A multi-dimensional performance portrait-based follow-up task granularity dispatching system

Abstract

This invention discloses a granular task dispatching system and method based on multi-dimensional performance profiling. The system includes a server and a mobile terminal connected to the server. The server receives and standardizes customer bills, meter readings, tasks, personnel, regions, environments, historical dispatches, historical executions, and terminal feedback data through a data access module, obtaining multiple feature types. A multi-layered task graph is constructed using a graph granularity module, generating task granularity results. A task difficulty index and corrected performance profile are generated using a difficulty profiling module. A ternary matching module calculates the ternary matching score between task clusters, executors, and regions. A constraint dispatching module generates a dispatching scheme with a confidence lower bound on the completion probability and sends it to the mobile terminal. A feedback update module updates relevant parameters based on the credibility of execution feedback. This invention improves the accuracy, stability, and traceability of task dispatching.

Description

Technical Field

[0001] This invention relates to the field of computer-aided task scheduling and resource allocation technology, and in particular to a task dispatching system based on multi-dimensional performance profiling with granularity. Background Technology

[0002] In the process of electricity marketing and customer service management, field tasks such as meter reading, overdue payment collection, anomaly investigation, and on-site verification are typically characterized by a large number of tasks, wide distribution, high timeliness requirements, and complex on-site conditions. These tasks often involve multiple types of data, including customer bills, meter status, task location, personnel involved, service area, weather and traffic conditions, and on-site feedback. There are strong correlations between different types of data, and the results of task execution are easily affected by factors such as customer outreach, regional environment, personnel status, and on-site anomalies.

[0003] Existing task assignment methods typically focus on allocating tasks based on fixed areas, fixed grids, personnel load, task distance, or historical completion status. While this approach can meet general task flow needs, it is prone to problems such as unreasonable task splitting and merging, duplicate execution routes, biased personnel ability evaluation, and uneven allocation of high-risk tasks in scenarios with large task scale, complex task types, or significant regional differences. These issues can negatively impact task execution efficiency and task completion quality.

[0004] Furthermore, the outcome of field tasks is not solely determined by the capabilities of the personnel performing the tasks. It is also influenced by factors such as task difficulty, customer cooperation, meter anomalies, familiarity with the area, historical task assignment tendencies, and weather and traffic conditions. Evaluating personnel capabilities solely based on historical completion rates, task quantity, or payment collection results may lead to distorted assessments of personnel abilities, thereby affecting the accuracy of subsequent task allocation.

[0005] Furthermore, in actual execution, mobile terminals typically collect feedback data such as location, images, readings, access records, time information, and operation logs. However, the authenticity, completeness, and consistency of this feedback data vary. Without effective verification of execution feedback, low-quality or abnormal feedback can easily affect subsequent task evaluation and dispatch strategies, reducing the system's continuous optimization capabilities.

[0006] Therefore, how to improve the rationality of task allocation, the accuracy of personnel matching, the stability of dispatch results, and the traceability of the execution process in complex copy-and-collection business scenarios has become an urgent problem to be solved. Summary of the Invention

[0007] To address the shortcomings of existing technologies, the present invention aims to provide a task dispatching system and method based on multi-dimensional performance profiles, which improves the accuracy, stability, and traceability of task granularity determination, personnel matching, risk-constrained task dispatching, and execution feedback updates.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a task granularity dispatching system based on multi-dimensional performance profiling, comprising a server and a mobile terminal communicatively connected to the server, wherein the server includes: The data access module is used to receive and standardize customer billing data, meter data, task data, personnel data, regional data, environmental data, historical dispatch data, historical execution data, and terminal feedback data to obtain task characteristics, personnel characteristics, regional characteristics, environmental characteristics, and feedback characteristics. The graph granularity module, connected to the data access module, is used to construct a multi-level task graph containing multi-level task nodes and associated edges based on the task characteristics, generate candidate task clusters, and calculate granularity benefit values ​​according to spatial aggregation, route saving, time window consistency, difficulty dispersion, risk concentration and uncertainty penalty to obtain task granularity results. The difficulty profile module connects the graph granularity module and the data access module. It is used to generate a task difficulty index based on the task granularity result and the task characteristics, and to generate a corrected performance profile based on the task difficulty index, the personnel characteristics, the dispatch tendency probability determined by the historical dispatch data, and the historical execution data. The ternary matching module connects the graph granularity module, the difficulty profiling module, and the data access module. It is used to calculate the ternary matching score with task clusters, executors, and dispatch areas as nodes and candidate dispatch relationships as hyperedges. The constraint dispatch module connects the ternary matching module and the mobile terminal. It is used to generate a dispatch plan with a confidence lower bound on the completion probability based on the task granularity result, the task difficulty index, the correction efficiency profile, the ternary matching score, personnel characteristics, task characteristics, and environmental characteristics, and send it to the mobile terminal. The feedback update module, which connects the mobile terminal, the data access module, and the constraint dispatch module, is used to generate execution feedback credibility based on the feedback features, and use it as a sample weight to update the task difficulty parameter, the correction performance profile, and the dispatch parameter.

[0009] Furthermore, the data access module includes: The data partitioning unit is used to divide customer billing data, meter data, task data, personnel data, regional data, environmental data, historical dispatch data, historical execution data, and terminal feedback data into mutually unique data partitions according to data source identifier, business primary key, and timestamp. The same original data is written to only one data partition after deduplication. The standardization unit, connected to the data partitioning unit, is used to perform unified encoding, missing data marking, anomaly marking, and normalization processing on the data in each data partition to obtain the task characteristics, personnel characteristics, regional characteristics, environmental characteristics, and feedback characteristics.

[0010] Furthermore, the spectral granularity module includes: A node construction unit, connected to the standardization unit, is used to construct household-level task nodes, meter box-level task nodes, building-level task nodes, community-level task nodes, grid-level task nodes, and route segment-level task nodes according to the task characteristics. An edge construction unit, connected to the node construction unit, is used to generate the associated edges based on spatial adjacency, task window relationship, task type relationship, customer reach relationship, historical execution relationship, and risk association relationship; The region verification unit, connected to the edge construction unit, is used to verify the region affiliation of task nodes in the multi-layer dispatch task graph based on the region characteristics, so that each task node belongs to only one dispatch region under the same dispatch region level determined by the region characteristics.

[0011] Furthermore, the spectral granularity module also includes: A candidate cluster generation unit, connected to the edge construction unit, is used to generate candidate task clusters based on the task nodes and associated edges in the multi-layer task graph. The revenue calculation unit, connected to the candidate cluster generation unit, is used to calculate the granular revenue value of the candidate task cluster based on spatial aggregation degree, route saving rate, time window consistency degree, task type consistency degree, estimated difficulty dispersion determined by the task characteristics, risk concentration degree, and uncertainty penalty value. The granularity decision unit, connected to the revenue calculation unit, is used to generate the task granularity result, including at least one of the following: merged task clusters, split tasks, review tasks, and dual-person execution tasks, based on the comparison results between the granularity revenue value and the merging threshold, splitting threshold, and review threshold.

[0012] Furthermore, the difficulty profiling module includes: The task object determination unit, connected to the granularity decision unit, is used to determine the task object to be evaluated based on the task granularity result, wherein the task object to be evaluated is a single task or a cluster of tasks. The difficulty index extraction unit, connected to the standardization unit, is used to extract the following from the task characteristics and the regional characteristics: overdue payment age, overdue amount, difficulty in reaching customers, historical commitment default rate determined by the historical execution data, meter anomaly degree, geographical dispersion, execution window limitation, historical complaint or refusal risk rate determined by the task characteristics, the regional characteristics and the historical execution data, and data uncertainty. The difficulty calculation unit is connected to the task object determination unit and the difficulty index extraction unit, respectively, and is used to generate the task difficulty index of the task object to be evaluated based on the index extracted by the difficulty index extraction unit. When the task to be evaluated is a task cluster, the difficulty calculation unit generates the task difficulty index of the task cluster based on the task difficulty index of each task within the task cluster and the importance weight determined by the amount of outstanding payment, task risk level, task deadline, and meter anomaly marker in the task characteristics.

[0013] Furthermore, the difficulty profiling module also includes: A baseline probability generation unit, connected to the difficulty calculation unit, is used to generate a baseline success probability for a task based on the task difficulty index, the customer reach probability determined by the task characteristics and the historical execution data, the regional convenience determined by the regional characteristics, the task urgency determined by the task characteristics, and the environmental impact determined by the environmental characteristics. The deviation correction unit, connected to the baseline probability generation unit and the standardization unit, is used to generate the correction contribution value of the executor based on the task baseline success probability, the dispatch tendency probability determined by the historical dispatch data, and the historical execution data. The image generation unit, connected to the deviation correction unit, is used to generate the correction effectiveness image based on the correction contribution value.

[0014] Furthermore, the ternary matching module includes: The hypergraph construction unit connects the graph granularity module, the difficulty profiling module, and the data access module. It is used to take the task clusters in the task granularity results as task cluster nodes, the executors corresponding to the personnel characteristics as personnel nodes, the non-overlapping dispatch areas determined by the regional characteristics under the same dispatch area level as regional nodes, and the candidate dispatch relationship as a hyperedge connecting a task cluster node, a personnel node, and a regional node. The matching calculation unit, connected to the hypergraph construction unit, is used to calculate the ternary matching score based on the task difficulty index, the correction efficiency profile, the regional features, the personnel load determined by the personnel features, the route adaptability determined by the task features and the regional features, the time window matching degree determined by the task features and the personnel features, and the historical complaint risk determined by the historical execution data.

[0015] Furthermore, the constraint dispatch module includes: The probability prediction unit is connected to the ternary matching module and the difficulty profiling module. It is used to predict the average completion probability based on the ternary matching score, the correction efficiency profile, the task difficulty index, and the customer reach probability determined by the task characteristics and the historical execution data. An uncertainty calculation unit, connected to the completion probability prediction unit, is used to determine the prediction standard deviation based on the number of historical samples determined by the historical execution data, the degree of data missing determined by the task characteristics, the risk concentration determined by the revenue calculation unit, and the volatility of the completion probability prediction model used by the completion probability prediction unit, and to generate the completion probability confidence lower bound based on the average completion probability and the prediction standard deviation. The dispatch decision unit, connected to the uncertainty calculation unit, is used to generate the dispatch scheme under the conditions of satisfying the task time window determined by the task characteristics, the personnel load and personnel permissions determined by the personnel characteristics, the terminal status determined by the feedback characteristics or the mobile terminal, and the minimum confidence threshold constraint for high-risk tasks.

[0016] Furthermore, the feedback update module includes: The feedback credibility calculation unit, connected to the mobile terminal and the data access module, is used to calculate the consistency of location, time, image, OCR reading, touch record, and operation log based on the feedback features, and generate the execution feedback credibility based on the consistency of location, time, image, OCR reading, touch record, and operation log. The review marking unit, connected to the feedback credibility calculation unit, is used to mark the corresponding task as a task to be reviewed when the execution feedback credibility is lower than the credibility threshold. The parameter update unit, connected to the feedback credibility calculation unit, the difficulty profiling module, and the constraint dispatching module, is used to update the task difficulty parameter in the difficulty profiling module, the correction performance profile, and the dispatching parameter in the constraint dispatching module, using the execution feedback credibility as the sample weight, when the execution feedback credibility is not lower than the credibility threshold.

[0017] A method for granular task dispatching based on multi-dimensional performance profiling, applied to the aforementioned granular task dispatching system based on multi-dimensional performance profiling, includes: Step S1: The data access module receives customer billing data, meter data, task data, personnel data, regional data, environmental data, historical dispatch data, historical execution data, and terminal feedback data, and performs standardization processing to obtain task characteristics, personnel characteristics, regional characteristics, environmental characteristics, and feedback characteristics. Step S2: The graph granularity module constructs a multi-level copying and dispatching task graph containing multi-level task nodes and associated edges based on the task characteristics. Based on the multi-level copying and dispatching task graph, candidate task clusters are generated. The granularity benefit value is calculated according to spatial aggregation degree, route saving rate, time window consistency, estimated difficulty dispersion, risk concentration and uncertainty penalty value to obtain the task granularity result. Step S3: The difficulty profile module generates a task difficulty index based on the task granularity result and the task characteristics, and generates a corrected performance profile based on the task difficulty index, the personnel characteristics, the dispatch tendency probability determined by the historical dispatch data, and the historical execution data. Step S4: The ternary matching module uses the task cluster, executor, and dispatch area in the task granularity result as nodes, the candidate dispatch relationship as a hyperedge, and calculates the ternary matching score based on the task difficulty index, the correction efficiency profile, and the regional features. Step S5: The task dispatching module generates a task dispatching scheme containing a confidence lower bound on the completion probability based on the task granularity result, the task difficulty index, the correction efficiency profile, the ternary matching score, the personnel characteristics, the task characteristics, and the environmental characteristics, and sends the task dispatching scheme to the mobile terminal. Step S6: The feedback update module generates execution feedback credibility based on the feedback features, and uses the execution feedback credibility as the sample weight to update the task difficulty parameter in the difficulty profile module, the correction performance profile, and the dispatch parameter in the constraint dispatch module.

[0018] The beneficial effects of this invention are: This invention constructs a multi-layered task map and calculates granularity benefit values. It can dynamically determine task granularity based on task spatial distribution, task time window, risk concentration, and data uncertainty, thereby reducing duplicate visits and route waste caused by unreasonable task splitting or merging.

[0019] This invention generates a corrected performance profile by using task difficulty index, historical dispatch tendency probability, and historical execution data. This can reduce the impact of task difficulty and historical dispatch deviation on personnel performance evaluation and improve the accuracy of personnel performance assessment.

[0020] This invention constructs a ternary matching relationship with task clusters, personnel, and regions as nodes, which can comprehensively consider the correlation between task characteristics, personnel capabilities, and regional context, thereby improving the rationality of task assignment matching.

[0021] This invention generates a dispatch scheme with a confidence lower bound on the probability of completion, which can take into account the uncertainty of task completion results during the dispatch process and improve the stability of high-risk task allocation.

[0022] This invention generates execution feedback credibility based on feedback characteristics, and uses the execution feedback credibility as sample weights to update task difficulty parameters, correct performance profiles, and dispatch parameters. This reduces the impact of abnormal feedback on system updates and improves the traceability and continuous optimization capabilities of the dispatch process. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the granular task dispatching system based on multi-dimensional performance profiling in this invention. Figure 2 This is a flowchart of the steps of the task granularity dispatching method based on multi-dimensional performance profile in this invention.

[0024] Attached reference numerals: 1. Data access module; 11. Data partitioning unit; 12. Standardization unit; 2. Graph granularity module; 21. Node construction unit; 22. Edge construction unit; 23. Region verification unit; 24. Candidate cluster generation unit; 25. Revenue calculation unit; 26. Granularity decision unit; 3. Difficulty profiling module; 31. Task object determination unit; 32. Difficulty index extraction unit; 33. Difficulty calculation unit; 34. Baseline probability generation unit; 35. Deviation correction unit; 36. Profiling generation unit; 4. Triad matching module; 41. Hypergraph construction unit; 42. Matching calculation unit; 5. Constraint dispatching module; 51. Completion probability prediction unit; 52. Uncertainty calculation unit; 53. Dispatch decision unit; 6. Feedback update module; 61. Feedback credibility calculation unit; 62. Verification marking unit; 63. Parameter update unit. Detailed Implementation

[0025] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0026] Example 1, as Figure 1 As shown, this embodiment provides a granular dispatching system for meter reading and collection tasks based on multi-dimensional performance profiling. This system is applied to remote meter reading failure verification, overdue payment follow-up, abnormal electricity consumption review, and high-risk customer visits within the power supply business network. The system includes a server and a mobile terminal communicating with the server.

[0027] The server includes a data access module 1, a graph granularity module 2, a difficulty profiling module 3, a ternary matching module 4, a constraint dispatching module 5, and a feedback update module 6.

[0028] This embodiment mainly illustrates the overall deployment of the system, the collaborative relationship between modules, key thresholds, and the closed-loop working principle; the specific implementation methods of data partitioning, standardized feature table, task graph construction, granular calculation, and correction performance profile generation can adopt the method in embodiment two; the specific implementation methods of ternary hypergraph matching, uncertainty constraint dispatching, execution feedback verification, and online updating can adopt the method in embodiment three.

[0029] Among them, the data access module 1 is used to receive customer billing data, meter data, task data, personnel data, regional data, environmental data, historical dispatch data, historical execution data and terminal feedback data, and standardize them to obtain task characteristics, personnel characteristics, regional characteristics, environmental characteristics and feedback characteristics.

[0030] The graph granularity module 2 connects to the data access module 1 and is used to construct a multi-level task graph containing multi-level task nodes and associated edges based on task characteristics, generate candidate task clusters, and calculate granularity benefit values ​​according to spatial aggregation, route saving, time window consistency, difficulty dispersion, risk concentration and uncertainty penalty to obtain task granularity results.

[0031] The difficulty profile module 3 connects the graph granularity module 2 and the data access module 1. It is used to generate a task difficulty index based on the task granularity results and task characteristics, and to generate a correction performance profile based on the task difficulty index, personnel characteristics, the order dispatch tendency probability determined by historical order dispatch data, and historical execution data.

[0032] The ternary matching module 4 connects the graph granularity module 2, the difficulty profiling module 3, and the data access module 1. It is used to calculate the ternary matching score with task clusters, executors, and dispatch areas as nodes and candidate dispatch relationships as hyperedges. The constraint dispatch module 5 connects the ternary matching module 4 and the mobile terminal. It is used to generate a dispatch plan with a confidence lower bound on the completion probability based on the task granularity results, task difficulty index, correction efficiency profile, ternary matching score, personnel characteristics, task characteristics, and environmental characteristics, and send it to the mobile terminal.

[0033] The feedback update module 6 connects to the mobile terminal, the data access module 1, and the constraint dispatch module 5. It is used to generate execution feedback credibility based on feedback characteristics, and to update the task difficulty parameters, correction performance profile, and dispatch parameters using the execution feedback credibility as sample weights.

[0034] In this embodiment, the system is deployed in an environment consisting of a cloud server, an edge server, a business database, a graph database, a time-series database, an object storage server, and a mobile work terminal. The cloud server is used to perform task graph construction, task granularity calculation, task difficulty model training, personnel performance profile updates, ternary hypergraph matching, and uncertainty-constrained task assignment. The edge server is deployed in branch offices or regional grid centers to cache local task data, personnel status data, map data, and task assignment rules. The mobile work terminal is used to receive task assignments, perform navigation, collect positioning trajectories, take photos of the site, take pictures of meters, identify meter readings, record contact information, and upload execution feedback.

[0035] Specific hardware selection options include: The cloud servers utilize dual Intel Xeon Silver 4314 processors, 128GB of RAM, 2TB NVMe SSDs, and an Ubuntu Server 22.04 operating system; the edge servers employ 8-core x86 processors, 32GB of RAM, 1TB SSDs, and a Linux operating system; mobile work terminals use Android systems and are equipped with 4G / 5G communication modules, BeiDou / GPS positioning modules, cameras with a resolution of 12 megapixels or higher, NFC modules, and local encrypted storage modules. The relational database uses PostgreSQL to store basic customer, billing, task, and personnel data; the graph database uses Neo4j to store multi-layered toll collection task maps and task-personnel-region ternary hypergraphs; the time-series database uses TimescaleDB to store location trajectories, task status changes, and terminal operation logs; object storage employs an S3-compatible object storage system to save on-site photos, meter images, and electronic vouchers; and the message queue uses Kafka to transmit task events, dispatch events, and feedback events.

[0036] Data access module 1 accesses data from the marketing business system, electricity consumption information collection system, mobile operation terminal system, customer service system, and map service system. The accessed data includes at least customer billing data, meter data, task data, personnel data, regional data, environmental data, historical order dispatch data, historical execution data, and terminal feedback data. Data access module 1 partitions the data according to data source identifier, business primary key, and timestamp, and performs data cleaning, deduplication, unified encoding, missing data marking, anomaly marking, and normalization. For data with the same business primary key, the same data source identifier, and a timestamp difference less than a preset repetition time window, the system retains the data record with the later update time and higher field completeness. In this embodiment, the preset repetition time window is 30 seconds. Through the above processing, multi-source heterogeneous data is transformed into unified task characteristics, personnel characteristics, regional characteristics, environmental characteristics, and feedback characteristics, providing a data foundation for subsequent map construction, granular calculation, profile generation, and order dispatch decision-making.

[0037] The graph granularity module 2 constructs a multi-layered meter reading and collection task graph based on task address, meter location, user ID, building information, grid information, and route information. Multi-level task nodes include household-level, meter box-level, building-level, community-level, grid-level, and route segment-level task nodes; associated edges include spatial adjacency edges, task time window edges, task type edges, customer reach edges, historical execution edges, and risk-related edges. The graph granularity module 2 verifies the regional affiliation of task nodes based on regional characteristics, ensuring that each task node belongs to only one dispatch region within the same dispatch region level determined by regional characteristics, thereby preventing the same task from repeatedly entering multiple candidate task clusters within the same dispatch region level.

[0038] The granularity module 2 generates candidate task clusters in the multi-layered task map and evaluates these clusters at different granularity levels. Evaluation factors include spatial aggregation, route saving rate, time window consistency, task type consistency, dispersion of estimated difficulty determined by task characteristics, risk concentration, and uncertainty penalty value. In this embodiment, the merging threshold is 0.68, the splitting threshold is 0.35, and the review threshold is 0.75. When the granularity benefit value of a candidate task cluster is greater than or equal to 0.68, the system generates a merged task cluster; when the granularity benefit value is less than or equal to 0.35, the system splits the candidate task cluster into finer-grained tasks; when the risk concentration of a candidate task cluster is greater than or equal to 0.75, the system sets the candidate task cluster as a review task or a two-person execution task. Therefore, the system no longer assigns tasks based on individual households, buildings, or grids. Instead, it dynamically determines task splitting, merging, or upgrade review strategies based on task data status, reducing duplicate routes and invalid task assignments, and lowering the probability that high-risk tasks cannot be effectively completed after being excessively merged.

[0039] The Difficulty Profile Module 3 generates a task difficulty index based on task granularity results and task characteristics. Evaluation factors for the task difficulty index include the age of outstanding payments, the amount of outstanding payments, difficulty in reaching customers, historical failure to fulfill commitments, degree of meter anomalies, geographical dispersion, limitations of the execution window, risk of historical complaints or refusal to visit, and the degree of data uncertainty. When the task to be evaluated is a task cluster, the Difficulty Profile Module 3 generates the task difficulty index for the task cluster based on the task difficulty index of each task within the cluster and the importance weight determined by the amount of outstanding payments, task risk level, task deadline, and meter anomaly markers in the task characteristics. Through the task difficulty index, the system can avoid directly interpreting a high completion rate for low-difficulty tasks as strong personnel ability, and also avoid directly interpreting the failure of high-difficulty tasks as weak personnel ability.

[0040] The Difficulty Profile Module 3 further generates a corrected performance profile based on the task difficulty index, personnel characteristics, the dispatch tendency probability determined by historical dispatch data, and historical execution data. The corrected performance profile includes task type capability, regional familiarity, customer outreach capability, abnormal meter verification capability, evidence compliance capability, timeliness stability, risk task acceptance capability, and load stability. The system first generates a baseline task success probability, then combines the dispatch tendency probability determined by historical dispatch data and historical execution data to generate a corrected contribution value for the executor. Finally, based on the corrected contribution value, it generates a corrected performance profile for the executor under different task types and regional contexts. Through this process, the system can reduce the impact of differences in task difficulty, regional differences, and historical dispatch deviations on personnel capability evaluation, making the personnel profile closer to their actual execution capabilities.

[0041] The ternary matching module 4 constructs a task-personnel-region ternary hypergraph. In this hypergraph, task clusters from the task granularity results serve as task cluster nodes, executors corresponding to personnel features serve as personnel nodes, and non-overlapping dispatch regions determined by region features at the same dispatch region level serve as region nodes. Candidate dispatch relationships serve as hyperedges connecting a task cluster node, a personnel node, and a region node. The ternary matching module 4 calculates the ternary matching score based on task difficulty index, correction performance profile, region features, personnel load, route adaptability, time window matching degree, and historical complaint risk. This ternary matching method can simultaneously express the high-order coupling relationship between task clusters, executors, and region context, solving the problem that ordinary task-personnel binary matching cannot express region familiarity, route adaptability, region risk, and task cluster structure.

[0042] The task dispatch module 5 generates a task dispatch plan with a confidence lower bound on the completion probability based on task granularity results, task difficulty index, corrected performance profile, ternary matching score, personnel characteristics, task characteristics, and environmental characteristics. The confidence lower bound on the completion probability is determined based on the average completion probability and the prediction standard deviation. The average completion probability is determined by the ternary matching score, corrected performance profile, task difficulty index, and customer reach probability. The prediction standard deviation is determined by the number of historical samples, the degree of data missingness, risk concentration, and the volatility of the completion probability prediction model. In this embodiment, the confidence coefficient for ordinary tasks is 1.0, the confidence coefficient for high-risk tasks is 1.6, and the minimum confidence threshold for high-risk tasks is 0.72. The task dispatch decision also satisfies constraints such as task time window, personnel load, personnel permissions, terminal status, and the minimum confidence threshold for high-risk tasks. Specifically, the threshold for the estimated daily task duration for personnel is 7.5 hours, the upper limit for personnel load is 90% of the available working hours for the day, and tasks with high real-time requirements are not assigned when a terminal is offline for more than 10 minutes and has not cached task packages. By introducing a lower bound on the completion probability confidence level, the system can avoid assigning high-risk tasks to personnel with large fluctuations in prediction results based solely on the average completion probability, thereby improving the stability of high-risk task assignment.

[0043] After the dispatch module 5 generates a dispatch plan, it sends the plan to the mobile terminal via the dispatch output and terminal interaction function. The mobile terminal displays the task route, task priority, task time window, evidence collection requirements, and anomaly reporting entry. Upon receiving the dispatch task, the mobile terminal executes map navigation, task arrival confirmation, on-site image collection, meter image collection, OCR reading recognition, call and SMS record association, user signature, and anomaly reporting. In weak network or offline scenarios, the mobile terminal caches dispatched tasks and evidence collection results. After network recovery, the mobile terminal synchronizes the cached data to the cloud based on the task number, personnel number, terminal number, and timestamp. The cloud verifies duplicate submissions, conflicting states, and timestamp anomalies, using data with higher execution feedback credibility as valid records, thus ensuring continuous task execution and traceable feedback data under weak network or offline conditions.

[0044] The feedback update module 6 generates execution feedback credibility based on feedback characteristics and uses this credibility as a sample weight to update the task difficulty parameter, correction performance profile, and dispatch parameter. The feedback update module 6 generates execution feedback credibility based on location consistency, time consistency, image consistency, OCR reading consistency, reach record consistency, and operation log consistency. In this embodiment, the credibility threshold is 0.70. When the execution feedback credibility is below 0.70, the feedback update module 6 marks the corresponding task as a task awaiting review; when the execution feedback credibility is not below 0.70, the feedback update module 6 uses the execution feedback credibility as a sample weight to update the task difficulty parameter in the difficulty profile module 3, the correction performance profile, and the dispatch parameter in the constraint dispatch module 5. If the location consistency is below 0.40 and the image consistency is below 0.50, even if the mobile terminal submits a task completion status, the system will still mark the task as suspected invalid completion and trigger a review. Through this processing, low-credibility feedback will not directly affect the personnel profile and dispatch model as a positive sample, reducing issues such as false completion, invalid visits, inconsistent evidence, and duplicate reviews.

[0045] The working principle of this embodiment is as follows: The system first accesses data from the marketing business system, electricity information collection system, mobile operation terminal system, customer service system, and map service system, and forms task characteristics, personnel characteristics, regional characteristics, environmental characteristics, and feedback characteristics. Subsequently, the graph granularity module 2 constructs a multi-layered meter reading and collection task graph based on task address, meter location, user number, building information, grid information, and route information. Candidate task clusters are formed in the multi-layered meter reading and collection task graph, and the task splitting, merging, or escalation review strategy is determined based on the granularity benefit value. Then, the difficulty profile module 3 generates a task difficulty index based on the overdue payment age, overdue amount, difficulty in reaching customers, degree of meter anomaly, regional complexity, time window limitation, and data uncertainty. Then, based on the task difficulty index, the task baseline success probability, the historical dispatch tendency probability, and the historical execution results, a correction effectiveness profile of the executor is generated under different task types and regional contexts. Then, the ternary matching module 4 uses the task cluster, the executor, and the dispatch area as nodes, and the candidate dispatch relationship as a hyperedge, to calculate the ternary matching score of the executor executing a specific task cluster in a specific area. Then, the constraint dispatch module 5 combines the task granularity results, task difficulty index, correction efficiency profile, ternary matching score, personnel status and task time window to generate a dispatch plan; Finally, the mobile terminal transmits its location trajectory, on-site images, meter images, OCR readings, call and SMS records, user signature information, and terminal operation logs. The feedback update module 6 calculates the credibility of the execution feedback and updates the task difficulty index, multi-dimensional performance profiles of executors, regional risk profiles, and dispatch model parameters based on the task execution results and the credibility of the execution feedback. The entire process forms a closed loop following the sequence of "task granularity determination—task difficulty calculation—personnel performance correction—ternary hypergraph matching—uncertainty constraint dispatch—feedback credibility verification—online update."

[0046] In specific power supply grid tests, test data included 90 consecutive days of customer billing data, meter data, task data, personnel data, regional data, environmental data, historical order dispatch data, historical execution data, and terminal feedback data. The test scope included 15 power supply grids, 120 personnel, approximately 8,000 historical tasks, and 1,200 pending tasks. Test terminals included 50 mobile work terminals. Test scenarios included three types of conditions: normal network, weak network, and offline caching followed by synchronization. The weak network test conditions were: uplink bandwidth not exceeding 512kbps, average latency not less than 300ms, and packet loss rate of 5% to 15%.

[0047] During testing, the system first assigns each user meter as a household-level task node, and then progressively forms candidate task clusters at the meter box level, building level, community level, and grid level. For ordinary overdue payment contact tasks, if the task locations are concentrated, the suitable contact time for customers is similar, the task type is consistent, and the risk concentration is low, the system merges multiple household-level tasks into a community-level or route-segment-level task cluster. For tasks with high overdue payments, historical complaints, and meter anomalies, the system splits these tasks into household-level review tasks or sets them as tasks to be executed by two people. Subsequently, the system selects executors based on the task difficulty index and correction efficiency profile; ordinary overdue payment contact task clusters are preferentially assigned to executors with high correction efficiency and familiarity with the target area, while meter anomaly review tasks are preferentially assigned to executors with high meter anomaly verification capabilities or those capable of review.

[0048] After completing a task, the personnel upload their location trajectory, on-site photos, meter images, OCR readings, call logs, and operation logs via their mobile terminals. If the location trajectory enters the task's geofence, the dwell time meets the requirements, the on-site photos match the meter information, and the OCR readings are consistent with historical readings, the feedback update module 6 identifies the task as a high-confidence sample and uses it to update the personnel profile and task difficulty model. If the location does not enter the geofence and the photos do not match the task address, the feedback update module 6 marks the task as potentially invalid and triggers a review. Thus, this embodiment can unify task granularity dynamic modeling, task difficulty correction, personnel performance evaluation, high-order matching, robust dispatching, and execution feedback credibility verification into a single closed-loop process, improving the accuracy, stability, and traceability of task execution resource allocation.

[0049] Example 2: Based on the system architecture of Example 1, this example describes the specific implementation of the data access module 1, the graph granularity module 2, and the difficulty profiling module 3. In this example, the data access module 1 includes a data partitioning unit 11 and a standardization unit 12; the graph granularity module 2 includes a node construction unit 21, an edge construction unit 22, a region verification unit 23, a candidate cluster generation unit 24, a revenue calculation unit 25, and a granularity decision-making unit 26; the difficulty profiling module 3 includes a task object determination unit 31, a difficulty index extraction unit 32, a difficulty calculation unit 33, a baseline probability generation unit 34, a deviation correction unit 35, and a profiling generation unit 36.

[0050] Data partitioning unit 11 divides customer billing data, meter data, task data, personnel data, regional data, environmental data, historical dispatch data, historical execution data, and terminal feedback data into mutually unique data partitions according to data source identifier, business primary key, and timestamp. Furthermore, the same original data is written to only one data partition after deduplication.

[0051] The above data includes customer billing data, which comprises basic customer data and billing data. Basic customer data includes customer ID, customer type, address, building, unit, floor, meter box number, grid affiliation, and contact information validity. Billing data includes billing period, outstanding amount, outstanding aging, number of consecutive outstanding payments, historical payment dates, promised payment records, and promised payment default records. Meter data includes meter number, meter location, historical readings, abnormal readings, remote data acquisition failure records, estimated reading flags, and meter alarm records. Task data includes task ID, task type, task window, task status, and task risk flags for tasks such as meter reading, anomaly verification, outstanding payment contact, and review. Personnel data includes personnel ID, organization affiliation, skill tags, executable task types, current task load, current location, terminal online status, historical completion records, historical evidence qualification records, and historical complaint records. Regional data includes grid boundaries, community boundaries, building locations, road connectivity information, regional historical completion rate, regional historical complaint rate, and regional route complexity. Environmental data includes weather, holidays, traffic congestion levels, and task execution dates and times. Terminal feedback data includes location trajectory, arrival time, dwell time, on-site photos, meter photos, OCR readings, call and SMS records, user confirmation information, execution notes, and terminal operation logs.

[0052] Standardization unit 12 connects to data partitioning unit 11 and is used to clean, deduplicate, uniformly encode, mark missing values, mark outliers, and normalize the data in each data partition to obtain task features, personnel features, regional features, environmental features, and feedback features. For numerical fields, standardization unit 12 uses maximum-minimum normalization or quantile normalization; for categorical fields, standardization unit 12 uses one-hot encoding or label encoding; for address and location fields, standardization unit 12 uses geocoding to generate latitude, longitude, and grid numbers; for sensitive fields, standardization unit 12 uses anonymization or access control methods. Through this processing, data from different sources, in different formats, and with different time granularities are converted into unified features, reducing the impact of inconsistent multi-source data on subsequent map construction and model calculation.

[0053] Standardization Unit 12 generates task feature tables, personnel feature tables, regional feature tables, and feedback feature tables, and forms environmental feature records. The task feature table includes task number, task type, customer number, meter number, location, task time window, overdue payment age, overdue amount, meter anomaly flag, customer contact record, and task risk flag. The personnel feature table includes personnel number, skill tag, current location, current load, regional familiarity, task type capability, evidence qualification rate, complaint rate, and terminal online status. The regional feature table includes regional number, regional route complexity, regional historical contact success rate, regional historical complaint rate, and regional task density. The feedback feature table includes task number, personnel number, location consistency, time consistency, image consistency, OCR consistency, contact record consistency, and operation log consistency. These feature tables serve as inputs for subsequent task map construction, task granularity calculation, performance profile generation, and dispatch model training. The environmental feature record describes the impact of weather, holidays, traffic congestion, and task execution time on task execution.

[0054] Node building unit 21 connects to standardization unit 12 and is used to construct multi-level meter reading and collection task graphs based on task characteristics, including household-level task nodes, meter box-level task nodes, building-level task nodes, community-level task nodes, grid-level task nodes, and route segment-level task nodes. Household-level task nodes correspond to the specific tasks of a single customer or a single meter; meter box-level task nodes are aggregated from multiple household-level tasks within the same meter box or adjacent meter boxes; building-level task nodes are aggregated from multiple meter box-level tasks within the same building; community-level task nodes are aggregated from multiple building-level tasks within the same community or park; grid-level task nodes are aggregated from multiple community-level tasks within the same grid; and route segment-level task nodes are aggregated from multiple task nodes arranged consecutively along the expected execution route. Through this multi-level task node structure, the system can switch task granularity between household, meter box, building, community, grid, and route segment levels, providing a structural basis for task merging, splitting, or verification.

[0055] Edge construction unit 22 connects to node construction unit 21 and is used to generate associated edges based on spatial adjacency, task time window, task type, customer reach, historical execution, and risk association. Spatial adjacency edges represent the distance or path cost between two task nodes; time window similarity edges represent the degree of overlap in the suitable execution periods of two task nodes; task type similarity edges represent whether two task nodes belong to the same or continuously executable task types; customer reach similarity edges represent whether the reach periods, answer probabilities, door opening probabilities, or historical performance behaviors of two customers are similar; historical execution association edges represent whether two task nodes have been frequently executed continuously or by the same person in the past; and risk transmission edges represent whether there is a correlation between complaints, refusal to visit, meter abnormalities, or high arrears risks in the same area, the same building, or the same customer group. The region verification unit 23 connects to the edge construction unit 22, and is used to verify the region affiliation of task nodes in the multi-layer dispatch task graph according to the region characteristics, so that each task node belongs to only one dispatch region under the same dispatch region level determined by the region characteristics, thereby avoiding the same task being repeatedly clustered or dispatched due to overlapping dispatch regions.

[0056] The candidate cluster generation unit 24 connects to the edge construction unit 22, and is used to generate candidate task clusters based on task nodes and associated edges in the multi-layered meter reading and collection task graph. Candidate task clusters can be household-level, meter box-level, building-level, community-level, grid-level, or route segment-level task clusters. When generating candidate task clusters, the candidate cluster generation unit 24 first filters task nodes with similar locations based on spatial adjacency edges, then determines whether a task is suitable for continuous execution based on time window similarity edges and task type similarity edges, and combines customer reach similarity edges, historical execution associated edges, and risk transmission edges to determine the internal association strength of the candidate task cluster. In this way, candidate task clusters are not simply generated according to administrative regions or fixed grids, but are jointly determined by the multi-dimensional association relationships between task nodes, which helps improve the rationality of task aggregation.

[0057] In the formulas of this embodiment and subsequent embodiments, i is fixed to represent the task sequence number, j is fixed to represent the executor sequence number, q is fixed to represent the task cluster sequence number, and r is fixed to represent the dispatch area sequence number. "Fixed" represents the granularity level sequence number, and "k" represents the task type sequence number. All input indicators used in the formulas are dimensionless quantities. For fields in the original data that have physical dimensions, including path distance, route length, amount, time duration, task duration, dwell time, number of times, sample size, and timestamp difference, the system first converts them into dimensionless scores within the 0-1 range using max-min normalization, quantile normalization, proportionalization, or saturation mapping methods before substituting them into the formulas for calculation.

[0058] For any primitive index with physical dimensions Standardized units can generate dimensionless indices in the following manner. : ; in, This represents a dimensionless normalized index; Indicates the original indicators; and These represent the minimum and maximum values ​​of the corresponding indicators within a preset statistical period, respectively. This represents a positive smoothing constant to prevent the denominator from being zero. For metrics such as frequency or sample size, the system can also use this. The saturation mapping method is converted into a dimensionless index, where This represents the reference scale constant for the corresponding indicator.

[0059] The logical normalization function is defined as: ; in, Used to map input values ​​to the interval between 0 and 1. The smoothing nonnegativity function is defined as: ; in, This is used to ensure that the output value is non-negative. , The inputs to ln(*), exp(*), tanh(*), and the square root function are all dimensionless quantities. Unless otherwise specified, the weight parameters in each formula are dimensionless non-negative numbers, the bias terms are dimensionless real numbers, and the smoothing constant and numerical stability constant are dimensionless positive numbers.

[0060] The revenue calculation unit 25 is connected to the candidate cluster generation unit 24 and is used to calculate the granular revenue value of the candidate task clusters based on spatial aggregation degree, route saving rate, time window consistency degree, task type consistency degree, estimated difficulty dispersion determined by task characteristics, risk concentration degree, and uncertainty penalty value. To ensure that the granular revenue value simultaneously reflects positive aggregation revenue and negative risk penalty, the granular revenue value in this embodiment is calculated in the following manner: ; in, Indicates that the q-th candidate task cluster is in the th order. Granularity-level revenue value at different granularity levels; Indicates spatial aggregation degree; Indicates the route savings rate; Indicates time window consistency; Indicates the consistency of task types; This indicates the dispersion of the estimated difficulty. Indicates risk concentration; Indicates the uncertainty penalty value; This represents the granularity-based revenue bias term; to This indicates the weight of each granularity of the revenue calculation.

[0061] Among them, spatial aggregation degree The route saving rate is obtained by normalizing the path distance between task locations within a task cluster. The consistency of time windows is obtained by scaling up the differences in route length. Task type consistency is obtained from the overlap ratio of time windows. Difficulty dispersion is estimated based on task type compatibility score. Risk concentration and uncertainty penalty value All scores are dimensionless and range from 0 to 1. Therefore, the numerator, denominator, and logical normalization function input in the granularity profit value formula are all dimensionless, resulting in a unified formula dimension.

[0062] In this formula, the numerator represents the positive benefits brought by task aggregation, and the denominator represents the penalties brought by discrete difficulty, concentrated risk, and data uncertainty. This is used to suppress the marginal gain when spatial aggregation is too high, and to prevent highly concentrated tasks from being merged without limit; Used to smooth out route savings rates and reduce the sudden increase in the impact of abnormal route estimates on granular revenue values; This is designed to amplify the punitive effect in high-risk, concentrated scenarios. When risks such as complaints, refusal to visit, meter anomalies, or high arrears accumulate, the system is more inclined to split or review them. Therefore, granular reward values ​​can balance route efficiency and risk control, avoiding the excessive merging of high-risk tasks simply because they are geographically close.

[0063] Spatial aggregation is determined based on the average path distance between the locations of tasks within a task cluster; the more concentrated the tasks, the greater the spatial aggregation. Route saving rate is determined based on the difference between the estimated route length for merging the task cluster and the estimated route length for executing each task separately. Time window consistency is determined based on the overlap ratio of suitable execution periods for each task within a task cluster. Task type consistency is determined based on whether the task types within a task cluster are suitable for continuous execution. Estimated difficulty dispersion is determined based on the dispersion of estimated difficulty values ​​formed by the outstanding payment aging, outstanding amount, meter anomaly markers, customer contact records, and task risk markers for each task within a task cluster. Risk concentration is determined based on the concentration of complaint risk, refusal to visit risk, meter anomaly risk, and high outstanding payment risk within the task cluster. Uncertainty penalty value is determined based on the proportion of missing addresses, invalid contact information, location drift, insufficient historical data, and abnormal task status.

[0064] The granularity decision unit 26 is connected to the revenue calculation unit 25 and is used to generate task granularity results, including at least one of merged task clusters, split tasks, review tasks, and dual-person execution tasks, based on the comparison results of the granularity revenue value with the merging threshold, splitting threshold, and review threshold. In this embodiment, the merging threshold is 0.68, the splitting threshold is 0.35, and the review threshold is 0.75; in other embodiments, the merging threshold can be between 0.60 and 0.75, the splitting threshold can be between 0.25 and 0.40, and the review threshold can be between 0.70 and 0.85. The system can dynamically update the above thresholds based on the execution results of the most recent 30, 60, or 90 days. When the granularity revenue value is greater than the merging threshold, the candidate task cluster is merged into a task cluster of the corresponding granularity level; when the granularity revenue value is less than the splitting threshold, the candidate task cluster is split into tasks with finer granularity; when the risk concentration is greater than the review threshold, the candidate task cluster is set as a review task or a dual-person execution task. Through the above processing, the system does not assign tasks by household, building, or grid in a fixed manner, but automatically determines the task granularity based on the task data status. When the spatial aggregation, route saving rate, and time window consistency are high and the risk concentration is low, the system tends to merge tasks; when the task difficulty varies greatly, the risk concentration is high, or the data uncertainty is high, the system tends to split tasks or set up review tasks.

[0065] The task object determination unit 31, connected to the granularity decision-making unit 26, is used to determine the task object to be evaluated based on the task granularity results. The task object to be evaluated can be a single task or a cluster of tasks. The difficulty index extraction unit 32, connected to the standardization unit 12, is used to extract the following from task characteristics and regional characteristics: overdue payment age, overdue amount, difficulty in reaching customers, historical commitment default rate determined by historical execution data, meter anomaly degree, geographical dispersion, execution window limitation, historical complaint or refusal risk rate determined by task characteristics, regional characteristics, and historical execution data, and data uncertainty. The difficulty calculation unit 33, connected to both the task object determination unit 31 and the difficulty index extraction unit 32, is used to generate a task difficulty index for the task object to be evaluated based on the indicators extracted by the difficulty index extraction unit 32.

[0066] In this embodiment, the task difficulty index is calculated as follows: ; in, Indicates the first The task difficulty index for each task; This represents the normalized value of the outstanding payment aging. This represents the normalized value of the outstanding amount; Indicates the difficulty of reaching customers; Indicates the degree of meter abnormality; Indicates geographical dispersion; Indicates the execution window limitation; Indicates the degree of uncertainty in the data; Indicates the degree of breach of historical commitments; Indicates the level of risk of historical complaints or refusal to visit; Indicates the task difficulty bias; to This indicates the weighting of the difficulty calculation for each task. Among them, the normalized value of the overdue payment aging is... Normalized value of outstanding amount, derived from the number of days or billing periods in arrears. Geographic dispersion is derived from the amount owed. Derived from path distance, execution window is limited. The difficulty index is derived from the time window length and the urgency of the deadline. All the converted indicators are dimensionless scores. Similarly, customer reach difficulty, meter anomaly level, data uncertainty, historical commitment default rate, and historical complaint or refusal risk are also dimensionless scores ranging from 0 to 1. Therefore, the fractional term, square root term, and logical normalization function input in the task difficulty index formula are all dimensionless quantities.

[0067] In this formula, the fractional part is used to calculate the basic difficulty of the task, and the denominator contains... This indicates the certainty of the data. The higher the degree of uncertainty of the data, the smaller the denominator, and the easier it is for the task difficulty index to increase. This is used to represent the interactive impact when the risks of broken promises and refusal to respond to complaints coexist. Through this design, the task difficulty index can simultaneously reflect factors such as billing, outreach, metering, geography, time window, broken promises, refusal to respond to complaints, and data uncertainty, avoiding the distortion of task difficulty judgment caused by a single indicator.

[0068] When multiple tasks form a task cluster, the difficulty calculation unit 33 generates a task cluster difficulty index based on the task difficulty index of each task within the cluster and the importance weight determined by the outstanding amount, task risk level, task deadline, and meter anomaly marker in the task characteristics. The task cluster difficulty index is calculated as follows: ; ; in, This represents the task difficulty index of the q-th task cluster; This represents the set of tasks within the q-th task cluster; This represents the importance weight of the i-th task; This represents a smoothing constant to prevent the denominator from being zero; This indicates the weighting factor for the amount owed; This indicates the weighting factor for the task risk level; Indicates the weighting factor for the urgency of the task deadline; Indicates the weighting factor for meter anomalies; Indicates a weighting factor for data uncertainty; to This represents the weighting parameters for the importance of each item.

[0069] Among them, importance weight Dimensionless weights; weighting factor for outstanding amount. Task risk level weighting factor 1. Task deadline urgency weighting factor Meter anomaly weighting factor and data uncertainty weighting factor All scores are normalized dimensionless scores. Since the input to the exponential function is a dimensionless quantity, and the task difficulty index... Importance weights All are dimensionless quantities, task cluster difficulty index It is also a dimensionless quantity.

[0070] This formula uses an exponential function to increase the importance of tasks with high outstanding amounts, high task risk levels, near-deadline times, and meter anomalies. It also reduces the amplifying effect of uncertain data tasks on the difficulty of the task cluster by adjusting the denominator, thus preventing the average treatment of critical tasks within the cluster. Therefore, the task cluster difficulty index can more accurately reflect the impact of critical tasks within the cluster on the overall execution difficulty.

[0071] The baseline probability generation unit 34, connected to the difficulty calculation unit 33, generates a baseline success probability for the task based on the task difficulty index, the customer reach probability determined by task characteristics and historical execution data, the regional convenience determined by regional characteristics, the task urgency determined by task characteristics, and the environmental impact determined by environmental characteristics. The baseline success probability represents the basic likelihood of the task being effectively completed under the corresponding regional, time-of-day, and environmental conditions, without considering the capabilities of specific personnel. The customer reach probability is calculated based on the customer's historical call answering time, SMS delivery status, historical successful on-site visits, historical payment periods, and the validity of contact information. Regional convenience is calculated based on road connectivity, parking convenience, building density, historical average arrival time, and regional route complexity.

[0072] In this embodiment, the baseline success probability of the task is calculated as follows: ; in, This represents the baseline success probability of the i-th task; Indicates the probability of customer reach; Indicates regional convenience; Indicates the urgency of the task; Indicates the degree of environmental impact; This represents the baseline probability bias term; to This represents the parameters of the baseline probability model. Among them, customer reach probability... Regional convenience Task Difficulty Index urgency of the task and environmental impact All values ​​are dimensionless scores or probability values. Task urgency is derived from the remaining executable time, task deadline, and task window length, while environmental impact is derived from weather, traffic congestion, holidays, and execution time. Therefore, the numerator, denominator, and output of the task baseline success probability formula are all dimensionless quantities.

[0073] In this formula, customer reach probability and regional convenience are included as positive factors in the numerator, while task difficulty index, task urgency, and environmental impact are included as inhibitory factors in the denominator. This separates the influence of the task's own conditions and the external environment on the task outcome before evaluating the performers' capabilities. Through this process, subsequent personnel performance profiling will not simply attribute the success of low-difficulty tasks to personnel capability, nor will it simply attribute the failure of high-difficulty tasks to insufficient personnel capability.

[0074] The deviation correction unit 35 connects the baseline probability generation unit 34 and the standardization unit 12. It generates the correction contribution value for each executor based on the baseline success probability of the task, the assignment tendency probability determined by historical assignment data, and historical execution data. The execution result score is calculated based on factors such as whether the task was completed, whether effective contact was achieved, whether effective payment or verification results were generated, whether qualified evidence was uploaded, and whether a complaint was generated. The score ranges from 0 to 1. The historical assignment tendency probability represents the probability that a task will be assigned to a specific executor under a historical assignment strategy. It can be trained based on the executor's region, skill tags, current location, task type, and historical assignment rules.

[0075] In this embodiment, the correction contribution value is calculated as follows: ; in, This represents the correction contribution value of the j-th executor to the i-th task; Indicates the score of the execution result; Indicates the probability of historical order placement tendency; This represents the lower limit of the probability of order dispatch tendency; in this embodiment, it is set to 0.02. The value represents the stability constant, which is 0.005 in this embodiment and can be 0.001 to 0.01 in other embodiments; max(*) represents the maximum value function.

[0076] Among them, the execution result score Baseline success probability of the task Historical order placement probability Lower limit of order dispatch tendency probability and stability constant Both are dimensionless quantities. Therefore, the numerator and denominator of the fraction are dimensionless, the hyperbolic tangent function input is dimensionless, and the correction contribution value... The score is dimensionless.

[0077] This formula weakens the bias of historical task allocation tendencies in personnel evaluation through fractional methods and limits the impact of outliers on the correction contribution value through a hyperbolic tangent function, thus avoiding drastic fluctuations in the correction performance profile caused by extreme execution results. Therefore, the impact of historical task allocation bias, differences in task difficulty, and occasional outliers on personnel performance evaluation can be reduced.

[0078] The profile generation unit 36 ​​connects to the deviation correction unit 35 and is used to generate a correction performance profile based on the correction contribution value. The correction performance profile includes at least the following: task type capability, regional familiarity, customer outreach capability, abnormal meter verification capability, evidence compliance capability, timeliness stability, risk task acceptance capability, and load stability. For the correction performance value of the j-th executor under the k-th task type, the profile generation unit updates it as follows: ; in, This represents the corrected performance value of the j-th executor under the k-th task type after the update; This indicates the correction performance value before the update; This represents the image update step size, which is 0.15 in this embodiment, but can be 0.05 to 0.30 in other embodiments; This represents the set of task samples corresponding to the k-th task type; This indicates whether the i-th task belongs to the k-th task type; if it does, set it to 1, otherwise set it to 0. This represents the reliability of the execution feedback for the i-th task; The smoothing constant is denoted as 0.0001 in this embodiment. The correction efficiency value, profile update step size, task type indicator, execution feedback credibility, correction contribution value, and smoothing constant are all dimensionless quantities. This formula is essentially a weighted average update based on execution feedback credibility, maintaining the same dimensionless scoring caliber for the correction efficiency values ​​before and after the update.

[0079] This formula weights samples based on the reliability of feedback, ensuring that high-reliability samples have a greater impact on personnel profile updates and low-reliability samples have a smaller impact, thereby reducing the contamination of performance profiles by anomalous feedback. Through this update method, the true performance of personnel in different task types and regional contexts can be continuously corrected, improving the accuracy of subsequent task assignment matching.

[0080] The working principle of this embodiment is as follows: Data partitioning unit 11 first divides customer billing data, meter data, task data, personnel data, regional data, environmental data, historical dispatch data, historical execution data, and terminal feedback data from different sources into mutually exclusive data partitions. Standardization unit 12 then transforms the data in each data partition into task characteristics, personnel characteristics, regional characteristics, environmental characteristics, and feedback characteristics. Node construction unit 21 and edge construction unit 22 form a multi-layered collection and reminder task graph based on task characteristics. Regional verification unit 23 ensures that task nodes under the same dispatch region level have unique affiliation. Candidate cluster generation unit 24 generates candidate task clusters at different granularity levels based on the multi-layered collection and reminder task graph. Revenue calculation unit 25 calculates the results based on spatial... Aggregation, route saving, time window consistency, task type consistency, estimated difficulty dispersion, risk concentration, and uncertainty penalty generate granular benefit values. Based on these values, the granular decision-making unit 26 determines whether to merge, split, review, or have two people execute the task. The task object determination unit 31 then determines the task object to be evaluated based on the task granularity results. The difficulty index extraction unit 32 and the difficulty calculation unit 33 generate a task difficulty index or a task cluster difficulty index. The baseline probability generation unit 34 generates a task baseline success probability. The deviation correction unit 35 generates a correction contribution value based on the task baseline success probability, the dispatch tendency probability, and historical execution data. Based on this, the profile generation unit 36 ​​forms a profile of the executor's correction effectiveness under different task types and regional contexts.

[0081] Through the above implementation method of this embodiment, the data access module 1 can provide a data foundation that is consistent in scope, non-repeating in partitions, and traceable for subsequent calculations; the map granularity module 2 can set the task granularity as a computable variable, reducing route waste and task failure caused by fixed granularity; the difficulty profile module 3 can separate task difficulty, regional factors, and historical dispatch deviations from personnel performance, avoiding misjudging high completion rates of low-difficulty tasks as strong personnel ability, and also avoiding misjudging failure results of high-difficulty tasks as weak personnel ability, thereby improving the accuracy of personnel performance profile correction.

[0082] Example 3, based on Examples 1 and 2, describes the specific implementation of the ternary matching module 4, the constraint dispatching module 5, and the feedback update module 6. In this example, the ternary matching module 4 includes a hypergraph construction unit 41 and a matching calculation unit 42; the constraint dispatching module 5 includes a completion probability prediction unit 51, an uncertainty calculation unit 52, and a dispatching decision unit 53; the feedback update module 6 includes a feedback credibility calculation unit 61, a verification marking unit 62, and a parameter update unit 63.

[0083] In this embodiment, the ternary matching score, average completion probability, prediction standard deviation, lower bound of completion probability confidence, dispatch objective function, and execution feedback credibility are all based on the dimensionless rules in Embodiment 2. The original fields entering each formula, such as path distance, route cost, time length, task duration, sample size, number of complaints, number of rejections, dwell time, and reading change difference, are all first converted into dimensionless scores or dimensionless probability values ​​within the range of 0 to 1.

[0084] The hypergraph construction unit 41 connects the graph granularity module 2, the difficulty profiling module 3, and the data access module 1. It is used to take the task clusters in the task granularity results as task cluster nodes, the executors corresponding to the personnel characteristics as personnel nodes, the non-overlapping dispatch areas determined by the regional characteristics under the same dispatch area level as region nodes, and the candidate dispatch relationship as the hyperedge connecting a task cluster node, a personnel node, and a region node.

[0085] In this embodiment, the ternary hypergraph is represented as: ; in, This represents a task-personnel-region ternary hypermap; This represents a set of nodes, which includes task cluster nodes, personnel nodes, and region nodes. Let represent a set of hyperedges, where each hyperedge connects a task cluster node, a personnel node, and a region node. The candidate dispatch relationship between the q-th task cluster, the j-th executor, and the r-th dispatch region forms a hyperedge, which is used to express the higher-order matching relationship of an executor executing a specific task cluster in a specific dispatch region.

[0086] The characteristics of task cluster nodes include task type distribution, task cluster difficulty, task time window, risk concentration, data uncertainty, and task granularity level. The characteristics of personnel nodes include correction performance profile, current location, current load, evidence pass rate, complaint rate, and terminal online status. The characteristics of region nodes include region convenience, region risk, region historical completion rate, region route complexity, and region historical reach success rate. Hyperedge characteristics include personnel capability matching degree, region familiarity, route adaptability, task time window matching degree, task difficulty handling capacity, current load, historical evidence pass rate, and historical complaint risk.

[0087] The matching calculation unit 42 is connected to the hypergraph construction unit 41 and is used to calculate the ternary matching score based on the task difficulty index, correction efficiency profile, regional features, personnel load determined by personnel features, route adaptability determined by task features and regional features, time window matching degree determined by task features and personnel features, and historical complaint risk determined by historical execution data.

[0088] In this embodiment, the ternary matching model includes a node encoding layer, a hyperedge feature combination layer, a nonlinear scoring layer, and a parameter training layer. The node encoding layer encodes task cluster nodes, personnel nodes, and region nodes into dimensionless feature vectors. The hyperedge feature combination layer concatenates the features of task cluster nodes, personnel nodes, region nodes, and candidate dispatch relationships into a hyperedge feature vector. The nonlinear scoring layer outputs the ternary matching score based on the hyperedge feature vector. The parameter training layer trains matching score weights based on historical candidate dispatch relationships. Training samples are formed from historical candidate dispatch relationships, and training labels include whether the task was completed on time and effectively, whether the execution feedback was reliable, and whether a complaint was generated. When the task was completed on time and effectively, the execution feedback was reliable, and no complaint was generated, the training label is a positive sample; when the task was not completed on time, the execution feedback was unreliable, or a complaint was generated, the training label is a negative sample.

[0089] The ternary matching score is calculated as follows: ; in, This represents the ternary matching score of the j-th executor executing the q-th task cluster in the r-th dispatch area; Indicates the degree of matching between personnel capabilities; Indicates familiarity with the area; Indicates route suitability; Indicates the window matching degree of the task; Indicates risk suitability; Indicates the current load; Indicates the risk of historical complaints; This indicates the matching score bias term; to This indicates the matching score weight. Among them, the personnel ability matching degree... Familiarity with the region Route adaptability Task window matching degree Risk suitability Current load and historical complaint risks All scores are dimensionless. The current workload is derived from the ratio of the estimated duration of assigned tasks to the available working hours, and the historical complaint risk is derived from the number of historical complaints, the regional task load, and the severity of the complaints. Therefore, the numerator, denominator, and logical normalization function input in the ternary matching scoring formula are all dimensionless quantities.

[0090] In this formula, the numerator represents the positive matching factors between task clusters, executors, and dispatch areas, while the denominator represents the suppressive factors brought about by current load and historical complaint risks. Suppressing the amplification of abnormal personnel competence matching, through Expressing diminishing returns in regional familiarity, through By amplifying the high-load penalty, the problem of tasks being over-concentrated on a small number of personnel is reduced. Through this ternary matching model, the system can simultaneously consider the coupling relationship between task clusters, personnel, and dispatch areas, reducing matching biases caused by dispatching tasks based solely on distance or a single capability label.

[0091] Before calculating the training loss, the triple matching score is first numerically stabilized: ; in, This represents the ternary matching score after numerical stabilization processing; This represents a positive numerical stability constant, which can be taken as [value missing] in this embodiment. min(*) represents the function that takes the minimum value, and max(*) represents the function that takes the maximum value.

[0092] The training loss for the ternary matching model is: ; in, This represents the training loss of the ternary matching model; This represents the set of historical candidate order dispatch relationship samples; This indicates the number of historical candidate order dispatch relationship samples; This represents the training label for the three-element matching, and its value is either 0 or 1. Represents the sample smoothing constant; Represents the regularization coefficient; denoted as the weight of the h-th matching score. The training label for the ternary matching is 0 or 1. Both the ternary matching score and the numerically stabilized ternary matching score are dimensionless probability values. The cross-entropy loss is also dimensionless. The weight parameters in the regularization term are dimensionless, therefore the overall training loss is dimensionless.

[0093] This formula reduces the impact of sample size differences on training loss by normalizing the number of samples, avoids computational anomalies when the logarithmic function is close to 0 or 1 by numerical stabilization, and suppresses excessive weights by fractional regularization to avoid overfitting of the ternary matching model to historical order assignment bias.

[0094] The completion probability prediction unit 51 connects the ternary matching module 4 and the difficulty profiling module 3. It predicts the average completion probability based on the ternary matching score, the corrected performance profile, the task difficulty index, and the customer reach probability determined by task features and historical execution data. The completion probability prediction model employed by the completion probability prediction unit 51 includes an input layer, a feature interaction layer, a probability output layer, and a fluctuation estimation layer. The input layer receives the ternary matching score, the corrected performance profile, the task cluster difficulty, the customer reach probability, the task time window, and environmental features. The feature interaction layer calculates the nonlinear relationship between task clusters, executors, and customer reach. The probability output layer outputs the average completion probability. The fluctuation estimation layer outputs the degree of fluctuation in the completion probability prediction model.

[0095] After the task granularity results are determined, the risk concentration corresponding to the q-th task cluster is denoted as: ; in, This represents the risk concentration used by the q-th task cluster during the task dispatching phase; This represents the granularity level corresponding to the q-th task cluster after granularity decision-making; Indicates that the q-th candidate task cluster is in the th order. Risk concentration at each granular level. Through this connection, the risk concentration obtained in the granularity return calculation stage can serve as input for uncertainty calculation and order dispatch decision-making stages, avoiding a lack of data sources for the same technical meaning at different stages. Because... The dimensionless risk concentration score generated during the granularity return calculation stage. It directly inherits the dimensionless score, so the two have the same dimensions.

[0096] The average completion probability is calculated as follows: ; in, This represents the average probability of the j-th executor completing the q-th task cluster; This represents the corrected performance value of the j-th executor relative to the corresponding task type of the q-th task cluster; This represents the customer reach probability of the customer set corresponding to the q-th task cluster; This represents the regional convenience of the r-th dispatch area; This represents the average completion probability bias term; to This represents the weighting of the average completion probability calculation for each item. Among them, the ternary matching score, corrected efficiency value, customer reach probability, regional convenience, task cluster difficulty index, and current load are all dimensionless values. The task cluster difficulty index is obtained from the task cluster difficulty formula in Example 2, and the current load is obtained by proportionalizing the task duration. Therefore, the numerator, denominator, and output in the average completion probability formula are all dimensionless quantities.

[0097] This formula uses the numerator to express the positive effects of ternary matching, personnel profiling, customer outreach, and regional convenience, and the denominator to express the inhibitory effects of task cluster difficulty and current workload, making the completion probability prediction more consistent with the field task execution logic. In this way, the system can comprehensively determine whether a task cluster is suitable for execution by the assigned personnel in the assigned area before dispatching a task.

[0098] Uncertainty calculation unit 52 is connected to completion probability prediction unit 51. It is used to determine the prediction standard deviation based on the number of historical samples determined by historical execution data, the degree of data missing determined by task characteristics, the risk concentration determined by revenue calculation unit 25, and the volatility of the completion probability prediction model adopted by completion probability prediction unit 51. It also generates a lower bound of completion probability confidence based on the average completion probability and prediction standard deviation.

[0099] The standard deviation of prediction is calculated as follows: ; in, This represents the prediction standard deviation of the j-th executor performing the q-th task cluster; Indicates the degree of data missing; This represents the risk concentration corresponding to the q-th task cluster; This indicates the degree of fluctuation in the probability prediction model. The historical sample abundance score represents the combination of similar personnel, similar task clusters, and similar dispatch areas. This historical sample abundance score is obtained by saturating the number of historical samples, and can be specifically represented as: ; in, This represents the number of historical samples combining similar personnel, similar task clusters, and similar dispatch areas. This represents a reference value for the number of historical samples. The score is a dimensionless score in the range of 0 to 1. This represents the bias term of the predicted standard deviation; This indicates the weights used to calculate the standard deviation of the forecast.

[0100] In this formula, the historical sample abundance score is in the denominator; the higher the sample abundance, the lower the prediction standard deviation. Data missingness and risk concentration are expressed as squared terms in the numerator, amplifying the uncertainty in scenarios with poor data quality or risk clustering. The volatility of the completion probability prediction model reflects the model's predictive stability for the current candidate order assignment relationships. The prediction standard deviation is limited to the range of 0 to 1, matching the dimension of the average completion probability. Through this uncertainty estimation, the system can identify candidate order assignment relationships with high average completion probabilities but unstable prediction results. Specifically, data missingness, risk concentration, volatility of the completion probability prediction model, and historical sample abundance score are all dimensionless scores. The fractional input, squared term, square root term, and Softplus function input in the prediction standard deviation formula are all dimensionless, and the prediction standard deviation is limited to the range of 0 to 1, thus allowing for operations with the same dimension as the average completion probability.

[0101] The lower bound of the completion probability confidence is calculated as follows: ; in, This represents the lower confidence bound of the probability that the j-th executor will complete the q-th task cluster; The confidence coefficient is denoted as 1.0 to 2.0. For ordinary tasks, the confidence coefficient can be 1.0 to 1.3, and for high-risk tasks, it can be 1.5 to 2.0. The average completion probability, prediction standard deviation, and confidence coefficient are all dimensionless quantities, as is the fractional penalty term. Therefore, the lower bound of the completion probability confidence level has the same dimension as the average completion probability, both being dimensionless probability values ​​within the interval 0 to 1.

[0102] This formula limits the extreme amplification of uncertainty penalties through a fractional form, while retaining the downward adjustment effect of the forecast standard deviation on the completion probability. The larger the forecast standard deviation, the lower the confidence lower bound of the completion probability, thereby preventing high-risk tasks from being assigned to executors with volatile outcomes and improving the stability of high-risk task assignment.

[0103] The dispatch decision unit 53 is connected to the uncertainty calculation unit 52, and is used to generate a dispatch plan under the conditions of satisfying the task time window determined by task characteristics, personnel load and personnel permissions determined by personnel characteristics, terminal status determined by feedback characteristics or mobile terminals, and the minimum confidence threshold constraint for high-risk tasks. The dispatch objective function is as follows: ; in, Represents the set of candidate executors; Represents the set of tasks to be assigned; This indicates whether the q-th task cluster should be assigned to the j-th executor; it is set to 1 if assigned and 0 if not assigned. Indicates the value of the task; Indicates route cost; Indicates the risk of timeout; This indicates a complaint, refusal to visit, or compliance risk; This indicates a penalty for unfairness in personnel workload; to This indicates the weight of each task assignment objective. Among them, the task value... Route cost Risk of exceeding time limit Complaints, refusal to visit, or compliance risks and penalties for unfairness in personnel workload All scores are dimensionless scores after normalization.

[0104] Route cost is derived from route distance, estimated travel time, and traffic congestion level; timeout risk is derived from the deviation between the estimated arrival time and the task window; and personnel load fairness penalty is derived from the deviation between the current personnel load and the average load. Therefore, both the numerator and denominator of the dispatch objective function are dimensionless quantities, and there is no issue of directly adding different physical dimensions.

[0105] The objective function uses the numerator to represent the order dispatch revenue and the denominator to represent the order dispatch cost, aiming to achieve a balance between completion probability, task value, matching degree, route cost, timeout risk, compliance risk, and load fairness in the order dispatch scheme. All evaluation indicators have been normalized and their dimensions are matched.

[0106] Dispatch constraints include: each task cluster can be assigned to a maximum of one primary executor within the same dispatch cycle; the estimated task duration for each executor within the dispatch cycle must not exceed the working duration threshold; the lower bound of the completion probability confidence for high-risk tasks must not be lower than the minimum confidence threshold for high-risk tasks; review tasks are assigned to executors with review capability tags, or configured for dual-person execution; the proportion of high-risk tasks continuously undertaken by the same executor must not exceed the risk load threshold; the estimated arrival time of the task must be within the task execution time window; personnel permissions cover the data range and task type involved in the task; offline personnel are not assigned tasks with high real-time requirements unless the terminal has cached the task package for the corresponding dispatch area.

[0107] In this embodiment, the minimum confidence threshold for high-risk tasks is 0.72, while in other embodiments it can be 0.65 to 0.80; the risk load threshold is 0.40, while in other embodiments it can be 0.30 to 0.50; the working time threshold can be configured to 6 to 8 hours according to staff shifts. The dispatch model recalculates the dispatch plan every preset period; when events such as customers not being home, staff being offline, traffic congestion, on-site anomalies, or task timeout warnings occur, the system triggers partial redispatch.

[0108] In an optional implementation, the constraint dispatch module 5 employs a combination of rule constraints and reinforcement learning. Rule constraints first filter candidate dispatch relationships that do not meet the minimum confidence thresholds for permissions, task time windows, personnel load, terminal status, and high-risk tasks. The reinforcement learning strategy then selects dispatch, split, merge, postpone, transfer, or escalate review actions from the remaining candidate dispatch relationships. The reinforcement learning model includes a state encoding network, a policy network, and a value network. The state encoding network takes into account the task cluster set, personnel status, task difficulty, task time window, customer reach probability, regional risk, ternary matching score, and historical feedback. The policy network outputs the selection probabilities of dispatch, split, merge, postpone, transfer, and escalate review actions. The value network outputs a long-term revenue estimate of the current dispatch status. The reward function includes rewards for effective task completion, rewards for valid evidence, route cost penalties, timeout penalties, complaint penalties, invalid visit penalties, and unfair load penalties. This strategy can further improve multi-cycle dispatch revenue while meeting hard constraints.

[0109] The feedback credibility calculation unit 61 is connected to the mobile terminal and the data access module 1. It is used to calculate the consistency of location, time, image, OCR reading, touch record, and operation log based on the feedback characteristics, and generate the execution feedback credibility based on the consistency of location, time, image, OCR reading, touch record, and operation log.

[0110] Location consistency is used to determine whether the executor's location trajectory enters the task geofence; time consistency is used to determine whether the arrival time and dwell time meet the task requirements; image consistency is used to determine whether the on-site photos, meter photos and task address, meter number or historical images match; OCR reading consistency is used to determine whether the meter reading recognition results are consistent with historical reading changes and anomaly rules; contact record consistency is used to determine whether the telephone, SMS, electronic notification or on-site record is consistent with the task status; operation log consistency is used to determine whether the order of operations such as task reception, navigation, arrival, taking pictures, and submission is reasonable.

[0111] The reliability of the execution feedback is calculated as follows: ; in, This represents the reliability of the execution feedback for the i-th task; This indicates consistency in positioning; Indicates time consistency; Indicates image consistency; This indicates consistency in OCR readings; This indicates consistency in the reach records; This indicates consistency in the operation logs; This represents a score indicating the reasonableness of reading changes based on historical reading change rules. This indicates the feedback credibility bias term; to This represents the weighting of each feedback credibility calculation. Among them, location consistency, time consistency, image consistency, OCR reading consistency, reach record consistency, operation log consistency, and reading change reasonableness score are all dimensionless scores ranging from 0 to 1. Location consistency is derived from the matching degree between the location trajectory and the task geofence; time consistency is derived from arrival time and dwell time; and OCR reading consistency and reading change reasonableness score are derived from the proportion of reading differences. Therefore, the numerator, denominator, and output of the execution feedback credibility formula are all dimensionless quantities.

[0112] In this formula, the numerator represents the consistency contribution of multimodal feedback, and the denominator represents the penalty term when the OCR reading is inconsistent with the historical reading change rule; This is used to express the synergistic relationship between location consistency and image consistency. When the location trajectory enters the task geofence and the image matches the task address, the reliability of the execution feedback is improved; conversely, when the OCR reading deviates significantly from the historical reading change rules, the reliability of the execution feedback is reduced. In this way, the system can reduce the impact of false presences, image mismatches, abnormal readings, and abnormal operation logs on the judgment of task completion status.

[0113] The review marking unit 62 is connected to the feedback credibility calculation unit 61, and is used to mark the corresponding task as a task to be reviewed when the execution feedback credibility is lower than the credibility threshold. The parameter update unit 63 is connected to the feedback credibility calculation unit 61, the difficulty profile module 3, and the constraint dispatch module 5, and is used to update the task difficulty parameter, the correction performance profile, and the dispatch parameter in the constraint dispatch module 5, using the execution feedback credibility as the sample weight, when the execution feedback credibility is not lower than the credibility threshold.

[0114] In this embodiment, the confidence threshold is set to 0.70; in other embodiments, it can be set to 0.65 to 0.80. When the confidence level of the execution feedback is lower than the confidence threshold, the system marks the task as pending review and reduces the positive update weight of the task sample for the personnel profile. When the confidence level of the execution feedback is not lower than the confidence threshold and the task execution result is valid, the system uses the task sample as a high-confidence sample to update the personnel profile, task difficulty, and area profile. When the location consistency is lower than 0.40 and the image consistency is lower than 0.50, even if the executor submits a task completion status, the system will still mark the task as suspected invalid completion and trigger the review process.

[0115] The online update mechanism is as follows: When a task is successfully executed and the credibility of the execution feedback is higher than the credibility threshold, the system treats the task as a valid sample and updates the executor's corresponding task type ability, regional familiarity, evidence compliance ability, and timeliness stability based on the task difficulty and execution result. When a task fails but the credibility of the execution feedback is higher than the credibility threshold, the system updates the task difficulty and customer reach model based on the reason for failure; if the reason for failure is that the customer is not at home, the difficulty of customer reach is increased; if the reason for failure is that the contact information is invalid, the data uncertainty is increased; if the reason for failure is that the meter is abnormal, the degree of meter abnormality is increased; if the reason for failure is that the address is incorrect, the address data abnormality mark is increased. When the credibility of the execution feedback is lower than the credibility threshold, the system does not update the personnel profile for the task as a positive sample and adds the task to the review task pool. If the same executor has consecutive low credibility feedback, the system increases the review ratio of the executor's subsequent tasks and reduces the evidence compliance ability score. When the same dispatch area has consecutive customer reach failures, refusal to visit, complaints, or meter abnormalities within a preset period, the system updates the regional risk profile and increases the risk concentration and review probability of tasks in that area in subsequent dispatches. The system can perform online updates daily, weekly, or per dispatch cycle; for dispatch areas with a large sample size, a shorter update cycle is used; for dispatch areas with a small sample size, a longer update cycle is used and the prediction uncertainty is increased.

[0116] The working principle of this embodiment is as follows: The hypergraph construction unit 41 first constructs a task-person-region ternary hypergraph based on task granularity results, personnel characteristics, and region characteristics, and represents the candidate dispatch relationship as a hyperedge connecting task cluster nodes, personnel nodes, and region nodes; the matching calculation unit 42 calculates the ternary matching score based on task difficulty index, corrected performance profile, region characteristics, personnel load, route adaptability, time window matching degree, and historical complaint risk; the completion probability prediction unit 51 further uses the ternary matching score, corrected performance profile, task difficulty index, and customer reach probability to predict the average completion probability; the uncertainty calculation unit 52 calculates the ternary matching score based on historical samples. The prediction standard deviation is determined by the quantity, data missingness, risk concentration, and volatility of the completion probability prediction model, and the lower bound of the completion probability confidence is obtained. The dispatch decision unit 53 generates a dispatch plan under constraints such as task time window, personnel load, personnel permissions, terminal status, minimum confidence threshold for high-risk tasks, and risk load threshold. After the task is executed, the feedback credibility calculation unit 61 calculates the execution feedback credibility based on the multimodal feedback returned by the mobile terminal, the verification and marking unit 62 identifies low-credibility tasks, and the parameter update unit 63 updates the task difficulty parameter, corrected performance profile, and dispatch parameter by using high-credibility feedback as sample weights.

[0117] Through the above implementation method of this embodiment, the ternary matching module 4 can express the high-order coupling relationship between task clusters, executors and dispatch areas, and improve the accuracy of dispatch matching; the constraint dispatch module 5 can use the lower bound of the completion probability confidence to suppress high-risk tasks from being assigned to executors with large fluctuations in prediction results, and improve the stability of dispatch; the feedback update module 6 can filter effective samples based on the credibility of multimodal execution feedback, and avoid low-credibility feedback from polluting the task difficulty model, personnel profile and dispatch parameters, thereby forming a traceable, correctable and continuously optimized dispatch closed loop.

[0118] A method for granular task dispatching based on multi-dimensional performance profiling is proposed and applied to the above-mentioned granular task dispatching system based on multi-dimensional performance profiling. Figure 2 As shown, it includes: Step S1: Data access module 1 receives customer billing data, meter data, task data, personnel data, regional data, environmental data, historical dispatch data, historical execution data, and terminal feedback data, and performs standardization processing to obtain task characteristics, personnel characteristics, regional characteristics, environmental characteristics, and feedback characteristics. Step S2: The graph granularity module 2 constructs a multi-level copying and dispatching task graph containing multi-level task nodes and associated edges based on task characteristics. Based on the multi-level copying and dispatching task graph, candidate task clusters are generated. The granularity benefit value is calculated based on spatial aggregation degree, route saving rate, time window consistency, estimated difficulty dispersion, risk concentration and uncertainty penalty value to obtain the task granularity result. Step S3: The difficulty profile module 3 generates a task difficulty index based on the task granularity results and task characteristics, and generates a correction performance profile based on the task difficulty index, personnel characteristics, the order dispatch tendency probability determined by historical order dispatch data, and historical execution data. Step S4: The ternary matching module 4 uses the task cluster, executor and dispatch area in the task granularity result as nodes, the candidate dispatch relationship as the hyperedge, and calculates the ternary matching score based on the task difficulty index, correction efficiency profile and regional features. Step S5: The task dispatching module 5 generates a task dispatching plan containing a confidence lower bound on the completion probability based on the task granularity result, task difficulty index, correction efficiency profile, ternary matching score, personnel characteristics, task characteristics and environmental characteristics, and sends the task dispatching plan to the mobile terminal. In step S6, the feedback update module 6 generates the execution feedback credibility based on the feedback features, and uses the execution feedback credibility as the sample weight to update the task difficulty parameter, the correction performance profile, and the dispatch parameter in the constraint dispatch module 5.

[0119] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principle of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A task dispatching system with granularity based on multi-dimensional performance profiling, characterized in that, Includes a server and a mobile terminal communicating with the server, the server comprising: The data access module is used to receive and standardize customer billing data, meter data, task data, personnel data, regional data, environmental data, historical dispatch data, historical execution data, and terminal feedback data to obtain task characteristics, personnel characteristics, regional characteristics, environmental characteristics, and feedback characteristics. The graph granularity module, connected to the data access module, is used to construct a multi-level task graph containing multi-level task nodes and associated edges based on the task characteristics, generate candidate task clusters, and calculate granularity benefit values ​​according to spatial aggregation, route saving, time window consistency, difficulty dispersion, risk concentration and uncertainty penalty to obtain task granularity results. The difficulty profile module connects the graph granularity module and the data access module. It is used to generate a task difficulty index based on the task granularity result and the task characteristics, and to generate a corrected performance profile based on the task difficulty index, the personnel characteristics, the dispatch tendency probability determined by the historical dispatch data, and the historical execution data. The ternary matching module connects the graph granularity module, the difficulty profiling module, and the data access module. It is used to calculate the ternary matching score with the executor and the dispatch area as nodes and the candidate dispatch relationship as the hyperedge. The constraint dispatch module connects the ternary matching module and the mobile terminal. It is used to generate a dispatch plan with a confidence lower bound on the completion probability based on the task granularity result, the task difficulty index, the correction efficiency profile, the ternary matching score, personnel characteristics, task characteristics, and environmental characteristics, and send it to the mobile terminal. The feedback update module, which connects the mobile terminal, the data access module, and the constraint dispatch module, is used to generate execution feedback credibility based on the feedback features, and use it as a sample weight to update the task difficulty parameter, the correction performance profile, and the dispatch parameter.

2. The task granularity dispatching system based on multi-dimensional performance profiling as described in claim 1, characterized in that, The data access module includes: The data partitioning unit is used to divide customer billing data, meter data, task data, personnel data, regional data, environmental data, historical dispatch data, historical execution data, and terminal feedback data into mutually unique data partitions according to data source identifier, business primary key, and timestamp. The same original data is written to only one data partition after deduplication. The standardization unit, connected to the data partitioning unit, is used to perform unified encoding, missing data marking, anomaly marking, and normalization processing on the data in each data partition to obtain the task characteristics, personnel characteristics, regional characteristics, environmental characteristics, and feedback characteristics.

3. The task granularity dispatching system based on multi-dimensional performance profiling as described in claim 2, characterized in that, The spectral granularity module includes: A node construction unit, connected to the standardization unit, is used to construct household-level task nodes, meter box-level task nodes, building-level task nodes, community-level task nodes, grid-level task nodes, and route segment-level task nodes according to the task characteristics. An edge construction unit, connected to the node construction unit, is used to generate the associated edges based on spatial adjacency, task window relationship, task type relationship, customer reach relationship, historical execution relationship, and risk association relationship; The region verification unit, connected to the edge construction unit, is used to verify the region affiliation of task nodes in the multi-layer dispatch task graph based on the region characteristics, so that each task node belongs to only one dispatch region under the same dispatch region level determined by the region characteristics.

4. The task granularity dispatching system based on multi-dimensional performance profiling as described in claim 3, characterized in that, The spectral granularity module also includes: A candidate cluster generation unit, connected to the edge construction unit, is used to generate candidate task clusters based on the task nodes and associated edges in the multi-layer task graph. The revenue calculation unit, connected to the candidate cluster generation unit, is used to calculate the granular revenue value of the candidate task cluster based on spatial aggregation degree, route saving rate, time window consistency degree, task type consistency degree, estimated difficulty dispersion determined by the task characteristics, risk concentration degree, and uncertainty penalty value. The granularity decision unit, connected to the revenue calculation unit, is used to generate the task granularity result, including at least one of the following: merged task clusters, split tasks, review tasks, and dual-person execution tasks, based on the comparison results between the granularity revenue value and the merging threshold, splitting threshold, and review threshold.

5. The task granularity dispatching system based on multi-dimensional performance profiling according to claim 4, characterized in that, The difficulty profiling module includes: The task object determination unit, connected to the granularity decision unit, is used to determine the task object to be evaluated based on the task granularity result, wherein the task object to be evaluated is a single task or a cluster of tasks. The difficulty index extraction unit, connected to the standardization unit, is used to extract the following from the task characteristics and the regional characteristics: overdue payment age, overdue amount, difficulty in reaching customers, historical commitment default rate determined by the historical execution data, meter anomaly degree, geographical dispersion, execution window limitation, historical complaint or refusal risk rate determined by the task characteristics, the regional characteristics and the historical execution data, and data uncertainty. The difficulty calculation unit is connected to the task object determination unit and the difficulty index extraction unit, respectively, and is used to generate the task difficulty index of the task object to be evaluated based on the index extracted by the difficulty index extraction unit. When the task to be evaluated is a task cluster, the difficulty calculation unit generates the task difficulty index of the task cluster based on the task difficulty index of each task within the task cluster and the importance weight determined by the amount of outstanding payment, task risk level, task deadline, and meter anomaly marker in the task characteristics.

6. The task granularity dispatching system based on multi-dimensional performance profiling as described in claim 5, characterized in that, The difficulty profiling module also includes: A baseline probability generation unit, connected to the difficulty calculation unit, is used to generate a baseline success probability for a task based on the task difficulty index, the customer reach probability determined by the task characteristics and the historical execution data, the regional convenience determined by the regional characteristics, the task urgency determined by the task characteristics, and the environmental impact determined by the environmental characteristics. The deviation correction unit, connected to the baseline probability generation unit and the standardization unit, is used to generate the correction contribution value of the executor based on the task baseline success probability, the dispatch tendency probability determined by the historical dispatch data, and the historical execution data. The image generation unit, connected to the deviation correction unit, is used to generate the correction effectiveness image based on the correction contribution value.

7. The task granularity dispatching system based on multi-dimensional performance profiling as described in claim 6, characterized in that, The ternary matching module includes: The hypergraph construction unit connects the graph granularity module, the difficulty profiling module, and the data access module. It is used to take the task clusters in the task granularity results as task cluster nodes, the executors corresponding to the personnel characteristics as personnel nodes, the non-overlapping dispatch areas determined by the regional characteristics under the same dispatch area level as regional nodes, and the candidate dispatch relationship as a hyperedge connecting a task cluster node, a personnel node, and a regional node. The matching calculation unit, connected to the hypergraph construction unit, is used to calculate the ternary matching score based on the task difficulty index, the correction efficiency profile, the regional features, the personnel load determined by the personnel features, the route adaptability determined by the task features and the regional features, the time window matching degree determined by the task features and the personnel features, and the historical complaint risk determined by the historical execution data.

8. The task granularity dispatching system based on multi-dimensional performance profiling according to claim 7, characterized in that, The constraint dispatch module includes: The probability prediction unit is connected to the ternary matching module and the difficulty profiling module. It is used to predict the average completion probability based on the ternary matching score, the correction efficiency profile, the task difficulty index, and the customer reach probability determined by the task characteristics and the historical execution data. An uncertainty calculation unit, connected to the completion probability prediction unit, is used to determine the prediction standard deviation based on the number of historical samples determined by the historical execution data, the degree of data missing determined by the task characteristics, the risk concentration determined by the revenue calculation unit, and the volatility of the completion probability prediction model used by the completion probability prediction unit, and to generate the completion probability confidence lower bound based on the average completion probability and the prediction standard deviation. The dispatch decision unit, connected to the uncertainty calculation unit, is used to generate the dispatch scheme under the conditions of satisfying the task time window determined by the task characteristics, the personnel load and personnel permissions determined by the personnel characteristics, the terminal status determined by the feedback characteristics or the mobile terminal, and the minimum confidence threshold constraint for high-risk tasks.

9. The task granularity dispatching system based on multi-dimensional performance profiling as described in claim 8, characterized in that, The feedback update module includes: The feedback credibility calculation unit, connected to the mobile terminal and the data access module, is used to calculate the consistency of location, time, image, OCR reading, touch record, and operation log based on the feedback features, and generate the execution feedback credibility based on the consistency of location, time, image, OCR reading, touch record, and operation log. The review marking unit, connected to the feedback credibility calculation unit, is used to mark the corresponding task as a task to be reviewed when the execution feedback credibility is lower than the credibility threshold. The parameter update unit, connected to the feedback credibility calculation unit, the difficulty profiling module, and the constraint dispatching module, is used to update the task difficulty parameter in the difficulty profiling module, the correction performance profile, and the dispatching parameter in the constraint dispatching module, using the execution feedback credibility as the sample weight, when the execution feedback credibility is not lower than the credibility threshold.

10. A method for granular task dispatching based on multi-dimensional performance profiles, applied to the granular task dispatching system based on multi-dimensional performance profiles as described in any one of claims 1 to 9, characterized in that, include: Step S1: The data access module receives customer billing data, meter data, task data, personnel data, regional data, environmental data, historical dispatch data, historical execution data, and terminal feedback data, and performs standardization processing to obtain task characteristics, personnel characteristics, regional characteristics, environmental characteristics, and feedback characteristics. Step S2: The graph granularity module constructs a multi-level copying and dispatching task graph containing multi-level task nodes and associated edges based on the task characteristics. Based on the multi-level copying and dispatching task graph, candidate task clusters are generated. The granularity benefit value is calculated according to spatial aggregation degree, route saving rate, time window consistency, estimated difficulty dispersion, risk concentration and uncertainty penalty value to obtain the task granularity result. Step S3: The difficulty profile module generates a task difficulty index based on the task granularity result and the task characteristics, and generates a corrected performance profile based on the task difficulty index, the personnel characteristics, the dispatch tendency probability determined by the historical dispatch data, and the historical execution data. Step S4: The ternary matching module uses the task cluster, executor, and dispatch area in the task granularity result as nodes, the candidate dispatch relationship as a hyperedge, and calculates the ternary matching score based on the task difficulty index, the correction efficiency profile, and the regional features. Step S5: The task dispatching module generates a task dispatching scheme containing a confidence lower bound on the completion probability based on the task granularity result, the task difficulty index, the correction efficiency profile, the ternary matching score, the personnel characteristics, the task characteristics, and the environmental characteristics, and sends the task dispatching scheme to the mobile terminal. Step S6: The feedback update module generates execution feedback credibility based on the feedback features, and uses the execution feedback credibility as the sample weight to update the task difficulty parameter in the difficulty profile module, the correction performance profile, and the dispatch parameter in the constraint dispatch module.

Images