A sensor scheduling method for large-scale WSN multi-target tracking
By adopting time-delay broadcast dynamic clustering and Q learning methods in large-scale WSN, the target tracking requirements are calculated and resource scheduling is performed, which solves the problem of uneven resource allocation in multi-objective tracking tasks, and improves network performance and energy efficiency.
Patent Information
- Application Number
- CN202210228020.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-03-08
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2042-03-08
AI Technical Summary
Large-scale WSN faces node resource allocation and scheduling problems in multi-target tracking tasks, resulting in unbalanced resource allocation and affecting network performance.
The sensor scheduling method based on time-delay broadcast dynamic clustering is adopted to calculate the tracking requirements of each target through the cluster head, and the node resource scheduling is used to ensure reasonable resource allocation.
The balanced allocation of requirements for tracking of each goal is achieved, network resource utilization and energy efficiency are improved, and the risk of uneven task allocation is reduced.
Smart Images

Figure CN114727415B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of wireless sensor networks, and in particular to a sensor scheduling method for large-scale WSN multi-target tracking. Background Art
[0002] WSN is a network system that integrates monitoring, control and wireless communication functions. It has a wide range of application prospects in many fields such as environmental monitoring, smart home, military operations and disaster relief. As a key technology used in many applications, target tracking is essentially to estimate the motion state of the target and analyze the relevant attributes of the target. The target tracking system based on large-scale WSN has the characteristics of low price, strong robustness and concealment, and has high practical value, especially in unmanned areas, it can complete long-term and reliable regional monitoring tasks.
[0003] Large-scale WSNs are composed of a large number of sensor nodes. Due to dense deployment, a large number of nodes may participate in tracking tasks at the same time. However, a large number of nodes will work redundantly, which will generate a lot of unnecessary energy consumption. Cluster-based network structures can effectively deal with this problem because their topological structure is easy to manage, and can reduce data redundancy to a certain extent, achieving higher energy efficiency.
[0004] Compared with single target tracking, the network faces greater challenges in performing multi-target tracking tasks. Because when multiple targets appear in the monitoring area at the same time, the network will face the problem of node resource allocation and scheduling, that is, which targets should be given priority to obtain more sensor resources, which targets can obtain fewer resources when the sensor energy or quantity is insufficient, etc. The reasonable allocation of node resources plays an extremely critical role in the performance optimization of the entire multi-target tracking network. Summary of the invention
[0005] The technical problem to be solved by the present invention is to provide a sensor scheduling method for large-scale WSN multi-target tracking, which can reasonably allocate resources according to the tracking priority of the target.
[0006] The technical solution adopted by the present invention to solve the technical problem is: to provide a sensor scheduling method for large-scale WSN multi-target tracking, comprising the following steps:
[0007] (1) Nodes that perceive the presence of a target are dynamically clustered based on delayed broadcasting, and nodes within the cluster perform target perception work;
[0008] (2) When there are multiple targets in the cluster, at each moment, the cluster head calculates the tracking requirement of each target based on the perception data of the nodes in the cluster;
[0009] (3) The cluster head feeds back the tracking requirements of each target in the cluster to the nodes in the cluster, and the nodes in the cluster select actions based on the Q-value function and the exploration and development strategy.
[0010] The delay value of the node in step (1) is related to the remaining energy of the node and the distance to the target. The more remaining energy a node has and the closer it is to the target, the smaller the delay value of the node.
[0011] When the nodes in the cluster perform target sensing in step (1), at each moment, the nodes in the cluster send sensing data to the cluster head and report the distance from themselves to the target and the remaining energy information.
[0012] The step (1) also includes that after the nodes that perceive the presence of the target are dynamically clustered based on the delayed broadcast, the cluster head will determine the nodes in the cluster that are farther away from itself than a preset distance when receiving a confirmation reply message from the nodes in the cluster, and mark the nodes in the cluster that are farther away than the preset distance as boundary nodes; when a target moves near the boundary node, the cluster head will send a first message, and the boundary node that receives the first message will broadcast a second message, and the dormant node that receives the second message will re-perform the dynamic clustering process to form a new cluster; if there are still other targets in the original cluster, the new cluster and the original cluster will coexist, and the original member nodes in the overlapping area of the two clusters will still belong to the original cluster and send perception data to the original cluster head; if there are no other targets in the original cluster, the original cluster will be dissolved, and the original member nodes in the overlapping area of the new cluster and the original cluster will join the new cluster.
[0013] In step (2), the cluster head calculates the tracking requirement of each target according to the target importance k, target deviation d, node sparsity s and node energy level e; wherein the target importance k is determined by the cluster head according to the attributes of the target itself, k∈(0,1); the target deviation d is determined according to Calculated, the node sparsity s is based on Calculated, the node energy level e is based on Calculated, where Dis(i,CH) represents the distance from node i tracking a target to cluster head CH, Dis(i,CH) max Indicates the maximum distance from all nodes tracking a target to the cluster head CH, Dis(i,CH) min represents the minimum distance from all nodes tracking a target to the cluster head CH, n represents the number of nodes tracking a target, n max Indicates the number of nodes among all targets that are tracked by the most nodes, n min represents the number of nodes of the target that has the least number of nodes tracking among all targets, e residual (i) represents the residual energy of node i, e residual,max represents the residual energy of the node with the largest residual energy among all nodes, e residual,minRepresents the residual energy of the node with the least residual energy among all nodes.
[0014] In step (3), the node performs trial-and-error learning from the wireless sensor network environment at time t and selects to execute action a according to the current state S. t , then observe the subsequent state S′ and receive the delayed reward R, and then according to Update Q value. At time t, node i determines the current state according to the current cluster tracking information and its own working state. Node i selects the task to be performed based on the exploration and development strategy. At time t+1, as a result of performing the task, the state changes, and node i receives a delayed reward R. The node learns the optimal strategy π through a trial and error process, with the purpose of selecting the task that is most beneficial to itself and balancing the target demand in the next state. * (s), where Q t+1 (s t ,a t ) represents the updated Q value, Q t (s t ,a t ) represents the Q value at time t, s t represents the target state at time t, β∈[0,1] represents the learning rate coefficient, γ∈[0,1] represents the discount factor, Indicates that node i tracks target o j The learning effect weight, m is the number of targets, Represents the state value function of the node at time t+1.
[0015] The delayed reward R is defined as Among them, λ∈(0,1), Indicates that node i is in tracking state, Indicates that node i is in sleep state, e i represents the remaining energy of node i after executing the task, e0 represents the initial energy of node i, and r s represents the perception radius of node i, Represents the target from node i to the selected tracking target distance, n0 represents the number of sensor nodes required to reliably track the target. Indicates the number of nodes whose Q value is greater than that of node i among all nodes of the tracking target m i represents the size of the target set, Indicates the number of targets in the target set that meet the tracking requirements.
[0016] The node i tracks the target o j The learning effect weight Among them, m′ represents the number of targets that the node can observe, Indicates the target o j The tracking requirement.
[0017] The state value function of the node at time t+1 Defined as:
[0018] The exploration and development strategy in step (3) selects random exploration actions according to the exploration probability ε, where the exploration probability ε is defined as: ε = min(ε max ,ε min +(I max -I min ) / I max ), where ε max and ε min Respectively represent the upper and lower bounds of the exploration coefficient, I max and I min They represent the maximum and minimum values of the target tracking requirement respectively.
[0019] Beneficial Effects
[0020] Due to the adoption of the above technical solution, the present invention has the following advantages and positive effects compared with the prior art:
[0021] First, this paper studies how to calculate the tracking priority of each target based on the target's own attributes and the status of the nodes around the target in the scenario where multiple targets compete for sensor node resources, and proposes the concept of tracking demand. Tracking demand is the basis and important basis for sensor node allocation and scheduling tracking tasks, and can comprehensively and objectively reflect the relative value of resources that should be allocated to each target.
[0022] Secondly, the present invention designs a sensor scheduling method based on Q-learning, in which each sensor is set to act as an agent, and can learn behaviors through continuous trial and error in a dynamic environment. The agent selects one from a series of possible actions and will receive a reward. Through this learning process, an optimal scheduling strategy can be constructed. Each node can decide the next adjustment action based on the real-time status, so under this framework, the nodes can be scheduled in real time according to the target tracking status. Q-learning is a model-independent reinforcement learning algorithm that can directly learn strategies from data and directly optimize an iteratively calculable Q function, which is more suitable for event-driven systems.
[0023] At the same time, the present invention improves the improved ε-greedy strategy for exploration and development. Considering that our goal is to reasonably allocate the tracking tasks of multiple targets, we hope to ensure that the tracking requirements I of each target are as balanced as possible. When the task allocation is not balanced enough in the early stage of training, the tracking requirements of each target vary greatly, so more exploration is needed. With trial and error learning, the task allocation of multiple targets gradually becomes reasonable, and ε should be reduced in order to take more greedy actions. In this way, it is possible to better converge to the optimal strategy for balancing the tracking requirements of each target. BRIEF DESCRIPTION OF THE DRAWINGS
[0024] Figure 1 is a flow chart of an embodiment of the present invention;
[0025] Figure 2 It is a scene graph of an embodiment of the present invention. DETAILED DESCRIPTION
[0026] The present invention will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms fall within the scope limited by the appended claims of the application equally.
[0027] The embodiments of the present invention relate to a sensor scheduling method for large-scale WSN multi-target tracking, such as Figure 1 As shown, the following steps are included: nodes that perceive the presence of a target are dynamically clustered based on delayed broadcasting, and nodes in the cluster perform target perception work; when multiple targets appear in the cluster, at each moment, the cluster head calculates the tracking requirement of each target based on the perception data of the nodes in the cluster; the cluster head feeds back the tracking requirement of each target in the cluster to the nodes in the cluster, and the nodes in the cluster select actions based on the Q value function and the exploration and development strategy. The scenario of this implementation method is as follows Figure 2 As shown, the specific steps are as follows:
[0028] 1. Network Deployment
[0029] A large number of sensor nodes are randomly deployed in the monitoring area of WSN, and they remain in a dormant state after deployment. The sensing radius of each node is r s , the communication radius is r c To ensure that all nodes that perceive the target can communicate with each other, set r c ≥2r s Multiple targets appear in the monitoring area and move. The total duration of the target movement is divided into finite time intervals, called moments t. In each moment, the target is static, and the position of the target in two consecutive rounds is not necessarily the same.
[0030] 2. Dynamic Clustering
[0031] According to the location of the target, the nodes in the network are organized into clusters using a dynamic scheme. s All nodes are waiting for a delay t delay After that, in r c Broadcast a "Gather CM" message within the range, t delay The duration of the node is related to the distance from the node to the target Dis(s,o) and the remaining energy of the node e residual related:
[0032] t delay ∝(Dis(s,o),e residual )
[0033] The first node to send a "Gather CM" message serves as the cluster head. The node that receives the message replies with a confirmation message to the cluster head and becomes a cluster member. At each moment, cluster members need to send sensing data to the cluster head and report their distance to the target and remaining energy information. After receiving the sensing data, the cluster head will fuse it and obtain the target location and report it to the sink. In addition, the cluster head will provide them with feedback on the target tracking status based on the information collected from cluster members. Cluster members can use this feedback information to assist in making scheduling decisions at the next moment.
[0034] 3. Boundary Processing
[0035] After the target moves within the cluster for a period of time, it will gradually move to the edge. Therefore, it is necessary to handle the target tracking task in a cluster migration manner. First, it can be determined that the range of the cluster is approximately r with the cluster head as the center. c When the cluster head receives a confirmation reply message from a member, it will determine that the distance to itself is greater than 0.9r c members, marking and recording them as "boundary nodes". When a target moves near a boundary node, the cluster head will send a "TargetOut" message, and the boundary nodes that receive the message will broadcast a "Cluster Construction" message. The dormant nodes that receive the message will re-perform the dynamic clustering process in step 2. At the same time, there are now two situations in the original cluster: 1) there are still other targets 2) there are no targets. For situation 1), the new cluster and the original cluster will coexist. Although the original member nodes in the overlapping area of the two clusters receive the "CM Gather" message, they still belong to the original cluster and send data to the original cluster head. For situation 2), if the original cluster head finds that no members send data to it, it will send a "Stop Work" message to its members, after which the original cluster will disband, and the original member nodes in the overlapping area of the two clusters will join the new cluster.
[0036] 4. Target Tracking Demand Estimation
[0037] When multiple targets appear in the monitoring area of the same cluster in a certain period of time, if the tracking tasks of each node are not properly allocated, it is easy to collect insufficient tracking information for some targets or even lose tracking. The priority of target tracking task allocation is measured by considering four factors:
[0038] Target importance k. The system attaches different importance to different types of targets, so this factor reflects the properties of the target itself. After the target is identified by the sensor, the cluster head will give a parameter value k∈(0,1) that is positively correlated with the system's importance to it. In addition to the target importance k, the other three factors need to take into account the relative state and trend of the target and the environment.
[0039] Target deviation d. This factor reflects the degree to which the position of the target in the cluster deviates from the center of the cluster. According to step three, when the target moves to the vicinity of the boundary position deviating from the center of the cluster, the cluster migration process may be triggered. At this time, compared with targets with lower deviations, targets located in the boundary area require more reliable tracking results. Because once a node or link fails, the cluster head cannot get timely updates on the target position. In this case, on the one hand, if the target moves out of the current cluster and the new cluster is not activated in time, it is likely to cause the target to be lost; on the other hand, if the target does not move out of the current cluster and the new cluster is mistakenly activated, it will also cause unnecessary energy consumption. The deviation of a certain target can be measured by the distance Dis(i,CH) from its tracking node i to the cluster head CH.
[0040] Node sparsity s. This factor reflects the density of target tracking nodes. Due to the randomness of node deployment and the exhaustion of node energy during the tracking process, there will be areas with sparse node distribution in the monitoring area. When the target moves to such areas, if the node resources are not properly allocated, it is very likely that insufficient tracking information will be collected. In order to ensure the tracking reliability of node sparse areas, a higher tracking priority should be set for the targets in the area. The tracking node sparsity s of a target can be measured by the number of tracking nodes n.
[0041] Node energy level e. This factor reflects the energy level of the target tracking node. As the tracking progresses, some nodes may undertake tracking tasks multiple times, resulting in greater energy consumption. In order to seek energy balance, resources should be allocated to targets with lower energy levels of surrounding nodes first to alleviate the work pressure of low-energy nodes. For a target tracking node energy level, the remaining energy e can be used to calculate the energy level of the target tracking node. residual To measure.
[0042] Since the dimensions and magnitudes of the factors that affect the tracking priority are different, the raw data needs to be preprocessed. The quantification method is as follows:
[0043]
[0044]
[0045]
[0046] Where Dis(i,CH) represents the distance from node i to cluster head CH, n represents the number of nodes tracking a target, e residual (i) represents the remaining energy of node i, and the maximum and minimum values marked by max and min are respectively. Based on the linear weighting idea, the tracking demand function I of the target can be established as follows:
[0047] I=ω1k+ω2d+ω3s+ω4e
[0048] Among them, ω1+ω2+ω3+ω4=1. At each moment, the cluster head will calculate the tracking demand of each target based on the collected perception data and node information, which is also an important consideration for the multi-target task allocation strategy.
[0049] This implementation method studies how to calculate the tracking priority of each target based on the target's own attributes and the status of nodes around the target in a scenario where multiple targets compete for sensor node resources, and proposes the concept of tracking demand. Tracking demand is the basis and important basis for allocating and scheduling tracking tasks for sensor nodes, and can comprehensively and objectively reflect the relative value of resources that should be allocated to each target.
[0050] 5. Sensor Scheduling Method Based on Q-Learning
[0051] Some basic concepts of this method are defined as follows:
[0052] Agent and Environment: Each node and environment in WSN is mapped to an agent and environment in reinforcement learning.
[0053] State Space: In this model, the target set is represented as O = {o1,o2,...,o M}, where M is the number of targets. The state space is determined by the cluster head CH's estimate of the demand for each target, expressed as in
[0054] ActionSpace: The action space A is the set of all actions of a node at round t, which can be expressed as in N is the number of nodes. It is used to refer to the node i selecting the tracking target o at time t j . In particular, a t =0 means the node is dormant at time t.
[0055] Delayed Reward: Delayed reward represents the network's evaluation of the action performed by the agent in the previous round, and uses a weighted function R to weigh energy efficiency and tracking quality. In each round of transmission, the cluster head will make a delayed reward evaluation based on the actions selected by the cluster members and the state of the environment.
[0056] During the scheduling process, each node tries to find the optimal action strategy π in each state * The purpose is to achieve the reasonable allocation of multiple target tracking tasks and timely node dormancy to save energy. The sensor node learns from the wireless sensor network environment through trial and error at time t and selects the execution action a according to the current state S. t , then observe the subsequent state S′ and receive the delayed reward R, and then update the Q value according to the following formula:
[0057]
[0058] β∈[0,1] represents the learning rate coefficient, which is used to control the learning speed. The larger the learning rate, the faster the convergence. However, the greater the possibility of oscillation. γ∈[0,1] is the discount factor, which indicates the foresight of the learning system. The smaller γ is, the more the sensor node relies on the delayed reward of recent work; otherwise, it will pay more attention to the impact of future states.
[0059] According to the definition of delayed reward R:
[0060]
[0061] The parameter λ∈(0,1). First, let’s look at what happens when a node chooses a target to track instead of sleeping, that is When, e i represents the remaining energy of node i after executing the task, e0 represents the initial energy of node i, and r s represents the perception radius of node i, Represents the target from node i to the selected tracking target distance, n0 represents the number of sensor nodes required to reliably track the target. Indicates the number of nodes whose Q value is greater than that of node i among all nodes of the tracking target Right now:
[0062]
[0063] It can be seen that in addition to energy and distance, the number of nodes tracking the target is an important factor affecting the reward value. When the Q value ranking of node i is within n0, it can be seen that the node receives positive feedback, and the higher the ranking, the higher the reward; when the Q value ranking of node i is outside n0, it can be seen that the node receives negative feedback, and the lower the ranking, the greater the negative reward.
[0064] When a node chooses to sleep, When , only energy efficiency performance is rewarded, and tracking performance is no longer rewarded. In addition, the size of the reward value is also related to the target set that sensor node i can perceive at time t. i represents the size of the target set, that is, the number of targets, and It indicates the number of targets in the target set that meet the tracking requirements, so there must be It can be seen from the formula that when there are still targets that do not meet the tracking requirements within the sensing range of node i, negative feedback will be generated. When no target appears or all targets meet the requirements, positive feedback will be generated.
[0065] In addition, the tracking requirement of each target will also be used to update the Q value. j Indicates the tracking target o j The learning effect weight is determined by the relative size of the tracking requirements of each target and is defined as follows:
[0066]
[0067] Where m′ represents the number of targets that the node can observe. Obviously, when the node cannot observe any target, the tracking learning effect can be considered to be 0. t+1 Selecting targets with greater tracking requirements will result in a larger corresponding value function, in order to encourage nodes to supplement targets that do not meet the tracking requirements. The state value function representing the node at time t+1 is defined as:
[0068]
[0069] At time t, sensor node i determines its current state based on the current cluster tracking information and its own working state. The sensor node selects the task to be performed based on the exploration-exploitation strategy. At time t+1, as a result of performing the task, the state changes and the node receives a delayed reward R. With the goal of selecting the task that is most beneficial to itself and balances the target demand in the next state, the node learns the optimal strategy π through a trial-and-error process * (s):
[0070]
[0071] The function is learned through an iterative method to obtain the optimal scheduling strategy.
[0072] This implementation method designs a sensor scheduling method based on Q-learning, and sets each sensor to act as an agent, which can learn behaviors through trial and error in a dynamic environment. The agent selects one from a series of possible actions and will receive a reward. Through this learning process, an optimal scheduling strategy can be constructed. Each node can decide the next adjustment action based on the real-time status, so under this framework, nodes can be scheduled in real time based on the target tracking status. Q-learning is a model-independent reinforcement learning algorithm that can learn strategies directly from data and directly optimize an iteratively calculable Q function, which is more suitable for event-driven systems.
[0073] 6. Exploration-Exploitation Strategy
[0074] An important aspect of the reinforcement learning framework is the trade-off between exploration and exploitation. Exploration involves randomly selecting actions that may not have high reward values in order to search for potentially better rewarding behaviors, while exploitation is about maximizing the agent's reward value based on what has been learned. The algorithm should be encouraged to perform more exploration at the beginning of the training period, and then slowly reduce exploration to exploit the learned information. With sufficient training, the algorithm should converge to its optimal policy. However, taking actions greedily before convergence may lead to suboptimal policies because the agent does not have the opportunity to sample state-actions that may lead to higher rewards. To avoid this, an ε-greedy policy can be adopted: the agent will greedily choose actions most of the time, but will choose random exploratory actions with probability ε. A heuristic definition of the exploration probability ε is as follows:
[0075] ε=min(ε max ,ε min +k×(S max -S) / S max )
[0076] ε max and ε min Respectively represent the upper and lower bounds of the exploration coefficient. maxRepresents the maximum number of states, and S represents the number of currently known states. In each round, the agent calculates ε and selects a random number between [0,1]. If the random number is less than or equal to ε, the agent will choose to perform a random action (exploration), otherwise the agent will choose the best action based on the Q value. Since the state is not a discrete random variable in this model, the definition of the exploration probability ε is improved accordingly. First of all, the state set S is not a discrete quantity in this model and cannot be measured in quantity. Considering that the goal is to reasonably allocate the tracking tasks of multiple targets, it is hoped that the tracking requirements I of each target are as balanced as possible. When the task allocation is not balanced enough in the early stage of training, the tracking requirements of each target vary greatly, so more exploration is needed. With trial and error learning, the task allocation of multiple targets gradually becomes reasonable, and ε should be reduced in order to take more greedy actions. According to the above analysis, the modified ε is defined as:
[0077] ε=min(ε max ,ε min +(I max -I min ) / I max )
[0078] I max and I min Respectively represent the maximum and minimum values of the target tracking demand. max -I min ) is large, indicating that the task allocation is less balanced, and exploration should be performed with a higher probability, otherwise more focus will be placed on development.
[0079] This implementation makes improvements to the improved ε-greedy strategy used for exploration and development. In this way, it is possible to better converge to an optimal strategy that balances the tracking requirements of each target.
Claims
1. A sensor scheduling method for large-scale WSN multi-target tracking, characterized in that: The following steps are involved: (1) Nodes that perceive the presence of a target are dynamically clustered based on delayed broadcasting, and nodes within the cluster perform target perception work; (2) When multiple targets appear in a cluster, at each moment, the cluster head calculates the tracking requirement of each target based on the perception data of the nodes in the cluster; the cluster head calculates the tracking requirement of each target based on the target importance k, target deviation d, node sparsity s and node energy level e, and the calculation method is: I = ω1k + ω2d + ω3s + ω4e, where ω1 + ω2 + ω3 + ω4 = 1; the target importance k represents the importance of the system to different types of targets, and the target deviation d is the degree to which the position of the target in the cluster deviates from the cluster center; (3) The cluster head feeds back the tracking requirements of each target in the cluster to the nodes in the cluster, and the nodes in the cluster select actions based on the Q-value function and the exploration and development strategy. At time t, the node learns from the wireless sensor network environment through trial and error, and selects to execute action a according to the current state S. t , then observe the subsequent state S′ and receive the delayed reward R, and then according to Update Q value. At time t, node i determines the current state according to the current tracking information in the cluster and its own working state. Node i selects the task to be performed based on the exploration and development strategy. At time t+1, as a result of performing the task, the state changes, and node i receives a delayed reward R. The purpose is to select the task that is most beneficial to itself and the tracking demand of the target in the next state. The node learns the optimal strategy π through a trial and error process. * (s), where Q t+1 (s t ,a t ) represents the updated Q value, Q t (s t ,a t ) represents the Q value at time t, s t represents the target state at time t, β∈[0,1] represents the learning rate coefficient, γ∈[0,1] represents the discount factor, Indicates that node i tracks target o j The learning effect weight, m is the number of targets, Represents the state value function of the node at time t+1; the delayed reward R is defined as Among them, λ∈(0,1), Indicates that node i is in tracking state, Indicates that node i is in sleep state, e i represents the remaining energy of node i after executing the task, e0 represents the initial energy of node i, and r s represents the perception radius of node i, Represents the target from node i to the selected tracking target distance, n0 represents the number of sensor nodes required to reliably track the target. Indicates the number of nodes whose Q value is greater than that of node i among all nodes of the tracking target m i represents the size of the target set, represents the number of targets in the target set that meet the tracking requirements; the exploration and development strategy selects random exploration actions according to the exploration probability ε, and the exploration probability ε is defined as: ε=min(ε max ,ε min +(I max -I min ) / I max ), where ε max and ε min Respectively represent the upper and lower bounds of the exploration coefficient, I max and I min They represent the maximum and minimum values of the target tracking requirement respectively.
2. The sensor scheduling method for large-scale WSN multi-target tracking according to claim 1 is characterized in that: The delay value of the node in step (1) is related to the remaining energy of the node and the distance to the target. The more remaining energy a node has and the closer it is to the target, the smaller the delay value of the node.
3. The sensor scheduling method for large-scale WSN multi-target tracking according to claim 1 is characterized in that: When the nodes in the cluster perform target sensing in step (1), at each moment, the nodes in the cluster send sensing data to the cluster head and report the distance from themselves to the target and the remaining energy information.
4. The sensor scheduling method for large-scale WSN multi-target tracking according to claim 1 is characterized in that: The step (1) also includes that after the nodes that sense the presence of the target dynamically form clusters based on the time delay broadcast, the cluster head, upon receiving a confirmation reply message from the nodes in the cluster, will determine the nodes in the cluster that are farther away from itself than a preset distance, and mark the nodes in the cluster that are farther away than the preset distance as boundary nodes; when a target moves near the boundary nodes, the cluster head will send a first message, the boundary nodes that receive the first message will broadcast a second message, and the dormant nodes that receive the second message will re-perform the dynamic clustering process to form a new cluster; If there are still other targets in the original cluster, the new cluster and the original cluster will coexist, and the original member nodes in the overlapping area of the two clusters will still belong to the original cluster and send sensing data to the original cluster head; If there is no other target in the original cluster, the original cluster is dissolved, and the original member nodes in the overlapping area between the new cluster and the original cluster are added to the new cluster.
5. The sensor scheduling method for large-scale WSN multi-target tracking according to claim 1 is characterized in that: In step (2), the target importance degree k is determined by the cluster head according to the attributes of the target itself, k∈(0,1); the target deviation degree d is determined according to Calculated, the node sparsity s is based on Calculated, the node energy level e is based on Calculated, where Dis(i,CH) represents the distance from node i tracking a target to cluster head CH, Dis(i,CH) max Indicates the maximum distance from all nodes tracking a target to the cluster head CH, Dis(i,CH) min represents the minimum distance from all nodes tracking a target to the cluster head CH, n represents the number of nodes tracking a target, n max Indicates the number of nodes among all targets that are tracked by the most nodes, n min represents the number of nodes of the target that has the least number of nodes tracking among all targets, e residual (i) represents the residual energy of node i, e residual,max represents the residual energy of the node with the largest residual energy among all nodes, e residual,min Represents the residual energy of the node with the least residual energy among all nodes.
6. The sensor scheduling method for large-scale WSN multi-target tracking according to claim 1, characterized in that: The node i tracks the target o j The learning effect weight Among them, m′ represents the number of targets that the node can observe, Indicates the target o j The tracking requirement.
7. The sensor scheduling method for large-scale WSN multi-target tracking according to claim 1, characterized in that: The state value function of the node at time t+1 Defined as:
