Trusted task allocation method for sparse mobile crowd sensing network

Abstract

The invention discloses a trusted task allocation method for a sparse mobile crowd sensing network. The trusted task allocation method is used for realizing high-quality data collection under the condition that workers report unreliable data. According to the method, firstly, contribution of tasks to data inference is evaluated based on spatio-temporal correlation, and task importance is determined accordingly; secondly, a task execution sequence of credible workers is optimized by adopting a discrete heuristic optimization algorithm, and it is ensured that perception data of important tasks has a high credible level; and finally, a task execution sequence of unknown trust workers is determined based on a confidence upper bound mechanism, tasks are preferentially allocated to potential workers which can be efficiently identified as credible, and malicious workers are eliminated in time. According to the method, collaborative optimization of task allocation cost and perception data quality is realized while the interference of malicious workers is tolerated, so that low-cost credible coverage and high-quality data set construction are realized in a multi-perception period, and the method is suitable for application scenes such as smart cities, environment monitoring and traffic management.

Description

Technical Field

[0001] This invention relates to the fields of task allocation in sparse swarm sensing networks and high-quality data collection, and particularly to a reliable task allocation method for sparse mobile swarm sensing networks. Background Technology

[0002] Sparse mobile swarm intelligence sensing collects environmental and traffic data across urban spaces using numerous mobile terminals, providing fundamental data support for smart city operations and public services. However, due to geographical obstruction, uneven population distribution, and budget constraints, the platform often only covers partial areas, leading to data sparsity and impacting the quality of sensing services. To ensure the quality of sensing services, the platform prioritizes collecting data from a small number of key tasks within each collection cycle and uses the spatiotemporal correlations between task data to infer data from unsensed tasks, thereby obtaining a complete sensing dataset at a limited cost.

[0003] In existing sparse swarm sensing networks, task allocation typically follows a process of "calculating task importance—assigning multiple tasks to workers—data sensing—inference of unsensed data." However, in real-world scenarios, existing methods generally face three key challenges: First, they assume workers will report data truthfully. If malicious false reporting occurs, erroneous data will be written into the dataset used to train / update the inference model, contaminating the training samples and ultimately resulting in low-quality inference results. Second, existing task importance calculations are based on long-term historical sensing data, failing to adequately consider changes in data distribution over time. This leads to platforms prioritizing tasks based on "outdated importance," resulting in insufficient coverage of critical tasks and weakening the performance of inference models trained on sensing task data. Third, when constraints such as sensing budget, maximum worker workload, and priority coverage of important areas coexist, the feasible combinations for multi-task allocation are enormous. Simultaneously, the platform must verify and update worker credibility during the allocation process with the lowest possible additional cost. These multiple constraints make it difficult for existing allocation strategies to simultaneously balance coverage effectiveness and verification overhead. Therefore, there is an urgent need for a reliable task allocation method for sparse mobile swarm sensing networks that can tolerate malicious workers. Summary of the Invention

[0004] This invention provides a trusted task allocation method for sparse mobile swarm sensing networks. The method involves: first, calculating the temporal variation intensity of sensing data for each task and its correlation with other task data within a rolling time window, and then fusing these two data to form a region importance score that can be adaptively updated over time. This enables dynamic evaluation of task importance and determines the priority of sensing tasks accordingly. Next, a nearest neighbor candidate task set is constructed starting from each worker's current location, and multiple task execution candidate sequences are generated based on this set. Further, the platform divides workers into trusted workers and untrusted workers, performs discrete particle encoding on the candidate sequences of trusted workers, and uses discrete particle swarm optimization iteratively to obtain a low-cost trusted worker task allocation scheme with high coverage of important tasks under budget constraints. For untrusted workers, the platform, based on a confidence upper bound strategy, prioritizes high-trust workers covering important tasks while allocating necessary verification tasks to efficiently identify potential trusted workers. Finally, the platform incorporates sparsely collected data into a historical data matrix and constructs a mask matrix, training a deep matrix factorization model to infer unsensitized task data with high quality. This achieves low-cost trusted coverage and complete high-quality dataset construction for important regions in multiple cycles.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] (1) Step 1: The platform obtains worker information and senses task information.

[0007] Platform acquires worker collection Each worker The initial location information reported and the number of tasks that can be completed. The platform obtains the set of tasks reported by data requesters. Various sensory tasks Location information, perception budget B for a single perception cycle, and total number of perception cycles. and each perception cycle Number of tasks to be covered .

[0008] (2) Step 2, during the perception cycle In this process, the platform evaluates each sensing task based on spatiotemporal correlation. The contribution to data inference, and the determination of the perception task based on this. Importance score.

[0009] The platform is in length of Computational perception task within a time sliding window Time variation range :

[0010]

[0011] In the above formula, Representing the perception task During the perception period The perceived value. Furthermore, in lengths of... Within the time window, the platform calculates Compared with other sensory tasks Spatial correlation :

[0012]

[0013] in and For the perception task and The average value of the perceived data within the corresponding time window; and Representing the perception task and During the perception period The perceived value; This represents the total number of perception tasks.

[0014] The platform then obtains the results for each sensing task based on a weighted fusion of temporal variation magnitude and spatial correlation. Importance score :

[0015]

[0016] Among them, parameters This is used to balance the weights of temporal variation and spatial relevance in the calculation of regional importance. Through the aforementioned adaptive spatiotemporal sliding window model, the platform can dynamically update the task during the data perception process. The importance score provides a precise priority basis for subsequent task allocation.

[0017] (3) Step 3: Platform adoption The -greedy strategy generates a sequence of candidate sequences for task execution for each worker.

[0018] The platform first initializes the workers. The Candidate sequences for task execution The list is empty. Next, the platform uses workers... Starting from the initial position, select the nearest [location] to that position. Each perception task was included in the candidate set. Platform sets thresholds And randomly generate a value random numbers within the range ;when The platform selects candidates from the set. Select candidate sequences to be added to the task execution process. The task that brings the greatest utility gain :

[0019]

[0020]

[0021] in, Represents the task execution sequence The sum of the importance of the covered sensing areas; Cost weighting; and Representing workers In the Mobility cost coefficient and perceived cost coefficient in each time slice. and These represent the candidate sequences for workers to complete their tasks. The required distance to move and the number of tasks to complete. The platform will assign tasks. Add to task execution candidate sequence .when The platform selects candidates from the set. Randomly select tasks Add to task execution candidate sequence .

[0022] Subsequently, the platform updated the workers. The location is the selected task location, and the corresponding nearest location is updated. A candidate set is generated; the task selection and worker location update process is repeated until the number of selected tasks reaches a certain threshold. thereby acquiring workers A candidate sequence for task execution; the platform for workers Repeat the above process to generate it A candidate sequence for task execution.

[0023] (4) Step 4: The platform uses a discrete heuristic optimization algorithm to optimize the candidate sequence of task execution for trusted workers, so as to ensure that the perception data of important tasks has a high level of trust.

[0024] The platform categorizes known trustworthy workers into a trustworthy worker set, and the remaining workers into a trustworthy unknown worker set. The platform then... The Execution sequence of candidate tasks Encoded as discrete particle fragments . For a length of An array, where each element is... Index number of the corresponding task The particle fragments belonging to trusted workers are defined as trusted fragments; the platform combines the trusted fragments of each trusted worker to construct sparse particles. .in It is a list, where each element's index corresponds to a worker's index; each element is the index of the corresponding worker. A sequence of candidate tasks to be executed. In sparse particles... Only the element corresponding to the trusted worker index is assigned the value of the corresponding trusted fragment; the remaining elements are assigned an empty value. Based on the trusted worker... There are 10 candidate sequences for task execution, and the platform build size is [size missing]. sparse particle swarm .

[0025] The platform is for each sparse particle Maintain a "discrete velocity" and based on For particles Update. Speed ​​is determined by a contained An array of n elements, where each element can take values ​​ranging from 1 to 2. When the value is 0, it means that the corresponding particle fragment remains unchanged; when the value is 1, it means that the particle fragment is updated. The update formula is shown below.

[0026]

[0027] in, and To be taken from the interval Random numbers; Inertial weight; and For learning factors; This refers to the Sigmoid function, used to map calculated continuous values ​​to... Within the range, and based on the threshold Mapping continuous calculated values ​​to discrete values ​​respectively Values. For discrete sparse particles The platform has evolved from its first iteration to the second. During the next iteration, the locally optimal sparse particle is selected. This refers to the particle that has the highest efficiency across all historical iterations. Subsequently, the local optimal particle set... Select the particle with the highest utility, and denote it as... This particle is the entire sparse particle swarm. The globally optimal sparse particle. The platform calculates sparse particles. Its effects are as follows:

[0028]

[0029] in, Indicates the first Sparse discrete particles in a perception time period The utility value; Represents particle fragments The utility value.

[0030] Following that, the platform based on For particles Update:

[0031]

[0032] Particle fragments exist When the corresponding element takes the value of 1, it is updated according to the following formula.

[0033]

[0034] Specifically, by comparing the current sparse particles Particle fragments, locally optimal particles The corresponding particle fragments, and the globally optimal particle. Among the corresponding particle fragments, the fragment with the highest fitness is selected from the three. As an updated particle fragment, and in the particle China and Israel Replace the original segment Finally, after After several iterations, the globally optimal sparse particle with the highest utility value is obtained. This is to determine the task execution sequence of all trusted workers.

[0035] (5) Step 5: The platform determines the task execution sequence of trusted unknown workers based on the confidence upper bound mechanism, prioritizes the allocation of tasks to potential workers that can be efficiently identified as trustworthy, and promptly eliminates malicious workers.

[0036] The platform encodes candidate sequences of task execution by trusted unknown workers into discrete particle fragments and defines them as unknown fragments. The platform computes each unknown fragment. Its utility:

[0037]

[0038] Indicates the newly added set of overlay tasks All sensing units in the time slice The sum of importance. Due to the fragments Once added, at least one uncovered task must be added, therefore .like This indicates that the segment did not provide any coverage gain, so let it benefit. In formula (11), Indicates workers go through The credibility score after verification is used to quantify the worker's level of trustworthiness. The closer to 1, the more workers The more trustworthy the platform, the better. Perform the same task as a trusted worker, and update the trust level based on the test results:

[0039]

[0040] in, Update the step size to improve credibility; To verify the label. If the data reported by ordinary workers matches that of trusted workers, then take... Credibility increases; otherwise, take This leads to a decrease in credibility. Furthermore, the platform sets a credibility threshold. With malicious threshold ,when ; The platform will not assign tasks to malicious workers during subsequent sensing cycles.

[0041] In the utility function (formula (11)) The "exploration term" in the confidence upper bound algorithm is defined as follows:

[0042]

[0043] in For workers The cumulative number of verifications. In the utility function (formula (11)), Indicates workers The verification reward is defined as follows:

[0044]

[0045] in, Representing reliable sparse particles Compared to ordinary segments The number of duplicate tasks between them. Parameter Defined as:

[0046]

[0047] in, The new coverage area set is defined. Based on the utility function (Formula (11)), the platform selects the unknown segment with the maximum fitness gain:

[0048]

[0049] Subsequently, the collection of unknown fragments was updated. and the filtered unknown fragments Adding optimal sparse particles Chinese index until the required number of tasks are covered. At this point, a complete particle can be obtained. This is used to represent the final task execution sequence (task allocation result) for all workers.

[0050] (6) Step Six: The platform performs missing data inference operations, and performs inference operations on the missing data for this period. The completed perception data is then appended to the historical perception dataset.

[0051] The platform will first consider this sensing cycle. Various sensory tasks The collected sensing data is integrated into a sparse data vector. And construct a mask matrix Among them, The collected locations are assigned a value of 1, and the uncollected locations are assigned a value of 0. Next, the platform constructs a deep matrix factorization model and sets the rank parameter. Network Hidden Layers and historical perception dataset and Input a deep matrix factorization model. The historical perception dataset is also included. Historical datasets known to the platform and the process Complete sensing dataset acquired in one sensing cycle The results are derived from the combined data. The platform uses the reconstruction error calculated only for the collected data as the fitting term, and combines it with network weight regularization. With latent variable regularization Construct the training objective function:

[0052]

[0053] in The weight parameters are the network weight regularization term. The weight parameters are for latent variable regularization. The platform minimizes the objective function (Equation (17)) through iterative updates using the Adam optimizer until the required number of iterations is met. Output the complete data vector after completion. Finally, the platform will Added to historical perception dataset In the middle, thus obtaining the process Historical dataset after the first sensing cycle .

[0054] (7) Step 7: Based on the saved worker trust scores and historical perception dataset, the platform enters the next perception cycle. Repeat steps one through six.

[0055] Beneficial effects

[0056] Compared with existing technologies, this invention provides a reliable task allocation scheme for sparse mobile swarm sensing networks: The platform first constructs an adaptive spatiotemporal sliding window importance assessment mechanism based on historical observation data, comprehensively characterizing the temporal changes and spatial correlations of tasks and dynamically updating task priorities. This ensures that high-importance areas are covered and the value of key data collection is improved in scenarios where data is sparse and requirements change over time. In the task allocation phase, this invention... Greedy's "nearest neighbor optimal selection + random exploration" rule generates multiple candidate task sequences for workers, balancing the quality and diversity of candidate solutions and avoiding local optima caused by simple greed. Furthermore, this invention encodes the trustworthy worker sequence into trustworthy segments and constructs a sparse particle swarm consisting only of these segments. Discrete particle swarm optimization is used to quickly obtain trustworthy task allocation results with higher global utility in the combinatorial search space, thereby improving high-importance coverage and overall benefit under budget, movement, and perception cost constraints. When trustworthy worker coverage is insufficient, this invention constructs unknown worker segments that include additional coverage gain, trustworthiness weighting, and confidence upper bound exploration. The invention enhances the utility of verification rewards and continuously identifies potential trustworthy workers, eliminates malicious workers, and gradually completes effective fragments to meet coverage requirements through a "verification-update-threshold determination" mechanism, thereby reducing the risk of untrusted data injection and improving task completion efficiency. Finally, the invention incorporates the data collected during the perception cycle into the historical data to form a sparse perception dataset, which is then input into a deep matrix factorization model for data completion, resulting in a complete perception dataset. By saving the updated trust score and historical data, cross-cycle adaptive iteration is achieved, significantly improving the data credibility, key area coverage quality, and platform stability and robustness in sparse swarm intelligence perception. Attached Figure Description

[0057] To more clearly illustrate the technical solution of the present invention, the embodiments of the present invention will be further described below with reference to the accompanying drawings. The accompanying drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0058] Figure 1This is a schematic diagram of the system structure of the present invention;

[0059] Figure 2 This is a schematic diagram illustrating the calculation of regional importance based on an adaptive spatiotemporal sliding window in this invention;

[0060] Figure 3 This is a schematic diagram of the reliable sparse discrete particle encoding in this invention;

[0061] Figure 4 This is a schematic diagram of the particle repair process for illegal discrete particles in this invention;

[0062] Figure 5 The data inference quality of the method of the present invention is compared with that of other methods as the maximum number of tasks performed by workers changes;

[0063] Figure 6 This is a comparison of the perception overhead of the method of the present invention with that of other methods as the maximum number of tasks a worker can perform.

[0064] Figure 7 The data inference quality of the method of the present invention is compared with that of other methods as the number of workers changes;

[0065] Figure 8 This is a comparison of the perception overhead of the method of the present invention with that of other methods as the number of workers changes. Detailed Implementation

[0066] The technical solution of the present invention will be further described below with reference to specific embodiments. However, the scope of protection of the present invention is not limited to the following embodiments. Without departing from the concept of the present invention, all equivalent modifications or substitutions made by those skilled in the art should fall within the scope of protection of the present invention.

[0067] Example 1

[0068] This embodiment uses the China air quality dataset as an example. The 144 monitoring stations are considered as 144 sensing tasks, each task... It has fixed latitude and longitude coordinates; the example of the latitude and longitude range of the station is longitude. ,latitude The platform uses 1-hour intervals as a sensing time period, and employs data from the first 50 time periods as the platform's known historical sensing dataset. Subsequent... A time period is taken as the perception time period. Worker scale is taken as... The percentages of trusted workers, trusted unknown workers, and malicious workers are 0.5, 0.4, and 0.1 respectively. The maximum number of tasks each worker can complete is set to... Unit moving cost Unit perceived cost exist The data requester provides the total budget for each time slice. and requires coverage of no less than This is one task. The specific implementation steps are as follows:

[0069] (1) Step 1: The platform obtains worker information and senses task information.

[0070] Platform acquires worker collection Each worker The initial location information reported and the number of tasks that can be completed. Get the set of tasks reported by the data requester. Various sensory tasks Location information, perception budget for a single perception cycle total and each perception cycle Number of tasks to be covered .

[0071] (2) Step 2, during the perception cycle In this process, the platform evaluates each sensing task based on spatiotemporal correlation. The contribution to data inference, and the determination of the perception task based on this. Importance score.

[0072] The platform is in length of Computational perception task within a time sliding window Time variation factor :

[0073]

[0074] In the above formula, Representing the perception task During the perception period The observed values. Furthermore, in length... Within the time window, the platform calculates Compared with other sensory tasks Spatial correlation :

[0075]

[0076] in and For the perception task and The average value of the perceived data within the corresponding time window; and Representing the perception task and During the perception period The perceived value; This represents the total number of perception tasks.

[0077] The platform then obtains the results for each sensing task based on a weighted fusion of temporal variation magnitude and spatial correlation. Importance score :

[0078]

[0079] Among them, parameters This is used to balance the weights of temporal variation and spatial relevance in the calculation of regional importance. Through the aforementioned adaptive spatiotemporal sliding window model, the platform can dynamically update the task during the data perception process. The importance score provides a precise priority basis for subsequent task allocation.

[0080] (3) Step 3: Platform adoption The -greedy strategy generates a sequence of candidate sequences for task execution for each worker.

[0081] The platform first initializes the workers. The Candidate sequences for task execution The list is empty. Next, the platform uses workers... Starting from the initial position, select the nearest [location] to that position. Each perception task was included in the candidate set. Platform sets thresholds And randomly generate a value random numbers within the range ;when The platform selects candidates from the set. Select candidate sequences to be added to the task execution process. The task that brings the greatest utility gain :

[0082]

[0083]

[0084] in, Indicates a candidate sequence for task execution. The sum of the importance of the covered sensing areas; Cost weighting; and Representing workers In the Mobility cost coefficient and perceived cost coefficient in each time slice. and These represent the candidate sequences for workers to complete their tasks. The required distance to move and the number of tasks to complete. The platform will assign tasks. Add to task execution candidate sequence .when The platform selects candidates from the set. Randomly select tasks Add to task execution candidate sequence .

[0085] Subsequently, the platform updated the workers. The location is the selected task location, and the corresponding nearest location is updated. A candidate set is generated; the task selection and worker location update process is repeated until the number of selected tasks reaches a certain threshold. thereby acquiring workers A candidate sequence for task execution; the platform for workers Repeat the above process to generate it A candidate sequence for task execution.

[0086] (4) Step 4: The platform uses a discrete heuristic optimization algorithm to optimize the candidate sequence of task execution for trusted workers, so as to ensure that the perception data of important tasks has a high level of trust.

[0087] The platform categorizes known trustworthy workers into a trustworthy worker set, and the remaining workers into a trustworthy unknown worker set. The platform then... The Execution sequence of candidate tasks Encoded as discrete particle fragments . For a length of An array, where each element is... Index number of the corresponding task The particle fragments belonging to trusted workers are defined as trusted fragments; the platform combines the trusted fragments of each trusted worker to construct sparse particles. .in It is a list, where each element's index corresponds to a worker's index; each element is the index of the corresponding worker. A sequence of candidate tasks to be executed. In sparse particles... Only the element corresponding to the trusted worker index is assigned the value of the corresponding trusted fragment; the remaining elements are assigned an empty value. Based on the trusted worker... There are 10 candidate sequences for task execution, and the platform build size is [size missing]. sparse particle swarm .

[0088] The platform is for each sparse particle Maintain a "discrete velocity" and based on For particles Update. Speed ​​is determined by a contained An array of n elements, where each element can take values ​​ranging from 1 to 2. When the value is 0, it means that the corresponding particle fragment remains unchanged; when the value is 1, it means that the particle fragment is updated. The update formula is shown below.

[0089]

[0090] in, and To be taken from the interval Random numbers; Inertial weight; and For learning factors; This refers to the Sigmoid function, used to map calculated continuous values ​​to... Within the range, and based on the threshold Mapping continuous calculated values ​​to discrete values ​​respectively Values. For discrete sparse particles The platform has evolved from its first iteration to the second. During the next iteration, the locally optimal sparse particle is selected. This refers to the particle that has the highest efficiency across all historical iterations. Subsequently, the local optimal particle set... Select the particle with the highest utility, and denote it as... This particle is the entire sparse particle swarm. The globally optimal sparse particle. The platform calculates sparse particles. Its effects are as follows:

[0091]

[0092] in, Indicates the first Sparse discrete particles in a perception time period The utility value; Represents particle fragments The utility value.

[0093] Following that, the platform based on For particles Update:

[0094]

[0095] Particle fragments exist When the corresponding element takes the value of 1, it is updated according to the following formula.

[0096]

[0097] Specifically, by comparing the current sparse particles Particle fragments, locally optimal particles The corresponding particle fragments, and the globally optimal particle. Among the corresponding particle fragments, the fragment with the highest fitness is selected from the three. As an updated particle fragment, and in the particle China and Israel Replace the original segment Finally, after After several iterations, the globally optimal sparse particle with the highest utility value is obtained. This is to determine the task execution sequence of all trusted workers.

[0098] (5) Step 5: The platform determines the task execution sequence of trusted unknown workers based on the confidence upper bound mechanism, prioritizes the allocation of tasks to potential workers that can be efficiently identified as trustworthy, and promptly eliminates malicious workers.

[0099] The platform encodes candidate sequences of task execution by trusted unknown workers into discrete particle fragments and defines them as unknown fragments. The platform computes each unknown fragment. Its utility:

[0100]

[0101] Indicates the newly added set of overlay tasks All sensing units in the time slice The sum of importance. Due to the fragments Once added, at least one uncovered task must be added, therefore .like This indicates that the segment did not provide any coverage gain, so let it benefit. In formula (11), Indicates workers go through The credibility score after verification is used to quantify the worker's level of trustworthiness. The closer to 1, the more workers The more trustworthy the platform, the better. Perform the same task as a trusted worker, and update the trust level based on the test results:

[0102]

[0103] in, Update the step size to improve credibility; To verify the label. If the data reported by ordinary workers matches that of trusted workers, then take... Credibility increases; otherwise, take This leads to a decrease in credibility. Furthermore, the platform sets a credibility threshold. With malicious threshold ,when ; Time to determine workers These are malicious workers. The platform will not assign tasks to malicious workers in subsequent detection cycles.

[0104] In the utility function (formula (11)) The "exploration term" in the confidence upper bound algorithm is defined as follows:

[0105]

[0106] in For workers The cumulative number of verifications. In the utility function (formula (11)), Indicates workers The verification reward is defined as follows:

[0107]

[0108] in, Representing reliable sparse particles Compared to ordinary segments The number of duplicate tasks between them. Parameter Defined as:

[0109]

[0110] in, The new coverage area set is defined. Based on the utility function (Formula (11)), the platform selects the unknown segment with the maximum fitness gain:

[0111]

[0112] Subsequently, the collection of unknown fragments was updated. and the filtered unknown fragments Adding optimal sparse particles Chinese index until the required number of tasks are covered. At this point, a complete particle can be obtained. This is used to represent the final task execution sequence (task allocation result) for all workers.

[0113] (6) Step Six: The platform performs missing data inference operations, and performs inference operations on the missing data for this period. The completed perception data is then appended to the historical perception dataset.

[0114] The platform will first consider this sensing cycle. Various sensory tasks The collected sensing data is integrated into a sparse data vector. And construct a mask matrix Among them, The collected locations are assigned a value of 1, and the uncollected locations are assigned a value of 0. Next, the platform constructs a deep matrix factorization model and sets the rank parameter. Network Hidden Layers and historical perception dataset and Input a deep matrix factorization model. The historical perception dataset is also included. Historical datasets known to the platform and the process Complete sensing dataset acquired in one sensing cycle The results are derived from the combined data. The platform uses the reconstruction error calculated only for the collected data as the fitting term, and combines it with network weight regularization. With latent variable regularization Construct the training objective function:

[0115]

[0116] in The weight parameters are the network weight regularization term. The weight parameters are for latent variable regularization. The platform minimizes the objective function (Equation (17)) through iterative updates using the Adam optimizer until the required number of iterations is met. Output the complete data vector after completion. Finally, the platform will Added to historical perception dataset In the middle, thus obtaining the process Historical dataset after the first sensing cycle .

[0117] (7) Step 7: Based on the saved worker trust scores and historical perception dataset, the platform enters the next perception cycle. Repeat steps one through six.

[0118] Figure 1This is a schematic diagram of the system structure of the present invention. The present invention includes a region division module, a spatiotemporal sliding window region importance assessment module, a trusted worker task allocation module, a trusted unknown worker verification and task completion module, a sparse data construction module, and a deep matrix factorization data completion module. The operation flow of the present invention is as follows: The platform first performs gridding processing on the target region, transforming the data collection scenario into a structured spatiotemporal matrix form; then, based on historical observation data, the dynamic importance of the region is calculated through a spatiotemporal sliding window model, and this is used as the priority basis for task allocation; in the task scheduling stage, trusted workers obtain the optimal task set covering high-importance regions through discrete particle swarm optimization; for regions with insufficient coverage by trusted workers, trusted unknown workers supplement the collection through a confidence upper bound selection strategy and continuously identify potential trusted workers; finally, the sparse data is inferred and completed through a deep matrix factorization model, thereby forming a complete perception dataset and forming an adaptive loop.

[0119] Figure 2 This diagram illustrates the calculation of regional importance based on an adaptive spatiotemporal sliding window in this invention. The invention uses temporal and spatial sliding windows to dynamically model the importance of a perceived region. The temporal sliding window section demonstrates the calculation of the time variation factor using the observation difference between consecutive time slices. The process is used to characterize the changing trends of regional data over time; the spatial sliding window section shows how spatial factors are obtained through correlation calculations based on perception unit targets. This is used to characterize the spatial correlation of regional data. The figure also shows the two factors through weighting parameters. Integration forms the ultimate regional importance The process enables spatiotemporal adaptive modeling of regional importance, thus accurately identifying important regions even under dynamic environmental changes.

[0120] Figure 3 This is a schematic diagram of the trusted sparse discrete particle encoding in this invention. This invention constructs the task sequence of trusted workers as sparse discrete particles and combines fragments from multiple workers into a particle swarm. Each particle fragment consists of tasks that the worker can execute within the current time slice, and the particle structure illustrates the sparse representation characteristics of the fragment in the task space. In the particle swarm, different particles represent different task combination methods, and the differences between particles are reflected in different task orders, different coverage areas, and different movement costs. This structure provides the encoding foundation for subsequent discrete particle swarm optimization.

[0121] Figure 4This diagram illustrates the particle repair process for illegal discrete particles in this invention. This invention addresses "illegal particles" generated during the update process of discrete particle swarms that do not meet task constraints, workload limits, or reachability conditions, and designs a particle repair process. The process shown includes: a uniqueness constraint check, i.e., detecting whether the same task is repeatedly executed by multiple workers; a workload limit check, i.e., determining whether the number of tasks undertaken by a worker exceeds a preset upper limit; and a budget constraint check, i.e., determining whether the total cost of all workers executing tasks exceeds the budget. After detecting an illegal particle, Figure 4 The invention demonstrates corresponding repair strategies, such as task deletion and task addition operations, to ensure that the updated particles satisfy all constraints again. Through this repair mechanism, the invention ensures that the particle swarm optimization process always searches within the feasible solution space, improving overall optimization efficiency and stability.

[0122] Figure 5 This is a comparison graph showing how the data inference quality of the method of the present invention changes with the maximum number of tasks performed by workers. (See figure.) Figure 5 As shown, the data inference quality of the method of this invention is compared with existing task allocation strategies (greedy algorithm, distance-first algorithm, and genetic algorithm) as the maximum number of tasks for workers gradually increases from a small value. The figure shows that the method of this invention consistently has a lower inference error (RMSE) than the comparison methods. Furthermore, as the number of executable tasks increases, the method of this invention can more effectively allocate high-value tasks, further reducing the data inference error. This figure reflects that the task allocation strategy of this invention, based on trusted worker optimization and regional importance modeling, can maintain stable and high-quality data inference performance even in information-sparse scenarios.

[0123] Figure 6 This is a comparison graph showing the change in perception overhead of the method of the present invention as the maximum number of tasks a worker can perform. For example... Figure 6 As shown, the differences in total perception overhead (including worker movement distance, path length, or time cost) between the present invention and other methods were compared under different scenarios with varying maximum number of tasks per worker. The results show that the present invention, while maintaining high data quality, can reduce the total movement cost and task execution overhead of workers, and the overhead growth rate is significantly lower than that of the comparative methods. This is because the present invention effectively reduces the allocation of redundant and low-value tasks through particle swarm optimization and task repair mechanisms, thereby improving resource utilization and further demonstrating its high execution efficiency in budget-constrained scenarios.

[0124] Figure 7 A comparative graph showing how the quality inferred from the data of the method of the present invention changes with the number of workers. For example... Figure 7As shown, different task allocation methods exhibit varying performance in inference accuracy as the number of workers gradually increases. The figure demonstrates that the method of this invention maintains good inference performance even with a small number of workers, and further reduces inference error as the number of workers increases, significantly outperforming the comparative methods. This indicates that the trusted worker selection, UCB verification, and deep matrix factorization completion of this invention work together to ensure the system maintains stable and robust inference capabilities even with significant changes in the number of workers.

[0125] Figure 8 This is a comparison graph showing the perceptual overhead of the method of the present invention as a function of the number of workers. For example... Figure 8 As shown, under conditions of continuously increasing worker numbers, the method of this invention can more effectively control perception overhead compared to other task allocation strategies. Although more workers mean a larger scheduling space, this invention utilizes mechanisms such as trusted worker priority scheduling, particle repair, and reduced redundant execution to ensure that the increase in task execution costs is significantly lower than that of the control method. This demonstrates that this invention still maintains high cost efficiency and scalability in large-scale worker scenarios, making it more suitable for deploying crowdsourced perception systems in real-world scenarios.

[0126] In summary, this invention achieves systematic optimization of the entire task allocation process in complex environments with malicious workers by constructing a collaborative mechanism that integrates adaptive spatiotemporal region importance assessment, priority allocation of trusted workers, verification and supplementation of unknown workers, and sparse data inference and completion. This invention not only significantly reduces inference errors and maintains stable data quality under conditions of uncertain worker behavior and missing data, but also effectively controls perception overhead under budget constraints and scalability limitations, improving the feasibility of task allocation and the overall execution efficiency of the system. Therefore, it provides higher reliability, adaptability, and scalability for sparse swarm intelligence sensing networks in real-world scenarios.

Claims

1. A reliable task allocation method for sparse mobile swarm sensing networks, characterized in that, Includes the following steps: (1) Step 1: The platform obtains worker information and senses task information. Platform acquires worker collection Each worker The initial location information reported and the number of tasks that can be completed. The platform obtains the set of tasks reported by data requesters. Various sensory tasks Location information, perception budget B for a single perception cycle, and total number of perception cycles. and each perception cycle Number of tasks to be covered . (2) Step 2, during the perception cycle In this process, the platform evaluates each sensing task based on spatiotemporal correlation. The contribution to data inference, and the determination of the perception task based on this. Importance score. The platform is in length of Computational perception task within a time sliding window Time variation range : In the above formula, Representing the perception task During the perception period The perceived value. Furthermore, in lengths of... Within the time window, the platform calculates Compared with other sensory tasks Spatial correlation : in and For the perception task and The average value of the perceived data within the corresponding time window; and Representing the perception task and During the perception period The perceived value; This represents the total number of perception tasks. The platform then obtains the results for each sensing task based on a weighted fusion of temporal variation magnitude and spatial correlation. Importance score : Among them, parameters This is used to balance the weights of temporal variation and spatial relevance in the calculation of regional importance. Through the aforementioned adaptive spatiotemporal sliding window model, the platform can dynamically update the task during the data perception process. The importance score provides a precise priority basis for subsequent task allocation. (3) Step 3: Platform adoption The -greedy strategy generates a sequence of candidate sequences for task execution for each worker. The platform first initializes the workers. The Candidate sequences for task execution The list is empty. Next, the platform uses workers... Starting from the initial position, select the nearest [location] to that position. Each perception task was included in the candidate set. Platform sets thresholds And randomly generate a value random numbers within the range ;when The platform selects candidates from the set. Select candidate sequences to be added to the task execution process. The task that brings the greatest utility gain : in, Indicates a candidate sequence for task execution. The sum of the importance of the covered sensing areas; Cost weighting; and Representing workers In the Mobility cost coefficient and perceived cost coefficient in each time slice. and These represent the candidate sequences for workers to complete their tasks. The required distance to move and the number of tasks to complete. The platform will assign tasks. Add to task execution candidate sequence .when The platform selects candidates from the set. Randomly select tasks Add to task execution candidate sequence . Subsequently, the platform updated the workers. The location is the selected task location, and the corresponding nearest location is updated. A candidate set is generated; the task selection and worker location update process is repeated until the number of selected tasks reaches a certain threshold. thereby acquiring workers A candidate sequence for task execution; the platform for workers Repeat the above process to generate it A candidate sequence for task execution. (4) Step 4: The platform uses a discrete heuristic optimization algorithm to optimize the candidate sequence of task execution for trusted workers, so as to ensure that the perception data of important tasks has a high level of trust. The platform categorizes known trustworthy workers into a trustworthy worker set, and the remaining workers into a trustworthy unknown worker set. The platform then... The Execution sequence of candidate tasks Encoded as discrete particle fragments . For a length of An array, where each element is... Index number of the corresponding task The particle fragments belonging to trusted workers are defined as trusted fragments; the platform combines the trusted fragments of each trusted worker to construct sparse particles. .in It is a list, where each element's index corresponds to a worker's index; each element is the index of the corresponding worker. A sequence of candidate tasks to be executed. In sparse particles... Only the element corresponding to the trusted worker index is assigned the value of the corresponding trusted fragment; the remaining elements are assigned an empty value. Based on the trusted worker... There are 10 candidate sequences for task execution, and the platform build size is [size missing]. sparse particle swarm . The platform is for each sparse particle Maintain a "discrete velocity" and based on For particles Update. Speed ​​is determined by a contained An array of n elements, where each element can take values ​​ranging from 1 to 2. When the value is 0, it means that the corresponding particle fragment remains unchanged; when the value is 1, it means that the particle fragment is updated. The update formula is shown below. in, and To be taken from the interval Random numbers; Inertial weight; and For learning factors; This refers to the Sigmoid function, used to map calculated continuous values ​​to... Within the range, and based on the threshold Mapping continuous calculated values ​​to discrete values ​​respectively Values. For discrete sparse particles The platform has evolved from its first iteration to the second. During the next iteration, the locally optimal sparse particle is selected. This refers to the particle that has the highest efficiency across all historical iterations. Subsequently, the local optimal particle set... Select the particle with the highest utility, and denote it as... This particle is the entire sparse particle swarm. The globally optimal sparse particle. The platform calculates sparse particles. Its effects are as follows: in, Indicates the first Sparse discrete particles in a perception time period The utility value; Represents particle fragments The utility value. Following that, the platform based on For particles Update: Particle fragments exist When the corresponding element takes the value of 1, it is updated according to the following formula. Specifically, by comparing the current sparse particles Particle fragments, locally optimal particles The corresponding particle fragments, and the globally optimal particle. Among the corresponding particle fragments, the fragment with the highest fitness is selected from the three. As an updated particle fragment, and in the particle China and Israel Replace the original segment Finally, after After several iterations, the globally optimal sparse particle with the highest utility value is obtained. This is to determine the task execution sequence of all trusted workers. (5) Step 5: The platform determines the task execution sequence of trusted unknown workers based on the confidence upper bound mechanism, prioritizes the allocation of tasks to potential workers that can be efficiently identified as trustworthy, and promptly eliminates malicious workers. The platform encodes candidate sequences of task execution by trusted unknown workers into discrete particle fragments and defines them as unknown fragments. The platform computes each unknown fragment. Its utility: Indicates the newly added set of overlay tasks All sensing units in the time slice The sum of importance. Due to the fragments Once added, at least one uncovered task must be added, therefore .like This indicates that the segment did not provide any coverage gain, so let it benefit. In formula (11), Indicates workers go through The credibility score after verification is used to quantify the worker's level of trustworthiness. The closer to 1, the more workers The more trustworthy the platform, the better. Perform the same task as a trusted worker, and update the trust level based on the test results: in, Update the step size to improve credibility; To verify the label. If the data reported by ordinary workers matches that of trusted workers, then take... Credibility increases; otherwise, take This leads to a decrease in credibility. Furthermore, the platform sets a credibility threshold. With malicious threshold ,when ; The platform will not assign tasks to malicious workers during subsequent sensing cycles. In the utility function (formula (11)) The "exploration term" in the confidence upper bound algorithm is defined as follows: in For workers The cumulative number of verifications. In the utility function (formula (11)), Indicates workers The verification reward is defined as follows: in, Representing reliable sparse particles Compared to ordinary segments The number of duplicate tasks between them. Parameter Defined as: in, The new coverage area set is defined. Based on the utility function (Formula (11)), the platform selects the unknown segment with the maximum fitness gain: Subsequently, the collection of unknown fragments was updated. and the filtered unknown fragments Adding optimal sparse particles Chinese index until the required number of tasks are covered. At this point, a complete particle can be obtained. This is used to represent the final task execution sequence (task allocation result) for all workers. (6) Step Six: The platform performs missing data inference operations, and performs inference operations on the missing data for this period. The completed perception data is then appended to the historical perception dataset. The platform will first consider this sensing cycle. Various sensory tasks The collected sensing data is integrated into a sparse data vector. And construct a mask matrix Among them, The collected locations are assigned a value of 1, and the uncollected locations are assigned a value of 0. Next, the platform constructs a deep matrix factorization model and sets the rank parameter. Network Hidden Layers and historical perception dataset and Input a deep matrix factorization model. The historical perception dataset is also included. Historical datasets known to the platform and the process Complete sensing dataset acquired in one sensing cycle The results are derived from the combined data. The platform uses the reconstruction error calculated only for the collected data as the fitting term, and combines it with network weight regularization. With latent variable regularization Construct the training objective function: in The weight parameters are the network weight regularization term. The weight parameters are for latent variable regularization. The platform minimizes the objective function (Equation (17)) through iterative updates using the Adam optimizer until the required number of iterations is met. Output the complete data vector after completion. Finally, the platform will Added to historical perception dataset In the middle, thus obtaining the process Historical dataset after the first sensing cycle . (7) Step 7: Based on the saved worker trust scores and historical perception dataset, the platform enters the next perception cycle. Repeat steps one through six.

Images