Multi-dimensional digital twin driven tobacco industry chain collaborative optimization decision method and system

Abstract

The application discloses a multi-dimensional digital twin driven tobacco industry chain collaborative optimization decision method and system, belonging to the technical field of tobacco intelligent production. By gathering multi-source heterogeneous data of tobacco production in all links, a multi-dimensional digital twin body including five sub-models of plot resources, tobacco growth, process flow, equipment and facilities, and production management is constructed after space-time and semantic processing. Based on the twin simulation environment, an optimization system containing decision variables, multi-dimensional constraints and multi-objective functions is built, and the Pareto optimal production decision vector is solved by evolutionary algorithm iteration. At the same time, cross-system service registration, analysis and arrangement are completed, the standardized operation calling sequence is generated, and each intelligent module is driven to cooperatively execute agricultural operations. The operation data is fed back to the twin system in real time, realizing model updating and dynamic iterative optimization of production strategy. The whole-link intelligent, collaborative and precise optimization decision of tobacco is realized, and the product quality is effectively improved, and the production cost and resource consumption are reduced.

Description

Technical Field

[0001] This invention relates to the fields of agricultural systems engineering and digital twin technology, and in particular to a multi-dimensional digital twin-driven collaborative optimization decision-making method and system for the tobacco industry chain. Background Technology

[0002] With the deep integration of information technologies such as the Internet of Things, big data, and artificial intelligence with modern agriculture, the level of intelligence in all aspects of tobacco production has been continuously improving. The industry has formed a series of single-point intelligent technology solutions, such as precise sensing, intelligent bud suppression, variable crop protection, intelligent harvesting, adaptive curing, reliable traceability, and collaborative scheduling. These solutions can effectively address the pain points in each sub-stage of operation and provide basic technical support for the development of smart tobacco. However, existing technologies are all independently developed and deployed for single production stages, resulting in different system architectures and fragmented technical systems. The lack of a unified and standardized integrated architecture limits the overall production efficiency and the overall level of intelligence, and there are many technical shortcomings.

[0003] First, there is a significant problem of data and model silos. The data standards, transmission and storage formats of various intelligent systems are inconsistent, and the business model interfaces are closed and incompatible, resulting in fragmented and isolated core data related to the entire tobacco lifecycle, including growth, environment, operations, resources, costs, and quality. Existing technologies such as big data analytics and single-stage agricultural machinery collaboration have not built a unified data governance system spanning the entire tobacco production chain, failing to create a unified data view across the entire domain. This leads to incomplete and inconsistent data support for upper-level intelligent decision-making.

[0004] Secondly, there is the problem of local optima in single-stage decision-making, resulting in overall efficiency loss. Existing intelligent systems optimize only around their own stage objectives; for example, harvesting systems prioritize operational efficiency, and curing systems focus on the quality of a single curing operation. A global collaborative optimization mechanism centered on the final yield, quality, economic benefits, and resource consumption of tobacco leaves has not been established. Extreme optimization of a single stage easily leads to cross-process conflicts and internal efficiency losses. For instance, blindly increasing harvesting efficiency can damage tobacco quality, directly affecting subsequent curing and processing effects and overall profits, making it difficult to achieve optimal comprehensive benefits throughout the entire process. Thirdly, cross-stage production collaboration is highly dependent on manual labor, with insufficient automation and agility. Currently, the coordinated scheduling of harvesting, curing, plant protection, and resource allocation across processes and systems in tobacco production mainly relies on manual judgment and coordination by management personnel. This results in problems such as slow response, low accuracy, high cost, and low fault tolerance, failing to meet the needs of large-scale precision production. Furthermore, existing agricultural machinery collaborative technologies are limited to basic operational functions such as field path planning and equipment collision avoidance, without involving cross-process global optimization decision-making and coordinated control, thus failing to meet the needs of full-chain collaborative production. Finally, existing production control lacks forward-looking simulation and pre-decision-making capabilities. Current tobacco intelligent systems mostly focus on post-event data recording and real-time status monitoring, enabling only passive review and post-event adjustments. They cannot conduct pre-event virtual simulations, effect comparisons, or risk predictions for various production decision-making schemes. Production managers find it difficult to select the optimal production strategy through virtual trial and error, resulting in highly subjective decisions that lack scientific rigor and foresight, thus failing to achieve predictive intelligent control.

[0005] Digital twin technology, as a core technology for realizing real-time mapping, interactive fusion, and simulation between physical entities and virtual information, has been widely applied in complex scenarios such as complex industrial equipment and full lifecycle management, possessing technical advantages in multi-dimensional modeling, dynamic simulation, iterative optimization, and global collaboration. In modern agriculture, digital twins have been initially applied to closed, controllable, and short-cycle production scenarios such as greenhouse cultivation and aquaculture. However, for complex production systems involving tobacco fields—open environments, long growth cycles, deep coupling of multiple processes, and multi-stage linkages—a mature integrated digital twin application solution has not yet been developed. It cannot achieve full-link dynamic simulation, cross-stage collaborative optimization, and fully automated intelligent decision-making, representing a technological gap in the current field of smart tobacco agriculture. Therefore, there is an urgent need to design a digital twin-based intelligent decision-making and collaborative operation method for the entire tobacco process to address the industry pain points of existing technologies, such as data silos, locally optimal decision-making, low collaborative efficiency, and a lack of forward-looking simulation capabilities. Summary of the Invention

[0006] To break down the barriers between intelligent systems in various stages of smart tobacco production and solve the problems of data silos, fragmented decision-making, and inefficient collaboration, this paper provides a multi-dimensional digital twin-driven collaborative optimization decision-making method and system for the tobacco industry chain. This method can perform unified modeling, simulation optimization, and automatic collaborative decision-making across the entire tobacco production chain.

[0007] In a first aspect, the present invention provides a full-link decision-making method based on multi-dimensional digital twins and cross-link optimization, the method comprising: Multi-source heterogeneous data from various stages of tobacco production are acquired, and after spatiotemporal alignment and semantic association of the multi-source heterogeneous data, it is written into the full-element data lake of the data layer. The full-element data lake is used to store multi-source heterogeneous data from various stages of the entire tobacco production chain. Based on the multi-source heterogeneous data, multiple interconnected twin models are constructed and mapped in real time with the physical tobacco production system. These multiple twin models constitute a multidimensional dynamic digital twin, and the multidimensional dynamic digital twins interact with each other through a unified data bus. Based on the multidimensional dynamic digital twin, the optimization algorithm is run in a preset simulation environment based on a preset set of decision variables, a set of constraints, and a multi-objective function to perform simulation and output the optimal decision strategy that satisfies the set of constraints. Based on the optimal decision-making strategy, a series of cross-system task instructions are obtained. The cross-system task instructions are matched with the execution units of each link through the service registration and discovery mechanism to generate a cross-unit service call sequence. The job tasks are executed according to the service call sequence, and the execution results and generated running data are fed back to the data layer, driving the synchronous update of the twin model and the optimization of the optimal decision strategy.

[0008] Furthermore, the construction of multiple interconnected twin models based on the multi-source heterogeneous data, which are synchronously mapped in real time to the physical tobacco production system, includes: The land parcel resource twin model is constructed based on the geographic data, soil data, meteorological data, and pipeline status data in the full-element data. Based on the tobacco variety parameters and environmental perception data in the full-element data, the tobacco growth twin model is constructed in combination with the preset development model, and the tobacco growth status is output through the tobacco growth twin model. The process flow twin model is constructed based on the work tasks and work task requirements in the full-element data; A twin model of the equipment and facilities is constructed based on the location, operational status, and operational capabilities of the equipment in the full-element data. The production management twin model is constructed based on the personnel, tasks, material flow, and cost elements in the total data.

[0009] Furthermore, based on the multidimensional dynamic digital twin, the simulation is performed in a preset simulation environment using an optimization algorithm based on a preset set of decision variables, a set of constraints, and a multi-objective function, outputting the optimal decision strategy that satisfies the set of constraints, including: Obtain the controllable decision variables for each production stage, and generate the decision variable set based on the controllable decision variables; Obtain agronomic constraints, resource constraints, equipment capability constraints, and time window constraints, and generate the constraint condition set based on the agronomic constraints, resource constraints, equipment capability constraints, and time window constraints; Obtain production targets, quality targets, cost targets, and resource consumption targets, and generate the multi-objective function based on the production targets, quality targets, cost targets, and resource consumption targets; In the simulation environment provided by the multidimensional dynamic digital twin, a preset evolutionary algorithm is used to search the set of decision variables to determine the target decision vector that enables the multi-objective function to reach Pareto optimality while satisfying the set of constraints. The target decision vector is output as the optimal end-to-end production strategy.

[0010] Furthermore, in the simulation environment provided by the multidimensional dynamic digital twin, the step of using a preset evolutionary algorithm to search the set of decision variables and determine the target decision vector that enables the multi-objective function to achieve Pareto optimality while satisfying the set of constraints includes: In the preset simulation environment, multiple initial decision vectors are randomly generated, and an initial population is formed based on the multiple initial decision vectors; For each initial decision vector in the initial population, the entire process from the current moment to the end of the production season is simulated in the multidimensional dynamic digital twin to obtain the comprehensive target value corresponding to each initial decision vector; Based on the comprehensive objective value, the fitness of each initial decision vector is calculated using the multi-objective function, and all solutions of the current population are divided into Pareto front solutions and dominant solutions according to the fitness. Evolutionary operations such as selection, crossover, and mutation are performed on the current population to complete the population iteration and generate a new generation of population; If the new generation population satisfies the preset convergence condition, the Pareto front solution set in the final iteration population is extracted as the Pareto optimal objective decision vector, wherein the termination condition includes at least one of reaching the maximum number of iterations and the Pareto front solution set being stable with no optimization gain.

[0011] Furthermore, the step of obtaining a series of cross-system task instructions based on the optimal decision-making strategy, matching these cross-system task instructions with execution units at each stage through a service registration and discovery mechanism, and generating a cross-unit service call sequence includes: Based on the optimal strategy, a series of cross-system task instructions are obtained. After receiving the cross-system task instructions, the unified data bus sends a parsing request to the service orchestration engine. The orchestration engine sends a service discovery request to the service registry. Based on the service discovery request, the service registry queries the corresponding service interface address, invocation method, and service status. Based on the service interface address, invocation method, and service status, the cross-unit service invocation sequence is generated.

[0012] Furthermore, the step of executing job tasks according to the service call sequence and feeding back the execution results and generated running data to the data layer to drive the synchronous update of the twin model and the optimization of the optimal decision-making strategy includes: Send invocation instructions to the corresponding stages according to the service invocation sequence; If the service interface address of the corresponding link is found to be inconsistent, then the corresponding link is converted in terms of protocol and data format. Receive the execution status and result data returned by the corresponding link, and feed back the execution status and result data to the full-element data lake for storage; Extract real-time abnormal events from the execution status and result data, and drive the synchronous update of the twin model and the optimization of the optimal decision strategy based on the abnormal events.

[0013] Furthermore, the step of constructing the tobacco growth twin model based on the variety parameters and environmental perception data in the total data, combined with a preset developmental model, and outputting the tobacco growth status through the tobacco growth twin model includes: The extracted variety parameters and environmental perception multi-source heterogeneous data are preprocessed to generate a standardized model input dataset; The standardized dataset is imported into the preset development model in real time to complete the parameter initialization and dynamic iterative update of the development model, so that the updated development model can be adapted to the tobacco variety attributes and real-time growth environment of the current field. Based on the updated developmental model, the growth process of tobacco throughout its entire growth period is dynamically extrapolated, plant growth parameters and tobacco maturity index are quantitatively calculated, and the actual growth status of tobacco in the field is mapped in real time. Based on real-time growth status and maturity index, and by binding field resource information, growth time series data and process-related parameters, a tobacco growth twin model that is bidirectionally synchronized with the physical tobacco growth status in the field is constructed, and the tobacco growth status is output through the tobacco growth twin model.

[0014] Furthermore, the step of constructing the process flow twin model based on the operational logic and process parameters in the full-element data includes: Based on the tobacco growth status, a preset process model library is queried to determine the target operation time window and target operation parameters; Send an equipment query request to the equipment facility twin model to obtain a list of target equipment available within the target operation time window; Send a resource assessment request to the production management twin model to obtain resource confirmation information for the target operation time window and the target equipment list; The operation logic, equipment status, and process parameters are generated based on the target operation time window, the target operation parameters, the target equipment list, and the resource confirmation information. Based on the aforementioned operational logic, equipment status, and process parameters, a twin model of the process flow is constructed.

[0015] Secondly, the present invention provides a multi-dimensional digital twin-driven collaborative optimization decision-making system for the tobacco industry chain, the system comprising: a data layer, an optimization layer, a collaboration layer, and an execution layer; The data layer includes a full-element data lake, a multi-dimensional dynamic digital twin construction module, and a twin sub-model library. The full-element data lake is used to store multi-source heterogeneous data from all stages of the tobacco production chain. The multi-dimensional dynamic digital twin construction module creates a virtual twin sub-model library that is synchronously mapped with the entire tobacco production chain based on the multi-source heterogeneous data. The twin sub-model library includes land resource twin sub-models, tobacco growth twin sub-models, process flow twin sub-models, equipment and facility twin sub-models, and production management twin sub-models. The optimization layer includes a cross-link optimization decision engine, an optimization algorithm library, and a strategy simulation sandbox. The cross-link optimization decision engine is used to define the decision variables, constraints, and multi-objective functions of the entire tobacco production chain, call the algorithms in the optimization algorithm library, and complete the iterative simulation and selection of the scheme in the strategy simulation sandbox to obtain the optimal decision strategy. The collaboration layer includes a unified data bus, a service registry, and a service orchestration engine. The service registry is used to register and manage the service interfaces and operating status of the entire tobacco production chain. The service orchestration engine is used to parse the optimal decision strategy and decompose it to generate a standardized service call sequence. The execution layer includes multiple links covering the entire tobacco production chain. Each link is connected to the unified service bus as an independent service node, receives collaborative instructions to execute tasks, and feeds back the execution results and generated operation data to the data layer, driving the synchronous update of the twin model and the optimization of the optimal decision strategy.

[0016] The technical solutions provided by the embodiments of this invention can include the following beneficial effects: Addressing the industry drawbacks of traditional tobacco production, such as reliance on manual experience, fragmented production processes, low data utilization, delayed and one-sided decision-making, poor system coordination, and inability to dynamically iterate and optimize, this invention constructs a full-link intelligent decision-making system based on multi-dimensional digital twins and cross-process optimization. By aggregating multi-source heterogeneous data from all stages of tobacco production and completing standardized governance, a full-element data lake is built to break down data silos. Simultaneously, five interconnected and coupled digital twin sub-models are created, encompassing land resources, tobacco growth, technological processes, equipment and facilities, and production management, achieving real-time, two-way, and precise mapping between physical tobacco fields and virtual models. The system relies on a multi-objective evolutionary algorithm combined with multi-dimensional constraints to solve Pareto optimal production decisions, overcoming the limitations of human experience-based decision-making. It then achieves automated collaborative scheduling of heterogeneous businesses across systems through service registration and discovery and intelligent service orchestration technologies. Combined with production execution data feedback, it forms a closed-loop optimization mechanism of "simulation decision-making - execution implementation - data feedback - model iteration". It can dynamically adapt to various complex production scenarios, effectively achieving multi-objective collaborative optimization of tobacco production to improve quality, increase output, reduce costs, and reduce consumption. It comprehensively improves the digitalization, refinement, and intelligence level of tobacco production throughout its entire growth cycle and the entire industry chain, and has excellent scenario adaptability and engineering implementation value. Attached Figure Description

[0017] Figure 1 This is a flowchart of the multi-dimensional digital twin-driven collaborative optimization decision-making method for the tobacco industry chain according to the present invention.

[0018] Figure 2 This is a schematic diagram illustrating the relationships and data flow between the sub-models of the multidimensional dynamic digital twin of the present invention.

[0019] Figure 3 This is a schematic diagram of the iterative optimization process of the evolutionary algorithm of this invention. Detailed Implementation

[0020] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] like Figure 1As shown, this embodiment of a full-link decision-making method based on multi-dimensional digital twins and cross-link optimization specifically includes: Step S1: Obtain multi-source heterogeneous data from each stage of tobacco production, and write the multi-source heterogeneous data into the full-element data lake of the data layer after performing spatiotemporal alignment and semantic association. The full-element data lake is used to store multi-source heterogeneous data from each stage of the entire tobacco production chain. Here, multi-source heterogeneous data refers to a collection of data originating from different subsystems, different acquisition devices, and different storage structures. In one embodiment, multi-source heterogeneous data may include environmental data collected by field IoT sensors, such as soil temperature and humidity, electrical conductivity, light intensity, air temperature and humidity, wind speed, and rainfall; image data acquired by satellite remote sensing and UAV aerial photography, such as normalized difference vegetation index layers, land surface temperature layers, and high-resolution orthophotos; operational data reported by agricultural machinery, such as tractor location, fuel consumption, operating speed, and implement posture; and business data from production management systems, such as personnel schedules, agricultural input inventory tables, order contracts, and cost ledgers. Because these data differ in acquisition frequency, coordinate system, field definitions, and encoding formats, they cannot be directly used for subsequent twin modeling and decision-making calculations, and therefore require unified processing.

[0022] Specifically, this can be further broken down into the following sub-steps: First, connect to various data sources through a pre-defined data access gateway. This gateway supports multiple communication methods, including message queue protocols, file transfer protocols, and structured query protocols. Configure corresponding access adapters for different data sources and cache the raw data in the raw data area of ​​the full-element data lake. Second, perform time alignment on the raw data. Due to inconsistent sampling frequencies of different sensors (e.g., weather stations report every ten minutes, soil moisture sensors report every hour, and agricultural machinery positioning data reports every second), all data needs to be resampled and aligned according to a unified time benchmark. Linear interpolation or the nearest neighbor method can be used to fill in sparsely sampled data onto a unified time grid. Spatial alignment of the raw data is performed, projecting all data with geographical attributes onto the same coordinate reference system, such as a provincial coordinate system, and spatial clipping according to plot boundaries, so that each data point can be assigned to a specific plot unit; semantic association is performed, establishing a unified semantic ontology dictionary to map fields with different names but the same meaning from different sources. For example, a field returned by one sensor is named soil moisture, while a field returned by another sensor is named water content, and both are mapped to the same concept node in the semantic ontology dictionary; at the same time, the relationships between plots, tobacco, equipment, personnel, and tasks are established to form full-element data that can be directly accessed by the twin.

[0023] For example, in an application scenario at a tobacco production base, the multi-source heterogeneous data includes 240 environmental sensors deployed in 30 greenhouses, location and flow data from 12 self-propelled sprinkler irrigation machines, weekly multispectral images from drones, and work order data from the production management system. After the above processing steps, all data is organized into a full-element data table with plot number, tobacco batch number, and timestamp as the primary key. Each record contains the environmental status, tobacco status, equipment status, and management status of the plot at that moment. This organization method provides clean, aligned, and correlated input for subsequent digital twin modeling, avoiding repeated processing of data format issues during the modeling process.

[0024] Step S2: Based on the multi-source heterogeneous data, construct multiple interconnected twin models that are synchronously mapped with the physical tobacco production system in real time. The multiple twin models constitute a multidimensional dynamic digital twin, and the multidimensional dynamic digital twins interact with each other through a unified data bus. In this embodiment, as Figure 2 The diagram illustrates the relationships and data flows between the sub-models of the multidimensional dynamic digital twin. A land resource twin model is constructed based on geographical data, soil data, meteorological data, and pipeline status data from the comprehensive data set. A tobacco growth twin model is constructed based on tobacco variety parameters and environmental perception data from the comprehensive data set, combined with a pre-defined development model, and outputs the tobacco growth status through this model. A process flow twin model is constructed based on the operational tasks and requirements from the comprehensive data set. An equipment and facility twin model is constructed based on the equipment location, operational status, and operational capabilities from the comprehensive data set. A production management twin model is constructed based on personnel, tasks, material flow, and cost elements from the comprehensive data set.

[0025] A digital twin is a mirror image of a physical entity in digital space. A multidimensional dynamic digital twin emphasizes depicting a physical object from multiple dimensions and dynamically updating it as the physical object's state changes. In this embodiment, the multidimensional dynamic digital twin is composed of five sub-models, each responsible for the twin representation of one dimension. The sub-models communicate with each other via an event bus, collectively supporting overall decision-making.

[0026] Based on the geographic, soil, meteorological, and pipeline status data from the comprehensive data set, a land resource twin model is constructed. Specifically, the land resource twin model uses land parcels as the basic unit. Each parcel object includes attributes such as parcel number, boundary polygon, area, slope, soil texture, organic matter content, pH, irrigation pipeline route, water source access point, and historical planting records. The parcel can be further divided into several grid units. Each grid unit stores the real-time soil temperature, humidity, and meteorological conditions at that location, making the land resource twin model a raster layer with spatiotemporal attributes. For example, a parcel of 30 mu (approximately 2 hectares) can be divided into approximately 200 grid units using a 10-meter by 10-meter grid. Each grid unit independently maintains a set of sensor mappings, supporting the decision-making needs of precision irrigation and precision fertilization.

[0027] Based on the tobacco variety parameters and environmental sensing data in the comprehensive data set, a tobacco growth twin model is constructed using a pre-defined developmental model. This model outputs the tobacco growth status. The developmental model can be an effective accumulated temperature model based on accumulated temperature, or a comprehensive developmental model combining light, temperature, and water. Variety parameters include the basic growth period days for the tobacco variety, accumulated temperature requirements for each growth stage, photoperiod sensitivity, and water and fertilizer requirement curves. Environmental sensing data includes real-time temperature, soil moisture, and canopy photosynthetically active radiation collected by sensors. The tobacco growth twin model uses field batches as the basic unit. Each batch includes attributes such as variety, sowing date, current growth stage, accumulated temperature, plant height, leaf area index, and estimated yield.

[0028] Furthermore, a tobacco growth twin model was constructed, employing a dynamic iterative update mechanism to achieve real-time synchronization with the physical tobacco field. The model uses the planting batch as the smallest management unit, continuously accessing real-time environmental perception data collected by field sensors, linking inherent varietal parameters with the crop development model, and periodically correcting core status indicators such as cumulative accumulated temperature, growth progress, plant height, and leaf area index.

[0029] The model is internally structured into three layers: a data access layer, a growth projection layer, and a state output layer. The data access layer is responsible for aggregating heterogeneous sensing data such as real-time temperature and humidity, photosynthetic radiation, and soil water and fertilizer, and completing standardized preprocessing. The growth projection layer relies on a preset effective accumulated temperature model or a light, temperature, and water integrated development model, combined with the growth characteristics of tobacco varieties, to simulate and project the real-time growth status of the crop. The state output layer uniformly encapsulates the results data such as the current growth stage of tobacco, growth characteristics, water and fertilizer gap, and estimated yield, and pushes them to other twin models through a unified data bus.

[0030] Meanwhile, the model supports historical growth data review and future growth prediction. By combining historical planting records and weather forecast data, it can predict subsequent growth stages of tobacco and the needs for water, fertilizer, and plant protection operations, providing accurate data support for subsequent process matching and operation scheduling. The model status is dynamically updated every second / minute according to the field physical environment and crop growth, ensuring a high degree of consistency between digital mapping and physical entities.

[0031] The process flow twin model is constructed based on the work tasks and work task requirements in the full-element data; In this embodiment, a preset process model library is queried based on the tobacco growth status to determine the target operation time window and target operation parameters; an equipment query request is sent to the equipment facility twin model to obtain a list of target equipment available within the target operation time window; a resource assessment request is sent to the production management twin model to obtain resource confirmation information for the target operation time window and the target equipment list; operation logic, equipment status, and process parameters are generated based on the target operation time window, the target operation parameters, the target equipment list, and the resource confirmation information; and a process flow twin model is constructed based on the operation logic, equipment status, and process parameters.

[0032] In this embodiment, the process flow twin sub-model is the core digital mirror model of the entire tobacco production process. It focuses on all processes of field production operations, relies on real operation tasks and task-related requirements, and links with other sub-model data of the multi-dimensional digital twin to achieve full-dimensional digital mapping, simulation, and dynamic adaptation of production technology, operation process, equipment scheduling, and resource allocation. It receives the output data of various twin models of land resources, tobacco growth, equipment facilities, and production management, forming a closed-loop digital control system for process operations.

[0033] In the specific implementation process, the system first relies on the real-time tobacco growth status output by the pre-constructed tobacco growth twin model, including core growth indicators such as the tobacco's growth stage, plant height, leaf area index, water and fertilizer deficit, and risk of pests and diseases. It then searches the system's pre-set standardized tobacco process model library. This library pre-packages standard process specifications, operational adaptation conditions, parameter thresholds, and time adaptation rules for all operations throughout the tobacco growth cycle, including seedling raising, transplanting, irrigation, fertilization, plant protection, and harvesting. The system uses a growth feature matching algorithm to accurately match the corresponding production process, locking in the target operation time window and target operation process parameters suitable for the current tobacco growth status. The target operation time window distinguishes between optimal operation periods, operable periods, and prohibited operation periods. The target operation parameters include refined process indicators such as operation depth, operation rate, water and fertilizer application, pesticide ratio, and operation coverage.

[0034] After determining the process parameters and operation time window, the system initiates a targeted equipment query request to the equipment and facility twin model. The request information includes key conditions such as the target operation type, operation site location, target operation time window, and operation scale. The equipment and facility twin model verifies the operating status, location information, endurance, equipment load, fault warning status, and time slot occupancy of each field operation equipment in real time. It then filters out a list of target equipment that has no operation conflicts, is in normal status, and can match the current operation requirements within the target operation time window. Simultaneously, it returns supplementary information such as the rated operation parameters and suitable operating conditions of each target equipment.

[0035] To ensure the feasibility and implementability of the operational plan, the system further sends a resource assessment request to the production management twin model, submitting the target operation time window, target operation type, target equipment list, and operation coverage area as core assessment parameters. This model, combining manpower scheduling plans, agricultural input inventory levels, water and electricity supply capacity, operation scheduling priorities, and historical operation log data, conducts a comprehensive resource verification and feasibility assessment, ultimately outputting resource confirmation information, including whether manpower allocation is sufficient, whether agricultural input resources are matched, whether time-slot scheduling is feasible, whether resource conflicts exist, and corresponding resource optimization and adjustment suggestions.

[0036] The system integrates all the aforementioned core data, taking operational compliance, optimal resource allocation, and improved production quality and efficiency as its core principles, to generate standardized operational logic, equipment operating status parameters, and a complete set of process execution parameters. The operational logic includes the sequence of operational procedures, the order of operations per plot, equipment scheduling logic, abnormal operating condition handling mechanisms, and operation start / stop conditions; equipment status includes the preset operating speed, operating speed, working width, start / stop sequence, and load threshold of the operating equipment; and process parameters include directly executable parameters such as precise irrigation volume, fertilizer ratio and dosage, plant protection agent concentration, operating temperature and humidity adaptation thresholds, and operation interval cycles.

[0037] Finally, based on the generated operation logic, equipment status, and process parameters, and combined with real-time operation task requirements, a process flow twin model is constructed. This model uses a single operation task as the smallest modeling unit, binding it to the corresponding plot number, tobacco planting batch, operating equipment, and resource configuration scheme. Internally, the model encapsulates three core databases: a process rule library, an operation scheduling logic library, and an equipment parameter adaptation library. Simultaneously, the process flow twin model interacts bidirectionally with the plot resource, tobacco growth, equipment facilities, and production management twin models in real time through the system's unified data bus. It can dynamically receive status change data from the physical production scenario, updating process schemes and operation flows in real time. It supports pre-operation simulation and pre-playback, real-time monitoring during operation, and post-operation process review, achieving full-cycle digital, visual, and precise control of the tobacco production process, providing core process model support for standardized and intelligent tobacco production operations.

[0038] Based on the equipment's location, operational status, and operational capabilities from the comprehensive data, a twin model of the equipment and facilities is constructed. First, the comprehensive data of the equipment is aggregated and standardized, collecting three core basic data categories covering all field operation equipment. These include: Equipment location data (real-time GPS coordinates, operational coverage radius, fixed deployment points, activity range, and spatial distribution information of each irrigation device, plant protection drone, fertilization machinery, and harvesting equipment, accurately mapping the spatial layout of equipment in tobacco fields); Equipment operational status data (dynamic operating conditions such as equipment start / stop status, real-time operating power, remaining battery / fuel, operational load, fault alarm signals, idle / occupied status, and historical start / stop sequences); and Equipment operational capability data (inherent performance parameters such as rated operating width, operating speed, operating area per unit time, maximum water and fertilizer output flow rate, pesticide spraying accuracy, equipment-compatible operation type, operating condition thresholds, extreme operating parameters, and equipment wear and tear coefficient).

[0039] After data preprocessing, independent digital twin sub-objects are constructed using a single piece of equipment as the smallest modeling unit. Each digital equipment object is uniquely assigned a device ID and bound to the physical equipment, forming a comprehensive digital ledger for all equipment. The model constructs multi-dimensional attribute profiles for each equipment object, covering four dimensions: spatial location attributes, real-time status attributes, operational performance attributes, and equipment lifecycle attributes. This comprehensively records the equipment's static parameters and dynamic operational data, achieving a digital mapping of all attributes of the physical equipment. Simultaneously, the model incorporates a time-series update mechanism, synchronizing real-time changes in the physical equipment's location, status switching, and load variations at a second-level acquisition frequency. This ensures real-time synchronization between the digital model and the physical entity's status, location, and operational capabilities, eliminating spatiotemporal mapping discrepancies.

[0040] Building upon this foundation, the model incorporates algorithms for assessing equipment operational capabilities and predicting working conditions. By combining real-time equipment status and spatial location, it dynamically calculates the equipment's current available operational capacity. For mobile equipment, it can predict the area it can cover, the duration of continuous operation, and the area it can complete, based on the equipment's real-time location, remaining battery life, and operating speed. For fixed equipment, it can verify whether it meets the current site's irrigation, fertilization, and monitoring requirements based on the equipment's operating status and output parameters. Simultaneously, the model can identify abnormal operating conditions in real time, such as equipment malfunctions, overload operation, insufficient battery life, and location deviation. It automatically marks unusable equipment and outputs status tags such as idle, busy, faulty, and awaiting maintenance, providing accurate data for selecting operational equipment.

[0041] Once the equipment and facility twin model is constructed, it can respond to equipment query requests from the process flow twin model. Based on the target operation type, operation site location, operation time window, and operation scale requirements, it automatically traverses all digital equipment resources in the entire domain, filters out the optimal equipment list that matches spatial location, adapts to operational capabilities, is available during the time period, and is in normal working condition, and simultaneously outputs the corresponding equipment's operable parameters, suitable operating conditions, and operational constraints. This model participates in the entire production process scheduling and control, realizing digital management, intelligent screening, and precise scheduling of tobacco field equipment resources, ensuring efficient, rational, and orderly allocation of equipment resources for field production operations.

[0042] The production management twin model is constructed based on the personnel, tasks, material flow, and cost elements in the total data.

[0043] In this embodiment, the production management twin model belongs to the resource scheduling and control decision-making layer model of the multi-dimensional dynamic digital twin. It mainly realizes the digital mirror mapping and dynamic control of manpower, materials, tasks, and costs throughout the entire tobacco field production process. Unlike the physical entity twin models of plots, crops, and equipment, this model focuses on production soft resources and management processes. It takes personnel configuration information, work task lists, material flow trajectories, and production cost elements from the full-element data as the core drivers to realize the simulation and deduction of production resource matching, task timing scheduling, material supply and demand balance, and dynamic accounting of production costs. It can be linked in real time with the plot resources, tobacco growth, equipment facilities, and process flow twin models through a unified data bus, providing resource feasibility verification and management decision support for the implementation of process solutions and the execution of equipment scheduling.

[0044] In the specific implementation process, the system first collects and analyzes data on all dimensions of production management elements, completing the standardized storage and mapping binding of four core resource data. Among them, personnel data includes the number of field workers, job division, skill level, shift schedule, on-duty status, available working hours, and personnel workload; task data includes the type of task to be performed, the scope of the work area, the time sequence requirements, the quality standards, task priority, the schedule, and the relationship between upstream and downstream processes; material flow data includes the inventory balance, location, requisition records, transportation and circulation paths, consumption rate, and replenishment cycle of production materials such as seeds, fertilizers, pesticides, and irrigation water; and cost element data includes the unit price, pricing rules, and historical cost ledger data of various sub-costs such as labor costs, material procurement costs, equipment energy consumption costs, operation and maintenance costs, and management costs.

[0045] The model uses a single batch production task as the smallest management modeling unit, building a digital mirror system integrating personnel, tasks, materials, and costs. Regarding personnel, the model maps the on-duty status and workload of each worker in real time, automatically verifying manpower matching based on task operation periods. It determines whether the skills, number, and working hours of currently on-duty personnel meet the target operation requirements, avoiding issues such as manpower shortages, job mismatches, and overwork. Regarding tasks, the model breaks down all field operation tasks into time sequences and sorts them by process, binding them to corresponding operation time windows, operation plots, and process requirements. It establishes a task execution progress ledger, synchronizing the physical scene's task status in real time, including pending, in progress, completed, and abnormally stalled states.

[0046] Regarding the material flow dimension, the model fully replicates the dynamic flow of production materials. It intelligently calculates material requirements based on the upcoming workload, compares current inventory levels with material consumption needs in real time, and accurately predicts material shortages, replenishment points, and material allocation paths. This enables digital traceability of the entire process of material requisition, deployment, consumption, and replenishment, ensuring on-demand supply of materials for irrigation, fertilization, and plant protection operations, and preventing material shortages from delaying operations or material redundancy from wasting resources. Regarding the cost dimension, the model relies on preset cost accounting rules and combines real-time data such as man-hours, material consumption, equipment energy consumption, and operation duration to dynamically accumulate and calculate the production costs of a single operation, a single plot, and a single batch of planting, forming a real-time dynamic cost ledger. This enables the monitoring, traceability, and prediction of production costs.

[0047] Furthermore, the production management twin model possesses core capabilities for resource assessment and conflict verification. It can receive resource assessment requests initiated by the process flow twin model and conduct multi-dimensional comprehensive resource verification based on factors such as the target operation time window, target operation type, list of equipment to be scheduled, and operation scale. On the one hand, it verifies whether the manpower scheduling matches the time period, whether the material inventory is sufficient, and whether the material flow timeliness meets the operation schedule. On the other hand, it verifies resource occupation conflicts, equipment scheduling conflicts, and manpower time conflicts under multiple overlapping tasks. Finally, it outputs standardized resource confirmation information, including resource matching results, existing resource conflicts, material shortage warnings, manpower optimization plans, material allocation suggestions, and cost estimation results.

[0048] Meanwhile, the model continuously maintains dynamic synchronization with the physical production management system, updating its status in real time based on daily personnel shift changes, material inflow and outflow, task execution progress, and cost consumption changes. Within the overall digital twin system, this model serves as a bridge between upper-level decision-making and lower-level operational execution. It provides resource feasibility verification for process flow plans and also optimizes operational timing, material delivery, and personnel scheduling strategies, achieving optimal resource allocation, efficient process control, and precise cost control in tobacco production operations. This supports a closed-loop intelligent and refined production management system for tobacco fields.

[0049] Understandably, although each twin model is described independently, the five sub-models run concurrently at runtime and exchange messages through a shared event bus. For example, if the land resource twin model detects that the soil moisture of a certain plot has fallen below the threshold, it will issue a water shortage event. The tobacco growth twin model will then adjust its growth rate accordingly, the process flow twin model will generate candidate irrigation operation events, the equipment and facility twin model will reserve sprinkler resources, and the production management twin model will budget water and labor costs, forming a collaborative closed loop across the sub-models.

[0050] Step S3: Based on the multidimensional dynamic digital twin, perform simulation in a preset simulation environment using an optimization algorithm based on a preset set of decision variables, a set of constraints, and a multi-objective function, and output the optimal decision strategy that satisfies the set of constraints.

[0051] In this embodiment, controllable decision variables for each production stage are obtained, and a set of decision variables is generated based on these controllable decision variables; agronomic constraints, resource constraints, equipment capability constraints, and time window constraints are obtained, and a set of constraint conditions is generated based on these constraints; output targets, quality targets, cost targets, and resource consumption targets are obtained, and a multi-objective function is generated based on these targets; in the simulation environment provided by the multi-dimensional dynamic digital twin, a preset evolutionary algorithm is used to search the set of decision variables to determine the target decision vector that enables the multi-objective function to achieve Pareto optimality while satisfying the set of constraint conditions; the target decision vector is output as the optimal end-to-end production strategy.

[0052] Specifically, based on the five existing twin models of land resources, tobacco growth, equipment and facilities, production management, and technological processes, a virtual simulation environment covering the entire growth cycle, all production factors, and the entire operational process of tobacco is built. This simulation environment can completely replicate the resource status, crop growth, equipment condition, and management conditions of a physical tobacco field. It supports virtual simulation of production operation processes, parameter calculation, scheme comparison, and optimization iteration, providing a high-fidelity, reproducible, and iterative digital simulation platform for production decision optimization. This step constructs standardized decision variables, multi-dimensional constraints, and multi-objective optimization functions, and uses a preset evolutionary algorithm to complete global optimization, ultimately outputting the optimal production decision strategy across the entire chain that is adapted to the current tobacco field conditions.

[0053] First, a set of controllable decision variables for the system is constructed. Combining the operational characteristics of each stage of tobacco production, controllable decision variables for each stage are extracted, covering four categories of adjustable parameters: time, process, equipment, and resources. These mainly include: time-related decision variables such as start-up time, duration, and interval of operations at each growth stage; process-related decision variables such as irrigation frequency, irrigation volume per irrigation, fertilizer ratio, fertilizer application rate, pesticide concentration, and operational intensity; equipment-related decision variables such as equipment selection, equipment travel speed, operating width, and equipment start-up / shutdown sequence; and resource scheduling decision variables such as personnel scheduling, material input, material allocation time periods, and energy consumption allocation ratio. All independently controllable and iteratively simulated parameters are standardized, coded, and integrated to form a complete set of decision variables, serving as the global optimization space for the optimization algorithm.

[0054] Secondly, a comprehensive and closely aligned set of constraints is constructed to ensure that the optimized production plan fully complies with tobacco agronomical patterns, on-site resource conditions, equipment operational capabilities, and production time requirements, eliminating decisions that exceed regulations, limits, or are unfeasible. These constraints specifically include four core categories: First, agronomical constraints, strictly adhering to the accumulated temperature requirements, water and fertilizer sensitivity periods, growth rhythms, pest and disease control windows, and standardized planting specifications for each growth stage of tobacco, ensuring that all operational parameters and timing meet the requirements for normal tobacco growth and development; second, resource constraints, combining real-time data from the production management twin model to constrain resource boundaries such as the number of personnel on duty, total agricultural input inventory, upper limits of water and electricity supply, and production cost limits; third, equipment capability constraints, based on the rated parameters and real-time operating conditions of the equipment in the equipment twin model, to constrain hardware capability boundaries such as maximum operating width, fastest operating speed, maximum water and fertilizer output flow, equipment load threshold, and continuous operating time; and fourth, time window constraints, based on the tobacco growth status and meteorological conditions, locking in the optimal operating time, operable time, and prohibited operating time, ensuring that all operations are completed within the effective time window. All hard boundary conditions are integrated to construct a set of constraints, which serve as the feasible region restriction rules for algorithm optimization.

[0055] Furthermore, a multi-objective function is constructed that balances production efficiency and production costs to achieve synergistic optimization of multi-dimensional production objectives. This embodiment establishes four core optimization objectives, forming a multi-objective function system: First, the yield objective, focusing on maximizing tobacco final yield, effective leaf count, and yield per plant, constructing a yield gain objective function; second, the quality objective, focusing on minimizing tobacco leaf uniformity, maturity, chemical composition coordination, and pest and disease incidence, constructing a quality optimization objective function; third, the cost objective, focusing on minimizing labor costs, material costs, equipment energy consumption, and operation and maintenance costs, constructing a production cost control objective function; and fourth, the resource consumption objective, focusing on optimizing irrigation water consumption, fertilizer and pesticide input, and energy consumption, constructing a resource reduction objective function. These four objectives are coupled and partially mutually constraining, forming a typical multi-objective optimization solution scenario.

[0056] Finally, in the simulation environment provided by the multidimensional dynamic digital twin, a preset evolutionary algorithm is used to search the set of decision variables. This preset evolutionary algorithm includes mainstream multi-objective evolutionary algorithms such as NSGA-II and MOPSO. The objective decision vector that enables the multi-objective function to achieve Pareto optimality while satisfying the set of constraints is determined. See details... Figure 3The flowchart illustrates the iterative optimization process of the evolutionary algorithm. In the preset simulation environment, multiple initial decision vectors are randomly generated, forming an initial population. For each initial decision vector in the initial population, the entire process from the current moment to the end of the harvest season is simulated in the multidimensional dynamic digital twin, obtaining the comprehensive objective value corresponding to each initial decision vector. Based on the comprehensive objective value, the fitness of each initial decision vector is calculated using the multi-objective function. Based on the fitness, all solutions in the current population are divided into Pareto front solutions and dominant solutions. The fitness is calculated by substituting the original comprehensive objective value output from the twin simulation into a preset multi-objective function set to obtain a multidimensional objective vector, which is the individual fitness. The fitness vectors of all individuals in the population are compared pairwise, following the Pareto dominance criterion: if all objectives of solution A are no worse than those of solution B, and at least one objective is significantly better than solution B → A dominates B, then B... The solution is classified as a dominant solution (inferior solution); if no individual can dominate the solution, it is classified as a Pareto front solution (high-quality non-dominated solution). Based on the above rules, the entire population is divided into two categories: Pareto front solution set and dominant solution set, providing a quantitative basis for evolutionary selection. Selection, crossover, and mutation evolutionary operations are performed on the current population to iterate and generate a new generation of population. If the new generation of population meets the preset convergence conditions, the Pareto front solution set in the final iterative population is extracted as the Pareto optimal objective decision vector. The termination conditions include at least one of reaching the maximum number of iterations and the Pareto front solution set being stable with no optimization gain. During the simulation, the system uses the real-time twin mapping state as the initial simulation condition. Within the constraints of the entire set of constraints, virtual operation simulations are performed on each combination of decision variables. The multi-objective function values ​​corresponding to different production schemes are calculated in batches. Through selection, crossover, and mutation iterative optimization, inferior solutions are continuously eliminated, high-quality feasible solutions are retained, and the multi-objective collaborative optimal solution set is gradually approached. This embodiment adopts the Pareto optimality evaluation mechanism, which no longer pursues the ultimate optimality of a single objective, but seeks the global equilibrium optimal solution among output improvement, quality improvement, cost reduction and resource conservation, and selects the objective decision vector with the highest comprehensive benefit, the strongest adaptability and the best implementation.

[0057] Finally, the system performs structured analysis and encapsulation of the target decision vector, and outputs the optimal end-to-end production strategy. This strategy includes a precise operation timing scheme, the optimal combination of process parameters, equipment scheduling scheme, human and material allocation scheme, and cost control scheme. It can be directly distributed to the process flow twin sub-model to complete the process update, guide the physical tobacco field to carry out standardized, precise, and intelligent production operations, and achieve multi-objective collaborative optimal management and control of tobacco production to "improve quality, increase production, reduce costs, and reduce consumption".

[0058] Step S4: Based on the optimal decision-making strategy, obtain a series of cross-system task instructions, and match the cross-system task instructions with the execution units of each link through the service registration and discovery mechanism to generate a cross-unit service call sequence.

[0059] In this embodiment, a series of cross-system task instructions are obtained based on the optimal strategy scheme. After receiving the cross-system task instructions, the unified data bus sends a parsing request to the service orchestration engine. The orchestration engine sends a service discovery request to the service registry. Based on the service discovery request, the service registry queries the corresponding service interface address, invocation method, and service status. Based on the service interface address, invocation method, and service status, the cross-unit service invocation sequence is generated.

[0060] In this embodiment, the optimal decision-making strategy output through digital twin simulation optimization is a standardized and structured production control scheme, covering time-series operation plans, process parameter configurations, equipment scheduling rules, resource allocation schemes, and cost control rules. This optimal strategy cannot directly drive the execution of the underlying hardware and business systems. It needs to be decomposed into cross-system task instructions that can be recognized and executed by each business unit and equipment unit through task parsing, service discovery, service orchestration, and call adaptation. Relying on the platform service registration and discovery mechanism, multi-unit service matching and linkage scheduling are completed, ultimately forming an orderly, implementable, and executable cross-unit service call sequence, realizing a closed-loop connection from digital decision-making to physical execution.

[0061] In practical implementation, the system first decomposes and encapsulates the optimal decision-making strategy into structured instructions. Combining the multi-dimensional decision results included in the optimal strategy, such as operation sequence, process parameters, equipment scheduling, human and material allocation, and energy consumption control, the system automatically generates a series of cross-system task instructions covering planting resource management, crop growth regulation, field equipment operations, production resource scheduling, and process execution, according to business links and execution entity boundaries. All task instructions uniformly follow the platform's data interaction specifications, including task type, execution location, execution time period, target process parameters, execution equipment number, resource allocation quota, quality threshold, and anomaly handling rules, achieving unified instruction formats for multi-source heterogeneous execution units.

[0062] After generating cross-system task instructions, the unified data bus receives all standardized task instructions and performs legality verification, timing sorting, and classification aggregation on the instructions. It removes conflicting instructions and completes instructions with missing parameters. Then, it sends instruction parsing and task orchestration requests to the platform service orchestration engine. As the top-level scheduling core, the service orchestration engine is responsible for parsing the business logic, execution dependencies, sequence of operations, parallel relationships, and unit affiliation of cross-system task instructions. It clarifies the business execution stage, required service type, hardware execution unit, and resource support unit corresponding to each instruction, providing a logical basis for accurate matching of subsequent services.

[0063] Furthermore, based on the parsed task service requirements, the service orchestration engine initiates a real-time service discovery request to the platform's service registry. The service registry pre-stores registration information for all business services, device services, and resource services in the system, including each service's unique service identifier, service interface address, protocol invocation method, service permissions, service load status, online running status, and available invocation time period. The service registry responds to service discovery requests in real time, accurately retrieving and matching the corresponding service interface address, standardized invocation method, and current service running status based on the task's business type and execution requirements. Simultaneously, it automatically filters out offline, faulty, high-load, and unavailable abnormal services to ensure the reliability of the scheduling service.

[0064] Finally, based on the valid service interface addresses, standard calling methods, and real-time service status obtained from the query, the service orchestration engine performs global orchestration of all cross-system tasks, combined with the pre- and post-dependencies between tasks, parallel operation rules, and time-sequenced execution windows. By integrating multi-unit service resources, streamlining task execution logic, and solidifying the interface call order, a complete, orderly, and conflict-free cross-unit service call sequence is generated. This call sequence clarifies the service call order, call timing, call parameters, and return verification rules for each business unit, equipment unit, and control unit. It can directly drive each execution unit to complete precise tobacco field production operations step by step, in parallel, or collaboratively, realizing the precise implementation and automated closed-loop execution of the digital twin's optimal decision-making strategy to the physical production execution end.

[0065] In this embodiment, a virtual simulation environment covering the entire tobacco growth cycle, all production factors, and the entire operational process is built based on five pre-constructed twin models: land resources, tobacco growth, equipment and facilities, production management, and technological processes. This simulation environment can completely replicate the resource status, crop growth, equipment condition, and management conditions of a physical tobacco field. It supports virtual simulation of production operation processes, parameter calculations, scheme comparisons, and iterative optimization, providing a high-fidelity, reproducible, and iterative digital simulation platform for production decision optimization. This step constructs standardized decision variables, multi-dimensional constraints, and multi-objective optimization functions, employs an intelligent evolutionary algorithm to achieve global optimization, and ultimately outputs the optimal production decision strategy across the entire chain that is adapted to the current tobacco field conditions.

[0066] In the specific implementation process, the first step is to construct a set of controllable decision variables for the system. Combining the operational characteristics of each stage of tobacco production, controllable decision variables for each production stage are extracted, covering four major categories of adjustable parameters: time, process, equipment, and resources. These mainly include: time-related decision variables such as start-up time, duration, and interval of operations at each growth stage; process-related decision variables such as irrigation frequency, single irrigation volume, fertilizer ratio, fertilizer application rate, pesticide concentration, and operational intensity; equipment-related decision variables such as equipment selection, equipment travel speed, operating width, and equipment start-up and shutdown sequence; and resource scheduling decision variables such as personnel scheduling, material input, material allocation time periods, and energy consumption allocation ratio. All independently controllable and iteratively simulated parameters are standardized, coded, and integrated to form a complete set of decision variables, serving as the global optimization space for the optimization algorithm.

[0067] Secondly, a comprehensive and closely aligned set of constraints is constructed to ensure that the optimized production plan fully complies with tobacco agronomical patterns, on-site resource conditions, equipment operational capabilities, and production time requirements, eliminating decisions that exceed regulations, limits, or are unfeasible. These constraints specifically include four core categories: First, agronomical constraints, strictly adhering to the accumulated temperature requirements, water and fertilizer sensitivity periods, growth rhythms, pest and disease control windows, and standardized planting specifications for each growth stage of tobacco, ensuring that all operational parameters and timing meet the requirements for normal tobacco growth and development; second, resource constraints, combining real-time data from the production management twin model to constrain resource boundaries such as the number of personnel on duty, total agricultural input inventory, upper limits of water and electricity supply, and production cost limits; third, equipment capability constraints, based on the rated parameters and real-time operating conditions of the equipment in the equipment twin model, to constrain hardware capability boundaries such as maximum operating width, fastest operating speed, maximum water and fertilizer output flow, equipment load threshold, and continuous operating time; and fourth, time window constraints, based on the tobacco growth status and meteorological conditions, locking in the optimal operating time, operable time, and prohibited operating time, ensuring that all operations are completed within the effective time window. All hard boundary conditions are integrated to construct a set of constraints, which serve as the feasible region restriction rules for algorithm optimization.

[0068] Furthermore, a multi-objective function is constructed that balances production efficiency and production costs to achieve synergistic optimization of multi-dimensional production objectives. This embodiment establishes four core optimization objectives, forming a multi-objective function system: First, the yield objective, focusing on maximizing tobacco final yield, effective leaf count, and yield per plant, constructing a yield gain objective function; second, the quality objective, focusing on minimizing tobacco leaf uniformity, maturity, chemical composition coordination, and pest and disease incidence, constructing a quality optimization objective function; third, the cost objective, focusing on minimizing labor costs, material costs, equipment energy consumption, and operation and maintenance costs, constructing a production cost control objective function; and fourth, the resource consumption objective, focusing on optimizing irrigation water consumption, fertilizer and pesticide input, and energy consumption, constructing a resource reduction objective function. These four objectives are coupled and partially mutually constraining, forming a typical multi-objective optimization solution scenario.

[0069] Finally, in a high-fidelity simulation environment constructed from a multi-dimensional dynamic digital twin, a pre-defined evolutionary algorithm is used to perform global iterative optimization of the decision variable set. During the simulation, the system uses the real-time twin mapping state as the initial simulation condition. Within the constraints of the entire set of constraints, it performs virtual operation simulations for each combination of decision variables, batch calculates the multi-objective function values ​​corresponding to different production schemes, and continuously eliminates inferior schemes and retains high-quality feasible schemes through selection, crossover, and mutation iterative optimization, gradually approaching the multi-objective collaborative optimal solution set. This embodiment adopts a Pareto optimality evaluation mechanism, no longer pursuing the ultimate optimality of a single objective, but seeking a globally balanced optimal solution among output improvement, quality improvement, cost reduction, and resource conservation, and selecting the objective decision vector with the highest comprehensive benefit, strongest adaptability, and best implementation.

[0070] Finally, the system performs structured analysis and encapsulation of the target decision vector, and outputs the optimal end-to-end production strategy. This strategy includes a precise operation timing scheme, the optimal combination of process parameters, equipment scheduling scheme, human and material allocation scheme, and cost control scheme. It can be directly distributed to the process flow twin sub-model to complete the process update, guide the physical tobacco field to carry out standardized, precise, and intelligent production operations, and achieve multi-objective collaborative optimal management and control of tobacco production to "improve quality, increase production, reduce costs, and reduce consumption".

[0071] Step S5 involves executing the job task according to the service call sequence and feeding back the execution results and generated running data to the data layer, driving the synchronous update of the digital twin model and the optimization of the optimal decision-making strategy. This step establishes a closed loop between the physical execution end and the digital twin end, enabling decisions to be implemented as actual tasks and allowing the actual results of the tasks to correct the state of the digital twin, forming a continuous rolling decision-making and execution cycle.

[0072] In this embodiment, a call instruction is sent to the corresponding stage according to the service call sequence; if the service interface address of the corresponding stage is found to be inconsistent, the corresponding stage is converted in terms of protocol and data format; the execution status and result data returned by the corresponding stage are received, and the execution status and result data are fed back to the full-element data lake for storage; real-time abnormal events are extracted from the execution status and result data, and the abnormal events drive the synchronous update of the twin model and the optimization of the optimal decision strategy.

[0073] Specifically, the system sends call instructions to the intelligent system of the process according to the service call sequence. The message routing module pushes the instructions to the target system according to the call timing marked in the sequence. The push method can be synchronous blocking call or asynchronous message delivery. For instructions with high real-time requirements, such as opening and closing irrigation valves, synchronous calls are used, and the message routing module considers the call complete only after receiving a response from the target system. For instructions with lower real-time requirements, such as reserving harvesting vehicles, asynchronous message delivery is used, and the target system writes the result back to the designated callback channel after processing. The message routing module records an independent execution document for each call, which stores the call number, initiation time, target system, parameter snapshot, current status, most recent heartbeat timestamp, and final result.

[0074] If inconsistencies are detected in the service interfaces of the intelligent systems at each stage, protocol and data format conversion is performed on the corresponding stage. In agricultural production scenarios, intelligent systems at each stage are built by different vendors at different times, often resulting in differences in interface specifications. For example, some systems use remote procedure calls based on structured text, some use binary messages based on industrial fieldbus, and some use cloud message subscription models. The execution monitoring and adapter provide pre-built conversion plugins for each type of protocol. When the instruction format issued by the orchestration engine is inconsistent with the target system's expected format, the adapter performs transparent conversion before and after the call.

[0075] Specifically, protocol conversion involves semantic alignment at the communication layer, such as converting structured task objects used within the unified service bus into the field sequences and encoding formats required by the target system. Data format conversion involves semantic alignment at the content layer. For example, the unified service bus expresses irrigation flow rate in milliliters per second, while a certain manufacturer's system receives the same parameter in gallons per minute; the adapter performs unit conversion during the conversion. Similarly, the unified service bus uses acres as the area unit, while a certain agricultural machinery scheduling system uses hectares; the adapter performs area conversion during the conversion. Preferably, the adapter can also perform field completion. For example, an older system requires the input of the operator's employee number field, but the instruction issued by the orchestration engine does not include this field; the adapter automatically completes it based on the shift schedule for that plot.

[0076] In one possible implementation, the execution monitoring and adapter also assume the responsibility of security verification, performing signature verification and permission checks on each incoming and outgoing call to prevent unauthorized calls and man-in-the-middle tampering. Simultaneously, the adapter records snapshots of the input and output for each transformation, facilitating later auditing and fault backtracking.

[0077] The system receives the execution status and result data returned by the corresponding stages and stores them in the full-element data lake. The execution status includes start time, stage progress, completion time, success status, and error code. Result data is related to the job type; for example, irrigation results include actual irrigation duration, actual water volume, and flow distribution in each zone; fertilization results include actual application rates for each formula and application speed curves; plant protection results include flight path coverage, spraying volume per unit area, and images of key operation points; and harvesting results include actual harvest quantity, grading results, and loss rate. The execution monitoring and adapter validate the returned data and encapsulate it into standardized execution receipt objects. The full-element data lake receives the actual operation data from the execution end and associates it with corresponding task instructions, decision vectors, and twin states for archiving, forming a traceable, end-to-end data archive. Archived data is indexed by plot, batch, job type, and time dimension for easy subsequent dimensional queries. It should be noted that this archiving process not only serves as a review of the current production season, but also provides valuable real samples for the twin calibration and decision model iteration of subsequent production seasons.

[0078] The system acquires real-time anomalies in the execution status and result data, and uses these anomalies to drive the synchronous update of the twin model and the optimization of the optimal decision-making strategy. Real-time anomalies refer to situations that significantly deviate from the original plan during execution. For example, in irrigation operations, the actual water discharge for a certain plot is only 60% of the planned value, possibly due to insufficient water pressure or pipeline blockage; another example is a harvester malfunctioning and unable to continue operation; yet another example is a sudden outbreak of pests and diseases discovered during inspections.

[0079] Upon receiving execution receipts, the monitoring and adapter perform rule matching. The rule set includes pre-defined anomaly criteria, such as completion rate below a threshold, error status code, unit workload deviating from the planned value by more than a preset percentage, and anomaly features identified in uploaded images. Once an anomaly is matched, a real-time anomaly event is immediately generated and broadcast to the relevant twin models via the event bus. Upon receiving the anomaly event, the relevant twin models update their own status. For example, the equipment and facility twin model changes the status of a faulty harvester from idle to under maintenance and marks the estimated repair time; the land resource twin model marks plots with insufficient water supply as water-deficient and lowers the predicted soil moisture value; the tobacco growth twin model corrects the predicted water stress index and growth rate of tobacco in that plot accordingly. This triggers a local re-optimization operation. Local re-optimization is similar to global optimization in its algorithm mechanism, but the search scope and time window are significantly narrowed, only re-searching the decision variables of the affected plots or equipment for the next few days. Local re-optimization involves resimulating the current state of the affected sub-model to obtain a set of adjusted decision suggestions. For example, for plots with insufficient water supply, adjustment strategies might include supplemental irrigation during off-peak electricity hours, dispatching other idle sprinkler systems to assist, or adjusting the timing of subsequent fertilization to match the corrected water conditions. These adjustment strategies are injected into the service orchestration process as new cross-system task instructions. The orchestration engine then determines whether to interrupt the currently executing sequence or add it as a supplementary sequence. Preferably, adjustment strategies that do not conflict with the currently executing task are directly added as subsequent execution nodes; for adjustment strategies that directly conflict with the currently executing task, a cancellation or pause instruction is first issued, and a new instruction is issued only after the target system confirms the state change.

[0080] In another possible implementation, the handling process for equipment malfunctions during the harvesting phase is similar to the above. At 10:00 AM on a harvesting morning, a harvester, while operating on the third plot, reports an abnormal hydraulic system pressure. The monitoring and adapter processes generate an equipment malfunction event based on this, and the equipment facility twin model switches the machine's status to "under maintenance," estimating a repair time of four hours based on the maintenance knowledge base. The production management twin model then reassesses whether the harvesting tasks for the remaining plots that day can be completed within the maturity window, finding that completing the latter two plots as originally planned poses a risk. The adjustment strategy generated by the local re-optimization includes three new instructions: first, to relocate the backup harvester originally scheduled to another area in the afternoon back to the current area; second, to postpone the harvesting of one plot to the following morning; and third, to simultaneously notify the cold chain logistics system to adjust the scheduled transport times for that plot accordingly. The adjustment strategy is organized by the orchestration engine and then executed, with the entire response process completed within fifteen minutes of the malfunction event.

[0081] Once all adjustment strategies have been implemented, the occurrence, response, and outcome of this anomaly are recorded as a complete handling record in the full-element data lake and made available for management review in the strategy evaluation view. Managers can use this record to review the root causes of the event, such as whether insufficient water pressure was due to a temporary reduction in water release from an upstream reservoir, thereby prompting infrastructure upgrades or the construction of backup water sources. Understandably, this closed loop from anomaly to handling record not only improves the robustness of the current production run but also continuously accumulates experience data over long-term operation, providing increasingly rich sample support for optimization in subsequent production seasons.

[0082] The system leverages real-time operational data and anomaly information to drive iterative optimization and continuous updates of optimal decision-making strategies. By combining the actual results of the current round of operations, identifying shortcomings, and monitoring changes in on-site conditions, the system re-adjusts constraint thresholds and multi-objective optimization weights. In the updated digital twin simulation environment, it conducts a new round of multi-objective optimization calculations, iteratively updating operational time windows, process parameters, equipment scheduling schemes, and resource allocation strategies to generate optimal production decision-making schemes adapted to the latest field conditions. Through continuous operation execution, data feedback, model updates, and strategy iterations, a self-aware, self-correcting, and self-optimizing intelligent production closed loop is constructed, achieving dynamic and precise control over the entire tobacco production process.

[0083] The above description is merely a preferred embodiment of one or more embodiments of this specification and is not intended to limit the scope of one or more embodiments of this specification. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of one or more embodiments of this specification should be included within the protection scope of one or more embodiments of this specification.

Claims

1. A multi-dimensional digital twin-driven collaborative optimization decision-making method for the tobacco industry chain, characterized in that, The method includes: Multi-source heterogeneous data from various stages of tobacco production are acquired, and after spatiotemporal alignment and semantic association of the multi-source heterogeneous data, it is written into the full-element data lake of the data layer. The full-element data lake is used to store multi-source heterogeneous data from various stages of the entire tobacco production chain. Based on the multi-source heterogeneous data, multiple interconnected twin models are constructed and mapped in real time with the physical tobacco production system. These multiple twin models constitute a multidimensional dynamic digital twin, and the multidimensional dynamic digital twins interact with each other through a unified data bus. Based on the multidimensional dynamic digital twin, the optimization algorithm is run in a preset simulation environment based on a preset set of decision variables, a set of constraints, and a multi-objective function to perform simulation and output the optimal decision strategy that satisfies the set of constraints. Based on the optimal decision-making strategy, a series of cross-system task instructions are obtained. The cross-system task instructions are matched with the execution units of each link through the service registration and discovery mechanism to generate a cross-unit service call sequence. The job tasks are executed according to the service call sequence, and the execution results and generated running data are fed back to the data layer to drive the synchronous update of the twin model and the optimization of the optimal decision strategy.

2. The method as described in claim 1, characterized in that, The construction of multiple interconnected twin models based on the multi-source heterogeneous data, which are synchronously mapped in real time to the physical tobacco production system, includes: A land parcel resource twin model is constructed based on the geographic data, soil data, meteorological data, and pipeline status data in the aforementioned full-element data. Based on the tobacco variety parameters and environmental perception data in the full-element data, a tobacco growth twin model is constructed in combination with a preset development model, and the tobacco growth status is output through the tobacco growth twin model. A process flow twin model is constructed based on the work tasks and work task requirements in the full-element data; Based on the location, operational status, and operational capabilities of the equipment in the comprehensive data, construct a twin model of the equipment and facilities; A production management twin model is constructed based on the personnel, tasks, material flow, and cost elements in the full-element data.

3. The method as described in claim 1, characterized in that, The process involves using the multidimensional dynamic digital twin to simulate an optimization algorithm within a preset simulation environment, based on a preset set of decision variables, a set of constraints, and a multi-objective function. The algorithm outputs the optimal decision strategy that satisfies the set of constraints, including: Obtain the controllable decision variables for each production stage, and generate the decision variable set based on the controllable decision variables; Obtain agronomic constraints, resource constraints, equipment capability constraints, and time window constraints, and generate the constraint condition set based on the agronomic constraints, resource constraints, equipment capability constraints, and time window constraints; Obtain production targets, quality targets, cost targets, and resource consumption targets, and generate the multi-objective function based on the production targets, quality targets, cost targets, and resource consumption targets; In the simulation environment provided by the multidimensional dynamic digital twin, a preset evolutionary algorithm is used to search the set of decision variables to determine the target decision vector that enables the multi-objective function to reach Pareto optimality while satisfying the set of constraints. The target decision vector is output as the optimal end-to-end production strategy.

4. The method as described in claim 3, characterized in that, In the simulation environment provided by the multidimensional dynamic digital twin, a preset evolutionary algorithm is used to search the set of decision variables to determine the objective decision vector that enables the multi-objective function to achieve Pareto optimality while satisfying the set of constraints, including: In the preset simulation environment, multiple initial decision vectors are randomly generated, and an initial population is formed based on the multiple initial decision vectors; For each initial decision vector in the initial population, the entire process from the current moment to the end of the production season is simulated in the multidimensional dynamic digital twin to obtain the comprehensive target value corresponding to each initial decision vector; Based on the comprehensive objective value, the fitness of each initial decision vector is calculated using the multi-objective function, and all solutions of the current population are divided into Pareto front solutions and dominant solutions according to the fitness. Evolutionary operations such as selection, crossover, and mutation are performed on the current population to complete the population iteration and generate a new generation of population; If the new generation population satisfies the preset convergence condition, the Pareto front solution set in the final iteration population is extracted as the Pareto optimal objective decision vector, wherein the termination condition includes at least one of reaching the maximum number of iterations and the Pareto front solution set being stable with no optimization gain.

5. The method as described in claim 1, characterized in that, The process involves obtaining a series of cross-system task instructions based on the optimal decision-making strategy, matching these instructions with execution units at each stage through a service registration and discovery mechanism, and generating a cross-unit service call sequence, including: Based on the optimal strategy, a series of cross-system task instructions are obtained. After receiving the cross-system task instructions, the unified data bus sends a parsing request to the service orchestration engine. The orchestration engine sends a service discovery request to the service registry. Based on the service discovery request, the service registry queries the corresponding service interface address, invocation method, and service status. Based on the service interface address, invocation method, and service status, the cross-unit service invocation sequence is generated.

6. The method as described in claim 1, characterized in that, The step of executing job tasks according to the service call sequence and feeding back the execution results and generated running data to the data layer to drive the synchronous update of the twin model and the optimization of the optimal decision-making strategy includes: Send invocation instructions to the corresponding stages according to the service invocation sequence; If the service interface address of the corresponding link is found to be inconsistent, then the corresponding link is converted in terms of protocol and data format. Receive the execution status and result data returned by the corresponding link, and feed back the execution status and result data to the full-element data lake for storage; Extract real-time abnormal events from the execution status and result data, and drive the synchronous update of the twin model and the optimization of the optimal decision strategy based on the abnormal events.

7. The method as described in claim 2, characterized in that, The process involves constructing a tobacco growth twin model based on variety parameters and environmental perception data from the total data, combined with a pre-defined developmental model. The tobacco growth twin model then outputs the tobacco growth status, including: The extracted variety parameters and environmental perception multi-source heterogeneous data are preprocessed to generate a standardized model input dataset; The standardized dataset is imported into the preset development model in real time to complete the parameter initialization and dynamic iterative update of the development model, so that the updated development model can be adapted to the tobacco variety attributes and real-time growth environment of the current field. Based on the updated developmental model, the growth process of tobacco throughout its entire growth period is dynamically extrapolated, plant growth parameters and tobacco maturity index are quantitatively calculated, and the actual growth status of tobacco in the field is mapped in real time. Based on real-time growth status and maturity index, and by binding field resource information, growth time series data and process-related parameters, a tobacco growth twin model that is bidirectionally synchronized with the physical tobacco growth status in the field is constructed, and the tobacco growth status is output through the tobacco growth twin model.

8. The method as described in claim 2, characterized in that, The construction of the process flow twin model based on the operational logic and process parameters in the full-element data includes: Based on the tobacco growth status, a preset process model library is queried to determine the target operation time window and target operation parameters; Send an equipment query request to the equipment facility twin model to obtain a list of target equipment available within the target operation time window; Send a resource assessment request to the production management twin model to obtain resource confirmation information for the target operation time window and the target equipment list; The operation logic, equipment status, and process parameters are generated based on the target operation time window, the target operation parameters, the target equipment list, and the resource confirmation information. Based on the aforementioned operational logic, equipment status, and process parameters, a twin model of the process flow is constructed.

9. A multi-dimensional digital twin-driven collaborative optimization decision-making system for the tobacco industry chain, characterized in that, The system includes: a data layer, an optimization layer, a collaboration layer, and an execution layer; The data layer includes a full-element data lake, a multi-dimensional dynamic digital twin construction module, and a twin sub-model library. The full-element data lake is used to store multi-source heterogeneous data from all stages of the tobacco production chain. The multi-dimensional dynamic digital twin construction module creates a virtual twin sub-model library that is synchronously mapped with the entire tobacco production chain based on the multi-source heterogeneous data. The twin sub-model library includes land resource twin sub-models, tobacco growth twin sub-models, process flow twin sub-models, equipment and facility twin sub-models, and production management twin sub-models. The optimization layer includes a cross-link optimization decision engine, an optimization algorithm library, and a strategy simulation sandbox. The cross-link optimization decision engine is used to define the decision variables, constraints, and multi-objective functions of the entire tobacco production chain, call the algorithms in the optimization algorithm library, and complete the iterative simulation and selection of the scheme in the strategy simulation sandbox to obtain the optimal decision strategy. The collaboration layer includes a unified data bus, a service registry, and a service orchestration engine. The service registry is used to register and manage the service interfaces and operating status of the entire tobacco production chain. The service orchestration engine is used to parse the optimal decision strategy and decompose it to generate a standardized service call sequence. The execution layer includes multiple links covering the entire tobacco production chain. Each link is connected to the unified service bus as an independent service node, receives collaborative instructions to execute tasks, and feeds back the execution results and generated operation data to the data layer, driving the synchronous update of the twin model and the optimization of the optimal decision strategy.

10. The multi-dimensional digital twin-driven collaborative optimization decision-making system for the tobacco industry chain according to claim 9, wherein the twin sub-model library includes land resource twin sub-models, tobacco growth twin sub-models, process flow twin sub-models, equipment and facility twin sub-models, and production management twin sub-models, and each sub-model achieves dynamic causal linkage based on a unified data bus, characterized in that, The land resource twin model integrates geographic data, soil data, meteorological data, and pipeline status data; The tobacco growth twin model, based on the tobacco variety parameters and environmental perception data, simulates and predicts tobacco growth and maturity in real time, generating tobacco growth status. The process flow twin sub-model matches corresponding work tasks based on tobacco growth status events, generates work task requirements, and initiates equipment adaptation queries to the equipment and facility sub-model. The equipment and facility twin model maps the equipment's location, operating status, operating capacity, and available time periods in real time to obtain the matching degree between the operation task and the equipment. The production management twin model summarizes the environmental parameters, tobacco growth status, work task requirements, and the matching degree between the work tasks and equipment in real time, completes the calculation and evaluation of personnel, materials, costs, and task progress, and outputs the final executable production task.

Images