Crowd sensing online capability assessment method for mixed workers

Abstract

The invention discloses a hybrid worker-oriented crowd sensing online capability assessment method and device, a medium and equipment. The method comprises the following steps: constructing a hybrid worker-oriented mobile crowd sensing system model; constructing an optimization function for maximizing long-term cumulative utility as a target based on a mobile crowd sensing system model; the time-space trajectory prediction module based on Transform predicts the success probability of the opportunity worker completing the task in the future time window; evaluating the same-stage capabilities of the mixed workers by using an urgency degree weighted capability scoring framework; the weight adaptive adjustment module based on deep reinforcement learning processes the real-time state of the dynamic crowd sensing system model to output the ability evaluation score of the mixed workers, and distributes the current task to the worker with the highest score; according to the invention, the problems of low efficiency, poor cost control and poor environmental adaptability caused by adopting a two-stage distribution strategy in the prior art can be solved.

Description

Technical Field

[0001] This application relates to the field of mobile crowd sensing technology, and in particular to a method, apparatus, medium and device for assessing the online crowd sensing capabilities of hybrid workers. Background Technology

[0002] Driven by the Internet of Things (IoT) and 5G / 6G technologies, Mobile Crowdsensing (MCS) has become a mainstream data collection paradigm, utilizing personal smart devices (such as smartphones) to achieve large-scale environmental perception. In the task allocation process of MCS, users performing tasks (i.e., "workers") are typically divided into two categories: Participatory Workers and Opportunistic Workers. Participatory Workers are willing to actively change their movement paths to reach the task location, ensuring high task completion reliability but incurring additional movement costs. In contrast, Opportunistic Workers only complete tasks encountered along their path without altering their original routes; their costs are extremely low, but their uncertainty is high.

[0003] An ideal MCS system should combine the advantages of both—the reliability of participatory workers and the low cost of opportunistic workers. However, most existing research focuses on only a single type of worker or employs a flawed "two-stage task allocation strategy." This common strategy first assigns tasks to low-cost opportunistic workers and then recruits more expensive participatory workers for the remaining tasks.

[0004] This sequential, greedy allocation method has inherent suboptimal characteristics. First, it lacks a global perspective, failing to make a unified comparison and trade-off between the two types of workers at the initial decision-making stage. For example, for an urgent task, a participatory worker who is nearby and requires only minimal incentive may be far more valuable than an opportunistic worker who will arrive several hours later and whose participation is uncertain. However, the two-stage model cannot make this global trade-off, often sacrificing efficiency and overall utility. Second, it ignores spatiotemporal synergy. The uncertainty of opportunistic workers may cause the task to fail within a valuable time window, which undoubtedly increases the difficulty and cost of recruiting participatory workers in the second stage. Summary of the Invention

[0005] The main objective of this application is to provide a method, apparatus, medium, and device for assessing the online collective perception capabilities of hybrid workers, aiming to solve the technical problems of methods for assessing the online collective perception capabilities of hybrid workers.

[0006] To achieve the above objectives, this application provides an online assessment method for swarm intelligence perception capabilities of hybrid workers, comprising: constructing a mobile swarm intelligence perception system model for hybrid workers, wherein the mobile swarm intelligence perception system model includes hybrid workers, and hybrid workers include participatory workers and opportunistic workers; constructing an optimization function based on the mobile swarm intelligence perception system model to maximize long-term cumulative utility; capturing long-range spatiotemporal dependencies in user movement patterns using a Transformer-based spatiotemporal trajectory prediction module to predict the success probability of opportunistic workers completing tasks within future time windows; quantifying multi-dimensional factors affecting task utility into unified scores using an urgency-weighted capability scoring framework to assess the capabilities of hybrid workers at the same stage; wherein the multi-dimensional factors include a basic capability similarity factor, a movement cost factor, an urgency factor, and a proximity factor, wherein for opportunistic workers, the value of the movement cost factor is equal to the success probability; and processing the real-time state of the mobile swarm intelligence perception system model using a pre-trained agent based on a deep reinforcement learning-based weight adaptive adjustment module to obtain the optimal weight combination of the multi-dimensional factors, outputting the capability assessment score of the hybrid workers, and assigning the current task to the worker with the highest score.

[0007] Optionally, the Transformer-based spatiotemporal trajectory prediction module includes N stacked encoder layers, a linear layer, and an activation layer. Each encoder layer includes a multi-head self-attention mechanism and a feedforward neural network. The input vector of the spatiotemporal trajectory prediction module includes a spatial embedding vector and a temporal embedding vector. The temporal embedding vector is used to map the geographic location grid ID to a high-dimensional vector. The spatial embedding vector is used to encode the location of the time series using the original Transformer. The spatial embedding vector and temporal embedding vector, which are processed and summed by each encoder layer of the spatiotemporal trajectory prediction module, are used to output the geographic grid probability distribution of opportunistic workers at future time points. Based on the geographic grid probability distribution of opportunistic workers at future time points, the success probability of opportunistic workers completing tasks within future time windows is predicted.

[0008] Optionally, the expression for the basic capability similarity factor is:

[0009]

[0010] Among them, f BCS θ represents the basic ability similarity. j,dev θ represents the worker's equipment capability. j,rep p represents the worker's credit rating i,dev p represents the worker's equipment quality requirements. i,rep This indicates the worker's reputation and quality requirements, α dev α represents the equipment coefficient. repIndicates credit rating;

[0011] The expression for the mobility cost factor is:

[0012]

[0013] Among them, f MCF W represents the cost of movement. P W represents a participatory worker set. O w represents a set of opportunistic workers. j A represents the worker. i C(w) represents the perceptual region of the task. j ,T i B represents the cost of movement. i This indicates the platform's budget limit for completing the task, t i Represents the task time window, Tr k The worker's trajectory;

[0014] The expression for the urgency factor is:

[0015]

[0016] Among them, f UF Indicates urgency, u i Indicates the urgency of the task, t exp_comp (w j ,T i ) indicates worker w j Execute task T i The expected completion time, t current Indicates the current time;

[0017] The expression for the proximity factor is:

[0018]

[0019] Among them, f PF d represents proximity. G (l j,t A i ) represents the shortest path distance on the road network graph G; D represents the average shortest path distance on the road network graph G. max This represents the maximum distance on the road network map G.

[0020] Optionally, the dynamic sensing task is a set of at least one task, wherein each task is a tuple, and the expression of the tuple is:

[0021]

[0022] r i =(rti ,ra i ,rs i )

[0023]

[0024] Among them, A i This represents the perception area of ​​the task, which is a geographical region. [t] i,s ,t i,e ] represents the effective time window of the task, t i,s Indicates the earliest start time, t i,e Indicates a hard deadline that must be met, r i This indicates the quality requirements for the perceived data, rt i Indicates accuracy index, ra i Represents the resolution metric, rs i Indicates range indicators, B represents the urgency coefficient of the task. i V(t) represents the upper limit of the budget set by the system to complete the task, and V(t) represents the value function.

[0025] Optionally, the agent pre-trained by the deep reinforcement learning-based weight adaptive adjustment module processes the real-time state of the swarm intelligent perception system model to obtain the optimal weight combination of multi-dimensional factors, including: constructing a Markov decision process, which includes a quintuple comprising a state space vector, an action space vector, a reward function vector, a state transition probability vector, and a discount factor vector; wherein, the state space vector includes the average urgency of the tasks to be processed, the task backlog rate, the proportion of participatory workers among available workers, the proportion of the system's remaining total budget, and the task density of each region; the action space vector includes the normalized values ​​of the basic ability similarity factor, the movement cost factor, the urgency factor, and the proximity factor; the reward function vector includes the total value of the tasks completed by the workers within the decision cycle and the penalty for failed tasks; training the agent based on the soft actor-critic algorithm, wherein the agent includes a policy network and two Q-value networks respectively connected to the policy network, wherein the policy network is used to execute the Markov decision process; and inputting the real-time state in the state space into the agent and outputting the optimal weight combination of multi-dimensional factors.

[0026] Optionally, the mobility costs include the mobility costs of participatory workers and the mobility costs of opportunistic workers;

[0027] The expression for the movement cost of participatory workers is as follows:

[0028] C(w j ,T i ) = c d ×dist(lj,t A i )+c t ×travel_time(l j,t A i )

[0029] Among them, c d Dist(l) represents the distance cost coefficient. j,t A i ) is the shortest path distance on the road network map, c t The time cost coefficient is represented by travel_time(l j,t A i This represents the time cost of workers moving between locations;

[0030] The expression for the movement cost of opportunity workers is:

[0031]

[0032] in, A represents the effective time window of the task. i The perceptual area of ​​the task, l j,t This indicates the worker's current geographical location.

[0033] Furthermore, to achieve the above objectives, this application also provides an online assessment device for swarm intelligence perception capabilities for hybrid workers, characterized by comprising: an objective function construction module for constructing a mobile swarm intelligence perception system model for hybrid workers, wherein the mobile swarm intelligence perception system model includes hybrid workers, and hybrid workers include participatory workers and opportunistic workers; constructing an optimization function based on the mobile swarm intelligence perception system model to maximize long-term cumulative utility; and a task prediction module for capturing long-range spatiotemporal dependencies in user movement patterns based on a Transformer-based spatiotemporal trajectory prediction module to predict the success rate of opportunistic workers in completing tasks within future time windows. The evaluation module is used to quantify multi-dimensional factors affecting task utility into a unified score using an urgency-weighted capability scoring framework to evaluate the capabilities of hybrid workers at the same stage. The multi-dimensional factors include a basic capability similarity factor, a movement cost factor, an urgency factor, and a proximity factor, wherein, for opportunistic workers, the value of the movement cost factor is equal to the success probability. The weight adjustment module is used to obtain the optimal weight combination of the multi-dimensional factors based on the real-time state of the agent processing swarm intelligent perception system model pre-trained by the deep reinforcement learning weight adaptive adjustment module, output the capability evaluation score of the hybrid workers, and assign the current task to the worker with the highest score.

[0034] To achieve the above objectives, this application also provides a computer-readable storage medium including instructions that, when executed on a computer, cause the computer to perform the online assessment method for crowd intelligence perception for hybrid workers provided in the above embodiments.

[0035] To achieve the above objectives, this application also provides an electronic device, which includes: at least one processor, a memory, and an input / output unit; wherein the memory is used to store a computer program, and the processor is used to call the computer program stored in the memory to execute the online assessment method for crowd perception capabilities of hybrid workers provided in any of the foregoing embodiments.

[0036] This application proposes a method, apparatus, medium, and device for assessing the online collective intelligence perception capabilities of hybrid workers. It constructs a mobile collective intelligence perception system model for hybrid workers, where the model includes hybrid workers, comprising participatory workers and opportunistic workers. Based on the mobile collective intelligence perception system model, it constructs an optimization function aimed at maximizing long-term cumulative utility. A Transformer-based spatiotemporal trajectory prediction module captures long-range spatiotemporal dependencies in user movement patterns to predict the success probability of opportunistic workers completing tasks within future time windows. Finally, it utilizes an urgency-weighted capability scoring framework to incorporate multi-dimensional factors affecting task utility. The system quantifies the capabilities of hybrid workers into a unified score to assess their abilities at the same stage. The multi-dimensional factors include a basic ability similarity factor, a movement cost factor, an urgency factor, and a proximity factor. For opportunistic workers, the movement cost factor is equal to the success probability. An agent pre-trained with a deep reinforcement learning-based weight adaptive adjustment module processes the real-time state of the swarm intelligence perception system model to obtain the optimal weight combination of the multi-dimensional factors, outputting the hybrid worker's capability assessment score. The current task is then assigned to the worker with the highest score. This solves the problem of an online swarm intelligence perception capability assessment method for hybrid workers and realizes a method for online swarm intelligence perception capability assessment for hybrid workers. Attached Figure Description

[0037] Figure 1 This is a flowchart illustrating an embodiment of the online assessment method for collective intelligence perception capabilities of hybrid workers in this application.

[0038] Figure 2 This is a schematic diagram of an embodiment of the online assessment method for collective intelligence perception of hybrid workers provided in this application.

[0039] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0040] It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit this application.

[0041] The technical solution to achieve the objective of this invention is: an online assessment method for the capabilities of a crowd-sensing worker based on hybrid workers, the steps of which are as follows:

[0042] S1: Construct a mobile crowd sensing system model for hybrid workers.

[0043] S2: Constructing the Optimization Objective Function for a Mobile Crowdsourcing Sensing System for Hybrid Workers: Based on the system model constructed in S1, a mathematical optimization model is established from the perspective of global system optimization, aiming to maximize the cumulative utility of tasks. Since this global optimization problem is difficult to solve directly in online dynamic scenarios (NP-hard), this step cleverly transforms it into a more feasible core sub-problem: designing a unified online capability scoring function that can quantify the value of any "worker-task" match, thereby decomposing the complex global optimization problem into a series of locally optimal decisions based on this score.

[0044] S3: Proposes a unified online competency assessment framework, aiming to achieve unified, efficient, and adaptive assessment of both participatory and opportunity-based workers. As a top-level design, it integrates three tightly coupled functional modules (S4, S5, and S6) to jointly complete the comprehensive assessment and dynamic decision-making of worker competency.

[0045] S4: Transformer-based Spatiotemporal Trajectory Prediction: Leveraging its powerful self-attention mechanism to capture long-range spatiotemporal dependencies in user movement patterns, this method accurately predicts the probability of opportunistic workers reaching the task area within future time windows. This probability is a key basis for subsequent unified evaluation of opportunistic worker capabilities.

[0046] S5: Construct an Urgency-Weighted Competency Scoring Framework (UWCS): This function quantifies several key dimensions that need to be considered in assessing a worker’s “value” for a particular task—namely, intrinsic competency fit, cost or risk, task timeliness, and geographical accessibility—into four standardized factors.

[0047] S6: Adaptive Weight Adjustment Based on Deep Reinforcement Learning: To enable the UWCS scoring framework to adapt to dynamically changing system environments, this step introduces an adaptive weight adjustment mechanism based on deep reinforcement learning (DRL). This mechanism models the weight adjustment problem as a Markov Decision Process (MDP) and trains an agent using advanced DRL algorithms (such as Soft Actor-Critic SAC). This agent can dynamically output the optimal combination of scoring factor weights based on the real-time macro-state of the system (such as task density and budget margin), thereby guiding the system to make decisions that maximize long-term cumulative rewards.

[0048] The existing technology has the following shortcomings in the task allocation of mixed workers: (1) The allocation strategy is suboptimal. The commonly used two-stage strategy is sequential and greedy, which cannot achieve the global optimal allocation; (2) There is a lack of unified evaluation standards. It is impossible to directly compare the "ability" or "value" of different types of workers in the same dimension, which leads to limited decision-making; (3) Poor adaptability. It adopts a fixed allocation logic and it is difficult to flexibly adjust the strategy according to the dynamically changing environment (such as task urgency and worker distribution).

[0049] This application provides a solution to address the inefficiencies, poor cost control, and poor environmental adaptability caused by the two-stage allocation strategy in existing technologies. The proposed method significantly improves response efficiency and cost-effectiveness while ensuring a high task completion rate by introducing Transformer-based trajectory prediction, a unified multi-dimensional ability scoring framework, and a deep reinforcement learning-based weight adaptation mechanism.

[0050] The following description, in conjunction with the accompanying drawings and embodiments, further illustrates the online assessment method for collective intelligence perception capabilities for hybrid workers.

[0051] Figure 1 The present invention provides a flowchart of a unified online capability assessment method for hybrid workers, comprising the following steps: First, a mobile swarm intelligence sensing system model for hybrid workers is constructed, and an optimization function is built based on this model to maximize long-term cumulative utility. On this basis, a unified online capability assessment framework for hybrid workers is proposed, using a Transformer-based spatiotemporal trajectory prediction model to accurately generate the probability of opportunistic workers reaching the task area. Next, an urgency-weighted UWCS capability scoring framework is constructed, quantifying multi-dimensional factors affecting task utility into unified scores, enabling simultaneous capability assessment for both types of workers. Finally, a deep reinforcement learning-based adaptive weight adjustment mechanism is integrated to guide the system to make decisions that maximize long-term cumulative rewards based on real-time status, thereby transforming rigid phased allocation into a unified, flexible, and adaptive closed-loop optimization process, significantly improving cost-effectiveness while ensuring high task completion rates.

[0052] Figure 2 This demonstrates a specific example of a closed-loop decision-making process from input to output when allocating tasks using our invention. In one specific embodiment, the method proposed in this invention transforms the originally rigid task allocation process into a unified, flexible, and dynamically adaptive closed-loop optimization process: When a new task arrives, the system first calls a Transformer-based prediction model (S4) to accurately calculate the task completion probability for each opportunistic worker; subsequently, the system calls a UWCS scoring function (S5) composed of multi-dimensional factors to calculate a unified ability score for all available workers, including participatory workers, while the weights of each factor in this scoring function are dynamically provided by a deep reinforcement learning agent (S6) running in the background based on the real-time system state; finally, the platform sorts all candidate workers according to their comprehensive ability scores and assigns the task to the worker with the highest score, thereby achieving a globally optimal decision.

[0053] Reference Figure 1 The first embodiment of this application provides a method for assessing the online collective sensing capabilities of hybrid workers. This method may include:

[0054] S101. Construct a mobile swarm intelligence sensing system model for hybrid workers, wherein the mobile swarm intelligence sensing system model includes hybrid workers, which include participatory workers and opportunistic workers; construct an optimization function based on the mobile swarm intelligence sensing system model to maximize long-term cumulative utility.

[0055] This step constructs a model of a mobile swarm intelligence sensing system for hybrid workers. It begins by formally defining a mobile swarm intelligence sensing system comprising a centralized platform, dynamic sensing tasks, and a large number of hybrid workers (including participatory and opportunistic workers). This involves precisely describing the various attributes of the tasks (such as location, time window, budget, and urgency) and the core characteristics of the two types of workers (such as the cost model for participatory workers and the zero marginal cost and path dependence for opportunistic workers), providing a unified mathematical foundation and operational framework for the design of all subsequent algorithms and strategies.

[0056] Specifically, this application considers a typical centralized mobile swarm intelligence sensing system, which can consist of an MCS server (platform), a set of dynamically emerging sensing tasks, and a large number of registered mobile workers. The platform, as the central hub of the system, acts as a bridge connecting task requesters and data providers (workers). Its core functions include: Task management: receiving sensing requests from the application layer, converting them into standardized sensing tasks, and managing the entire lifecycle of these tasks (e.g., pending assignment, assigned, completed, expired). Worker node management: maintaining configuration files for all registered worker nodes, including their type, historical reputation, device capabilities, real-time location, or predicted trajectory.

[0057] For sensing tasks: During system operation, a series of sensing tasks appear dynamically and randomly, represented as a set T = {T1, T2, T3, ... T}. n Each task T i ∈T is precisely defined by a tuple for its complex requirements:

[0058] T i =(A i ,[t i,s ,t i,e ],r i ,u i B i )

[0059] r i =(rt i ,ra i ,rs i )

[0060]

[0061] Among them, A i The perception area for a task is a geographical range rather than a single point coordinate. It defines the geographical area where the task needs to be performed, usually a polygonal or circular area rather than a single coordinate point. Workers must enter this area to perform perception.

[0062] [t i,s ,t i,e ] represents the effective time window of the task, t i,s For the earliest start time, t i,e This is a hard deadline that must be met; any execution outside of this window is invalid.

[0063] r iThis represents the quality requirements for the perceived data. It can be a vector containing indicators such as accuracy and resolution. This is a multi-dimensional vector that details the quality requirements of the task for the collected data, such as the accuracy of data sampling, image resolution, data integrity, and the range of temperature sensors.

[0064] This represents the urgency coefficient of the task. This coefficient reflects the rate at which the task's value decays over time. For example, the value of completing the task at time t can be modeled as... u i The larger the value, the faster the value decays, and the more urgent the task. The larger the value of this coefficient, the faster the task's value decays, and the more urgent the task is. The platform needs to find workers for it as soon as possible.

[0065] B i This indicates the maximum budget set by the platform to complete the task.

[0066] For hybrid workers: mobile workers are the actual executors of perception tasks, forming the basis of the system's perception capabilities. The set of workers in the system is W = {w1, w2, ..., w...} n Based on their participation patterns, they are divided into two mutually exclusive subsets: the participatory worker set W. P And the collection of opportunistic workers W O Satisfying W = W P ∪W O and Where, at any time t, worker w j The state can be represented as S j,t =(l j,t status j BCS j ), where l j,t It is its current geographical location, status j ∈{A,B,O} is its current state (e.g., idle, busy), BCS j It is a basic capability score that comprehensively reflects the quality of its equipment and its historical reputation.

[0067] Participatory workers are willing to proactively change their course to gain incentives. When assigned task T i At that time, the resulting movement cost C(w) j ,T i ) is a multidimensional function that requires platform compensation, proportional to its travel distance and time, where It is the shortest path distance on the road network map.

[0068] C(w j ,T i ) = c d×dist(l j,t A i )+c t ×travel_time(l j,t A i )

[0069] Opportunistic workers follow their predetermined travel plans, and their marginal movement cost to complete a task is zero. Their cost function is a piecewise function, meaning that they can only be considered potential task performers if their predicted trajectory intersects with the spatiotemporal scope of the task.

[0070]

[0071] From a system-wide perspective, the platform's ultimate goal is to maximize the cumulative utility of all completed tasks over a given period, while satisfying various resource constraints such as budget. Let... For decision variables, a value of 1 indicates that task T will be performed at time t. i Assigned to worker w j The optimization function of the objective can be formalized as:

[0072]

[0073] The constraints are:

[0074]

[0075] S102. The Transformer-based spatiotemporal trajectory prediction module captures long-range spatiotemporal dependencies in user movement patterns to predict the probability of opportunistic workers successfully completing tasks within future time windows.

[0076] The Transformer-based spatiotemporal trajectory prediction module addresses the core challenge of uncertain movement paths for opportunistic workers. This step employs a high-precision spatiotemporal trajectory prediction model based on the Transformer architecture. This model leverages its powerful self-attention mechanism to capture long-range spatiotemporal dependencies in user movement patterns, thereby accurately predicting the probability that opportunistic workers will arrive at the task area within future task time windows. This probability value provides a solid data foundation for subsequent unified evaluation.

[0077] In one embodiment of this application, the Transformer-based spatiotemporal trajectory prediction module includes N stacked encoder layers, a linear layer and an activation layer, and each encoder layer includes a multi-head self-attention mechanism and a feedforward neural network.

[0078] The input vectors of the spatiotemporal trajectory prediction module include spatial embedding vectors and temporal embedding vectors.

[0079] Temporal embedding vectors are used to map geographic location grid IDs to high-dimensional vectors;

[0080] Spatial embedding vectors are used to encode the position of time series using the original Transformer;

[0081] The spatial embedding vector and temporal embedding vector, which are processed and summed by each Encoder layer of the spatiotemporal trajectory prediction module, are used to output the geographic grid probability distribution of opportunistic workers at future time points. Based on the geographic grid probability distribution of opportunistic workers at future time points, the success probability of opportunistic workers completing tasks within future time windows is predicted.

[0082] Specifically, spatial embedding grids geographic regions and assigns a unique ID to each grid cell. An embedding layer maps each location ID to a high-dimensional vector. Temporal embedding, to enable the model to understand the temporal order of the sequence, employs the same positional encoding (PE) as the original Transformer. For the p-th position in the sequence, its encoding consists of sine and cosine functions of different frequencies, as follows. The final input vector is the sum of the spatial embedding and the temporal embedding: x p =e spat (l p )+PE(p).

[0083] in,

[0084]

[0085] The model architecture employs a standard Encoder-Decoder structure. The input vector sequence is fed into an N-layer stacked encoder. Each encoder layer contains a multi-head self-attention mechanism and a feed-forward network. The self-attention mechanism is its core, allowing the model to weigh the importance of all other positions in the sequence when calculating the representation of a specific position. Its calculation formula is as follows, where Q, K, and V are the query, key, and value matrices, respectively. This mechanism enables the model to process the entire sequence in parallel and directly establish global dependencies.

[0086]

[0087] After processing by multiple encoder layers, the final output sequence representation is passed through a linear layer and a softmax activation function to obtain the predicted sequence at the future time point t. pred The worker w kThe probability distribution of occurrences across all possible geographic grids is denoted as . Using this model, the chance worker w can be calculated. k trajectory Tr k During the task time window t i =[t i,s ,t i,e Enter mission area A at least once. i The probability, i.e. This probability is a key input for subsequent competency assessments.

[0088] S103. Using an urgency-weighted competency scoring framework, the multi-dimensional factors affecting task utility are quantified into a unified score to evaluate the competencies of mixed workers at the same stage. The multi-dimensional factors include basic competency similarity factor, mobility cost factor, urgency factor, and proximity factor. For opportunistic workers, the value of the mobility cost factor is equal to the success probability.

[0089] The top-level design and core of this invention is the Urgency-Weighted Capability Score (UWCS) framework. This framework aims to fundamentally replace the traditional two-stage allocation strategy by establishing a unified and adaptive evaluation system to achieve global, dynamic, and optimized task allocation decisions for all types of workers.

[0090] To comprehensively evaluate the marginal utility of any worker-task pair, this invention decomposes utility into four basic dimensions: quality, cost, time, and opportunity, and constructs the UWCS scoring system accordingly. This system quantifies these four dimensions through four core factors, and finally obtains a unified capability score through weighted summation.

[0091]

[0092] This step details the specific structure of the UWCS scoring function. This function quantifies several key dimensions affecting task utility—quality (basic competence), cost (mobility costs or risks), time (urgency), and opportunity (proximity)—into four standardized factors. By weighted summing of these four factors, a unified competence score is generated, thus enabling, for the first time, direct comparison and ranking of participatory and opportunistic workers with vastly different characteristics under the same standard and at the same stage.

[0093] When a new task arrives at the platform, the UWCS framework is activated. First, for each opportunistic worker in the system, the framework calls the Transformer's spatiotemporal trajectory prediction module to calculate the probability of completing the task within its effective timespace. Then, for all available workers, including participatory and opportunistic workers, the framework calls an urgency-weighted capability scoring framework. This module, based on a deep reinforcement learning-based weight adaptive adjustment module, calculates the final UWCS capability score for each worker-task pair based on the optimal weight combination provided in real-time by the current system state. Finally, the framework outputs a list of all candidate workers sorted in descending order of capability score, allowing the platform to make a globally optimal allocation decision. In this way, the UWCS framework transforms the originally isolated, sequential decision-making process into a parallel, integrated, and adaptive closed-loop optimization process, thus theoretically and practically guaranteeing the global efficiency and effectiveness of task allocation.

[0094] The expression for the basic capability similarity factor is as follows:

[0095]

[0096] Among them, f BCS θ represents the basic ability similarity. j,dev θ represents the worker's equipment capability. j,rep p represents the worker's credit rating i,dev p represents the worker's equipment quality requirements. i,rep This indicates the worker's reputation and quality requirements, α dev α represents the equipment coefficient. rep Indicates credit rating;

[0097] The expression for the mobility cost factor is:

[0098]

[0099] Among them, f MCF W represents the cost of movement. P W represents a participatory worker set. O w represents a set of opportunistic workers. j A represents the worker. i C(w) represents the perceptual region of the task. j ,T i B represents the cost of movement. i This indicates the platform's budget limit for completing the task, t i Represents the task time window, Tr k The worker's trajectory;

[0100] The expression for the urgency factor is:

[0101]

[0102] Among them, f UF Indicates urgency, u i Indicates the urgency of the task, t exp_comp (w j ,T i ) indicates worker w j Execute task T i The expected completion time, t current Indicates the current time;

[0103] The expression for the proximity factor is:

[0104]

[0105] Among them, f PF d represents proximity. G (l j,t A i ) represents the shortest path distance on the road network graph G; D represents the average shortest path distance on the road network graph G. max This represents the maximum distance on the road network map G.

[0106] S104. The agent pre-trained by the weight adaptive adjustment module based on deep reinforcement learning processes the real-time state of the dynamic swarm intelligent perception system model, obtains the optimal weight combination of multi-dimensional factors, outputs the ability assessment score of the hybrid worker, and assigns the current task to the worker with the highest score.

[0107] To enable the evaluation framework to intelligently adapt to dynamically changing system environments, this step introduces a weight adaptive adjustment mechanism based on deep reinforcement learning. This mechanism models the weight adjustment problem as a Markov Decision Process (MDP) and trains an agent using advanced DRL algorithms (such as Soft Actor-Critic SAC). This agent can dynamically output the optimal combination of scoring factor weights based on the real-time macro-state of the system (such as task density and budget margin), thereby guiding the entire task allocation system to make decisions that maximize long-term cumulative rewards.

[0108] In one embodiment of this application, the agent pre-trained by the deep reinforcement learning-based weight adaptive adjustment module processes the real-time state of the swarm intelligent perception system model to obtain the optimal weight combination of multi-dimensional factors, including:

[0109] Construct a Markov decision process, which consists of a quintuple, including a state space vector, an action space vector, a reward function vector, a state transition probability vector, and a discount factor vector.

[0110] The state space vector includes the average urgency of the tasks to be processed, the task backlog rate, the proportion of participatory workers among available workers, the proportion of the system's remaining total budget, and the task density of each region; the action space vector includes the normalized values ​​of the basic capability similarity factor, the movement cost factor, the urgency factor, and the proximity factor; and the reward function vector includes the total value of tasks completed by workers within the decision-making cycle and the penalty for failed tasks.

[0111] The agent is trained based on the soft actor-critic algorithm. The agent includes a policy network and two Q-value networks respectively connected to the policy network. The policy network is used to execute a Markov decision process.

[0112] The agent takes the real-time state in the state space as input and outputs the optimal weight combination of multi-dimensional factors.

[0113] For example, the factor weights (ω) in the UWCS scoring framework BCS ,ω MCF ,ω UF ,ω PF The optimal configuration of weights is not static but depends on the dynamically changing system state. This invention models this dynamic weight adjustment problem as a Markov Decision Process (MDP) and solves it using Deep Reinforcement Learning (DRL) methods. The MDP consists of tuples... definition.

[0114] The state space This is a vector that provides a concise description of the current macroscopic state of the system, designed to capture key supply and demand relationships and resource status. The vector includes the average urgency of pending tasks, task backlog rate, the proportion of participatory workers among available workers, the proportion of the system's remaining total budget, and a vector describing the task density in each region.

[0115]

[0116] Action a in action space t ∈A is a four-dimensional continuous vector representing a set of normalized weights.

[0117] a t =(ω BCS ,ω MCF ,ω UF ,ω PF ),st∑ω=1,ω≥0

[0118] Reward function R t =R(s) t ,a tDesigned to be directly aligned with the system's optimization objective, it aims to guide the agent to learn a policy that maximizes long-term cumulative utility. It consists of the total value of completing tasks within a decision period Δt and the penalty for failing tasks.

[0119]

[0120] For the MDP with a continuous action space as defined above, this invention employs the SoftActor-Critic (SAC) algorithm. SAC is an advanced off-policy DRL algorithm based on a maximum entropy framework. It learns collaboratively by training a policy network (Actor) and two Q-value networks (Critic). Through continuous interaction with the environment, the agent constantly updates its network parameters, ultimately learning an algorithm that can adapt to the real-time system state s. t Dynamically output the optimal weight combination a t strategy This gives the entire UWCS evaluation framework strong environmental adaptability and self-optimization capabilities.

[0121] Reference Figure 2 The second embodiment of this application provides a crowd-sensing online capability assessment device for hybrid workers, which solves the same technical problem as the method embodiment. The device may include: an objective function construction module for constructing a mobile swarm intelligence perception system model for hybrid workers, wherein the mobile swarm intelligence perception system model includes hybrid workers, which include participatory workers and opportunistic workers; an optimization function based on the mobile swarm intelligence perception system model for maximizing long-term cumulative utility; a task prediction module for capturing long-range spatiotemporal dependencies in user movement patterns using a Transformer-based spatiotemporal trajectory prediction module to predict the success probability of opportunistic workers completing tasks within future time windows; an evaluation module for quantifying multi-dimensional factors affecting task utility into uniform scores using an urgency-weighted capability scoring framework to evaluate the capabilities of hybrid workers at the same stage; the multi-dimensional factors include a basic capability similarity factor, a movement cost factor, an urgency factor, and a proximity factor, wherein for opportunistic workers, the value of the movement cost factor is equal to the success probability; and a weight adjustment module for processing the real-time state of the mobile swarm intelligence perception system model by an agent pre-trained based on a deep reinforcement learning-based weight adaptive adjustment module, obtaining the optimal weight combination of multi-dimensional factors, outputting the capability evaluation score of the hybrid workers, and assigning the current task to the worker with the highest score.

[0122] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A hybrid worker-oriented crowd-sensing online capability evaluation method, characterized in that, include: A mobile swarm intelligence sensing system model for hybrid workers is constructed, wherein the mobile swarm intelligence sensing system model includes hybrid workers, which include participatory workers and opportunistic workers; an optimization function is constructed based on the mobile swarm intelligence sensing system model to maximize long-term cumulative utility. The Transformer-based spatiotemporal trajectory prediction module captures long-range spatiotemporal dependencies in user movement patterns to predict the probability of opportunistic workers successfully completing tasks within future time windows. An urgency-weighted competency scoring framework is used to quantify multi-dimensional factors affecting task utility into uniform scores to assess the competencies of hybrid workers at the same stage. The multi-dimensional factors include basic competency similarity factors, mobility cost factors, urgency factors, and proximity factors. For opportunistic workers, the value of the mobility cost factor is equal to the probability of success. The agent, pre-trained using a weight adaptive adjustment module based on deep reinforcement learning, processes the real-time state of the swarm intelligent perception system model, obtains the optimal weight combination of multi-dimensional factors, outputs the ability assessment score of the hybrid workers, and assigns the current task to the worker with the highest score.

2. The mixed worker oriented crowd-sensing online capability evaluation method according to claim 1, wherein, The Transformer-based spatiotemporal trajectory prediction module includes N stacked encoder layers, a linear layer, and an activation layer. Each encoder layer includes a multi-head self-attention mechanism and a feedforward neural network. The input vectors of the spatiotemporal trajectory prediction module include spatial embedding vectors and temporal embedding vectors. Temporal embedding vectors are used to map geographic location grid IDs to high-dimensional vectors; Spatial embedding vectors are used to encode the position of time series using the original Transformer; The spatial embedding vector and temporal embedding vector, which are processed and summed by each Encoder layer of the spatiotemporal trajectory prediction module, are used to output the geographic grid probability distribution of opportunistic workers at future time points. Based on the geographic grid probability distribution of opportunistic workers at future time points, the success probability of opportunistic workers completing tasks within future time windows is predicted. 3.The hybrid worker-oriented crowd sensing online capability evaluation method of claim 1, wherein, The expression for the basic capability similarity factor is: wherein, represents a base ability similarity, represents a worker's equipment ability, represents a worker's reputation score, represents a worker's equipment quality requirement, represents a worker's reputation quality requirement, represents an equipment coefficient, represents a reputation coefficient; The expression for the mobility cost factor is: wherein, denotes a moving cost, denotes a set of participative workers, denotes a set of opportunistic workers, denotes a worker, denotes a perception area of a task, denotes a moving cost, denotes an upper budget limit set by the platform for completing the task, denotes a time window of a task, a trajectory of a worker; The expression for the urgency factor is: wherein, represents an urgency, represents a task urgency, represents a worker performs a task an expected completion time, represents a current time; The expression for the proximity factor is: wherein, denotes the proximity, denotes the shortest path distance on the road network graph G; denotes the average shortest path distance on the road network graph G, denotes the maximum distance on the road network graph G. 4.The hybrid worker-oriented crowd sensing online capability evaluation method of claim 1, wherein, The dynamic sensing task is a set of at least one task, wherein each task is a tuple, and the expression of the tuple is: wherein, represents a perception area of the task, the perception area being a geographical range, represents a validity time window of the task, represents an earliest start time, represents a hard deadline that must be completed, represents a quality requirement on the perception data, represents a precision indicator, represents a resolution indicator, represents a range indicator, represents an urgency coefficient of the task, represents an upper budget set by the system to complete the task, represents a value function. 5.The hybrid worker-oriented crowd sensing online capability evaluation method according to claim 1, wherein, The pre-trained agent processing swarm intelligence system model using the deep reinforcement learning-based weight adaptive adjustment module obtains the optimal weight combination of multi-dimensional factors, including: Construct a Markov decision process, which consists of a quintuple, including a state space vector, an action space vector, a reward function vector, a state transition probability vector, and a discount factor vector. The state space vector includes the average urgency of the tasks to be processed, the task backlog rate, the proportion of participatory workers among available workers, the proportion of the total remaining budget of the system, and the task density of each region. The action space vector includes the normalized values of the base ability similarity factor, the movement cost factor, the urgency factor, and the proximity factor. The reward function vector includes the total value of the tasks completed by the workers within the decision cycle and the penalty for failed tasks. The agent includes a policy network and two Q-value networks connected to the policy network, respectively. The policy network is used to execute a Markov decision process. The real-time state in the state space is input into the agent, and the optimal weight combination of the multi-dimensional factors is output.

6. The mixed worker oriented crowd-sensing online capability evaluation method according to claim 1, wherein, The movement cost includes the movement cost of the participatory workers and the movement cost of the opportunistic workers. The expression of the movement cost of the participatory workers is: wherein, denotes a distance cost coefficient, is the shortest path distance on the road network graph, denotes a time cost coefficient, denotes a worker movement time cost; The expression of the movement cost of the opportunistic workers is: wherein, represents an effective time window of the task, represents a perception area of the task, represents a current geographic location of the worker. 7.A hybrid worker-oriented crowd-sensing online capability evaluation device, characterized in that, The method comprises the following steps: A target function construction module is used to construct a mobile crowd sensing system model for mixed workers, wherein the mobile crowd sensing system model includes and the mixed workers include the participatory workers and the opportunistic workers. An optimization function is constructed based on the mobile crowd sensing system model to maximize the long-term cumulative utility. A task prediction module is used to capture long-range spatiotemporal dependencies in user movement patterns based on a Transformer-based spatiotemporal trajectory prediction module to predict the success probability of the opportunistic workers in completing tasks within a future time window. An evaluation module is used to quantify the multi-dimensional factors affecting task utility into a unified score using an urgency-weighted ability scoring framework to evaluate the same-stage ability of the mixed workers. The multi-dimensional factors include the base ability similarity factor, the movement cost factor, the urgency factor, and the proximity factor. For the opportunistic workers, the value of the movement cost factor is equal to the success probability. A weight adjustment module is used to obtain the optimal weight combination of the multi-dimensional factors by processing the real-time state of the mobile crowd sensing system model based on a pre-trained agent of the deep reinforcement learning-based weight self-adaptive adjustment module to output the ability evaluation score of the mixed workers and assign the current task to the worker with the highest score.

8. A computer-readable storage medium, characterized in that, The electronic device comprises:

9. An electronic device, comprising: At least one processor, a memory, and an input-output unit; The memory is used to store a computer program, and the processor is used to call the computer program stored in the memory to execute the online ability evaluation method for mixed workers of the crowd sensing according to any one of claims 1-6. The electronic device comprises: At least one processor, a memory, and an input-output unit; The memory is used to store a computer program, and the processor is used to call the computer program stored in the memory to execute the online ability evaluation method for mixed workers of the crowd sensing according to any one of claims 1-6.

Images