Energy-constrained tracking-oriented intelligent sensor network dynamic optimization deployment method and system

Abstract

The application discloses a kind of intelligent sensor network dynamic optimization deployment method and system for energy constraint tracking, it is related to sensor network dynamic optimization deployment technical field.The technical points of the present application include: obtaining target initial position;Real-time prediction is carried out to target action track using unscented Kalman filtering algorithm;According to the predicted target action track, the sensor node for tracking target is path planned using the reinforcement learning model based on deep deterministic policy gradient, i.e.dynamically optimizes deployment.The present application autonomously learns optimal scheduling strategy in continuous high-dimensional space based on the reinforcement learning model based on deep deterministic policy gradient, and combines with unscented Kalman filtering algorithm to predict track, effectively cope with nonlinear dynamic environment, and avoid the dependence on complex artificial rules.

Description

Technical Field

[0001] This invention relates to the field of dynamic optimization deployment technology for sensor networks, and in particular to a method and system for dynamic optimization deployment of intelligent sensor networks oriented towards energy-constrained tracking. Background Technology

[0002] With the rapid development of IoT technology, intelligent sensor networks have been widely applied in many fields such as environmental monitoring, smart transportation, and smart healthcare. Intelligent sensor networks consist of a large number of sensor nodes with sensing, computing, and communication capabilities, enabling real-time monitoring and intelligent response to the physical environment. In dynamic scenarios such as moving target tracking and disaster emergency monitoring, sensor networks need to adjust node layout in real time according to task requirements to maintain efficient and stable monitoring performance. Therefore, dynamic optimization deployment has become one of the key technologies for improving the task execution capabilities of intelligent sensor networks. Currently, dynamic deployment optimization methods for intelligent sensor networks are mainly divided into two categories: one focuses on improving network coverage and energy efficiency through multi-objective optimization algorithms, such as the Virtual Force-Driven Moth to a Flame (VF-IMFO) algorithm and the Multi-Objective Antlion Algorithm (MOALO); the other is task-oriented, focusing on node scheduling and resource allocation, such as target tracking methods combining deep learning and unscented Kalman filtering, hierarchical scheduling architectures based on edge intelligence (EI), and the fully decentralized multi-agent reinforcement learning (FDMARL) framework. However, most of these methods do not fully consider the probability distribution characteristics of the sensing data, lack effective utilization of the network space topology, and often lack sustainable scheduling mechanisms when dealing with node energy constraints, making it difficult to balance tracking accuracy and network energy efficiency in dynamic and complex environments.

[0003] Existing technologies suffer from the following defects and shortcomings: 1) Single optimization objective and lack of multi-dimensional collaborative mechanisms: Most existing methods only aim to improve network coverage or reduce system energy consumption, failing to effectively coordinate task tracking accuracy with network resource efficiency. This one-sided optimization makes it difficult to achieve a balance between perception quality and resource efficiency in real-world environments. Therefore, it is urgent to introduce a collaborative optimization mechanism that can integrate multi-source information and multi-dimensional objectives to achieve intelligent allocation and dynamic adjustment of system resources. 2) Lack of probabilistic driving and quantitative evaluation capabilities in dynamic strategies: Current dynamic adjustment methods mostly rely on deterministic indicators or heuristic rules, lacking effective quantification and probabilistic modeling of the degree of network state changes, resulting in poor adaptive deployment decision-making and limited adjustment accuracy. In particular, the lack of mining and analysis of the joint probability distribution characteristics of perception data makes it impossible to evaluate the actual effect of deployment adjustments from a statistical perspective, thus limiting further improvement in tracking accuracy and system stability. 3) Lack of systematic and sustainable scheduling mechanisms under energy constraints: In real-world scenarios where node energy is severely limited, existing strategies often fail to establish systematic energy replenishment and node replacement mechanisms. Once critical nodes run out of energy, task tracking can easily be interrupted, severely impacting the long-term stability and reliability of the network. Such methods typically fail to incorporate remaining node energy, coverage loss, and task requirements into a unified framework, lacking dynamic rescue and replacement strategies based on energy awareness. Therefore, a sustainable scheduling mechanism integrating energy status awareness and network function maintenance needs to be constructed to significantly enhance the system's robustness and task continuity during long-term operation. Summary of the Invention

[0004] To solve or alleviate one or more of the above-mentioned technical problems, the present invention provides a method and system for dynamic optimization deployment of intelligent sensor networks for energy-constrained tracking.

[0005] According to one aspect of the present invention, a method for dynamic optimization deployment of intelligent sensor networks oriented towards energy-constrained tracking is proposed, the method comprising:

[0006] Obtain the initial position of the target;

[0007] Real-time prediction of target trajectory using unscented Kalman filtering algorithm;

[0008] Based on the predicted target trajectory, a reinforcement learning model based on deep deterministic policy gradients is used to perform path planning for the sensor nodes tracking the target, i.e., dynamic optimization deployment; the reinforcement learning model based on deep deterministic policy gradients is constructed as follows:

[0009] The state vector includes the distances and relative angles between multiple tracking sensor nodes and the target, as well as the target's current position;

[0010] The motion vector includes the displacement of multiple tracking sensor nodes in a two-dimensional plane;

[0011] reward function R The design is based on a constrained Boltzmann machine and energy consumption as follows:

[0012] ;

[0013] in, , These represent the weighting coefficients for KL divergence and energy consumption, respectively. Indicates total energy consumption. , , representing the energy consumption of the i-th tracking sensor node. k This is the energy consumption coefficient. Let be the distance moved by the i-th tracking sensor node at time t; This represents the initial set of measurements from multiple tracking sensor nodes, including the distance to the target, the relative angle, and the target's current position. This represents the probability distribution of the visible layer in a restricted Boltzmann machine corresponding to the initial set of measurements; This represents the set of measurements taken after multiple tracking sensor nodes have moved. The KL divergence value represents the probability distribution of the visible layer in a restricted Boltzmann machine corresponding to the set of measurements after node movement. , express Energy value, express The energy value.

[0014] Furthermore, the real-time prediction of the target's trajectory using the unscented Kalman filter algorithm includes:

[0015] Establish the target's state vector, which includes the target's position and velocity components; establish the target's state transition equation as a uniform linear motion equation.

[0016] State prediction: Generate a set of Sigma points based on the target's current state and covariance matrix; substitute the generated Sigma points into the state transition equation to predict the state, and recombine the Sigma points after propagation through the state transition equation according to their weights to obtain the predicted mean and predicted covariance of the target's state at time t.

[0017] State Update: Substitute each Sigma point into the measurement model to calculate the predicted monitoring value; perform a weighted average of the predicted monitoring values ​​of all Sigma points to obtain the predicted monitoring value of the target; calculate the covariance matrix of the predicted monitoring value of the target; calculate the Kalman gain based on the covariance matrix of the predicted monitoring value of the target and the observation covariance matrix; use the Kalman gain to update the state estimate and state covariance matrix of the target; finally output the state estimate of the target, i.e., the predicted value of the target position.

[0018] Furthermore, the method also includes: during the target tracking process, calculating the remaining energy of the tracking sensor node in real time; when the remaining energy of the tracking sensor node is less than a preset threshold, determining whether there are other candidate sensor nodes based on whether the distance to the target is within the preset sensing radius of the sensor node; if so, selecting the optimal sensor node from the other candidate sensor nodes based on the remaining energy, coverage contribution, and distance to the target, to take over the tracking sensor node with low energy and continue to perform tracking.

[0019] Furthermore, the step of selecting the optimal sensor node from among other candidate sensor nodes based on remaining energy, coverage contribution, and distance to the target includes:

[0020] Calculate the overall score for each of the other sensor nodes using the following formula. :

[0021] ;

[0022] in, , , These represent the weighting coefficients of each indicator; Indicates the i-th candidate sensor node's relationship with the target. The monitoring value, , This represents the distance between the candidate sensor node and the target. This indicates the sensing radius of the preset sensor node; Let represent the remaining energy of the i-th candidate sensor node at time t. , Let represent the remaining energy of the i-th candidate sensor node at time t-1. This represents the energy consumption of the i-th candidate sensor node due to movement. , This is the energy consumption coefficient. Let be the moving distance of the i-th candidate sensor node at time t; This represents the coverage contribution of the i-th candidate sensor node. , This represents the total coverage area of ​​the sensor network. This represents the independent coverage area of ​​the i-th candidate sensor node;

[0023] The candidate sensor node with the highest overall score is selected as the optimal sensor node.

[0024] Furthermore, the method also includes: if no sensor node is within a preset sensing radius of the target, then it is determined that no active node can sense the target, and a rescue score is calculated for all sensor nodes not in tracking mode. :

[0025] = ;

[0026] in, , , These represent the weight coefficients of each item;

[0027] The sensor node with the highest rescue score was selected as the replacement tracking sensor node.

[0028] According to another aspect of the present invention, a dynamic optimization deployment system for intelligent sensor networks oriented towards energy-constrained tracking is proposed, for implementing the aforementioned dynamic optimization deployment method for intelligent sensor networks oriented towards energy-constrained tracking; the system includes:

[0029] The location acquisition module is configured to acquire the initial position of the target.

[0030] The target trajectory prediction module is configured to predict the target's trajectory in real time using an unscented Kalman filter algorithm.

[0031] The tracking node path planning module is configured to perform path planning for the sensor nodes of the tracking target based on the predicted target movement trajectory using a reinforcement learning model based on deep deterministic policy gradients, i.e., dynamic optimization deployment.

[0032] According to another aspect of the invention, an electronic device is also proposed, comprising a memory, a processor, and a computer program; wherein the computer program is stored in the memory and configured to be executed by the processor to implement the dynamic optimization deployment method for energy-constrained tracking smart sensor networks as described above.

[0033] According to another aspect of the invention, a computer-readable storage medium is also provided, on which a computer program is stored; the computer program is executed by a processor to implement the dynamic optimization deployment method for energy-constrained tracking smart sensor networks as described above.

[0034] The present invention has the following technical effects:

[0035] This invention proposes a dynamic optimization deployment method and system for intelligent sensor networks oriented towards energy-constrained tracking, which has the following advantages: 1) Coordination and balance of multi-dimensional optimization objectives: By designing a composite reward function that integrates KL divergence and energy consumption, multi-dimensional objectives such as task tracking accuracy, network energy consumption, and coverage performance are incorporated into a unified optimization framework, overcoming the shortcomings of existing methods that have a single optimization objective and are difficult to balance perception quality and resource efficiency; 2) Probabilistic driving and quantitative evaluation mechanism to improve decision-making accuracy and adaptability: The invention innovatively uses a restricted Boltzmann machine (RBM) to deeply fuse multi-node monitoring information, generating a joint probability distribution representing the network state, and uses KL divergence to accurately quantify the change trend of the probability distribution before and after node scheduling, thereby driving reinforcement learning decision-making, overcoming the problems of poor adaptability and limited adjustment accuracy of traditional deterministic strategies, and improving statistical accuracy. 3) A systematic energy-aware sustainable scheduling mechanism to enhance system robustness and continuity: For real-world scenarios where node energy is severely limited, a hierarchical scheduling strategy including node replacement and node rescue mechanisms is designed. This strategy comprehensively considers multiple factors such as the remaining energy of nodes, coverage contribution, and target distance, and realizes intelligent prediction and rapid replacement of low-energy nodes, effectively avoiding tracking interruptions caused by energy depletion, and ensuring the reliability, robustness, and task continuity of the network during long-term operation; 4) Decision advantages of deep reinforcement learning in complex dynamic environments: Based on the Deep Deterministic Policy Gradient (DDPG) algorithm, this invention autonomously learns the optimal scheduling strategy in a continuous high-dimensional space, combined with unscented Kalman filter (UKF) trajectory prediction, effectively coping with nonlinear dynamic environments and avoiding dependence on complex manual rules. Attached Figure Description

[0036] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0037] Figure 1 This is a schematic diagram illustrating the research scenario for the dynamic optimization deployment of the intelligent sensor network according to the present invention.

[0038] Figure 2 This is a flowchart of a dynamic optimization deployment method for intelligent sensor networks oriented towards energy-constrained tracking, as described in this invention.

[0039] Figure 3 This is a diagram illustrating the overall architecture of a dynamic optimization deployment method for intelligent sensor networks oriented towards energy-constrained tracking, as described in this invention.

[0040] Figure 4This is a structural block diagram of the DDPG algorithm in this invention.

[0041] Figure 5 This is a schematic diagram of node switching in the node replacement mechanism and node rescue mechanism of this invention.

[0042] Figure 6 This is a schematic diagram of the structure of a dynamic optimization deployment system for an intelligent sensor network oriented towards energy-constrained tracking, as described in this invention. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0044] This invention provides a dynamic optimization deployment method and system for intelligent sensor networks oriented towards energy-constrained tracking, aiming to solve the challenges of maintaining accuracy and energy constraints faced by intelligent sensor networks in dynamic target tracking. To achieve this goal, this invention proposes a dynamic deployment and scheduling mechanism for sensor nodes based on joint probability density variation and deep reinforcement learning. By integrating target trajectory prediction based on unscented Kalman filtering (UKF), a deep reinforcement learning decision-making mechanism based on deep deterministic policy gradient (DDPG) algorithm, a joint probability density-driven reward function design based on restricted Boltzmann machine (RBM) and Kullback-Leibler divergence (KL divergence), and an energy-aware node replacement and rescue mechanism, continuous high-precision target tracking of intelligent sensor networks under limited energy conditions is achieved.

[0045] Research scenarios for dynamic optimization deployment of intelligent sensor networks, such as Figure 1 As shown, at the initial moment the target enters the monitoring area, the sensor tracking nodes at the edge of the monitoring area capture relevant information such as the target's initial position and upload it to the control center in real time. After receiving the target's initial position information, the control center uses the UKF algorithm to predict the target's trajectory in real time and estimate the target's future position. The prediction results provide a basis for the scheduling of tracking nodes. The control center controls the tracking nodes based on the target trajectory prediction and the relevant information of the tracking nodes, and uses the DDPG algorithm to perform path planning for the tracking nodes, i.e., dynamic optimization deployment.

[0046] This invention proposes a method for dynamically optimizing the deployment of intelligent sensor networks for energy-constrained tracking, such as... Figures 2-3 As shown, the method includes:

[0047] S1. Obtain the initial position of the target;

[0048] S2. Real-time prediction of target trajectory using unscented Kalman filter algorithm;

[0049] S3. Based on the predicted target trajectory, use a reinforcement learning model based on deep deterministic policy gradient to perform path planning for the sensor nodes tracking the target, i.e., dynamic optimization deployment.

[0050] The method begins with S1. In S1, the initial position of the target is obtained.

[0051] According to an embodiment of the present invention, the control center receives information such as the initial position of the target captured by the sensor tracking nodes at the edge of the monitoring area.

[0052] Then, S2 is executed, using the unscented Kalman filter algorithm to predict the target's trajectory in real time.

[0053] According to an embodiment of the present invention, the UKF algorithm is used to predict the trajectory of a target. This algorithm can be divided into two main processes: prediction and update. In the prediction phase, a known target dynamic model (i.e., state transition equation) is used to predict the next state of the target based on the current state, and the prediction error covariance is calculated. In the update phase, the predicted target state is corrected based on observation data. The prediction result is adjusted by calculating the Kalman gain, and the system state and error covariance are updated. Through continuous iterative prediction and update, the Kalman filter adjusts the predicted value based on the measured value in each iteration, making the final state estimate closer to the true target value.

[0054] The specific process is as follows: First, establish the target's state vector. ,in Indicates the target location. This represents the velocity component. The target's motion model (state transition equation) is uniform linear motion, and the formula is: ,in For time step, Zero-mean Gaussian process noise, This represents the current state of the target. UKF predicts and updates the state by generating Sigma points, as follows:

[0055] ①Based on the current state With covariance matrix Generate a set of Sigma points: , , Where M represents the dimension of the state vector and t represents the time step; The adjustment parameter is used to adjust the distribution of Sigma points; Represents the root mean square of the state covariance matrix; This represents the first time step t based on the information at time t-1. Sigma point;

[0056] ② Substitute the generated Sigma points into the state transition equation to predict and update the state, and then apply the mean weights to the Sigma points after propagation through the state equation. Covariance weights Recombining, we obtain the predicted mean of the state at time t. and predicted covariance T is the transpose operation. It is the process noise covariance matrix;

[0057] ③ After predicting the target's state, measurements are updated. The predicted state is corrected based on the actual monitored values, resulting in a new state estimate. The core of this process is adjusting the predicted state using Kalman gain. Assume the monitored values ​​of the target obtained through the sensor are... Furthermore, the monitored values ​​and the target state are correlated through a measurement function. Association. Details are as follows:

[0058] First, substitute each Sigma point into the measurement model. The formula for calculating the predicted monitoring value is: The predicted monitoring value of the target is obtained by weighted averaging the predicted monitoring values ​​of all Sigma points. , It is the first The predicted monitoring values ​​of the Sigma points are obtained; after obtaining the predicted monitoring values ​​of the target, the covariance matrix of the predicted monitoring values ​​is calculated. ,in It is the covariance matrix of the measurement noise, representing the uncertainty in sensor measurements; it is used to predict the covariance matrix of the monitored values. and observation covariance matrix Calculate Kalman gain It is used to update the state estimate, and the formula is: ,in Used to map the state space to the measurement space; utilizing Kalman gain To update the target's status: Update the target's state covariance matrix using the following formula: , These are the monitoring values ​​of the target.

[0059] The final output is the target's state estimate, i.e., the predicted value of the target's location, which provides a basis for node scheduling.

[0060] Then, S3 is executed, which uses a reinforcement learning model based on deep deterministic policy gradients to perform path planning for the sensor nodes tracking the target, i.e., dynamic optimization deployment, based on the predicted target trajectory.

[0061] According to an embodiment of the present invention, the DDPG algorithm is used as the core, combining the advantages of deep learning and reinforcement learning. Based on the UKF prediction results of the target and the state of the sensor nodes, the optimal dynamic sensor node movement strategy is sought in a high-dimensional continuous action space. The system is modeled, with the control center and dynamic sensor nodes regarded as intelligent agents, the entire monitoring area as the environment, and the motion state as continuous motion. The state space and action space in the DDPG algorithm are set as continuous spaces.

[0062] The specific designs for the state space, action space, and reward function are as follows:

[0063] ① State space design: The state vector is multi-dimensional, including the distances and relative angles between multiple (e.g., three) moving nodes and the target, as well as the target's current position. The formula is: , where the distance (i=1, 2, 3) and relative angles Calculate using the Euclidean distance and azimuth formulas respectively. Formulas: , ; 、 Indicates the location of the moving sensor node. 、 Indicates the current location of the target.

[0064] ② Motion space design: The motion vector is defined as the displacement of the tracking sensor node in a two-dimensional plane, formula: , These represent the displacements of the three moving nodes in the x-direction. , These represent the displacements of the three moving nodes in the y-direction. To prevent the nodes from deviating from the target, a constraint is set: if the angle between the node's moving direction and the target direction is greater than 90°, the node's moving direction will be forcibly adjusted to the target direction.

[0065] ③ Reward Function Design: To characterize the change in probability density of sensor network measurements before and after node movement, a Restricted Boltzmann Machine (RBM) is used to fuse multi-node monitoring information, and the Kullback-Leibler (KL) divergence is used to quantify the distribution differences, thereby constructing the reward function. The specific process is as follows:

[0066] 1) At each time t, the system receives initial measurement information from n initial motion tracking nodes. n is the initial total number of motion tracking sensor nodes. Represents a real matrix. For signal acquisition duration, For the nth sensor node Vector of measurement values ​​over time, yes The measurement matrix of n initial motion tracking sensor nodes over a time period. It serves as the input to a two-dimensional restricted Boltzmann machine (RBM). The RBM's structure comprises a visible layer and a hidden layer, wherein the visible layer is composed of... The structure, the hidden layer is represented as , It is the number of neurons in the hidden layer in the spatial dimension. It represents the number of neurons in the hidden layer over time. RBMs are determined by an energy function. It describes the interaction between neurons in the visible and hidden layers; the lower the energy, the more stable the system state.

[0067] 2) Based on the energy function The joint probability distribution of neurons in the visible and hidden layers can be expressed as follows: ,formula: ,in The partition function represents the summation of the energy exponents over all possible combinations of visible and hidden layer states, and its function is to normalize the probability.

[0068] 3) Then, by marginalizing the hidden layer variable Y, the probability distribution of the visible layer can be obtained. ,formula: .

[0069] 4) After the RBM model is trained and the parameters are fixed, the partition function Z is a constant, therefore the visible layer data... The probability distribution characteristics can be obtained through its corresponding energy value. Characterization and evaluation, formula: The lower the value, the better the data. The closer the distribution fits to the RBM model, the better.

[0070] 5) Acquire new measurement data after the tracking node moves. Calculate its probability distribution To quantitatively assess the difference in the distribution of measurement data before and after node movement, KL divergence is used as a metric, with the formula: The larger the KL divergence value, the greater the improvement in data distribution after node adjustment, reflecting an improvement in tracking accuracy.

[0071] To further ensure the network's continuous operation, an energy consumption constraint is introduced into the reward function. The energy consumption of node i at time t is defined as: ,in k This is the energy consumption coefficient. Let be the distance traveled by the i-th tracking sensor node at time t; the total energy consumption of the system is... .

[0072] Finally, the reward function R Defined as:

[0073] ;

[0074] in, , These are the weighting coefficients for the KL divergence and energy consumption, respectively. In an ideal, energy-sufficient scenario, let... =1, =0.

[0075] Figure 4 The block diagram of the DDPG algorithm is shown. The DDPG algorithm consists of two key networks: the Actor Network and the Critic Network. Each of these networks includes an Online Network and a Target Network, which have identical structures and work together to optimize the policy reward function. During the training of the DDPG algorithm, the reward function guides the updates of the Actor-Critic Network. The Actor Network is responsible for generating actions based on the current state and is a deterministic policy network; the Online Policy Network generates actions based on the current state. Output Action The target policy network obtains parameters from the online policy network through a soft update mechanism; the critic network is a Q-value network whose goal is to predict the cumulative reward that can be obtained by taking a specific action in a certain state; the online Q-network calculates the Q-value based on the action and state, and the target Q-network is responsible for calculating the target Q-value.

[0076] The specific process is as follows: First, initialize the online actor network and online critic network, as well as the experience replay pool, and initialize the parameters of the target actor network and target critic network to be the same as the online network; initialize a random process (for action exploration, such as OU noise) to obtain the initial state. Then, at each time step, the actor network adjusts its settings according to the current state. Select Action ,formula: , This represents the OU noise at time t. Indicates an online actor network, This represents the parameters of the actor network; after the action is performed, the reward is observed. and the next state ; Transition sample Store in the experience replay pool; randomly sample from the experience replay pool. A sample of experiences ; Calculate the target Q value , This is the current observation reward. It is the attenuation factor. It is a network of target commentators. It is the target actor network, where c represents... The c-th sample in the set, Indicates the current state. Indicates the current action. Indicate the next state; minimize the loss using gradient descent. Update the parameters of the online critic network. Indicates the size of the empirical sample set. Represents an online critic network; through policy gradient (often manifested as minimizing loss in practice). Update the online actor network, and soft update the target actor network and target critic network.

[0077] In the algorithm deployment and application phase, the trained actor network model is fed into the actual sensor network. The specific application process is as follows: the control center collects real-time monitoring data from dynamic sensor nodes; this data is processed using a UKF filter (the same as during training) to predict the target's current position and calculate the real-time distance and relative angle between each node and the predicted target; these values ​​are combined into a state vector with the exact same format as during training and input into the trained actor network; based on the input state, the actor network instantly (forward propagation) calculates what it considers the optimal movement command, which is a vector containing six values: The control center sends the motion commands output by the actor network to the dynamic sensor nodes, which then move according to the commands, achieving continuous and high-precision tracking of the moving target.

[0078] Furthermore, such as Figure 5As shown, to address energy constraints in real-world deployments and prevent monitoring blind spots or target task loss, a node replacement and rescue mechanism is introduced. When a sensor node's energy falls below a set threshold, replacement or rescue conditions are triggered. First, based on whether the distance to the target is within the preset sensor node's sensing radius, it is determined whether other candidate sensor nodes exist. If so, the optimal sensor node is selected from the remaining energy, coverage contribution, and distance to the target to take over from the low-energy tracking sensor node and continue tracking, ensuring continuous system operation. The specific process is as follows:

[0079] 1) The energy of each moving node is updated based on its moving distance. The remaining energy of node i at time t is: ,in This indicates the energy consumption caused by the node's movement.

[0080] 2) When a node's energy falls below a preset threshold, a node replacement mechanism is triggered. The system selects candidate nodes from the currently active nodes of the perceived target and determines the optimal replacement node based on the following comprehensive scoring function: ,in Let represent the monitoring value of the i-th candidate sensor node for target P, calculated according to the following exponential decay function: , This represents the distance between the candidate sensor node and the target. This indicates the sensing radius of the preset sensor node; Let represent the remaining energy of the i-th candidate sensor node at time t. , Let represent the remaining energy of the i-th candidate sensor node at time t-1. This represents the energy consumption of the i-th candidate sensor node due to movement. , This is the energy consumption coefficient. Let be the moving distance of the i-th candidate sensor node at time t; This represents the coverage contribution of the i-th candidate sensor node. , This represents the total coverage area of ​​the sensor network. This represents the independent coverage area of ​​the i-th candidate sensor node; , , The weighting coefficients for each indicator are used; the candidate sensor node with the highest comprehensive score is selected as the optimal sensor node.

[0081] 3) If no sensor node is within the preset sensing radius of the target, meaning there are currently no available replacement nodes (i.e., no active node can sense the target), then the node rescue machine is activated; a rescue score is calculated for all nodes using the following formula: = ;in , , The weighting coefficient is used to select the sensor node with the highest rescue score as the replacement tracking sensor node. That is, priority is given to selecting nodes that are closer to the target, have higher remaining energy, and lower coverage contribution, so as to ensure that the rescue node can quickly approach the target and have less impact on the overall area coverage, while avoiding the selection of nodes with too low energy.

[0082] Once the node replacement and node rescue mechanisms are triggered, the system selects the optimal node (i.e., the node with the highest score) based on the corresponding rating and controls it to move toward the target to take over from the original low-energy node and continue to perform the tracking task.

[0083] This invention also proposes a dynamic optimization deployment system for intelligent sensor networks oriented towards energy-constrained tracking, used to implement the dynamic optimization deployment method for intelligent sensor networks oriented towards energy-constrained tracking described in the above embodiments; such as Figure 6 As shown, the system includes:

[0084] Location acquisition module 610, configured to acquire the initial position of the target;

[0085] The target trajectory prediction module 620 is configured to predict the target's trajectory in real time using an unscented Kalman filter algorithm;

[0086] The tracking node path planning module 630 is configured to perform path planning for the sensor nodes of the tracking target based on the predicted target trajectory using a reinforcement learning model based on deep deterministic policy gradients, i.e., dynamic optimization deployment.

[0087] For the undescribed parts of the intelligent sensor network dynamic optimization deployment system for energy-constrained tracking according to an embodiment of the present invention, please refer to the detailed description of the method embodiment above.

[0088] Another embodiment of the present invention provides an electronic device including at least one processor and a memory storing a computer program; when the computer program is read and executed by the processor, the electronic device performs a dynamic optimization deployment method for an energy-constrained tracking smart sensor network as described above.

[0089] Another embodiment of the present invention provides a readable storage medium storing a computer program, which, when read and executed by an electronic device, causes the electronic device to perform a dynamic optimization deployment method for energy-constrained tracking smart sensor networks as described above.

[0090] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A method for dynamic optimization deployment of smart sensor networks for energy-constrained tracking, characterized in that, The method comprises: acquiring an initial position of a target; predicting a moving track of the target in real time by using an unscented Kalman filter algorithm; planning a path for a sensor node tracking the target, i.e., dynamically optimizing deployment, by using a deep deterministic policy gradient-based reinforcement learning model according to the predicted moving track of the target; the deep deterministic policy gradient-based reinforcement learning model is constructed as follows: a state vector comprises distances between the target and multiple tracking sensor nodes, relative angles, and a current position of the target; an action vector comprises displacement amounts of the multiple tracking sensor nodes in a two-dimensional plane; a reward function R is designed based on a restricted Boltzmann machine and energy consumption as follows: ； wherein, , respectively represent the weight coefficients of KL divergence and energy consumption; represents the total energy consumption, , , represents the energy consumption of the i-th tracking sensor node, k is the energy consumption coefficient, is the moving distance of the i-th tracking sensor node at time t; represents the initial measurement value set of the plurality of tracking sensor nodes, including the distance from the target, the relative angle and the current position of the target; represents the probability distribution of the visible layer in the restricted Boltzmann machine corresponding to the initial measurement value set; represents the measurement value set after the movement of the plurality of tracking sensor nodes, represents the probability distribution of the visible layer in the restricted Boltzmann machine corresponding to the measurement value set after the movement of the nodes; the KL divergence value , represents the energy value of , represents the energy value of , is the symbol of the splicing operation.

2. The method of claim 1, wherein, the predicting the moving track of the target in real time by using the unscented Kalman filter algorithm comprises: establishing a state vector of the target, the state vector comprising a target position and a velocity component; and establishing a state transition equation of the target as a uniform linear motion equation; state prediction: generating a set of Sigma points according to a current state of the target and a covariance matrix; predicting a state by substituting the generated Sigma points into the state transition equation, and recombining the Sigma points propagated through the state transition equation according to weights to obtain a predicted mean value and a predicted covariance of the state of the target at time t; state update: calculating a predicted monitoring value by substituting each Sigma point into a measurement model; obtaining a predicted monitoring value of the target by weightedly averaging the predicted monitoring values of all the Sigma points; calculating a covariance matrix of the predicted monitoring value of the target; calculating a Kalman gain according to the covariance matrix of the predicted monitoring value of the target and an observation covariance matrix; updating a state estimation of the target and a state covariance matrix by using the Kalman gain; and finally outputting the state estimation of the target, i.e., a predicted value of the target position.

3. The method of claim 1, wherein, The method further comprises: in a target tracking process, calculating a residual energy of a tracking sensor node in real time; when the residual energy of the tracking sensor node is less than a preset threshold, determining whether there is another candidate sensor node according to whether a distance to the target is within a preset sensing radius of the sensor node; if there is, selecting an optimal sensor node from the other candidate sensor nodes according to the residual energy, a coverage contribution degree, and the distance to the target, to replace the tracking sensor node with low energy to continue tracking.

4. The method of claim 3, wherein, The selecting the optimal sensor node from the other candidate sensor nodes according to the residual energy, the coverage contribution degree, and the distance to the target comprises: The overall score of each sensor node is calculated according to the following formula : ; wherein, , , respectively represent the weight coefficient of each index; represents the monitoring value of the target by the ith candidate sensor node, , , represents the distance between the candidate sensor node and the target, represents the preset sensing radius of the sensor node; represents the residual energy of the ith candidate sensor node at time t, , represents the residual energy of the ith candidate sensor node at time t-1, represents the energy consumption of the ith candidate sensor node due to movement, , is the energy consumption coefficient, is the movement distance of the ith candidate sensor node at time t; represents the coverage contribution degree of the ith candidate sensor node, , represents the total coverage area of the sensor network, represents the independent coverage area of the ith candidate sensor node; selecting a candidate sensor node with a highest comprehensive score as the optimal sensor node.

5. The method of claim 4, wherein, The method further comprises: if no sensor node is within a preset sensing radius of the target, determining that no active node can sense the target, and then calculating a rescue score for all sensor nodes that are not in a tracking state : = ； wherein , , respectively represent the weight coefficients of each term; selecting a sensor node with a highest rescue score as the replacement tracking sensor node.

6. The system for implementing the method for dynamic optimization deployment of energy-constrained tracking-oriented smart sensor network according to any one of claims 1-5, characterized in that, The system comprises: a position acquisition module configured to acquire an initial position of a target; a target track prediction module configured to predict a moving track of the target in real time by using an unscented Kalman filter algorithm; a tracking node path planning module configured to plan a path for a sensor node tracking the target, i.e., dynamically optimize deployment, by using a deep deterministic policy gradient-based reinforcement learning model according to the predicted moving track of the target.

7. An electronic device, comprising: A computer program product comprising a memory, a processor, and a computer program; wherein the computer program is stored in the memory and configured to be executed by the processor to implement the method for dynamic optimization deployment of smart sensor network oriented energy-constrained tracking according to any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, A computer program product comprising a memory, a processor, and a computer program; wherein the computer program is stored in the memory and configured to be executed by the processor to implement the method for dynamic optimization deployment of smart sensor network oriented energy-constrained tracking according to any one of claims 1 to 5. A computer program product comprising a memory, a processor, and a computer program; wherein the computer program is stored in the memory and configured to be executed by the processor to implement the method for dynamic optimization deployment of smart sensor network oriented energy-constrained tracking according to any one of claims 1 to 5.

Images