Large-scale Internet of Things data collection system and method based on dual pre-segmentation

Abstract

The invention belongs to the technical field of robot control, and more specifically relates to a large-scale Internet of Things data collection system and method based on dual pre-segmentation. The method comprises the steps that sensor nodes in the Internet of Things are pre-segmented, all the sensor nodes in a monitoring area are divided into a plurality of clusters, and each cluster is managed by a cluster head node; dividing the monitoring area into a plurality of sub-areas according to the number of the unmanned aerial vehicles and the data distribution characteristics; designing a collision-free flight path for the unmanned aerial vehicle group based on deep reinforcement learning; the unmanned aerial vehicle moves in the detection area according to a flight path provided in the path planning process, accesses the sensor nodes and collects data; and recording the states of the accessed nodes and the non-accessed nodes at the same time. In a large-scale complex Internet of Things environment, the unmanned aerial vehicle with the number far smaller than that of sensor nodes can achieve collection of a large amount of data, the path of the unmanned aerial vehicle is reasonably planned to achieve collision-free planning, and the node data collection task is efficiently completed.

Description

Technical Field

[0001] The present invention belongs to the technical field of robot control, and more specifically, relates to a large-scale Internet of Things data collection system and method based on double pre-segmentation. Background Art

[0002] With the rapid development of the industrial Internet, the manufacturing industry is transforming towards digitalization and intelligence. Deploying IoT sensor nodes in industrial environments to form wireless sensor networks can achieve real-time monitoring and data collection of the production environment, helping managers make timely decisions to improve production efficiency and reduce losses. Due to their excellent maneuverability, drones have many advantages in assisting wireless sensor nodes in data collection, such as rapid deployment, reduced manpower and material consumption, and improved work efficiency. As the scale of nodes continues to expand, a single drone is limited by battery energy and coverage, and is insufficient in performing time-sensitive tasks in large or complex areas. Rationally scheduling multiple drones through deep reinforcement learning technology and achieving collaborative execution between drones is an effective way to solve the above problems, which has important research value and practical application significance.

[0003] However, as industrial sites expand further, a large number of sensor nodes are needed to achieve comprehensive monitoring. When the number of drones is far less than the number of sensor nodes, simply increasing the number of drones cannot complete the data collection task, and the increase in the number of drones will increase the risk of collision. Therefore, how to effectively allocate nodes to each drone and avoid collisions when drone groups collaborate becomes a major technical challenge.

[0004] The current traditional path planning methods have problems such as large amount of calculation and low efficiency in large-scale complex IoT environments, and the reinforcement learning method is plagued by the "curse of dimensionality". Therefore, in order to realize real-time data collection of sensor networks, reasonably allocate nodes and avoid collisions when drone swarms perform tasks, it is urgent to build an effective data collection method that combines node division, area division and path planning.

[0005] Chinese patent document CN117406753A discloses a global path planning method for unmanned agricultural machinery in irregular environments, including the following steps: step S1, data collection: collect farmland environment data, crop planting pattern data, agricultural machinery operation data and agricultural machinery energy consumption data; step S2, deep learning model training: use the data collected in step S1 to train the deep learning model, take the agricultural machinery energy consumption as the objective function, and train the model to learn the optimal path planning; step S3, farmland environment perception: use sensors and visual systems to perceive the current farmland environment and obtain farmland environment data; step S4, agricultural machinery collaborative operation path planning: use the trained deep learning model to plan the optimal path for agricultural machinery collaborative operation according to the current farmland environment data; step S5, agricultural machinery operation: perform agricultural machinery collaborative operation according to the planned optimal path, and at the same time, collect agricultural machinery operation data and energy consumption data for further training and optimization of the deep learning model. However, although the patent applies deep reinforcement learning for path planning, it still cannot solve the problem that in a large-scale and complex Internet of Things environment, it is impossible to achieve effective collaboration of machines. If the number of agricultural machinery is blindly increased, it may lead to collision risks.

[0006] In view of this, the present invention designs a large-scale Internet of Things data collection system and method based on double pre-segmentation, which solves a series of problems faced in large-scale wireless Internet of Things due to the expansion of industrial sites, numerous sensor nodes, relatively small number of drones and limited energy. Summary of the invention

[0007] The present invention aims to overcome at least one defect of the above-mentioned prior art and provide a large-scale Internet of Things data collection system based on double pre-segmentation.

[0008] The present invention also discloses a large-scale Internet of Things data collection method based on double pre-segmentation. In a large-scale and complex Internet of Things environment, drones with far fewer than the number of sensor nodes can collect a large amount of data, and reasonably plan the drone paths to achieve collision-free planning, thereby efficiently completing the node data collection task.

[0009] The detailed technical scheme of the present invention is as follows: A large-scale Internet of Things data collection system based on double pre-segmentation, the system comprising: a node pre-segmentation module, an area pre-segmentation module, a path planning module and a data collection module; The node pre-segmentation module is used to group and manage sensor nodes in the Internet of Things, dividing all sensor nodes in the monitoring area into multiple clusters, each cluster is managed by a cluster head node; the selection of the cluster head node is based on the Euclidean distance between nodes, ensuring that the distance between nodes in the cluster is minimized, thereby providing a basis for subsequent data transmission and path planning; The area pre-segmentation module is used to divide the monitoring area into multiple sub-areas according to the number of drones and data distribution characteristics. Specifically, the diagonal line of the monitoring area is used as the baseline, and the baseline segment is divided into several segments. The entire area is divided into multiple closed sub-areas according to these segmentation points. The corresponding drone is responsible for data collection in each sub-area to ensure balanced distribution of tasks. The path planning module is based on deep reinforcement learning technology, combined with convolutional neural networks and deep Q networks to design collision-free flight paths for drone swarms; the path planning module dynamically plans the optimal path through the local observation information of the drones to ensure that the drones can efficiently access sensor nodes and avoid collisions; The data collection module is used to perform the data collection task of the drone: according to the flight path provided by the path planning module, the drone moves in the monitoring area, visits sensor nodes and collects data; the data collection module is also responsible for recording the status of visited nodes and unvisited nodes to ensure that the data of all nodes can be efficiently collected.

[0010] Preferably, according to the present invention, the data communication connection relationship between the modules is as follows: the node pre-segmentation module first groups and manages the sensor nodes in the monitoring area, generates allocation information of the cluster head node and the nodes within the cluster, and passes this information to the path planning module; the area pre-segmentation module divides the monitoring area into multiple sub-areas according to the monitoring area and the number of drones, and passes the division results to the path planning module; the path planning module generates a collision-free flight path based on the node grouping information of the node pre-segmentation module and the sub-area division results of the area pre-segmentation module, combined with the real-time observation data of the drone, and passes the path information to the data collection module; the data collection module guides the drone to perform data collection tasks according to the flight path, and feeds the collected data back to the system for further analysis and processing.

[0011] A large-scale Internet of Things data collection method based on double pre-segmentation, the method comprising: S1, node pre-segmentation: pre-segment the sensor nodes in the Internet of Things, divide all sensor nodes in the monitoring area into multiple clusters, each cluster is managed by a cluster head node; S2, regional pre-segmentation: according to the number of drones and data distribution characteristics, the monitoring area is divided into multiple sub-areas; S3, Path Planning: Design collision-free flight paths for drone swarms based on deep reinforcement learning; S4. The drone moves within the detection area according to the flight path provided by the path planning process, visits sensor nodes and collects data; at the same time, it records the status of visited nodes and unvisited nodes.

[0012] Preferably, according to the present invention, step S1 specifically comprises the following steps: S1.1, Allocation: Set the number of cluster heads K, randomly select K nodes as initial cluster heads, and for each non-cluster head node, assign it to the cluster where the nearest cluster head is located; S1.2. Update cluster heads: For each cluster, calculate the sum of the distances from all nodes to each point in the cluster, and select the point with the smallest sum of distances as the new cluster head. The current cluster head of the cluster is , the set of nodes in the cluster is , the process of updating the cluster head can use the following formula: (1) In formula (1), represents the updated cluster head, Is a node and nodes The distance measure between them is the Euclidean distance; S1.3, repeat steps S1.1 and S1.2 until the cluster head node no longer changes or the predetermined number of iterations is reached; S1.4. Output cluster assignment and final cluster head node.

[0013] Preferably, according to the present invention, step S2 specifically includes the following steps: S2.1. Set the diagonal line of the monitoring area as the reference baseline; S2.2, segmentation point: find all equal points, denoted as , is the number of drones; S2.3, Regional division: Connect to the starting point and the end point, dividing the monitoring area into closed sub-regions, each of which contains a cluster head node; the kth equally divided point is connected to the connecting line segment between the starting point and the end point and Indicates The boundaries of the sub-areas are divided into non-overlapping sets to improve the efficiency of multi-target UAV group path planning and data collection; S2.4. Output: Each sub-area is the output result.

[0014] Preferably, according to the present invention, step S3 comprises the following steps: S3.1, obtain the pre-segmentation node and pre-segmentation area information, and the UAV collects local observation data according to the cluster head node in the given area; S3.2. Design drone status representation, drone actions and rewards: S3.3, convert the local observation results of the drone into a matrix, input it into the convolutional neural network for processing, reduce the data dimension, and extract features; S3.4, input the features extracted by the convolutional neural network into the deep Q network for training; S3.5. The trained deep Q-network is deployed on drones to perform data collection tasks in a surveillance environment.

[0015] Preferably, the design of drone status representation, drone actions and rewards are as follows: The drone state is the information observed locally by the drone and can be expressed as: (2) In formula (2), Indicates the current position coordinates of the drone. represents the location of obstacles in the environment, Indicates the distribution location of sensor nodes; The monitoring area is gridded, and the target area is constructed into a grid map with p×q rectangular areas. The length and width of each rectangular area are ,but: (3) In formula (3), Indicates the length of the target area, Indicates the width of the target area; Design eight drone actions based on drone local observation , corresponding to moving to 4 adjacent grids and 4 diagonal grids, represented by numbers: 4 adjacent directions (1, 2, 3, 4), 4 diagonal directions (5, 6, 7, 8); The reward includes a safety distance constraint , Action Execution Reward and node access rewards : (4) In formula (4), Indicates drone With obstacles The distance between Indicates drone With drones The distance between Represents the preset safety threshold. This constraint is used to encourage drones to maintain a safe distance between each other and between drones and obstacles. (5) In formula (5), Indicates the drone action. Indicates that the drone moves to 4 adjacent grids. This reward is used to encourage the drone to reduce unnecessary paths and reach the destination quickly; Set two collections, the visited node collection and the set of unvisited nodes ; Whenever a drone visits node n for the first time, the node will be Remove and add , and the drone gets certain rewards: (6).

[0016] Preferably, according to the present invention, the features extracted by the convolutional neural network are input into the deep Q network for training, specifically as follows: S3.4.1. Initialize the predicted Q value network and target Q value network ,in and are the parameters of the two networks respectively; initialize the experience replay pool D to store the experience of the drone's interaction with the environment; set the hyperparameters, learning rate , discount factor γ, exploration rate , experience pool D; S3.4.2. Interaction with the environment: At each time step t, the drone interacts with the environment according to its current state. Select Action , with exploration rate Choose actions randomly, or with probability Select the action with the largest current Q value , perform the action , observe the reward returned by the environment and the next state , and transfer the state sample Stored in the experience replay pool; S3.4.3. Randomly sample a batch of experience from the experience replay pool Put it into the network for training; S3.4.4. Calculate the target Q value : (7) In formula (7), is the immediate reward of the current step, is a discount factor that measures the importance of future rewards, Is to perform an action The next state after is the parameter of the target Q-value network, Indicates the action that can obtain the maximum Q value; S3.4.5. Calculate the loss function: DQN measures the predicted Q value through the mean square error loss function and target Q value The difference between them is used to judge the quality of training. The smaller the difference, the better the training. Gradient It gradually decreases and becomes stable, indicating that the model has converged and the training is completed: (8); S3.4.6. Update the parameters of the predicted Q-value network using gradient descent : (9) In formula (9), is the learning rate, is the loss function over parameters The gradient of S3.4.7. At certain time steps, the parameters of the Q-value network are predicted Parameters copied to the target Q network ; S3.4.8. As training progresses, gradually reduce the exploration rate , so that drones gradually transition from exploration-oriented to utilization-oriented; S3.4.9, repeat steps S3.4.2 to S3.4.8 until the model converges or reaches the predetermined number of training rounds. Finally, through repeated training, the parameters of the Q value network are predicted. Will gradually converge, the predicted Q value It can accurately reflect the long-term value of each action under different states, thereby guiding the drone to choose the optimal action.

[0017] Compared with the prior art, the present invention has the following beneficial effects: (1) The present invention provides node pre-segmentation, which effectively groups and manages IoT nodes through a specific algorithm, providing a basis for subsequent data transmission. Then, combined with regional pre-segmentation, the monitoring area is reasonably divided into multiple sub-areas according to the number of drones and data distribution characteristics, ensuring that the tasks in each sub-area are balanced and convenient for drones to perform tasks.

[0018] (2) The present invention is based on deep reinforcement learning technology and takes the observation information of the UAV as the basis. It combines the convolutional neural network (CNN) and the deep Q network (DQN) to achieve optimal planning of the flight path, enabling the UAV to select the optimal action according to the dynamic changes of the environment and mission requirements, avoid collisions and efficiently collect data.

[0019] (3) The present invention achieves collision-free path planning and efficiently completes the IoT data collection task by combining the dual pre-segmentation strategy and deep reinforcement learning technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a structural diagram of the Internet of Things data collection system described in the present invention.

[0021] Figure 2 It is a schematic diagram of node pre-segmentation of the present invention.

[0022] Figure 3 It is a schematic diagram of regional pre-segmentation of the present invention.

[0023] Figure 4 It is a schematic diagram of the Internet of Things data collection method of the present invention.

[0024] Figure 5 It is a rendering of the effect of the UAV collecting data in the detection area according to the flight path in an embodiment of the present invention. DETAILED DESCRIPTION

[0025] The present disclosure is further described below in conjunction with the accompanying drawings and embodiments.

[0026] Embodiment 1, Ginseng Figure 1 ,This embodiment provides a large-scale Internet of Things data collection system based on double pre-segmentation, the system includes: a node pre-segmentation module, an area pre-segmentation module, a path planning module and a data collection module; The node pre-segmentation module is responsible for grouping and managing sensor nodes in the Internet of Things. All sensor nodes in the monitoring area are divided into multiple clusters through the K-medoids algorithm, and each cluster is managed by a cluster head node. The selection of the cluster head node is based on the Euclidean distance between nodes to ensure that the distance between nodes in the cluster is minimized, thereby providing a basis for subsequent data transmission and path planning.

[0027] The area pre-segmentation module divides the monitoring area into multiple sub-areas according to the number of drones and data distribution characteristics. The specific method is to use the diagonal line of the monitoring area as the baseline, divide the baseline segment into several equal segments, and divide the entire area into multiple closed sub-areas according to these segmentation points. The corresponding drone is responsible for data collection in each node in each sub-area to ensure balanced distribution of tasks.

[0028] The path planning module is based on deep reinforcement learning technology, combined with convolutional neural network (CNN) and deep Q network (DQN) to design collision-free flight paths for drone swarms. The module dynamically plans the optimal path through the local observation information of the drones, such as the current position, obstacle location, sensor node distribution, etc., to ensure that the drones can efficiently access sensor nodes and avoid collisions.

[0029] The data collection module is responsible for executing the data collection task of the drone. According to the flight path provided by the path planning module, the drone moves in the monitoring area, visits sensor nodes and collects data. The module is also responsible for recording the status of visited and unvisited nodes to ensure that the data of all nodes can be efficiently collected.

[0030] The data communication connection relationship between the modules is as follows: the node pre-segmentation module first groups and manages the sensor nodes in the monitoring area, generates the allocation information of the cluster head node and the nodes within the cluster, and passes this information to the path planning module; the area pre-segmentation module divides the monitoring area into multiple sub-areas according to the monitoring area and the number of drones, and passes the division results to the path planning module; the path planning module generates a collision-free flight path based on the node grouping information of the node pre-segmentation module and the sub-area division results of the area pre-segmentation module, combined with the real-time observation data of the drone, and passes the path information to the data collection module; the data collection module guides the drone to perform data collection tasks according to the flight path, and feeds the collected data back to the system for further analysis and processing.

[0031] Embodiment 2, Ginseng Figure 4 This embodiment provides a large-scale IoT data collection method based on double pre-segmentation, the method comprising: S1, node pre-division: Figure 2 As shown in the figure, the sensor nodes in the Internet of Things are pre-divided, and all sensor nodes in the monitoring area are divided into multiple clusters. Each cluster is managed by a cluster head node, as follows: S1.1, Allocation: Set the number of cluster heads K, randomly select K nodes as initial cluster heads, and for each non-cluster head node, assign it to the cluster where the nearest cluster head is located; S1.2. Update cluster heads: For each cluster, calculate the sum of the distances from all nodes to each point in the cluster, and select the point with the smallest sum of distances as the new cluster head. The current cluster head of the cluster is , the set of nodes in the cluster is , the process of updating the cluster head can use the following formula: (1) In formula (1), represents the updated cluster head, Is a node and nodes The distance measure between them is the Euclidean distance; S1.3, repeat steps S1.1 and S1.2 until the cluster head node no longer changes or the predetermined number of iterations is reached; S1.4. Output cluster assignment and final cluster head node.

[0032] S2, area pre-division: such as Figure 3 As shown in the figure, according to the number of drones and data distribution characteristics, the monitoring area is divided into multiple sub-areas, as follows: S2.1. Set the diagonal line of the monitoring area as the reference baseline; S2.2, segmentation point: find all equal points, denoted as , is the number of drones; S2.3, Regional division: Connect to the starting point and the end point, dividing the monitoring area into closed sub-regions, each of which contains a cluster head node; the kth equally divided point is connected to the connecting line segment between the starting point and the end point and Indicates The boundaries of the sub-areas are divided into non-overlapping sets to improve the efficiency of multi-target UAV group path planning and data collection; S2.4. Output: Each sub-area is the output result.

[0033] S3. Path planning: Design collision-free flight paths for drone swarms based on deep reinforcement learning, as follows: S3.1, obtain the pre-segmentation node and pre-segmentation area information, and the UAV collects local observation data according to the cluster head node in the given area; S3.2. Design drone status representation, drone actions and rewards: The drone state is the information observed locally by the drone and can be expressed as: (2) In formula (2), Indicates the current position coordinates of the drone. represents the location of obstacles in the environment, Indicates the distribution location of sensor nodes; The monitoring area is gridded, and the target area is constructed into a grid map with p×q rectangular areas. The length and width of each rectangular area are ,but: (3) In formula (3), Indicates the length of the target area, Indicates the width of the target area; Design eight drone actions based on drone local observation , corresponding to moving to 4 adjacent grids and 4 diagonal grids, represented by numbers: 4 adjacent directions (1, 2, 3, 4), 4 diagonal directions (5, 6, 7, 8); The reward includes a safety distance constraint , Action Execution Reward and node access rewards : (4) In formula (4), Indicates drone With obstacles The distance between Indicates drone With drones The distance between Represents the preset safety threshold. This constraint is used to encourage drones to maintain a safe distance between each other and between drones and obstacles. (5) In formula (5), Indicates the drone action. Indicates that the drone moves to 4 adjacent grids. This reward is used to encourage the drone to reduce unnecessary paths and reach the destination quickly; Set two collections, the visited node collection and the set of unvisited nodes ; Whenever a drone visits node n for the first time, the node will be Remove and add , and the drone gets certain rewards: (6).

[0034] S3.3, convert the local observation results of the drone into a matrix, input it into the convolutional neural network CNN for processing, reduce the data dimension, and extract features; S3.4. Input the features extracted by the convolutional neural network CNN into the deep Q network DQN for training, as follows: S3.4.1. Initialize the predicted Q value network and target Q value network ,in and are the parameters of the two networks respectively; initialize the experience replay pool D to store the experience of the drone's interaction with the environment; set the hyperparameters, learning rate , discount factor γ, exploration rate , experience pool D; S3.4.2. Interaction with the environment: At each time step t, the drone interacts with the environment according to its current state. Select Action , with exploration rate Choose actions randomly, or with probability Select the action with the largest current Q value , perform the action , observe the reward returned by the environment and the next state , and transfer the state sample Stored in the experience replay pool; S3.4.3. Randomly sample a batch of experience from the experience replay pool Put it into the network for training; S3.4.4. Calculate the target Q value : (7) In formula (7), is the immediate reward of the current step, is a discount factor that measures the importance of future rewards, Is to perform an action The next state after is the parameter of the target Q-value network, Indicates the action that can obtain the maximum Q value; S3.4.5. Calculate the loss function: DQN measures the predicted Q value through the mean square error loss function and target Q value The difference between them is used to judge the quality of training. The smaller the difference, the better the training. Gradient It gradually decreases and becomes stable, indicating that the model has converged and the training is completed: (8); S3.4.6. Update the parameters of the predicted Q-value network using gradient descent : (9) In formula (9), is the learning rate, is the loss function over parameters The gradient of S3.4.7. At certain time steps, the parameters of the Q-value network are predicted Parameters copied to the target Q network ; S3.4.8. As training progresses, gradually reduce the exploration rate , so that drones gradually transition from exploration-oriented to utilization-oriented; S3.4.9, repeat steps S3.4.2 to S3.4.8 until the model converges or reaches the predetermined number of training rounds. Finally, through repeated training, the parameters of the Q value network are predicted. Will gradually converge, the predicted Q value It can accurately reflect the long-term value of each action under different states, thereby guiding the drone to choose the optimal action.

[0035] S3.5. The trained DQN network is deployed on drones to perform data collection tasks in monitoring environments.

[0036] S4, such as Figure 5 As shown, the UAV moves in three sub-areas according to the flight path provided by the path planning process, visits sensor nodes and collects data; at the same time, it records the status of visited nodes and unvisited nodes; The three sub-areas include sub-area 1, sub-area 2 and sub-area 3. Connect to the start and end points, equally divided Connect the line segments between the start point and the end point and represents the boundary between sub-region 1 and sub-region 2; Connect to the start and end points, equally divided Connect the line segments between the start point and the end point and Indicates the boundary between sub-area 2 and sub-area 3. Each sub-area contains a cluster head node, and the drone moves according to the flight path in each area.

[0037] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the technical solution of the present invention, and are not intended to limit the specific implementation methods of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the claims of the present invention shall be included in the protection scope of the claims of the present invention.

Claims

1. A large-scale Internet of Things data collection system based on double pre-segmentation, characterized in that: The system comprises: a node pre-segmentation module, an area pre-segmentation module, a path planning module and a data collection module; The node pre-segmentation module is used to group and manage sensor nodes in the Internet of Things, dividing all sensor nodes in the monitoring area into multiple clusters, each cluster is managed by a cluster head node; the selection of the cluster head node is based on the Euclidean distance between nodes, ensuring that the distance between nodes in the cluster is minimized, thereby providing a basis for subsequent data transmission and path planning; The area pre-segmentation module is used to divide the monitoring area into multiple sub-areas according to the number of drones and data distribution characteristics. Specifically, the diagonal line of the monitoring area is used as the baseline, and the baseline segment is divided into several segments. The entire area is divided into multiple closed sub-areas according to these segmentation points. The corresponding drone is responsible for data collection in each sub-area to ensure balanced distribution of tasks. The path planning module is based on deep reinforcement learning technology, combined with convolutional neural networks and deep Q networks to design collision-free flight paths for drone swarms; the path planning module dynamically plans the optimal path through the local observation information of the drones to ensure that the drones can efficiently access sensor nodes and avoid collisions; The data collection module is used to perform the data collection task of the drone: according to the flight path provided by the path planning module, the drone moves in the monitoring area, visits sensor nodes and collects data; the data collection module is also responsible for recording the status of visited nodes and unvisited nodes to ensure that the data of all nodes can be efficiently collected.

2. According to claim 1, a large-scale Internet of Things data collection system based on double pre-segmentation is characterized in that: The data communication connection relationship between the modules is as follows: the node pre-segmentation module first manages the sensor nodes in the monitoring area in groups, generates the allocation information of the cluster head node and the nodes within the cluster, and passes this information to the path planning module; The regional pre-segmentation module divides the monitoring area into multiple sub-areas according to the monitoring area and the number of drones, and passes the division results to the path planning module; the path planning module generates a collision-free flight path based on the node grouping information of the node pre-segmentation module and the sub-area division results of the regional pre-segmentation module, combined with the real-time observation data of the drone, and passes the path information to the data collection module; The data collection module guides the UAV to perform data collection tasks according to the flight path and feeds the collected data back to the system for further analysis and processing.

3. A method for collecting large-scale Internet of Things data based on double pre-segmentation, using the large-scale Internet of Things data collection system based on double pre-segmentation as claimed in claim 1 or claim 2, characterized in that: The method comprises: S1, node pre-segmentation: pre-segment the sensor nodes in the Internet of Things, divide all sensor nodes in the monitoring area into multiple clusters, each cluster is managed by a cluster head node; S2, regional pre-segmentation: according to the number of drones and data distribution characteristics, the monitoring area is divided into multiple sub-areas; S3, Path Planning: Design collision-free flight paths for drone swarms based on deep reinforcement learning; S4. The drone moves within the detection area according to the flight path provided by the path planning process, visits sensor nodes and collects data; at the same time, it records the status of visited nodes and unvisited nodes.

4. The method for collecting large-scale IoT data based on double pre-segmentation according to claim 3, characterized in that: The step S1 specifically includes the following steps: S1.1, Allocation: Set the number of cluster heads K, randomly select K nodes as initial cluster heads, and for each non-cluster head node, assign it to the cluster where the nearest cluster head is located; S1.

2. Update cluster heads: For each cluster, calculate the sum of the distances from all nodes to each point in the cluster, and select the point with the smallest sum of distances as the new cluster head. The current cluster head of the cluster is , the set of nodes in the cluster is , the process of updating the cluster head can use the following formula: (1) In formula (1), represents the updated cluster head, Is a node and nodes The distance measure between them is the Euclidean distance; S1.3, repeat steps S1.1 and S1.2 until the cluster head node no longer changes or the predetermined number of iterations is reached; S1.

4. Output cluster assignment and final cluster head node.

5. The method for collecting large-scale IoT data based on double pre-segmentation according to claim 3, characterized in that: The step S2 specifically includes the following steps: S2.

1. Set the diagonal line of the monitoring area as the reference baseline; S2.2, segmentation point: find all equal points, denoted as , is the number of drones; S2.3, Regional division: Connect to the starting point and the end point, dividing the monitoring area into closed sub-regions, each of which contains a cluster head node; the kth equally divided point is connected to the connecting line segment between the starting point and the end point and Indicates The boundaries of the sub-areas are divided into non-overlapping sets to improve the efficiency of multi-target UAV group path planning and data collection; S2.

4. Output: Each sub-area is the output result.

6. The method for collecting large-scale IoT data based on double pre-segmentation according to claim 3, characterized in that: The step S3 comprises the following steps: S3.1, obtain the pre-segmentation node and pre-segmentation area information, and the UAV collects local observation data according to the cluster head node in the given area; S3.

2. Design drone status representation, drone actions and rewards: S3.3, convert the local observation results of the drone into a matrix, input it into the convolutional neural network for processing, reduce the data dimension, and extract features; S3.4, input the features extracted by the convolutional neural network into the deep Q network for training; S3.

5. The trained deep Q network is deployed on drones.

7. The method for collecting large-scale IoT data based on double pre-segmentation according to claim 6, characterized in that: The design of drone status representation, drone actions and rewards are as follows: The drone state is the information observed locally by the drone and can be expressed as: （2） In formula (2), Indicates the current position coordinates of the drone. represents the location of obstacles in the environment, Indicates the distribution location of sensor nodes; The monitoring area is gridded, and the target area is constructed into a grid map with p×q rectangular areas. The length and width of each rectangular area are ,but: （3） In formula (3), Indicates the length of the target area, Indicates the width of the target area; Design eight drone actions based on drone local observation , corresponding to moving to 4 adjacent grids and 4 diagonal grids, represented by numbers: 4 adjacent directions (1, 2, 3, 4), 4 diagonal directions (5, 6, 7, 8); The reward includes a safety distance constraint , Action Execution Reward and node access rewards : （4） In formula (4), Indicates drone With obstacles The distance between Indicates drone With drones The distance between Represents the preset safety threshold. This constraint is used to encourage drones to maintain a safe distance between each other and between drones and obstacles. （5） In formula (5), Indicates the drone action. Indicates that the drone moves to 4 adjacent grids. This reward is used to encourage the drone to reduce unnecessary paths and reach the destination quickly; Set two collections, the visited node collection and the set of unvisited nodes ; Whenever a drone visits a node n for the first time, the node will be Remove and add , and the drone gets certain rewards: （6）。 8. The method for collecting large-scale IoT data based on double pre-segmentation according to claim 6, characterized in that: The features extracted by the convolutional neural network are input into the deep Q network for training, as follows: S3.4.

1. Initialize the predicted Q value network and target Q value network ,in and are the parameters of the two networks respectively; initialize the experience replay pool D to store the experience of the drone's interaction with the environment; set the hyperparameters, learning rate , discount factor γ, exploration rate , experience pool D; S3.4.

2. Interaction with the environment: At each time step t, the drone interacts with the environment according to its current state. Select Action , with exploration rate Choose actions randomly, or with probability Select the action with the largest current Q value , perform the action , observe the reward returned by the environment and the next state , and transfer the state sample Stored in the experience replay pool; S3.4.

3. Randomly sample a batch of experiences from the experience replay pool and put them into the network for training; S3.4.

4. Calculate the target Q value : （7） In formula (7), is the immediate reward of the current step, is the discount factor, Is to perform an action The next state after is the parameter of the target Q-value network, Indicates the action that can obtain the maximum Q value; S3.4.

5. Calculate the loss function: DQN measures the predicted Q value through the mean square error loss function and target Q value The difference between: （8）； S3.4.

6. Update the parameters of the predicted Q-value network using gradient descent : （9） In formula (9), is the learning rate, is the loss function over parameters The gradient of S3.4.

7. At certain time steps, the parameters of the Q-value network are predicted Parameters copied to the target Q network ; S3.4.

8. As the training progresses, the exploration rate ϵ is gradually reduced, so that the UAV gradually transitions from exploration-oriented to utilization-oriented. S3.4.9, repeat steps S3.4.2 to S3.4.8 until the model converges or reaches the predetermined number of training rounds. Finally, through repeated training, the parameters of the Q value network are predicted. Will gradually converge, the predicted Q value It can accurately reflect the long-term value of each action under different states, thereby guiding the drone to choose the optimal action.

Images