A multi-constrained intelligent environmental sanitation full-scene operation intelligent planning method

Abstract

The application provides a multi-constraint intelligent planning method for intelligent environmental sanitation full-scene operation, and is applied to the technical field of data processing. In the application, multi-source data such as GIS geography, task demand, device state and constraint configuration parameters are collected around intelligent environmental sanitation full-scene operation planning, and structured data sets are generated after standardization processing. Then, a multi-constraint directed association graph is generated by constructing a three-element association node and a sliding time window aggregation feature in layers. With the aid of a collaborative optimization graph neural network modeling, constraint feature embedding vectors are extracted, introduced into a hierarchical planning engine, and after dynamic weight adjustment, scene adaptation calibration and conflict processing, a planning basic model is formed. Finally, a dynamic operation planning scheme and execution details that adapt to the full scene are output by combining real-time data dynamic deduction and multi-objective optimization.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a multi-constraint intelligent planning method for smart sanitation operations across all scenarios. Background Technology

[0002] With the acceleration of urbanization and the upgrading of environmental sanitation needs, traditional sanitation operation mode is no longer suitable for the complex and ever-changing urban governance scenarios. Smart sanitation, empowered by technologies such as the Internet of Things, artificial intelligence, and big data, has become the core direction for improving sanitation operation efficiency and optimizing resource allocation.

[0003] However, existing intelligent sanitation operation planning technologies still have many problems that urgently need to be solved, as follows:

[0004] The sanitation operations involve a variety of information sources, including GIS geographic data, IoT equipment status data, personnel configuration files, and environmental meteorological data, which are stored in different systems. Due to incompatible equipment protocols and inconsistent data standards, "data silos" are formed, making it difficult to effectively align multi-dimensional constraints and providing comprehensive support for planning decisions. Existing planning schemes are mostly designed for single, fixed scenarios, lacking comprehensive consideration of dynamic factors such as operational urgency, regional complexity, and unexpected tasks. When facing unconventional scenarios, the efficiency and accuracy of adjusting planning schemes decrease significantly, making it difficult to meet the needs of operations across all scenarios.

[0005] The planning process failed to fully explore the potential relationships between geographical constraints, resource load constraints, and task requirement constraints. Planning relied solely on simple rules or single-objective algorithms, making it difficult to coordinate and balance multiple objectives such as operational efficiency, resource utilization, and cost control. This resulted in resource idleness or task omissions. Existing algorithm models lacked generalization ability in complex environments, failed to establish scenario adaptation and calibration mechanisms for different urban areas and operation types, and lacked the ability to visualize and dynamically adjust constraint conflicts. Consequently, the implementation of the planning scheme was ineffective and unable to adapt to the diverse needs of sanitation operations.

[0006] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0007] Other features and advantages of this application will become apparent from the following detailed description, or may be learned in part by practice of the invention.

[0008] According to one aspect of this application, a multi-constraint intelligent planning method for smart sanitation operations across all scenarios is provided, comprising: acquiring multi-source core data for the entire sanitation operation scenario, including GIS geographic data of the operation area, dynamic task requirement lists, IoT status data of sanitation equipment, personnel skill configuration files, real-time environmental meteorological monitoring data, as well as constraint priority matrices and dynamic weight coefficients; performing regional gridding processing on the GIS geographic data based on a geospatial mapping algorithm, analyzing equipment load status in conjunction with IoT data, quantifying configuration matching degree through a personnel skill-task adaptation model, converting environmental meteorological data into operation difficulty coefficients, and generating a multi-dimensional constraint-aligned structured planning dataset; and layering the structured planning dataset according to operation urgency and regional complexity, with each layer constructing a structure including region-resource-task... The system uses a sliding time window to aggregate real-time dynamic features from the ternary association nodes of tasks, generating a multi-constraint directed association graph for all scenarios. It then models the multi-constraint directed association graph based on a collaborative optimization graph neural network, mining potential relationships between constraints and generating constraint feature embedding vectors. These vectors are imported into a hierarchical planning engine, optimized using a dynamic weight adaptive adjustment mechanism, and a scenario adaptation calibration module is introduced. Simultaneously, a decision chain for adjustment is generated through constraint conflict visualization, forming a basic planning model that integrates geographical, resource, and task collaboration logic. Based on this basic planning model, combined with real-time traffic data streams and sudden task trigger signals, dynamic simulations are performed. A multi-objective optimization algorithm balances operational efficiency, resource utilization, and cost control indicators, generating a smart sanitation operation dynamic planning scheme and execution details suitable for all scenarios.

[0009] Another aspect of this application is a multi-constraint intelligent sanitation full-scenario operation planning system, the system being configured to execute the above-described multi-constraint intelligent sanitation full-scenario operation planning method by executing the executable instructions.

[0010] According to another aspect of this application, an electronic device includes: a first processor; and a memory for storing executable instructions of the first processor; wherein the first processor is configured to execute the above-described intelligent planning method for multi-constraint smart sanitation operations across all scenarios by executing the executable instructions.

[0011] According to another aspect of this application, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a second processor, implements the above-described intelligent planning method for multi-constraint smart sanitation operations across all scenarios.

[0012] This application provides a multi-constraint intelligent planning method for smart sanitation operations across all scenarios. This method collects multi-source data, including GIS geographic data, equipment status data, and personnel files, along with constraint parameters. After processing through geospatial mapping and load analysis, a structured dataset is generated. Then, it is layered according to task urgency and regional complexity, constructing a three-element association node system of region-resource-task. A multi-constraint directed association graph is generated using a sliding time window. Constraint feature embedding vectors are extracted through collaborative optimization graph neural network modeling and imported into a hierarchical planning engine. Through dynamic weight adjustment, scenario adaptation calibration, and conflict resolution, a basic planning model is formed. Finally, by combining real-time data dynamic deduction and multi-objective optimization, a dynamic planning scheme and execution details adapted to all scenarios are output.

[0013] This application breaks down data silos from multiple sources, achieving effective alignment of constraints such as geography, resources, and tasks. It provides comprehensive data support for planning, adapts to dynamic changes across all scenarios, and can flexibly respond to unexpected tasks, road conditions, and weather changes, improving the accuracy and efficiency of planning adjustments. It uncovers potential relationships between constraints, balancing operational efficiency, resource utilization, and costs, reducing resource idleness and task omissions, and optimizing sanitation operation results. It visualizes constraint conflicts, generating a clear adjustment decision chain, reducing the difficulty of planning implementation, and facilitating efficient smart sanitation operations.

[0014] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0015] Figure 1 The flowchart illustrates a method for intelligent planning of smart sanitation operations across multiple constraints according to an embodiment of this application.

[0016] Figure 2 The diagram shows a structural schematic of a multi-constraint intelligent sanitation operation planning system for all scenarios, provided in an embodiment of this application. Detailed Implementation

[0017] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0018] The following is combined with Figure 1 This application describes a multi-constraint intelligent planning method for smart sanitation operations across all scenarios, based on exemplary embodiments thereof. It should be noted that the application scenarios described below are merely illustrative for understanding the spirit and principles of this application, and the embodiments of this application are not limited in any way. Rather, the embodiments of this application can be applied to any applicable scenario.

[0019] In one implementation, Figure 1A schematic diagram of a multi-constraint intelligent planning method for all-scenario smart sanitation operations according to an embodiment of this application is shown.

[0020] S101 acquires multi-source core data for the entire sanitation operation scenario, including GIS geographic data of the operation area, dynamic task requirement list, IoT status data of sanitation equipment, personnel skill configuration files, real-time environmental meteorological monitoring data, as well as constraint priority matrix and dynamic weight coefficient.

[0021] In one implementation, multi-dimensional core data and constraint configuration parameters required for smart sanitation operation planning across all scenarios are comprehensively collected to ensure that the data covers all constraint dimensions such as operation space, task requirements, resource supply, and environmental impact. This provides complete and accurate basic data support for subsequent standardized processing, correlation modeling, and planning scheme generation, ensuring the full-scenario adaptability of the planning model and the scientific nature of decision-making.

[0022] The GIS geographic data for the work area was acquired through methods such as interface calls to the city's geographic information public service platform, high-precision satellite remote sensing image analysis, and synchronization with the municipal road basic database. The data was uniformly calibrated using the WGS-84 coordinate system to ensure the accuracy of spatial location information. This data covers road network data (road direction, length, width, grade, number of lanes), regional boundary data (administrative division, scope of work responsibility area), key geographic marker data (location and scope of densely populated areas such as bus stops, schools, hospitals, and parks), and topographic data (slope, altitude, presence of bridges / tunnels, etc.). By connecting to a city's GIS public service platform, road network data for a 30-square-kilometer work area in the city center was obtained, including the direction and dimensions of 12 main roads, 35 secondary roads, and 86 branch roads. Simultaneously, the boundary coordinates and 500-meter radius data of 28 schools, 15 hospitals, and 12 parks within the area were acquired. After coordinate system calibration, a standardized geographic dataset was formed.

[0023] A task application and assignment system was established, supporting data collection from multiple channels, including online input by sanitation management departments, integration with citizen complaint platforms, and automatic reporting by intelligent inspection equipment. The system cleanses the data to remove duplicate and invalid task information, and then integrates and generates a list according to a unified field format. This list includes the task number, task type (e.g., road sweeping, garbage collection, public toilet cleaning, emergency cleaning), task initiation time, required completion deadline, specific coordinates of the work area, task priority, work standards (e.g., road sweeping must meet standards of no visible garbage and no dust accumulation), and associated auxiliary needs (e.g., coordination with traffic control is required for large-scale garbage collection). The intelligent inspection vehicle is equipped with a visual recognition device that automatically reports an emergency cleaning task for a non-motorized vehicle lane on a main road where there are a large number of fallen leaves. The system simultaneously enters the task number, task type (emergency cleaning), initiation time (9:15 AM), required completion time (within 2 hours), work area coordinates (116.4°-116.42° E, 39.9°-39.92° N), priority (high), and work standard (100% leaf removal rate), forming a dynamic task requirement list.

[0024] IoT sensors (GPS positioning modules, power / fuel sensors, operational status sensors, and fault detection sensors) are deployed on sanitation vehicles, cleaning equipment, and garbage transfer station equipment. Data is transmitted in real time to a cloud database via 5G / 4G networks, and equipment status information is updated every minute. Basic equipment information includes equipment number, equipment type, rated operating capacity, and factory parameters; real-time operating status includes whether it is in operation, standby, or faulty; location information includes real-time latitude and longitude, driving speed; performance parameters include remaining power / fuel, driving range, operating efficiency, and operating time; and fault information includes fault type, fault occurrence time, and fault location. The IoT sensors on a certain model of electric sweeper upload real-time data including device number, device type (small road sweeper), current status (operating), real-time location (a branch road, 116.38°E, 39.89°N), speed (5km / h), remaining battery power (65%), range (approximately 8km), current operating efficiency (2000 square meters / hour), operating time (1.5 hours), and no fault information, thus forming an IoT status data record for the device.

[0025] A sanitation worker information management system was established, where managers input basic personnel information and skills certification data. This data is dynamically updated based on skills assessment records and task completion quality evaluations during operations, ensuring the timeliness and accuracy of the data. Basic personnel information includes personnel ID, name, age, health status, and on-duty status; skills information includes the type of work skills possessed, skill proficiency level, and skills certification certificate number; work experience includes years of experience, preferred work scenarios, and historical work areas; task execution records include past task completion rates, completion quality scores, and attendance. For example, the personnel information management system is used to input the personnel ID, name, age (45 years old), good health, and on-duty status (normal attendance); skills information includes possession of both road sweeping and garbage collection skills, both at an advanced proficiency level, and holding a sanitation work skills level certificate; work experience includes 8 years of experience, specializing in urban main roads, with historical work areas being sections 1-5 in the city center; task execution records show a 98% task completion rate and an average completion quality score of 92 points over the past 3 months, with no absences, forming a complete personnel skills configuration file.

[0026] By connecting to the public data interface of the national meteorological department, the sensor network of urban environmental monitoring stations, and the miniature meteorological monitoring equipment deployed in the work area, environmental meteorological data is collected and updated hourly after cross-verification of multi-source data. Meteorological parameters include real-time temperature, humidity, wind speed, wind direction, rainfall, snowfall, and visibility; air quality parameters include PM2.5 concentration, PM10 concentration, and dust concentration; special weather warnings include rainstorm warnings, typhoon warnings, high temperature warnings, and cold wave warnings. By connecting to the data interface of urban environmental monitoring stations, real-time environmental meteorological data of a certain work area is obtained, including temperature 28℃, humidity 65%, wind speed 3m / s, wind direction southeast, rainfall 0mm, visibility 10km, PM2.5 concentration 32μg / m³, PM10 concentration 55μg / m³, and no special weather warnings. Simultaneously, the dust concentration of 0.3mg / m³ collected by the miniature meteorological monitoring equipment in the area is combined to form complete real-time environmental meteorological monitoring data.

[0027] Based on industry standards for sanitation operations, urban management requirements, and public demand surveys, a team of sanitation management experts formulated rules for determining constraint priorities. They used the analytic hierarchy process (AHP) to determine the relative importance weights of each constraint type, constructed a matrix structure, and entered it into the system. Constraint types include geographical constraints, resource constraints, task constraints, and environmental constraints. Each constraint type includes specific constraint items, priority levels (high, medium, low) for each constraint item in different scenarios, and priority quantification values ​​(values ​​between 0 and 1, with higher values ​​indicating higher priority). A review team of five sanitation management experts used the AHP to determine a constraint priority matrix, where constraint types include geographical constraints (including road grade constraints and regional boundary constraints), resource constraints (including equipment supply constraints and personnel allocation constraints), task constraints (including task urgency constraints and operational standard constraints), and environmental constraints (including meteorological conditions constraints and air quality constraints). In core urban area scenarios, the priority quantification value for task urgency constraints is 0.9 (high), for road grade constraints it is 0.8 (high), and for equipment supply constraints it is 0.7 (medium), forming a constraint priority matrix adapted to different scenarios.

[0028] Based on the constraint priority matrix and considering the dynamic changes in the work scenario, an initial weight coefficient is determined using a combination of entropy weighting and expert scoring. This initial weight coefficient is then dynamically adjusted based on real-time work data feedback, establishing a weight coefficient update mechanism. Core data includes the initial weight coefficient for each constraint type, the dynamic weight value for each specific constraint item, the triggering conditions for weight adjustment (such as sudden task triggering, sudden changes in environmental conditions, and changes in resource supply), and the range of weight adjustment.

[0029] Historical operation data is processed using the entropy weight method to determine the initial weight coefficients for geographical constraints (0.3), resource constraints (0.3), task constraints (0.25), and environmental constraints (0.15). When an emergency cleaning task occurs in the operation area, a weight adjustment mechanism is triggered to dynamically adjust the weight coefficients for task constraints to 0.4, geographical constraints to 0.25, resource constraints to 0.3, and environmental constraints to 0.15, ensuring that the weight coefficients adapt to changes in the scenario.

[0030] S102 uses a geospatial mapping algorithm to perform regional gridding processing on GIS geographic data, combines IoT data analysis of equipment load status, quantifies configuration matching degree through personnel skill-task adaptation model, transforms environmental meteorological data into operational difficulty coefficients, and generates a multi-dimensional constraint-aligned structured planning dataset.

[0031] In one implementation, a geospatial mapping algorithm is used to perform dual processing on the GIS geographic data of the work area, including road hierarchization and grid refinement, to delineate work units and path boundaries. The geospatial mapping algorithm integrates road network topology analysis and spatial grid division technology, processing the data according to the logic of "hierarchical classification first, then grid splitting." First, roads are hierarchically divided according to their grade attributes (arterial roads, secondary roads, and local roads), clarifying the work priorities and standards for different road levels. Then, based on a spatial projection coordinate system, the area where each road level is located is divided into regular grids with fixed side lengths (dynamically configurable), ensuring that each grid contains complete road segments and work areas. Finally, through coordinate calibration and boundary fusion, the work unit boundaries of each grid and the path connection relationships between adjacent grids are clarified. The grid side length is set to 200-500 meters (dynamically adjusted according to the density of the work area), the coordinate system is WGS-84, and the road hierarchical division threshold is set according to the urban road classification standard (arterial road width ≥ 24 meters, secondary road width 16-24 meters, local road width ≤ 16 meters).

[0032] When processing GIS geographic data of a 30-square-kilometer urban area, the following steps were taken: 12 roads with a width of more than 24 meters were first classified as main roads; 35 roads with a width of 16-24 meters were classified as secondary roads; and 86 roads with a width of less than 16 meters were classified as branch roads. The area was then divided into 333 regular grids with a side length of 300 meters. Each grid corresponds to a work unit. After calibration, the boundary coordinates of a certain main road grid unit were determined to be 116.4°-116.43° east longitude and 39.9°-39.93° north latitude. The coordinates of the path connection points of four adjacent grids and the work direction were also determined.

[0033] By combining real-time data streams from the device's IoT infrastructure, a load status analysis model is used to extract core indicators including device endurance, operational efficiency, and fault warnings. The load status analysis model consists of a data preprocessing layer, a feature extraction layer, and an indicator output layer, connected in series. Data is preprocessed and feature extracted sequentially before outputting core indicators. First, the real-time data stream from the device's IoT infrastructure is cleaned to remove outliers and missing values. Then, the feature extraction layer extracts device operating status characteristics (power / fuel consumption rate, percentage of operating time, frequency of fault codes, etc.). Finally, based on preset thresholds and a statistical model, three core indicators—device endurance, operational efficiency, and fault warnings—are calculated. The data sampling frequency is 1 minute / time, outlier removal uses the 3σ principle, endurance calculation is based on the ratio of remaining power / fuel to unit operating energy consumption, and fault warning thresholds are determined according to device factory parameters and historical fault data statistics (e.g., a warning is triggered if a certain model of sweeper detects abnormal motor current three times consecutively). The model's input is the real-time data stream from the device's Internet of Things (device type, real-time power / fuel level, operating time, operating parameters, fault codes, etc.); the model's output is a set of core indicators of the device's load status, including remaining driving range, operating area per unit time, fault warning level, and triggering reason.

[0034] When processing the real-time IoT data of a certain electric sweeper, three power consumption data points that exceed the normal range are first removed; then, features such as hourly power consumption rate of 15%, working time percentage of 80%, and no fault codes are extracted; combined with its rated power and unit working energy consumption, the remaining range of 8 kilometers and the working area of ​​2000 square meters per unit time are calculated, and the fault warning level is no warning, thus forming the load status index data of the equipment.

[0035] Using a personnel skills-task adaptation model, the matching degree between personnel and tasks is quantified based on skill proficiency, work experience, and on-the-job status. The personnel skills-task adaptation model constructs an evaluation system based on the Analytic Hierarchy Process (AHP), including a target layer (matching degree), a criterion layer (skill proficiency, work experience, on-the-job status), and an indicator layer (specific quantitative indicators under each criterion). First, weights are assigned to the criterion layer (skill proficiency 0.5, work experience 0.3, on-the-job status 0.2). Then, the indicator layer is quantitatively scored (skill proficiency: advanced 100-80 points, intermediate 79-60 points, basic 59-0 points; work experience: 5 points per year; on-the-job status: normal 100 points, absenteeism 0 points). Finally, a weighted summation is used to calculate the total matching degree score, ranging from 0 to 100 points, with higher scores indicating a higher matching degree.

[0036] The weights of the criteria layer are determined through expert scoring. The quantitative scoring thresholds are set based on industry operating standards and corporate management requirements. The matching degree grading standards are: high matching 85-100 points, medium matching 60-84 points, and low matching 0-59 points. The model's input data consists of personnel skill configuration files (skill type, proficiency, years of service, on-the-job status) and task requirements (required skills, task difficulty, execution time limit). The model's output data consists of the total matching degree score and grading results between personnel and tasks.

[0037] A certain road cleaning task requires advanced road cleaning skills, with medium difficulty, and requires on-duty availability. A sanitation worker has advanced skill proficiency (85 points), 8 years of experience (40 points), and normal on-duty status (100 points). The matching degree is calculated by weighted summation as 85×0.5+40×0.3+100×0.2=76.5 points, which is classified as medium matching. This clarifies the suitability of the person for the task.

[0038] A meteorological factor weighting algorithm is used to convert environmental meteorological data, including wind speed, rainfall, and temperature, into operational difficulty coefficients to differentiate the complexity of operations in different scenarios. The algorithm determines the weight of each meteorological factor using the entropy weighting method, and then calculates the comprehensive difficulty coefficient by combining the graded quantification values. First, three core meteorological factors—wind speed, rainfall, and temperature—are selected, and their weights are determined according to their impact on sanitation operations using the entropy weighting method (wind speed 0.4, rainfall 0.35, temperature 0.25). Then, each factor is graded and quantified (wind speed ≤3m / s = 1 point, 3-6m / s = 2 points, >6m / s = 3 points; no rainfall = 1 point, light rain = 2 points, moderate rain and above = 3 points; temperature 15-25℃ = 1 point, 5-14℃ or 26-35℃ = 2 points, <5℃ or >35℃ = 3 points). Finally, the operational difficulty coefficient is obtained by weighted summation, with the coefficient ranging from 1 to 3 points; the higher the score, the greater the difficulty.

[0039] The meteorological factor weights were calculated using the entropy weight method by processing historical operational data from the past three years. The grading and quantification thresholds referenced meteorological industry standards and sanitation operation safety regulations. The difficulty coefficient grading standard was 1 point for low difficulty, 2 points for medium difficulty, and 3 points for high difficulty. Real-time meteorological data for a certain operational area were: wind speed 3 m / s, no rainfall, and temperature 28℃. These were quantified as 1 point, 1 point, and 2 points respectively according to the grading standard. The weighted summation of the operational difficulty coefficient yielded a score of 1×0.4 + 1×0.35 + 2×0.25 = 1.25 points, classifying it as low difficulty, thus providing a difficulty reference for operational planning.

[0040] The linkage constraint priority matrix and dynamic weight coefficients complete multi-source data calibration, ultimately generating a multi-dimensional constraint-aligned structured planning dataset with clear geographical boundaries, well-defined resource status, accurate task matching, and reasonable difficulty quantification. Based on the constraint priority matrix and dynamic weight coefficients, collaborative calibration is performed on four types of data: geographical, equipment, personnel, and environmental, eliminating data conflicts and unifying data formats and dimensions. First, the priority of each data type is determined according to the constraint priority matrix, and data calibration weights are allocated based on the dynamic weight coefficients. Then, conflicting data (such as insufficient equipment endurance within the geographical boundary of a certain work unit) is coordinated and corrected according to priority and weight (prioritizing high-priority constraint data). Finally, the calibrated data of all types are integrated according to a unified field format to form a structured planning dataset.

[0041] The data calibration weights and constraint priority matrices are kept consistent with the dynamic weight coefficients. The data field format is set according to the JSON standard, including four major modules: work unit information, resource status information, task matching information, and work difficulty information. After integration, a certain structured planning dataset contains the boundary coordinates and road levels of 333 work units, the battery life and operating efficiency of 86 sanitation equipment, the task matching degree of 120 sanitation workers, and the work difficulty coefficient of each unit. Furthermore, the calibration corrected two data points where the battery life of equipment did not match the scope of the work unit, ensuring that all data is consistent and can be directly used for subsequent modeling.

[0042] S103 stratifies the structured planning dataset according to task urgency and regional complexity. Each layer constructs a ternary association node including region-resource-task. Real-time dynamic features are aggregated through a sliding time window to generate a multi-constraint directed association graph for the entire scenario.

[0043] In one implementation, the structured planning dataset is processed and layered according to task urgency and regional complexity. Each layer contains regional nodes, resource nodes, and task nodes. Real-time road conditions, equipment status, and dynamic features of weather changes are aggregated using a sliding time window. Data on regional boundaries, resource supply, and task requirements at each layer are extracted to generate corresponding subsets. A hierarchical clustering algorithm and a sliding time window aggregation algorithm are employed, combined with the core constraint dimensions of the task scenario, to achieve data layering and feature extraction. First, the hierarchical dimensions and standards are clearly defined. The urgency of the task is divided into three levels: high (≤2 hours), medium (2-8 hours), and low (>8 hours) according to the task completion time. The regional complexity is divided into three levels: core urban area, suburban area, and remote area according to road density and population density. Then, the two-level dimensions are cross-layered (e.g., core urban area - high urgency, suburban area - medium urgency, etc.), and a separate graph structure is constructed for each layer. Finally, a sliding time window is set (window size is 30 minutes, step size is 15 minutes), and the dynamic characteristics of real-time road conditions, equipment status, and weather changes within the window are aggregated. Regional boundary, resource supply, and task demand data are extracted from the data of each layer to generate exclusive data subsets for each layer.

[0044] The stratified dimensional thresholds are set based on sanitation operation management regulations and urban area planning standards. The sliding time window size and step size are determined through historical data verification (ensuring the timeliness and completeness of feature aggregation). Dynamic feature aggregation adopts a weighted average method (distributing weights according to data update frequency). When processing the structured planning dataset, the core urban area road cleaning tasks that need to be completed within 2 hours are divided into the core urban area - high urgency level. Through a 30-minute sliding time window, dynamic features such as real-time road conditions (main road congestion index 0.3), equipment status (3 cleaning vehicles with remaining range ≥ 5 km), and weather changes (no rainfall, wind speed 2 m / s) within this level are aggregated. The boundary coordinates of this level, the supply information of 3 cleaning vehicles, and the demand data of 5 emergency cleaning tasks are extracted to generate the core urban area - high urgency data subset.

[0045] Using regions as core nodes, resource allocation as supply nodes, and task requirements as target nodes, node attribute information is generated by combining GIS geographic coordinate mapping and task type identification. Edge association information is generated based on resource suitability and task execution sequence. The core elements of the graph structure are constructed by integrating GIS geographic coordinate mapping technology, task type identification algorithms, and association strength quantification models. Based on the layered data subset, node types and core attributes are determined—regional nodes include boundary coordinates, regional type, and complexity level; resource nodes include equipment / personnel number, status parameters, and supply capacity; and task nodes include task number, type, urgency, and required completion deadline. Following the logic of "region as core, resources as supply, and tasks as targets," spatial relationships between nodes are confirmed through GIS geographic coordinate mapping, and the suitability relationship between resources and tasks is matched through task type identification. Finally, resource suitability (between 0 and 1, with 1 being complete suitability) and task execution sequence (sorted by task initiation time) are quantified to generate edge association information (including association type, strength value, and sequence order).

[0046] Resource suitability is calculated based on the degree of matching between equipment / personnel skills and task requirements, and is quantified using a cosine similarity algorithm; task execution sequence is sorted by Unix timestamp, and edge association strength is positively correlated with suitability and time urgency. For the core urban area - high urgency data subset, regional nodes (boundary coordinates 116.4°-116.45°E, 39.9°-39.95°N, type: core urban area, high complexity), 3 resource nodes (sweeper numbers C01-C03, remaining range 6-8 km, supply capacity 2000 square meters / hour), and 5 task nodes (sweeping tasks T01-T05, high urgency, required to be completed before 12:00). Through GIS coordinate mapping, it was confirmed that all 3 sweepers are within the regional node range, and the task type identification determined that the sweepers and tasks are fully compatible (compatibility 1.0). Sorted by task initiation time (T01 earliest, T05 latest), edge association information from resource nodes to task nodes was generated (association type: "supply-demand", strength value 1.0, time sequence T01-T05).

[0047] A directed graph containing all operation area nodes, multi-type resource nodes, and differentiated task nodes is constructed. It is divided into three views based on geographical constraints, resource constraints, and task constraints. A constraint-based collaborative attention mechanism is used to optimize the cross-view edge association strength, generating a target multi-constraint directed association graph. The core architecture of the multi-view directed graph structure, employing a constraint-based collaborative attention mechanism to optimize cross-view associations, includes a view partitioning module, an association strength calculation module, and an attention optimization module, all of which operate in series. First, the node and edge association information at a single level is divided into three views according to constraint type: a geographic constraint view (focusing on the spatial association of regional nodes and their spatial adaptation to resources / tasks), a resource constraint view (focusing on the supply capacity of resource nodes and their adaptation to regions / tasks), and a task constraint view (focusing on the requirements of task nodes and their matching with regions / resources). Then, a basic directed graph containing all node and edge information is constructed. Each view shares node attributes but independently stores the edge association information of the corresponding constraints. Finally, through a constraint collaborative attention mechanism, the attention weight of the edge association of each view is calculated (dynamically allocated based on the constraint priority matrix), optimizing the strength of cross-view edge association, strengthening the edge association corresponding to high-priority constraints, and generating a multi-constraint directed association graph.

[0048] View partitioning is strictly defined according to constraint type. The initial weight of the constraint collaborative attention mechanism is consistent with the constraint priority matrix (e.g., task constraint weight of 0.4 in the core urban area scenario). The edge association strength optimization range is 0-1.2 (original strength value × attention weight coefficient, coefficient ≤ 1.2). Based on the node and edge information of the core urban area - high urgency data subset, it is divided into a geographic constraint view (recording the spatial association between regional nodes and resource / task nodes), a resource constraint view (recording the supply association of 3 sweepers), and a task constraint view (recording the demand association of 5 urgent tasks). A basic directed graph containing 8 nodes (1 region, 3 resources, 5 tasks) and 15 edges is constructed. Through the constraint collaborative attention mechanism, an attention weight of 0.4 is assigned to the task constraint view, and the edge association strength corresponding to the 5 urgent tasks is optimized from 1.0 to 1.2, strengthening the association priority of task constraints, and finally generating a multi-constraint directed association graph of the core urban area - high urgency.

[0049] S104 models a multi-constraint directed association graph based on a collaborative optimization graph neural network, mines the potential relationships between constraints, and generates constraint feature embedding vectors.

[0050] In one implementation, based on the needs of constraint association mining and feature extraction objectives, the node attributes, edge association strength, and dynamic feature time-series change data of multi-constraint directed association graphs are structurally collected and standardized to generate basic graph structure data containing constraint types, association weights, and time-series trends. A structured data collection protocol and the Z-Score standardization algorithm are employed to ensure uniform data format and reasonable numerical ranges. The data collection scope is clearly defined—node attributes cover core fields such as region boundaries, resource status, and task requirements; edge association strength includes fitness and time-series urgency quantification values; and dynamic feature time-series change data is collected using a sliding time window (30 minutes, consistent with previous descriptions). A structured collection protocol is used to unify the storage format according to field types (numerical, string, coordinate). Numerical data (such as edge association strength and dynamic feature values) are standardized using the Z-Score algorithm to eliminate the influence of units, mapping the data to the [-1,1] interval. Finally, the basic graph structure data containing constraint types, association weights, and time-series trends is integrated to generate basic graph structure data.

[0051] Z-Score standardization uses global mean and standard deviation calculations. Data collection frequency is consistent with the sliding time window step size (15 minutes / time). String data is converted into numerical identifiers according to encoding rules (e.g., constraint type "geographic constraint" is encoded as 1). When processing the directed association graph of the core urban area with high urgency and multiple constraints, the following data are collected: boundary coordinates of regional nodes (numerical), remaining range of resource nodes (numerical), urgency of task nodes (numerical), edge association strength (initial value 1.0), and dynamic changes in equipment status within 30 minutes (3 sets of remaining range values: 8km, 7.5km, 7km). After storage according to a structured protocol, Z-Score standardization is performed on edge association strength (1.0) and remaining range (8km, etc.) to obtain a standardized association weight value of 0.85 and a standardized remaining range value of 0.92, 0.81, and 0.70. The constraint type "task constraint" is encoded as 3, and the data is integrated to generate a graph structure basic data containing constraint type encoding, standardized association weight, and time-series trend sequence (0.92→0.81→0.70).

[0052] This paper designs modeling rules for graph data based on the constraints and collaborative characteristics of sanitation operations, clarifies the association threshold of region-resource-task and the dynamic feature decay coefficient, and generates a multi-constraint graph modeling specification. Combining the constraints and collaborative characteristics of sanitation operations, key modeling parameters are determined through statistical analysis and expert experience, forming standardized modeling rules. First, the constraints and collaborative patterns of sanitation operations (such as spatial collaboration between region and resource, and adaptive collaboration between resource and task) are analyzed, clarifying the core modeling parameters—the association threshold of region-resource-task (set to 0.5; values ​​below this are considered weak associations and are weakened during modeling) and the dynamic feature decay coefficient (set to 0.95, decaying over time steps to reflect feature timeliness). Then, modeling specifications are formulated, clarifying the weak association data processing method (feature weights multiplied by 0.3), the dynamic feature time-series fusion rule (weighted summation based on decay coefficient), and the constraint type priority mapping rule (consistent with the constraint priority matrix). Finally, a multi-constraint graph modeling specification document is formed to guide subsequent neural network modeling.

[0053] The association threshold is determined by analyzing the minimum fit of "strongly correlated" cases in historical operation data. The dynamic feature decay coefficient is set based on the feature timeliness decay law (such as the linear decay of equipment endurance over time). The constraint type priority mapping rule directly adopts the constraint priority quantification value. Based on the statistical result that "the operation completion rate exceeds 90% when the resource and task fit is ≥0.5" in sanitation operations, the association threshold is set to 0.5. According to the law that the equipment endurance decays by 5% per hour, the dynamic feature decay coefficient is set to 0.95 (decaying by 2.5% every 15 minutes). The priority quantification value of the constraint type "task constraint" is specified as 0.4, and its feature weights are weighted according to this value during modeling. A multi-constraint graph modeling specification is formed, stipulating that when the standardized value of edge association strength is lower than 0.5, the feature weight is multiplied by 0.3, and the dynamic feature is updated according to "current value = previous value × 0.95".

[0054] Combining the collaborative optimization graph neural network architecture and feature embedding requirements, a multi-level modeling mechanism is established, comprising node feature enhancement, edge association strengthening, and cross-view feature fusion. Based on the collaborative optimization graph neural network architecture, a three-level serial modeling mechanism of "node feature enhancement - edge association strengthening - cross-view feature fusion" is designed, with the layers connected through feature vector transfer. The first level (node ​​feature enhancement): Graph convolutional layers are used to perform nonlinear transformations on node attribute features, introducing attention weights (dynamically allocated based on node importance) to enhance the feature expression of core nodes (such as high-urgency task nodes). The second level (edge ​​association strengthening): Edge convolutional layers are used to fuse edge association information with node features, strengthening the feature weights of strongly associated edges (association threshold ≥ 0.5) and weakening the influence of weakly associated edges. The third level (cross-view feature fusion): A multi-head attention mechanism is used to fuse features from the geographic, resource, and task views, allocating view attention weights according to constraint priority to generate cross-view fused features. The three levels are executed sequentially, forming a complete multi-level modeling process.

[0055] The graph convolutional and edge convolutional layers use the ReLU function for activation. The number of heads in the multi-head attention mechanism is set to 3. The view attention weights are consistent with the quantized values ​​of the constraint priority matrix (e.g., the task constraint view weight is 0.4). Node importance is calculated by weighting the number and strength of the edges associated with the node. The standardized graph structure data is input into the multi-level modeling mechanism. In the first level, the graph convolutional layer assigns an attention weight of 0.4 to the task node features (due to its high priority) to enhance their feature representation. In the second level, the edge convolutional layer strengthens the feature weight of edges with a standardized association strength of 0.85 (strong association) to 1.02, while weakening the weight of edges with a standardized association strength of 0.4 (weak association) to 0.12. In the third level, the multi-head attention mechanism assigns weights according to the geographic view (0.25), resource view (0.3), and task view (0.4), fusing the features of the three views to generate a cross-view fused feature vector.

[0056] The algorithm integrates and executes basic graph structure data, multi-constraint graph modeling specifications, and multi-level modeling mechanisms. It optimizes feature mappings through iterative training of neural networks, generating constraint feature embedding vectors that include core constraint features, associated semantic information, and temporal dynamic attributes. The collaboratively optimized graph neural network consists of an input layer, a multi-level modeling mechanism layer, a feature mapping layer, and an output layer, with each layer connected using a fully connected approach. Input the basic graph structure data into the collaborative optimization graph neural network in batches (batch size set to 32); initialize the model parameters according to the multi-constraint graph modeling specification (such as association threshold and decay coefficient embedded in the model layer); enhance, strengthen and fuse the data through a multi-level modeling mechanism layer; the feature mapping layer uses a fully connected layer to map high-dimensional intermediate features to a low-dimensional space (embedding vector dimension set to 128 dimensions); take constraint association prediction as the training objective (predict the constraint association type between nodes), use the cross-entropy loss function to calculate the loss value, and iteratively update the model parameters through the Adam optimizer (learning rate set to 0.001, iteration number set to 100 rounds); stop training when the loss value is lower than the preset threshold (0.01), and output the constraint feature embedding vector finally generated by the model, which contains the core constraint features, association semantic information and temporal dynamic attributes.

[0057] The batch size was 32, the learning rate was 0.001, the number of iterations was 100, the loss threshold was 0.01, the embedding vector dimension was 128, and the Adam optimizer had β1=0.9 and β2=0.999. The basic graph structure data from 32 batches was input into a collaborative optimization graph neural network. The association threshold was initialized to 0.5 and the decay coefficient to 0.95 according to the modeling specifications. After processing through a multi-level modeling mechanism, the high-dimensional features were mapped to 128-dimensional vectors through a feature mapping layer. With the goal of "predicting that the constraint association type corresponding to this vector is a task constraint," the network was iterated for 100 rounds until the loss value decreased to 0.008 (below the threshold). The output was a 128-dimensional constraint feature embedding vector, where the first 30 dimensions represent core features (e.g., task urgency), the middle 50 dimensions represent association semantics (e.g., resource-task adaptation relationship), and the last 48 dimensions represent temporal dynamic attributes (e.g., device state decay trend).

[0058] S105 embeds constraint features into vectors and imports them into a hierarchical planning engine. It then uses a dynamic weight adaptive adjustment mechanism for optimization, introduces a scenario adaptation calibration module, and generates an adjustment decision chain through constraint conflict visualization, thus forming a basic planning model that integrates geographical, resource, and task collaborative logic.

[0059] In one implementation, based on the requirements of constraint-based collaborative planning and the goal of full-scene adaptation, a core architecture of hierarchical planning engine + dynamic weight adjustment module + scene adaptation calibration unit is built to perform hierarchical parsing and collaborative logic learning on the embedded vectors of geographical, resource, and task constraint features. A parallel-serial hybrid architecture of "hierarchical planning engine + dynamic weight adjustment module + scene adaptation calibration unit" is adopted, with the hierarchical planning engine as the core processing unit, and the dynamic weight adjustment module and scene adaptation calibration unit as auxiliary optimization units. The three are connected via a data bus to achieve real-time data interaction and command transmission. First, the functional positioning of each module is clarified: the hierarchical planning engine is responsible for hierarchically parsing the embedded vectors of geographical, resource, and task constraint features (divided into three parsing layers according to constraint type), extracting the core semantics and related logic of each constraint; the dynamic weight adjustment module receives the parsed feature data in real time and dynamically adjusts the weight allocation based on the scene; the scene adaptation calibration unit calibrates the parsing results and weight adjustment effects, correcting deviations; then, through a collaborative logic learning algorithm, the potential collaborative relationships between the three types of constraints are explored to form preliminary planning logic.

[0060] The feature extraction dimension of the hierarchical parsing layer is consistent with the constraint feature embedding vector dimension (128 dimensions). The data interaction frequency between modules is 10 times / second. The collaborative logic learning adopts the gradient descent method to optimize the learning efficiency, and the learning rate is set to 0.002. The 128-dimensional geographic constraint feature embedding vector (including core features such as regional boundaries and road grades), resource constraint feature embedding vector (including core features such as equipment endurance and personnel matching degree), and task constraint feature embedding vector (including core features such as urgency and operation standards) are input into the core architecture. The hierarchical planning engine extracts the core semantics of various constraints through the three parsing layers. The dynamic weight adjustment module initializes the weight allocation (geographic 0.25, resource 0.3, task 0.45), and the scene adaptation calibration unit verifies the parsing results. Through the collaborative logic learning algorithm, the logic of "strong correlation between geographic constraints and resource constraints in the core urban area scenario" is discovered, forming a preliminary planning logic model.

[0061] The planning logic is designed based on a multi-constraint balancing strategy, clarifying the spatial adaptation dimension of geographical constraints, the supply matching dimension of resource constraints, the priority dimension of task constraints, and the synergistic ratio of these three, generating constraint-based collaborative planning rules. Based on the multi-constraint balancing strategy, the Analytic Hierarchy Process (AHP) and weighted summation algorithm are used to clarify each constraint dimension and its synergistic ratio. First, the core dimensions of each constraint are broken down—geographical constraints focus on spatial adaptation (regional boundary adaptation, path rationality), resource constraints focus on supply matching (equipment / personnel supply matching task demand), and task constraints focus on priority (urgency, operational standard priority). Then, the synergistic ratio of each dimension is determined using the AHP, and considering industry standards and expert experience, the ratios are set as follows: spatial adaptation dimension 0.3, supply matching dimension 0.35, and priority dimension 0.35. Finally, constraint-based collaborative planning rules are formulated, clarifying the decision-making logic under each dimension (e.g., in the spatial adaptation dimension, paths must avoid congested sections; in the supply matching dimension, high-matching resources are prioritized), forming a standardized rule document.

[0062] The synergy ratios for each dimension were determined through scoring by at least five industry experts and calculation using the analytic hierarchy process (AHP). Decision logic thresholds were set based on statistical analysis of historical best-case scenarios (e.g., a path congestion index threshold of 0.5; values ​​higher than this were considered unreasonable). Based on a multi-constraint balancing strategy, the spatial adaptation dimension of geographical constraints was defined as requiring "overlap between the work path and the regional boundary ≥ 95%" and "congestion index ≤ 0.5"; the supply matching dimension of resource constraints required "equipment endurance ≥ 1.2 times the required mileage of the work area" and "personnel matching score ≥ 60 points"; and the priority dimension of task constraints required "high-urgency tasks to be allocated resources first." The AHP determined the synergy ratios for these three dimensions to be: spatial adaptation 0.3, supply matching 0.35, and priority 0.35. These were integrated to form constraint-based synergy planning rules, stipulating that resource scheduling optimization is triggered when the resource matching score for high-urgency tasks falls below 60 points.

[0063] To adapt to different work areas, sanitation resources, and task urgency, a dynamic weight adaptive adjustment mechanism is established. In core urban areas, visual geography and resource supply constraints are prioritized; in emergency scenarios, task priority and resource scheduling constraints are emphasized; and in suburban scenarios, geographical scope and cost control constraints are considered. A scenario-weight mapping model is constructed, employing fuzzy logic algorithms to dynamically adapt weights. The model consists of four layers: a scenario input layer, a fuzzification processing layer, a rule inference layer, and a weight output layer, all running in series. First, the scenario categories and core influencing factors are clearly defined: for core urban area scenarios, the influencing factors are road density and population density; for emergency task scenarios, the influencing factors are task urgency and resource shortage; and for suburban operation scenarios, the influencing factors are area scope and operation cost. Next, weight adjustment rules are set for each scenario: for core urban area scenarios, the weights for geographical constraints are 0.3, resource constraints are 0.4, and task constraints are 0.3; for emergency task scenarios, the weights are 0.2, 0.3, and 0.5; and for suburban operation scenarios, the weights are 0.35, 0.25, and 0.4. Finally, a fuzzy logic algorithm is used to dynamically adjust the weights based on the real-time influencing factor values ​​of the scenario (e.g., road density in core urban areas is 0.8, and population density is 0.9), with an adjustment range of ±0.1.

[0064] The quantification range of the scene influence factor is 0-1. The initial weight value is set based on the constraint collaboration ratio. The membership function of the fuzzy logic algorithm adopts a triangular function. The weight adjustment trigger condition is that the change in influence factor is ≥0.2. For the core urban area scene, the initial weights are set as follows: Geographic 0.3, Resource 0.4, and Task 0.3. When the real-time monitoring shows that the road density in the core urban area increases from 0.7 to 0.9 (change of 0.2, triggering adjustment), the fuzzy logic algorithm adjusts the geographic constraint weight to 0.35, the resource constraint weight remains at 0.4, and the task constraint weight is adjusted to 0.25. For the emergency task scene, when the task urgency increases from 0.8 to 1.0 (triggering adjustment), the task constraint weight is adjusted from 0.5 to 0.6, the geographic constraint weight decreases to 0.15, and the resource constraint weight remains at 0.25 to ensure that the weights adapt to scene changes.

[0065] The core architecture, constraint-based collaborative planning rules, and dynamic weight adjustment mechanism are integrated and executed. An adjustment decision chain is generated through constraint conflict visualization to optimize planning parameter configuration, forming a basic planning model that integrates geographic, resource, and task collaborative logic and adapts to the needs of all operational scenarios. A pipeline-style integrated execution mode is adopted, sequentially executing the core architecture, constraint-based collaborative planning rules, and dynamic weight adjustment mechanism in the order of "feature analysis → weight allocation → rule verification → conflict handling → parameter optimization." First, constraint features are embedded into vectors and input into the core architecture for hierarchical analysis and collaborative logic learning. Then, real-time weights are allocated through the dynamic weight adjustment mechanism. The preliminary planning logic is verified based on the constraint-based collaborative planning rules, identifying constraint conflicts (such as mismatch between resource supply and task requirements, and conflicts between paths and regional boundaries). An adjustment decision chain is generated through constraint conflict visualization technology (using heatmaps to display conflict areas and intensity) (sorted by conflict priority, with high-priority conflicts handled first). Finally, the planning parameter configuration is optimized based on the decision chain (such as adjusting resource allocation schemes and correcting operational paths), iteratively executing until there are no conflicts or the conflict intensity is below a threshold (conflict intensity ≤ 0.3), forming the basic planning model.

[0066] The conflict intensity quantification range is 0-1, calculated based on the weighted calculation of the conflict's impact range and severity. The maximum number of iterations is 20, and the conflict threshold is set at 0.3. Integrating the core architecture, constraint-based collaborative planning rules, and dynamic weight adjustment mechanism, after inputting the constraint feature embedding vector, the core architecture parses out that "the resource matching degree of a high-urgency task in a certain core urban area is 55 points" (below the rule threshold of 60 points), and dynamically assigns weights as follows: geography 0.35, resources 0.4, and task 0.25; it identifies the conflict between resource constraints and task constraints (conflict intensity 0.6), and locates the conflict point as "insufficient personnel matching degree" through a visual heatmap; it generates an adjustment decision chain (prioritizing the scheduling of personnel with high matching degree); it optimizes the planning parameters, scheduling two personnel with a matching degree of 80 points to this task, recalculating the matching degree to 78 points (meeting the threshold), and reducing the conflict intensity to 0.2 (below the threshold); after three iterations of optimization, a basic planning model adapted to the high-urgency scenario in the core urban area is formed.

[0067] S106, based on the planning basic model combined with real-time traffic data stream and sudden task trigger signals, performs dynamic simulations and uses multi-objective optimization algorithms to balance operation efficiency, resource utilization, and cost control indicators, generating a smart sanitation operation dynamic planning scheme and execution details that are suitable for all scenarios.

[0068] In one implementation, a planning baseline model is used to dynamically simulate sanitation operation scenarios by combining real-time traffic data streams and sudden task triggering signals. A three-dimensional balance model of efficiency, resources, and cost is constructed through a multi-objective optimization algorithm. Input constraint features are embedded into vectors to simulate the adaptability of planning schemes under different scenarios. First, the planning baseline model is used as the simulation benchmark, and real-time traffic data streams (road congestion status, traffic efficiency) and sudden task triggering signals (task type, urgency, and scope of operation) are input. Then, the three-dimensional balance model defines indicators for each dimension—efficiency indicator is the operation completion time, resource indicator is the equipment / personnel utilization rate, and cost indicator is the sum of energy consumption and labor costs. Finally, constraint features are embedded into vectors and input into the model to simulate the adaptability of planning schemes under different scenarios (such as increased congestion and sudden task insertion), and output multiple sets of non-dominated optimal solutions (i.e., candidate schemes under different indicator weights).

[0069] The NSGA-Ⅲ algorithm has a population size of 100, 50 iterations, a crossover probability of 0.8, and a mutation probability of 0.1. The initial weights for efficiency, resource, and cost indicators are all 1 / 3, which can be dynamically adjusted according to scenario requirements. The fit metric ranges from 0 to 1, with higher scores indicating stronger solution fit. Based on the core urban area operation planning model, real-time traffic data streams (main road congestion index 0.7, secondary road 0.3) and emergency cleaning task trigger signals (high urgency, operation area 2 square kilometers) are input. A three-dimensional balance model defines operation completion time ≤ 2 hours (efficiency), equipment utilization ≥ 80% (resource), and cost ≤ 2000 yuan (cost). Constraint feature embedding vectors (including regional boundaries, equipment endurance, task urgency, etc.) are input. The solution fit is 0.75 under congestion scenarios and 0.68 after inserting emergency tasks, outputting 3 candidate solutions (focusing on efficiency, resource, and cost optimization respectively).

[0070] When extracting real-time data features during dynamic simulation, a traffic congestion index and an emergency task urgency coefficient are introduced. Feature engineering is used to extract key features from real-time data, and a weighted quantization algorithm is used to generate the traffic congestion index and the emergency task urgency coefficient. Feature extraction is performed on the real-time traffic data stream, selecting three core features: road speed, traffic volume, and congestion duration. The weights are determined using the entropy weight method (speed 0.4, traffic volume 0.35, congestion duration 0.25), and the traffic congestion index (range 0-1, 0 for smooth traffic, 1 for complete congestion) is calculated using weighted quantization. For emergency task trigger signals, three features are extracted: task urgency, work scope, and time requirement. Similarly, the weights are determined using the entropy weight method (urgency 0.5, work scope 0.3, time requirement 0.2), and the emergency task urgency coefficient (range 0-1, 0 for low urgency, 1 for extremely high urgency) is obtained using weighted quantization. These two coefficients are used as dynamic correction factors and incorporated into the subsequent planning and optimization process.

[0071] Feature weights were determined by processing nearly 6 months of historical road condition and emergency task data using the entropy weight method. The quantification thresholds for the index and coefficients were set based on industry standards and operational practices (e.g., a congestion index ≥ 0.6 was considered severe congestion, and an emergency coefficient ≥ 0.8 was considered a top-level emergency task). For real-time road condition data of a main road, the following parameters were extracted: traffic speed 20 km / h (50% below normal speed), traffic volume 800 vehicles / hour (above the threshold of 600 vehicles / hour), and congestion duration 30 minutes. A road congestion index of 0.72 (severe congestion) was obtained through weighted calculation. For a road flooding cleanup task after a sudden rainstorm, the following parameters were extracted: high urgency, operation area of ​​3 square kilometers, and time requirement of 1.5 hours. An emergency coefficient of 0.85 (top-level emergency) was obtained through weighted quantification. Both coefficients were incorporated into the planning optimization as correction factors.

[0072] When analyzing the dynamic characteristics of a work scenario, the resource supply status and task execution time sequence are correlated. During the balance optimization analysis, the dynamic change chain of constraints within a sliding time window is incorporated. Specifically, the resource supply status (equipment endurance, personnel on-duty changes) is correlated with the task execution time sequence (task initiation time, required completion deadline, execution order) to ensure that resource supply matches the task execution rhythm (e.g., high-urgency tasks correspond to priority scheduling of high-endurance equipment). A sliding time window is set (30 minutes, consistent with the previous setting, with a 15-minute step), and dynamic change data of constraints within the window (road condition updates, equipment status changes, weather adjustments) are aggregated to form a dynamic change chain of constraints. During the balance optimization analysis, key nodes in the change chain (e.g., increased congestion, equipment failure) are used as trigger conditions for planning adjustments, dynamically correcting paths, resource allocation, and task execution order.

[0073] The sliding time window size and step size are set based on the real-time data update frequency, and the trigger threshold for the constraint change chain is referenced from historical operation adjustment cases (e.g., a change in the road congestion index ≥ 0.2 triggers path adjustment). When analyzing the operation scenario in the core urban area, the range changes of 3 sweepers (8km→7km→6.5km) are associated with the execution sequence of 5 tasks (T01→T02→T03→T04→T05) to ensure that equipment with a range of 8km is allocated to T01 (Special Emergency). Through a 30-minute sliding time window, the constraint dynamic change chain is aggregated (road congestion index 0.5→0.7→0.72, insufficient range of equipment C02 triggers fault warning). The adjustment trigger conditions are set as a change in the congestion index ≥ 0.2 and equipment fault, and the execution order and operation path of T02-T05 are dynamically corrected.

[0074] After the planning module outputs a preliminary planning scheme containing multi-dimensional adaptation indicators, it verifies and adjusts it by comparing it with the historical best-case case library to generate a dynamic planning scheme and execution details for smart sanitation operations across all scenarios with dynamic scenario adaptation. The planning module outputs a preliminary planning scheme containing multi-dimensional adaptation indicators (job completion rate, resource utilization rate, cost control rate, conflict occurrence rate); it calls the historical best-case case library (stores the best job cases across all scenarios in the past year, including scenario type, constraints, planning scheme, and execution effect), uses the cosine similarity algorithm to calculate the similarity between the preliminary scheme and historical cases, and selects the top 5 cases with the highest similarity as references; it compares and analyzes the differences between the preliminary scheme and the reference cases, and adjusts for items where the adaptation indicators do not meet the standards (such as resource utilization rate below 80%) (such as reallocating resources and optimizing paths); after iterative verification and adjustment 2-3 times, it generates a dynamic planning scheme and execution details for smart sanitation operations across all scenarios with dynamic scenario adaptation (including job paths, resource allocation tables, task execution sequences, and emergency adjustment plans).

[0075] The historical case library uses a case selection threshold of ≥90 points (out of 100) for execution performance score and a cosine similarity threshold of 0.8 (cases above this value are considered highly similar). The maximum number of iterations is 5 (to ensure planning efficiency). The planning module outputs a preliminary plan with multi-dimensional adaptation indicators: 95% completion rate, 75% resource utilization rate (not meeting the target), 90% cost control rate, and 5% conflict occurrence rate. The historical case library is then used to select the best cases (similarity ≥0.85) of severe congestion and emergency tasks in the core urban area. Comparison reveals that low resource utilization is due to insufficient scheduling of two devices. The resource allocation plan is adjusted, allocating idle devices to secondary tasks, and the resource utilization rate is recalculated to increase to 88%. After two iterations, the final plan is generated, including optimized work paths (avoiding severely congested sections), a resource allocation table (clear division of labor among the three devices), task execution sequence (T01 is prioritized and completed within 1.5 hours), and an emergency adjustment plan (if congestion worsens, backup devices from surrounding areas will be scheduled).

[0076] In one implementation, such as Figure 2 As shown, this application also provides a multi-constraint intelligent sanitation operation planning system for all scenarios, including:

[0077] The data acquisition module 201 is used to acquire multi-source core data of the entire sanitation operation scenario, including GIS geographic data of the operation area, dynamic task requirement list, IoT status data of sanitation equipment, personnel skill configuration files, real-time environmental meteorological monitoring data, as well as constraint priority matrix and dynamic weight coefficient.

[0078] The structured dataset generation module 202 is used to perform regional gridding processing on GIS geographic data based on geospatial mapping algorithm, combine equipment load status analysis with IoT data, quantify configuration matching degree through personnel skill-task adaptation model, transform environmental meteorological data into operation difficulty coefficient, and generate a multi-dimensional constraint-aligned structured planning dataset.

[0079] The multi-constraint directed association graph construction module 203 is used to layer the structured planning dataset according to the urgency of the task and the complexity of the region. Each layer constructs a ternary association node including region-resource-task. Real-time dynamic features are aggregated through a sliding time window to generate a multi-constraint directed association graph for the whole scenario.

[0080] The constraint feature embedding vector generation module 204 is used to model a multi-constraint directed association graph based on a collaborative optimization graph neural network, mine potential association relationships between constraints, and generate constraint feature embedding vectors containing core constraint features, association semantic information, and temporal dynamic attributes.

[0081] The planning basic model construction module 205 is used to embed constraint features into vectors and import them into the hierarchical planning engine. It adopts a dynamic weight adaptive adjustment mechanism for optimization, introduces a scenario adaptation calibration module, and generates an adjustment decision chain through constraint conflict visualization to form a planning basic model that integrates geographical, resource, and task collaborative logic.

[0082] The dynamic planning scheme generation module 206 is used to dynamically extrapolate based on the planning basic model combined with real-time traffic data streams and sudden task trigger signals. Through multi-objective optimization algorithms, it balances operation efficiency, resource utilization, and cost control indicators to generate a smart sanitation operation dynamic planning scheme and execution details that are suitable for all scenarios.

[0083] The computer-readable storage medium provided in the above embodiments of this application and the intelligent planning method for multi-constraint smart sanitation full-scene operation provided in the embodiments of this application are based on the same inventive concept and have the same beneficial effects as the methods adopted, run or implemented by the application stored therein.

[0084] The various embodiments in this application are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments for the intelligent planning method, electronic device, electronic device, and readable storage medium for evaluating multi-constraint intelligent sanitation full-scenario operations are basically similar to the embodiments of the multi-constraint intelligent sanitation full-scenario operation intelligent planning method described above, and are therefore described simply. Relevant parts can be referred to in the descriptions of the embodiments of the multi-constraint intelligent sanitation full-scenario operation intelligent planning method described above.

Claims

1. A multi-constraint intelligent planning method for smart sanitation operations across all scenarios, characterized in that, include: Acquire multi-source core data for the entire sanitation operation scenario, including GIS geographic data of the operation area, dynamic task requirement list, IoT status data of sanitation equipment, personnel skill configuration files, real-time environmental meteorological data, as well as constraint priority matrix and dynamic weight coefficients. This process utilizes geospatial mapping algorithms to perform regional gridding on GIS geographic data. Combined with IoT data analysis of equipment load status, a personnel skill-task adaptation model quantifies configuration matching. Environmental meteorological data is transformed into operational difficulty coefficients, generating a multi-dimensional constraint-aligned structured planning dataset. This includes dual processing of the GIS geographic data for the work area using geospatial mapping algorithms, involving road hierarchization and grid refinement to delineate work units and path boundaries; extracting core indicators such as equipment endurance, operational efficiency, and fault warnings using a load status analysis model, based on real-time IoT data streams; quantifying the matching degree between personnel and tasks based on skill proficiency, work experience, and on-duty status using a personnel skill-task adaptation model; converting environmental meteorological data, including wind speed, rainfall, and temperature, into operational difficulty coefficients using a meteorological factor weighting algorithm to differentiate operational complexity in different scenarios; and calibrating multi-source data using a linked constraint priority matrix and dynamic weight coefficients. The final result is a multi-dimensional constraint-aligned structured planning dataset with clear geographic boundaries, well-defined resource status, accurate task matching, and reasonable difficulty quantification. The structured planning dataset is layered according to task urgency and regional complexity. Each layer constructs a ternary association node including region-resource-task. Real-time dynamic features are aggregated through a sliding time window to generate a multi-constraint directed association graph for the entire scenario. This includes processing the structured planning dataset, layering it according to task urgency and regional complexity, with each layer containing region nodes, resource nodes, and task nodes. Real-time dynamic features of road conditions, equipment status, and weather changes are aggregated through a sliding time window. Data on regional boundaries, resource supply, and task requirements at each level are extracted to generate corresponding subsets. Using the region as the core node, resource allocation as the supply node, and task requirements as the target node, node attribute information is generated by combining GIS geographic coordinate mapping and task type identification. Edge association information is generated according to resource adaptability and task execution sequence. A directed graph containing all operation area nodes, multi-type resource nodes, and differentiated task nodes is constructed. It is divided into three views according to geographic constraints, resource constraints, and task constraints. The cross-view edge association strength is optimized through a constraint collaborative attention mechanism to generate a target multi-constraint directed association graph. This paper models a multi-constraint directed graph based on a collaborative optimization graph neural network, mines potential relationships between constraints, and generates constraint feature embedding vectors. This includes structured collection and standardization of node attributes, edge association strength, and dynamic feature temporal changes in the multi-constraint directed graph based on constraint association mining needs and feature extraction objectives, generating basic graph structure data containing constraint types, association weights, and temporal trends. Modeling rules for the graph data are designed based on the collaborative characteristics of sanitation operation constraints, clarifying the association threshold of region-resource-task and the dynamic feature decay coefficient, generating multi-constraint graph modeling specifications. Combining the collaborative optimization graph neural network architecture and feature embedding requirements, a multi-level modeling mechanism is established, including node feature enhancement, edge association strengthening, and cross-view feature fusion. The basic graph structure data, multi-constraint graph modeling specifications, and multi-level modeling mechanisms are integrated and executed. Feature mapping is optimized through iterative neural network training to generate constraint feature embedding vectors containing core constraint features, association semantic information, and temporal dynamic attributes. The constraint features are embedded into a vector and imported into a hierarchical planning engine. A dynamic weight adaptive adjustment mechanism is used for optimization. A scenario adaptation calibration module is introduced. At the same time, an adjustment decision chain is generated through constraint conflict visualization, forming a basic planning model that integrates geographical, resource, and task coordination logic. Based on the planning model, combined with real-time traffic data streams and emergency task triggering signals, dynamic simulations are performed. Through multi-objective optimization algorithms, operational efficiency, resource utilization, and cost control indicators are balanced to generate a smart sanitation operation dynamic planning scheme and execution details that are suitable for all scenarios.

2. The method as described in claim 1, characterized in that, The constraint features are embedded into a vector and imported into a hierarchical planning engine. A dynamic weight adaptive adjustment mechanism is used for optimization, and a scenario adaptation calibration module is introduced. Simultaneously, constraint conflict visualization generates an adjustment decision chain, forming a basic planning model that integrates geographical, resource, and task coordination logic, including: Based on the requirements of constrained collaborative planning and the goal of full-scenario adaptation, a core architecture of hierarchical planning engine + dynamic weight adjustment module + scenario adaptation calibration unit is built to perform hierarchical parsing and collaborative logic learning on the embedded vectors of geographical, resource and task constraint features. The planning logic is designed based on a multi-constraint balancing strategy, which clarifies the spatial adaptation dimension of geographical constraints, the supply matching dimension of resource constraints, the priority dimension of task constraints, and the synergistic ratio of the three, and generates constraint synergistic planning rules. Based on the adaptation requirements of operation area type, sanitation resource stock, and task urgency, a dynamic weight adaptive adjustment mechanism is set up. In core urban areas, visual geography and resource supply constraints are emphasized; in emergency task scenarios, task priority and resource scheduling constraints are emphasized; and in suburban operation scenarios, geographical scope and cost control constraints are emphasized. The core architecture, constraint-coordinated planning rules, and dynamic weight adjustment mechanism are integrated and executed. By visualizing constraint conflicts, an adjustment decision chain is generated to optimize planning parameter configuration and form a basic planning model that integrates geographical, resource, and task coordination logic and adapts to the needs of all scenarios.

3. The method as described in claim 2, characterized in that, Based on a planning model combined with real-time traffic data streams and emergency task triggering signals, dynamic simulations are performed. A multi-objective optimization algorithm balances operational efficiency, resource utilization, and cost control indicators to generate a smart sanitation operation dynamic planning scheme and execution details suitable for all scenarios, including: Based on the planning basic model, combined with real-time traffic data stream and emergency task triggering signals, the sanitation operation scenario is dynamically simulated. A three-dimensional balance model of efficiency-resource-cost is constructed through multi-objective optimization algorithm. Input constraint feature embedding vectors are used to simulate the adaptability of planning schemes under different scenarios. When extracting real-time data features during dynamic simulation, the traffic congestion index and the emergency coefficient of sudden tasks are introduced. When analyzing the dynamic characteristics of the work scenario, we associate the resource supply status with the task execution time sequence, and when performing balance optimization analysis, we combine the dynamic change chain of constraints within the sliding time window. After the planning module outputs a preliminary planning scheme containing multi-dimensional adaptation indicators, it is verified and adjusted by comparing with the best historical operation case library to generate a smart sanitation full-scenario operation dynamic planning scheme and execution details with dynamic scenario adaptation.

4. A multi-constraint intelligent sanitation operation planning system for all scenarios, characterized in that, The system is configured to execute the intelligent planning method for multi-constraint smart sanitation operations across all scenarios as described in any one of claims 1 to 3 by executing executable instructions.

5. An electronic device, characterized in that, include: First processor; and memory for storing executable instructions of the first processor; Wherein, the first processor is configured to execute the intelligent planning method for multi-constraint smart sanitation full-scene operation as described in any one of claims 1 to 3 by executing the executable instructions.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the second processor, it implements the intelligent planning method for multi-constraint smart sanitation operations across all scenarios as described in any one of claims 1 to 3.

Images