A multi-participant selection method for heterogeneous crowd sensing

Abstract

The present invention relates to a multi-participant selection method for heterogeneous crowd intelligence perception. The specific process is as follows: for heterogeneous participants in the context of heterogeneous crowd intelligence perception: human participants, unmanned vehicles and fixed sensors, a unified definition and quantification method of perception capability and cost are first given. For the perception area, according to the distribution of POIs, obstacles and charging stations where perception data is to be collected, a sub-area division method based on GraphSAGE is proposed by comprehensively considering the time and space characteristics. Finally, combining the spatiotemporal characteristics of the sub-area and the attributes of the heterogeneous participants, a heterogeneous participant selection algorithm based on the proximal strategy optimization algorithm in deep reinforcement learning is proposed to solve the collaborative scheduling problem of the sub-area, with the goal of maximizing the perception data collection rate and the balance of the perception data and minimizing the perception data cost, so as to achieve effective collaboration of heterogeneous participants under heterogeneous crowd intelligence perception.

Description

Technical Field

[0001] The present invention relates to the field of mobile crowd sensing and reinforcement learning based on deep learning, and in particular to a multi-participant selection method in heterogeneous crowd sensing. Background Art

[0002] In recent years, with the rapid development of the Internet and mobile devices, the perception and computing capabilities of smart devices have become greatly surplus. Mobile Crowd Sensing has become a popular perception mode. It utilizes the surplus sensor perception and computing capabilities of mobile devices and the mobility of human beings to conduct conscious or unconscious collaboration through the mobile Internet, realizes perception task distribution and perception data collection, and completes large-scale and complex social perception tasks at a relatively low cost.

[0003] Although mobile crowd sensing has many advantages, due to the limitations of human activity patterns, the data quality in different time and space areas is always unbalanced. Large perception blind spots often appear in places where few people go and at night when few people are active. With the development of smart cities, a large number of heterogeneous perception tasks involve various fields of life, such as environmental monitoring, smart transportation, urban management, public safety, etc. These tasks are different in time and space dimensions, and some sensors built into the mobile devices carried by pedestrians who perform crowd sensing tasks may not meet the perception needs of a large number of new heterogeneous tasks.

[0004] With the development of technology, the main force to achieve high-quality perception in cities is not only people carrying smart mobile devices, but also drones and a large number of heterogeneous sensors deployed in urban areas. Professor Yu Zhiwen and others proposed the next generation of MCS (MCS2.0) - heterogeneous crowd sensing (HCS) by integrating these functions. By integrating human participants, drones and fixed sensor groups, crowd sensing with a larger deployment scale, wider perception range, more comprehensive perception capabilities, and more continuous perception cycles can be achieved.

[0005] Despite its attractive advantages, heterogeneous crowd sensing still faces two major challenges. First, due to the different types of participants, heterogeneous participants have different perception capabilities and perception costs. Therefore, how to unify their modeling is the prerequisite for selecting heterogeneous participants to complete perception tasks. Second, compared with mobile traditional crowd sensing, heterogeneous crowd sensing includes a larger scale of participants and a more complex perception environment. Therefore, how to coordinate and schedule according to the ability cost differences of heterogeneous participants and the characteristics of the perception environment to achieve complementary perception advantages of heterogeneous participants in human-machine-object fusion is different from the participant selection problem under MCS. The participant selection problem under MCS is to optimize the participant selection decision for a single type of participant and a simple perception environment.

[0006] Although some works mentioned the idea of ​​combining mobile crowdsensing with drones or sensors, they only focused on how to unilaterally schedule these nodes to assist in sensor data collection, without considering how to schedule human participants with different sensing paradigms, as well as the collaboration of unmanned vehicle nodes and fixed sensors to complete sensing tasks. And they do not consider the spatiotemporal characteristics of the sensing environment in real urban scenarios, for example, human activity patterns are time-dependent in the temporal domain. In the spatial domain, road traffic in the central city is complex and dense, while roads in the suburbs are wider and have fewer buildings. Therefore, there is a lack of a unified spatiotemporal feature-based heterogeneous participant selection and collaborative scheduling framework to solve the perception imbalance or blind spot problem in mobile crowdsensing. Summary of the invention

[0007] Technical issues to be solved

[0008] In order to avoid the shortcomings of the prior art, the present invention proposes a multi-participant selection method for heterogeneous crowd intelligence sensing.

[0009] Technical Solution

[0010] A multi-participant selection method for heterogeneous crowd intelligence perception, characterized by the following steps:

[0011] Step 1: Define the nodes of heterogeneous crowd sensing that include human participants (U = {u1,u2,...,u N}), unmanned vehicle node (V={v1,v2,...,v M}) and fixed sensor nodes (S = {s1,s2,...,s K}), where the human actor node contains the attributes in, represents the position of the human participant at time slot t, represents the perceptual ability of human participants, Represents the perceived cost of human participants; the driverless car node contains the attribute in, represents the position of the unmanned vehicle node at time slot t, represents the perception capability of the unmanned vehicle node, Represents the perception cost of the unmanned vehicle node; the fixed sensor node contains the attribute in, represents the position of the fixed sensor node at time slot t, represents the sensing capability of the fixed sensor node, represents the perception cost of fixed sensor nodes; considering the differences between the three types of heterogeneous participant nodes, human participants, unmanned vehicles and fixed sensors in performing perception tasks, the perception capability is defined as the perception range and mobility; the definitions of the perception capabilities of the three types of heterogeneous participants are given:

[0012]

[0013] in, They represent the perception radius of human participants, unmanned vehicle nodes and fixed sensor nodes respectively, They represent the maximum moving distance of human participants, unmanned vehicle nodes and fixed sensor nodes at time slot t respectively; POI data defined within the perception radius of heterogeneous participants can be collected, and the time period is divided into a set of perception time slots of equal length T = {t1, t2, ..., t n}, select appropriate heterogeneous participants to complete the sensing task in each time slot by designing a scheduling strategy;

[0014] Step 2: Define a region containing a set of POIs to be collected: P = {p1, p2, ..., p I}; Calculate the data collection rate, data collection fairness and data collection cost of POI at time slot t; Given the heterogeneous participant selection strategy π, the data collection rate of POI at time slot t is defined as follows:

[0015]

[0016] Among them, d t (π; p i ) represents the time slot t p under strategy π i Cumulative data collection rate;

[0017] To avoid perceptual imbalance or blind spots in the data collection process, Jain’s fairness index is used to evaluate the fairness of data collection of the POI set at time slot t; the data collection fairness of the POI at time slot t is defined as follows:

[0018]

[0019] The perception costs of heterogeneous participants are defined as mobility costs and data collection costs. For human participants, unmanned vehicle nodes, and fixed sensor nodes, their costs are defined as follows:

[0020]

[0021] in, is the weight calculated for the moving cost, is the data collection cost, The penalty for the unmanned vehicle node hitting an obstacle;

[0022] Step 3: Based on the perception area division method of spatiotemporal features, the perception area for each time slot t is defined as G t =(V t ,ε t ,X t );V t represents the set of POI points in the sensing area, ε t is the adjacency matrix, indicating the reachability between nodes; X t Represents node features; G t All nodes in are sorted in descending order by degree, and the first C nodes are selected as the central nodes of each sub-region; the breadth-first search algorithm BFS is used to extract h-hop nodes in the sub-region, and there are s nodes in each sub-region; C and s are used to maximize the original graph G t The coverage rate of

[0023] Step 4: Use the GraphSAGE algorithm to extract the features of the nodes in the sub-region; the input of the first layer is In each layer of the neural network, for each node First of all The embedding vectors of the adjacent nodes are transformed nonlinearly; max-pooling is used to calculate the nodes The feature representation of all adjacent nodes in the kth layer is as follows:

[0024]

[0025] in, Representation Node Representation Node The embedding of the neighboring node u in the k-th layer, Representation Node Feature representation of all adjacent nodes at layer k; and vertices Perform nonlinear transformation on the representation vectors of the k-1 layers and concatenate the results to obtain the representation vector of the central node of the kth layer:

[0026]

[0027] After the above GraphSAGE layer operation, each node of the input sub-region aggregates the features of all its h-hop neighbor nodes; the feature vector is then input into the multi-layer perceptron MLP, which maps it to a scalar and concatenates it with other attributes observed in STPPO;

[0028] Step 5: In time slot t, cluster sub-regions The features of all nodes in the region are converted into one-dimensional scalars as sub-region observations. part of; The inputs to the actor network include: the spatiotemporal characteristics of the sub-region, the data collection rate, the locations of all heterogeneous actors, and the selected set of heterogeneous actors in the sub-region; then, the sub-region actor network provides a The action distribution is obtained by assuming that the action is sampled using the above distribution. Given a sensing area G and a scheduling time slot T, the goal of this invention is to design an optimal scheduling strategy π to maximize the overall performance, which is defined as the weighted ratio of the overall data collection rate of the POI set and the collection cost of heterogeneous participants with comprehensive data fairness:

[0029]

[0030] Among them, ξ(π) represents the data collection rate of POI when adopting strategy π, F(π) represents the fairness of data collection when adopting strategy π, and Cost(π) represents the collection cost when adopting strategy π; the joint action a is obtained from all sub-areas t , calculate the total reward R t , transition to the next state s t+1 , and the rewards for each sub-region Each sub-area is in buffer zone B c Stores the observed Then all stored observations are input into the critic network to obtain the state value function of all states

[0031] Step 6: From buffer B of each sub-region c Randomly select L consecutive time slots of information sets for training; use To calculate the advantage function and c-loss, use PPO to update the critic network; for the participant network, a-loss is calculated by calculating the importance weights between the old participant network and the new participant network, and the PPO algorithm is used to update the participant network.

[0032] A computer system, characterized in that it comprises: one or more processors, and a computer-readable storage medium for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the above-mentioned method.

[0033] A computer-readable storage medium, characterized in that it stores computer-executable instructions, which are used to implement the above method when executed.

[0034] A computer program, characterized by comprising computer executable instructions, which are used to implement the above method when executed.

[0035] Beneficial Effects

[0036] The present invention proposes a multi-participant selection method for heterogeneous crowd intelligence perception, which solves the problem of uneven collection of perception data and even perception blind spots caused by the mobility patterns of human mobile participants in crowd intelligence perception. Based on the background of heterogeneous crowd intelligence perception, the invention first gives the definition of the perception ability and cost of heterogeneous participants. Then, in view of the problem that the perception area is large and the situation is complex, an algorithm for sub-region construction and spatiotemporal region feature extraction based on the GraphSAGE algorithm is proposed. Finally, combining the perception ability and cost of heterogeneous participants and the spatiotemporal characteristics of sub-regions, a selection algorithm for heterogeneous participants is proposed based on a deep reinforcement learning algorithm.

[0037] 1. By uniformly modeling the capabilities and costs of heterogeneous participants and combining the spatiotemporal characteristics of the region, the problem of coordinated scheduling of heterogeneous participants in each sub-region is solved by using sub-region construction and a heterogeneous participant selection algorithm based on deep reinforcement learning, thus achieving the overall optimization goal.

[0038] 2. A participant selection algorithm is designed for heterogeneous participants. Different from the traditional crowd-sensing participant selection algorithm, the algorithm includes heterogeneous participants with large differences in capabilities and costs. Aiming at the differences in capabilities and costs of heterogeneous participants and the spatiotemporal characteristics of the perception area, the present invention realizes the coordinated scheduling of heterogeneous participants in a complex perception environment through a sub-region construction algorithm based on spatiotemporal characteristics and a heterogeneous participant selection algorithm based on reinforcement learning, with the goal of maximizing the perception data collection rate, the fairness of perception data collection and minimizing the perception cost, thus achieving good system utility. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] The drawings are only for the purpose of illustrating particular embodiments and are not to be considered limiting of the present invention. Like reference symbols denote like components throughout the drawings.

[0040] Figure 1 It is a flow chart of a multi-participant selection method for heterogeneous crowd intelligence sensing in an example of the present invention;

[0041] Figure 2 This is a structural diagram of heterogeneous crowd intelligence perception;

[0042] Figure 3 Schematic diagram of the perception area division. DETAILED DESCRIPTION

[0043] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0044] There are 3 drawings in this invention, please refer to Figure 1 As shown, the specific steps of the present invention are as follows:

[0045] Step 1: The present invention defines heterogeneous crowd sensing including human participant nodes (U = {u1, u2, ..., u N}), unmanned vehicle node (V={v1,v2,...,v M}) and fixed sensor nodes (S = {s1,s2,...,s K}), where the human actor node contains the attributes in, represents the position of the human participant at time slot t, represents the perceptual ability of human participants, Represents the perceived cost of human participants; the driverless car node contains the attribute in, represents the position of the unmanned vehicle node at time slot t, represents the perception capability of the unmanned vehicle node, represents the perception cost of the unmanned vehicle node; the fixed sensor node contains the attributes in, represents the position of the fixed sensor node at time slot t, represents the sensing capability of the fixed sensor node, Represents the perception cost of fixed sensor nodes. For the three types of heterogeneous participant nodes, human participants, unmanned vehicles, and fixed sensors, the perception capability is defined as the perception range and mobility considering their differences in performing perception tasks. The definitions of the perception capabilities of the three types of heterogeneous participants are given as follows:

[0046]

[0047] in They represent the perception radius of human participants, unmanned vehicle nodes and fixed sensor nodes respectively, Respectively represent the maximum moving distance of human participants, unmanned vehicle nodes and fixed sensor nodes at time slot t. POI data defined within the perception radius of heterogeneous participants can be collected. The time period is divided into a set of perception time slots of equal length. By designing a scheduling strategy, appropriate heterogeneous participants are selected to complete the sensing task in each time slot.

[0048] Step 2: Define a region containing a set of POIs to be collected: P = {p1, p2, ..., p I}. Calculate the data collection rate, data collection fairness and data collection cost of POI at time slot t. Given the heterogeneous participant selection strategy π, the data collection rate of POI at time slot t is defined as follows:

[0049]

[0050] d t (π; p i ) represents the time slot t p under strategy π i The cumulative data collection rate.

[0051] To avoid perceptual imbalance or blind spots in the data collection process, the present invention uses Jain's fairness index to evaluate the data collection fairness of the POI set at time slot t. The data collection fairness of the POI at time slot t is defined as follows:

[0052]

[0053] The present invention defines the perception cost of heterogeneous participants as mobility cost and data collection cost. For human participants, unmanned vehicle nodes and fixed sensor nodes, their costs are defined as follows:

[0054]

[0055] in is the weight calculated for the moving cost, is the data collection cost, It is the penalty for the unmanned vehicle node to hit an obstacle.

[0056] Step 3: Perception area division method based on spatiotemporal features. The perception area for each time slot t is defined as G t =(V t ,ε t ,X t ). Where V t represents the set of POI points in the sensing area, ε t is the adjacency matrix, which represents the reachability between nodes. t Represents the node features. t All nodes in are sorted in descending order by degree, and the first C nodes are selected as the central nodes of each sub-region. The sub-region h-hop nodes are extracted using the breadth-first search (BFS) algorithm, and there are s nodes in each sub-region. C and s are used to maximize the original graph G tThen, we get a set of sub-regions

[0057] Step 4: Use the GraphSAGE algorithm to extract the features of the nodes in the sub-region. The input of the first layer is In each layer of the neural network, for each node First of all The embedding vectors of the adjacent nodes are transformed nonlinearly. Max-pooling is used to calculate the nodes The feature representation of all adjacent nodes in the kth layer is as follows:

[0058]

[0059] in, Representation Node Representation Node The embedding of the neighboring node u in the k-th layer, Representation Node The feature representation of all adjacent nodes in layer k. After that, and vertices Perform nonlinear transformation on the representation vectors of the k-1 layers and concatenate the results to obtain the representation vector of the central node of the kth layer:

[0060]

[0061] After the above GraphSAGE layer operation, each node in the input sub-region aggregates the features of all its h-hop neighbor nodes. The feature vector is then input into a multi-layer perceptron (MLP) to map it to a scalar and concatenate it with other attributes observed in STPPO.

[0062] Step 5: In time slot t, cluster sub-regions The features of all nodes in the region are converted into one-dimensional scalars as sub-region observations. part of. The inputs to the actor network include: the spatiotemporal characteristics of the sub-region, the data collection rate, the locations of all heterogeneous actors, and the selected set of heterogeneous actors in the sub-region. The sub-region actor network then provides a The action distribution is obtained by assuming a sensing area G and a scheduling time slot T, and the goal of this invention is to design an optimal scheduling strategy π to maximize the overall performance, which is defined as the weighted ratio of the overall data collection rate of the POI set and the collection cost of heterogeneous participants with comprehensive data fairness:

[0063]

[0064] Among them, ξ(π) represents the data collection rate of POI when adopting strategy π, F(π) represents the fairness of data collection when adopting strategy π, and Cost(π) represents the collection cost when adopting strategy π. Get the joint action a from all sub-areas t , calculate the total reward R t , transition to the next state s t+1 , and the rewards for each sub-region Each sub-area is in buffer zone B c Stores the observed Then all stored observations are input into the critic network to obtain the state value function of all states

[0065] Step 6: From buffer B of each sub-region c Randomly select L consecutive time slots of information sets for training. To calculate the advantage function and c-loss, PPO is used to update the critic network. For the participant network, a-loss is calculated by calculating the importance weights between the old participant network and the new participant network, and the PPO algorithm is used to update the participant network. Finally, the algorithm solves the problem of collaborative scheduling of heterogeneous participants in each sub-region and achieves the overall optimization goal.

[0066] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should be included in the protection scope of the present invention.

Claims

1. A multi-participant selection method for heterogeneous crowd intelligence perception, characterized in that Here are the steps: Step 1: Define the nodes of heterogeneous crowd sensing that include human participants (U = {u1,u2,...,u N }), unmanned vehicle node (V={v1,v2,...,v M }) and fixed sensor nodes (S = {s1,s2,...,s k }), where the human actor node contains the attributes in, represents the position of the human participant at time slot t, represents the perceptual ability of human participants, Represents the perceived cost of human participants; the driverless car node contains the attribute in, represents the position of the unmanned vehicle node at time slot t, represents the perception capability of the unmanned vehicle node, Represents the perception cost of the unmanned vehicle node; the fixed sensor node contains the attribute in, represents the position of the fixed sensor node at time slot t, represents the sensing capability of the fixed sensor node, represents the perception cost of fixed sensor nodes; considering the differences between the three types of heterogeneous participant nodes, human participants, unmanned vehicles and fixed sensors in performing perception tasks, the perception capability is defined as the perception range and mobility; the definitions of the perception capabilities of the three types of heterogeneous participants are given: in, They represent the perception radius of human participants, unmanned vehicle nodes and fixed sensor nodes respectively, Respectively represent the maximum moving distance of human participants, unmanned vehicle nodes and fixed sensor nodes at time slot t; POI data defined within the perception radius of heterogeneous participants can be collected, dividing the time period into a set of perception time slots of equal length. By designing a scheduling strategy, appropriate heterogeneous participants are selected to complete the sensing task in each time slot; Step 2: Define a region containing a set of POIs to be collected: P = {p1, p2, ..., p I }; Calculate the data collection rate, data collection fairness and data collection cost of POI at time slot t; Given the heterogeneous participant selection strategy π, the data collection rate of POI at time slot t is defined as follows: Among them, d t (π; p i ) represents the time slot t p under strategy π i Cumulative data collection rate; To avoid perceptual imbalance or blind spots in the data collection process, Jain’s fairness index is used to evaluate the fairness of data collection of the POI set at time slot t; the data collection fairness of the POI at time slot t is defined as follows: The perception costs of heterogeneous participants are defined as mobility costs and data collection costs. For human participants, unmanned vehicle nodes, and fixed sensor nodes, their costs are defined as follows: in, is the weight for calculating the moving cost, is the data collection cost, The penalty for the unmanned vehicle node hitting an obstacle; Step 3: Based on the perception area division method of spatiotemporal features, the perception area for each time slot t is defined as G t =(V t ,ε t ,X t );V t represents the set of POI points in the sensing area, ε t is the adjacency matrix, indicating the reachability between nodes; X t Represents node features; G t All nodes in are sorted in descending order by degree, and the first C nodes are selected as the central nodes of each sub-region; the breadth-first search algorithm BFS is used to extract h-hop nodes in the sub-region, and there are s nodes in each sub-region; C and s are used to maximize the original graph G t The coverage rate of Step 4: Use the GraphSAGE algorithm to extract the features of the nodes in the sub-region; the input of the first layer is In each layer of the neural network, for each node First of all The embedding vectors of the adjacent nodes are transformed nonlinearly; max-pooling is used to calculate the nodes The feature representation of all adjacent nodes in the kth layer is as follows: in, Representation Node Representation Node The embedding of the neighboring node u in the k-th layer, Representation Node Feature representation of all adjacent nodes at layer k; and vertices Perform nonlinear transformation on the representation vectors of the k-1 layers and concatenate the results to obtain the representation vector of the central node of the kth layer: After the above GraphSAGE layer operation, each node of the input sub-region aggregates the features of all its h-hop neighbor nodes; the feature vector is then input into the multi-layer perceptron MLP, which maps it to a scalar and concatenates it with other attributes observed in STPPO; Step 5: In time slot t, cluster sub-regions The features of all nodes in the region are converted into one-dimensional scalars as sub-region observations. part of; The inputs to the actor network include: the spatiotemporal characteristics of the sub-region, the data collection rate, the locations of all heterogeneous actors, and the selected set of heterogeneous actors in the sub-region; then, the sub-region actor network provides a The action distribution is obtained by assuming that the action is sampled using the above distribution. Given a sensing area G and a scheduling time slot T, the goal of this invention is to design an optimal scheduling strategy π to maximize the overall performance, which is defined as the weighted ratio of the overall data collection rate of the POI set and the collection cost of heterogeneous participants with comprehensive data fairness: Among them, ξ(π) represents the data collection rate of POI when adopting strategy π, F(π) represents the fairness of data collection when adopting strategy π, and Cost(π) represents the collection cost when adopting strategy π; the joint action a is obtained from all sub-areas t , calculate the total reward R t , transition to the next state s t+1 , and the rewards for each sub-area Each sub-area is in buffer zone B c Stores the observed Then all stored observations are input into the critic network to obtain the state value function of all states Step 6: From buffer B of each sub-region c Randomly select L consecutive time slots of information sets for training; use To calculate the advantage function and c-loss, use PPO to update the critic network; for the participant network, a-loss is calculated by calculating the importance weights between the old participant network and the new participant network, and the PPO algorithm is used to update the participant network.

2. A computer system, characterized in that include: One or more processors, and a computer-readable storage medium for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors are enabled to implement the method of claim 1.

3. A computer-readable storage medium, characterized in that Computer executable instructions are stored, and when the instructions are executed, they are used to implement the method of claim 1.

4. A computer program product, characterized in that The invention comprises computer executable instructions, which are used to implement the method of claim 1 when being executed.

Images