis rule to obtain the final decision, as shown in the following equation: ； In the formula, To assign the basic probability to decision A after fusion. , These are the basic probability assignments for guided decision-making and local decision-making, respectively. Basic probability allocation Based on the confidence level of the guidance strategy Computation, Local Decision Basic probability allocation Reliability based on local experience calculate.

9. A clustered weak network micro-cooperative system based on the method of any one of claims 1 to 8, characterized in that, The system includes: The contingency plan generation module is used to generate lightweight guidance contingency plans online based on an improved genetic algorithm, and to dynamically update them in combination with task requirements and network status. The scenario adjustment module is used to perceive the status of cluster nodes and the environment through dynamic weighted fuzzy comprehensive evaluation and dynamically adjust the execution strategy. The local collaboration module is used to achieve micro-collaborative execution with local intuition of nodes in weak network environments based on Q-learning with communication quality constraints. The weight fusion module is used to dynamically allocate weights based on the guidance confidence and local confidence, and output the optimal decision through DS evidence theory fusion.

10. The clustered weak network micro-cooperative system based on claim 9, characterized in that, The system also includes: The communication adapter module is used to monitor the network status in real time and provide network status feedback to each module.

Images