A knowledge model-based instruction-driven machine task planning method and system

Abstract

The application provides a kind of instruction driving machine task planning method and system based on knowledge model, and the specific knowledge of field is structuredly represented by knowledge model, and the object of class level, instance level of field knowledge and its logical relationship are represented;Further, the semantics of instruction is understood, and then the task planning problem represented by instruction is normalized, specifically including task planning object and its mutual constraint relationship;On this basis, further consider the space-time constraint between task objects and the measurement and evaluation of task efficiency.Finally, through atlas and visual graph display, give the pareto optimal scheme under multi-dimension, support intelligent or man-machine interactive decision, realize man-machine interaction in task planning, and output field-specific reliable scheme based on natural language.

Description

Technical Field

[0001] This invention relates to the fields of knowledge engineering and task engineering, and in particular to an instruction-driven ad hoc task planning method and system based on a knowledge model. Background Technology

[0002] Mission planning refers to the process of integrating multiple types of resources to achieve a common goal through synergy. It is widely used in emergency disaster relief, the Internet of Things, and the military. Traditional mission planning heavily relies on human decision-makers, with information often transmitted primarily through natural language text. The collaborative process depends on human understanding, leading to significant uncertainty. Furthermore, current mission planning scenarios are becoming increasingly complex, with various types of multifunctional and heterogeneous resources emerging, making it increasingly difficult for human computing speed to meet the demands of complex planning objects and environments.

[0003] In recent years, the development of information technology and artificial intelligence has provided new opportunities for ad-hoc task planning based on natural language. Artificial intelligence technology can process and analyze natural language instructions to a certain extent, offering new opportunities for instruction-driven multimodal needs. However, existing artificial intelligence technologies are often based on general knowledge training and cannot understand small-sample, specialized knowledge within specific domains, nor can they provide interpretable and reliable solutions. This poses a significant bottleneck for applications in fields with high reliability requirements.

[0004] Graph-based knowledge models, as an effective way of representing knowledge, can structurally and formally display knowledge concepts, entities, and their relationships in the form of graphs, which helps in the reliable application of knowledge. However, how to utilize knowledge models and combine them with artificial intelligence to achieve reliable and agile task planning driven by instructions has always been an urgent problem to be solved in the field of intelligent task engineering. Summary of the Invention

[0005] The main objective of this invention is to provide an instruction-driven agile task planning method based on a knowledge model.

[0006] Another objective of this invention is to propose an instruction-driven agile task planning system based on a knowledge model.

[0007] To achieve the above objectives, a first aspect of the present invention proposes an instruction-driven agile task planning method based on a knowledge model, comprising:

[0008] Preprocess the natural language instructions to extract task-related keywords; Based on a pre-built knowledge model, semantic understanding of keywords is performed, fuzzy references are parsed into specific instance objects and their activities, nouns are mapped to the instance layer, verbs are aligned to relational constraints, and a semantically consistent task solution space is generated. The task scheme space is filtered under spatiotemporal constraints to obtain a set of feasible task schemes. Based on examples, the performance of each task scheme in the feasible task scheme set is evaluated in multiple dimensions to select the Pareto optimal subset. The Pareto optimal subset is output in the form of a graph or visualization, supporting intelligent or human-computer interactive decision-making.

[0009] In one embodiment of the present invention, the natural language instructions are preprocessed, including: Handle situations where words in natural language instructions refer to the same denotative intent but cause symbolic differences; Based on general knowledge, keywords are extracted using large language models or natural language processing techniques and aligned to concepts or instances in the knowledge model. After completing the semantic extraction and alignment of keywords, the implicit operational logic of the instructions is transformed into a structured execution flow, and an executable flow architecture is generated using SysML or an equivalent system modeling language.

[0010] In one embodiment of the present invention, the knowledge model is used to characterize the specialized knowledge required for task planning, including: The knowledge model is a system architecture for specific real-world tasks, including the construction of the components and relationships of task planning objects, and the adjustment of the instance architecture based on natural language instructions; The nodes of the knowledge model represent entities, and the relationships between nodes represent the activities between entities. There are multiple types of relationships between entities, and a contextual understanding of entities is established through a complex relationship network. Obtain adjustment instructions for the corresponding knowledge model; The knowledge model includes organizational entities, resource entities, and threat entities. Organizational entities include functional organizations and task organizations. Functional organizations are organizations that manage resources with specific professional functions and have ownership of resource entities. Task organizations are organizations that are created on the spot for specific tasks and have the right to use resource entities. The existence of functional organizations and task organizations is predefined and hierarchical, and they are adjusted on the spot according to instructions, reflecting the decision-making level's need to adjust the form of resource organization.

[0011] In one embodiment of the present invention, the adjustment instruction is instruction information for adjusting the task knowledge model; The adjustment instructions are categorized by object, including instructions for adjusting at the knowledge model class level and instructions for adjusting at the knowledge model instance level. Adjustment instructions can be categorized by their intent into adjustments to organizational relationships and adjustments to collaborative relationships.

[0012] In one embodiment of the present invention, the classification of the resource entities is diverse to reflect the heterogeneity of resources at different levels; According to the system composition, it is divided into system, subsystem, and system components, which are related to each other through their compositional relationships; Based on their mutual deployment dependencies, they can be divided into payloads and platforms. Payloads and platforms are associated through deployment relationships, and these relationships reflect existence; that is, if a platform fails, the payloads deployed on it will also fail. The knowledge model includes the concept of attributes that quantitatively describe entities, and these attributes include those that reflect spatiotemporal characteristics. The knowledge model includes descriptions of the relationships between entities, including parent-child relationships, entity-entity relationships, activity-activity relationships, entity-activity relationships, and entity-attribute relationships. The specific forms of knowledge models include graph databases, ontology model databases, and table databases that reflect SPO relationships.

[0013] In one embodiment of the present invention, the spatiotemporal constraints are used to characterize the sequential information of task execution, including: Constructing a basic representation framework for spatiotemporal constraints: Defining task execution as an emergent process in which multiple resource instances perform activities and generate activity effects under specific spatiotemporal constraints, where constraints within the time frame are a prerequisite for the superposition of activity effects and the generation of an overall task effect; Define the time interval constraints of the activity process: describe the activity process as a time interval, and use 13 Allen interval algebras to represent the relationship between the time intervals of two activity processes, including prior to, subsequent to, meeting, being met, overlapping, being overlapped, containing, being contained, starting at, being started, ending at, being ended, and equal relationships; Establish an effectiveness mechanism under spatiotemporal constraints: measure the effectiveness of task execution within the spatiotemporal scope, including setting the precondition that the interval between the discovery of the target and the actual effect acting on the target should be less than the interval between the target's escape and the range of effect.

[0014] In one embodiment of the present invention, a multi-dimensional performance measurement and evaluation system is used to construct the relationship between the state of task execution resource instances and the expected task execution performance, including: Establish a multi-dimensional evaluation system that includes task-level performance indicators, activity-level performance indicators, resource instance status, and environmental and target characteristic parameters; By directly accessing resource instance sensors through information systems to obtain time-varying state parameters, and obtaining stable design parameters from product design specifications; by acquiring environmental parameters that affect task activities through environmental monitoring capabilities, and designing and determining parameter requirements based on the sensitivity of resources to environmental parameters; by acquiring design parameters of target objects through intelligence information systems, and by acquiring time-varying state parameters of targets through monitoring systems. To address the discrepancies between intelligence and surveillance information and objective reality, rapid adjustments are made to the solution based on the current resource instance status, environmental parameters, and target characteristic parameters, while meeting the task triggering conditions. When the information discrepancy effect occurs, a solution is quickly regenerated. By conducting multi-dimensional quantitative performance evaluations of all feasible task solutions, the optimal solution under the current conditions is selected.

[0015] In one embodiment of the present invention, the task-level performance indicators cover three categories of indicators: task completion quality, completion time, and resource consumption, and a human-computer interaction interface is reserved. The decision-maker can predefine or adjust them on the spot according to the urgency of the task and environmental situation information. The activity-level performance indicators include the quality, time, and resource consumption of each entity's resources in completing the activity. Their calculation relies on the analysis of physical laws or empirical value lookup table calculation.

[0016] In one embodiment of the present invention, the output task solution is constrained to satisfy feasibility and optimality in at least one evaluation dimension, and is displayed graphically, including: The graphical interface presents the spatiotemporal location, effectiveness range, and direction of action of resource instances, enabling decision-makers to intuitively grasp the overall picture of the plan and confirm its feasibility through visualization. It presents the start time, end time, and expected time of the execution effect of the corresponding activities for resource execution, forming a complete time process view; When execution deviations occur due to resource limitations or errors in understanding the target, the system automatically initiates the next round of task planning and iteration based on the monitored deviation information, thereby achieving dynamic optimization and continuous improvement of the solution.

[0017] To achieve the above objectives, a second aspect of the present invention proposes an instruction-driven ad hoc task planning system based on a knowledge model, comprising: The keyword extraction module is used to preprocess natural language instructions and extract keywords related to the task. The task solution space generation module is used to perform semantic understanding of keywords based on a pre-built knowledge model, resolve fuzzy references into specific instance objects and their activities, map nouns to the instance layer, align verbs to relational constraints, and generate a semantically consistent task solution space. The task scheme set generation module is used to filter the task scheme space under spatiotemporal constraints to obtain a feasible task scheme set; The optimal subset selection module is used to perform multi-dimensional performance measurement and evaluation on each task scheme in the feasible task scheme set based on the instance, so as to select the Pareto optimal subset. The interactive output module is used to output the Pareto optimal subset in the form of a graph or visualization, supporting intelligent or human-computer interactive decision-making.

[0018] In summary, the instruction-driven ad hoc task planning method based on a knowledge model of the present invention includes the following specific processes: Deep preprocessing and semantic parsing of natural language instructions. The core of this step lies in transforming the user's ambiguous and unstructured natural language instructions into structured semantic information that can be recognized and processed by a computer. Specifically, this includes performing syntactic checks and semantic extraction and alignment on the natural language. The instructions expressed in the natural language originate from the user and reflect the user's needs, including but not limited to organizational restructuring needs and resource collaboration needs. By aligning and mapping the extracted semantics with concepts in the background knowledge model, it is ensured that the user's intent is accurately and unambiguously transformed into planning elements that the system can understand, laying the foundation for subsequent automated task planning.

[0019] The explicit and formalized representation of domain knowledge involves constructing a knowledge model that accurately represents the specialized knowledge required for task planning within a specific domain. The knowledge model is divided into three layers: class, model, and instance. The class layer categorizes entities based on system composition, deployment dependencies, and functions; the model layer categorizes entities of the same type with different design parameter attributes; and the instance layer describes the actual existence of instance objects, which are the objects ultimately defined in task planning. This knowledge model provides rich and structured knowledge support for understanding instruction requirements, performing logical reasoning, and generating solutions. A spatiotemporal constraint model for resource execution activities is constructed to characterize the spatiotemporal dependencies and sequence of resources during task execution. Addressing the issue of task planning that only considers resource availability, the spatiotemporal constraint model answers when resources are available and how to coordinate across time dimensions to ensure effectiveness. In the time dimension, interval algebra theory is employed to define start, end, and duration time points / intervals for each activity, accurately characterizing the temporal relationships between activities and defining spatiotemporal constraint templates for feasible solutions. A constraint satisfaction problem solver is used to perform continuous spatiotemporal matching before and during solution generation, thereby ensuring that the final planned resource execution sequence is spatiotemporally consistent and executable. Construction and Quantitative Calculation of a Multi-Dimensional Performance Evaluation Index System. A multi-dimensional performance measurement and evaluation system is constructed to quantitatively assess the overall performance of task execution. This evaluation system considers the diversity of evaluation criteria and establishes a series of interrelated performance indicators from multiple key dimensions. Typical dimensions include: 1) Time performance, such as total task completion time, critical path duration, and time redundancy; 2) Resource performance, such as resource utilization, cost consumption, and human resource load balance; 3) Quality performance, such as task completion prediction, risk index, and robustness (ability to cope with interference). Interpretable generation of multiple solutions and performance-based intelligent ranking. Based on the output of the preceding steps, the system takes semantically parsed user instructions as input, matches and retrieves them in the task template library of the knowledge model, filters and combines them through a spatiotemporal constraint model, and uses graph search and planning algorithms to generate multiple logically feasible initial task planning solutions that meet basic constraints. This method automatically generates an execution logic chain for each solution and uses a multi-dimensional performance evaluation system to conduct a comprehensive quantitative evaluation of each initial solution, calculating the multi-dimensional evaluation performance. Finally, it extracts task planning solutions that meet the Pareto front under the multi-dimensional evaluation, thus transforming a complex planning problem into a data-driven, logically clear auxiliary decision-making process. This system provides visualized interaction and dynamic monitoring throughout the entire decision-making lifecycle. Task planning schemes are presented graphically, supporting user-interactive decision-making and providing status monitoring and dynamic adjustment capabilities during task execution. In the human-computer interaction phase, this method uses various views such as Gantt charts, network diagrams, timelines, and resource load diagrams to comprehensively display the execution path, resource scheduling, critical path, and performance comparison of each alternative scheme. After scheme execution, a dynamic monitoring mode is entered, comparing the planned task timeline with real-time feedback status data. Task progress, actual resource location and status, and potential deviations (such as delays or resource failures) are highlighted on the same visual interface. Once a significant deviation is detected, the system immediately issues an alert and, based on the latest status, triggers the aforementioned steps for rapid local replanning or global scheme iteration, generating new adjustment schemes for user decision-making. This forms a closed-loop management system of "planning-decision-monitoring-replanning," ensuring that tasks steadily progress towards their predetermined goals even in dynamically changing environments.

[0020] Based on the above, this application constructs a dedicated knowledge model for task planning, enabling structured representation of domain-specific knowledge and information. Furthermore, by building a spatiotemporal constraint and multi-dimensional performance evaluation index system, it details the state information of resource instances within a task scenario, as well as the overall efficiency of multiple resources completing the task through collaborative activities. This allows for the ad-hoc, interpretable task planning solutions based on natural language input and specific scenario knowledge, significantly improving the agility and intelligence of task planning. Finally, by visually displaying task solutions and structured representation of key timelines for task execution, it supports users in making final decisions based on recommended multiple options and monitoring the execution process. In case of deviations, it supports iterative adjustments and replanning, ensuring efficient and reliable task execution.

[0021] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0022] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 A flowchart illustrating an instruction-driven ad hoc task planning method based on a knowledge model, provided as an embodiment of the present invention; Figure 2 A technical principle diagram of an instruction-driven ad hoc task planning method based on a knowledge model provided in an embodiment of the present invention; Figure 3 The execution flowchart of an instruction-driven ad hoc task planning method based on a knowledge model is provided in an embodiment of the present invention. Figure 4 A structural diagram of an instruction-driven ad hoc task planning system based on a knowledge model provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the composition architecture of an instruction-driven ad hoc task planning system based on a knowledge model, provided as an embodiment of the present invention. Detailed Implementation

[0023] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0024] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0025] The following description, with reference to the accompanying drawings, describes an instruction-driven ad hoc task planning method and system based on a knowledge model, according to an embodiment of the present invention.

[0026] Example 1 This embodiment provides an instruction-driven ad hoc task planning method based on a knowledge model. For example... Figure 1 As shown, it includes: S1 preprocesses natural language instructions and extracts task-related keywords; S2 performs semantic understanding of keywords based on a pre-built knowledge model, resolves fuzzy references into specific instance objects and their activities, maps nouns to the instance layer, aligns verbs to relational constraints, and generates a semantically consistent task solution space. S3, under spatiotemporal constraints, filter the task scheme space to obtain a set of feasible task schemes; S4. Based on the examples, perform multi-dimensional performance measurement and evaluation on each task scheme in the feasible task scheme set to select the Pareto optimal subset. S5, output the Pareto optimal subset in the form of a graph or visualization, supporting intelligent or human-computer interactive decision-making.

[0027] Specifically, the instructions expressed in natural language are preprocessed. Preprocessing techniques include natural language processing (NLP) and large language model (LLM). The purpose is to extract and organize the language to obtain keywords related to task planning. Based on the established knowledge model, semantic understanding of keywords is performed. Specifically, semantic understanding refers to understanding and tracing back to the specific instance object and its activities that the information refers to for ambiguous information generated in human instructions; for noun words, the instance layer information they refer to is obtained by layer-by-layer analysis; for verb information, their relational information is aligned to form a task solution space that meets semantic requirements. Under spatiotemporal constraints, the task scheme space is filtered to select a set of task schemes that meet the characteristics of spatiotemporal constraints; The performance of each task scheme in the set is measured and evaluated in multiple dimensions. Based on the state of the instance layer object, the task execution performance under multiple dimensions is obtained, and the Pareto optimal task scheme set under each dimension is obtained by comparison. Feasible task solutions are presented in graphs or visualizations to support intelligent or human-computer interactive decision-making processes.

[0028] In one embodiment of the present invention, a natural language preprocessing tool performs syntactic layer checks and semantic extraction alignment on natural language based on a knowledge model, including: Preprocessing is based on general knowledge, and its implementation can be a large language model (LLM) with general artificial intelligence capabilities or natural language processing (NLP) technology. If necessary, to improve reliability, human operation in the loop can be performed. The human resource requirement is a general business knowledge level. After completing the semantic extraction and alignment of keywords, preprocessing also needs to formally model the instruction operation process to generate the process execution architecture. Formal modeling of the execution process can be done using system modeling languages ​​such as SysML. Preprocessing also includes handling lexical errors, abbreviations, omissions, etc., in the instructions that refer to the same referential intent but cause symbolic differences.

[0029] In one embodiment of the present invention, the established knowledge model is used to characterize the specialized knowledge required for task planning, including: The knowledge model expresses special and specific characteristics, and is a system architecture for specific real-world tasks. Specifically, it includes the construction of the task planning objects and their relationships, and the adjustment of the instance architecture based on natural language instructions. The nodes of the knowledge model represent entities, and the relationships between nodes represent the activities between entities. There can be multiple types of relationships between entities, and a contextual understanding of entities is established through a complex relationship network. Obtain adjustment instructions for the corresponding knowledge model. The adjustment instructions are instructions to adjust the task knowledge model. The adjustment instructions are categorized by object, including instructions to adjust at the knowledge model class level and instructions to adjust at the knowledge model instance level. The adjustment instructions are categorized by intent, including adjustments to organizational relationships and adjustments to mutual collaboration relationships. The knowledge model includes organizational entities, resource entities, and threat entities. Organizational entities include functional organizations and task organizations. Functional organizations are organizations that manage resources with specific professional functions and have ownership of resource entities. Task organizations are organizations that are created on the spot for specific tasks and have the right to use resource entities. The existence of functional organizations and task organizations is predefined and hierarchical, and they are adjusted on the spot according to instructions, reflecting the decision-making level's need to adjust the form of resource organization. Resource entities can be classified in diverse ways to reflect the heterogeneity of resources at different levels. Based on system composition, they can be divided into systems, subsystems, and system components, which are interconnected through compositional relationships. Based on mutual deployment dependencies, they can be divided into payloads and platforms, which are interconnected through deployment relationships, reflecting existence; that is, if a platform fails, the payloads deployed on it will also fail. Resource entities can also be classified according to their functions, such as detection and sensing entities, control entities, performance entities, and enabling entities. These classifications facilitate semantic understanding of effect-enabling instructions, allowing tracing back to the resource entities that can implement those effect instructions. The knowledge model includes attribute concepts that quantitatively describe entities. Attributes include those that reflect spatiotemporal characteristics, such as spatial coordinates, absolute time, relative time, velocity, and acceleration; those that reflect working status, such as fault, idle, and in use; those that reflect capacity; and those that reflect the boundaries of action, including the distance of action. The knowledge model includes descriptions of the relationships between entities, such as parent-child relationships, entity-entity relationships, activity-activity relationships, entity-activity relationships, and entity-attribute relationships.

[0030] The specific forms of knowledge models can include graph databases, ontology model databases, and table databases that reflect SPO relationships.

[0031] In one embodiment of the present invention, spatiotemporal constraints are used to characterize the sequential information of task execution, including: The execution of a task can be understood as the superposition and emergence of the effects of multiple resource execution activities. For a specific goal, resource instances can generate activity effects under certain spatiotemporal constraints. Constraints within the time range are the premise that activity effects can be superimposed and generate an overall task effect. The specific information described by spatiotemporal constraints includes the activity process of completing the task under specific spatial constraints, and the constraints of this process in the time dimension. The description of an activity process is a time interval. The relationship between two activity process time intervals is represented by 13 Allen interval algebras, including prior to, subsequent to, meeting, being met, overlapping, being overlapped, containing, being contained, starting at, being started, ending at, being ended, and equal. The mechanism of effectiveness in the spatiotemporal scope during task execution is reflected in the spatiotemporal constraints. For example, one prerequisite for measuring effectiveness is that the time interval between the discovery of the target and the actual effect acting on the target should be less than the time interval between the target's escape and the range of effect.

[0032] In one embodiment of the present invention, a multi-dimensional performance measurement and evaluation system is used to construct the relationship between the state of task execution resource instances and the expected task execution performance, including: The purpose of task performance evaluation is to quantitatively evaluate and select the optimal solution under the current conditions from all feasible combinations of task solutions. The performance evaluation dimensions of the solution are multi-faceted, such as the quality of completion, the time of completion, and the resource cost consumed. The trade-offs between multiple dimensions need to be predefined or set up with human-computer interaction interfaces and determined by decision-makers. The performance evaluation system includes task-level performance indicators, activity-level performance indicators, resource instance status, and environmental and target characteristic parameters; Task-level performance indicators include the quality of task completion, the time taken to complete the task, and the amount of resources required to complete the task. The evaluation of these three types of indicators reserves a human-computer interaction interface, which can be predefined or adjusted by decision-makers based on information such as the urgency of the current task and the environmental situation. Activity-level performance indicators include the quality of each entity's resources in completing the activity, the time required to complete the activity, and the amount of resources required to complete the task. Their calculation relies on the analysis of physical laws or on table lookup calculations based on empirical values. Resource instance status is used to calculate activity-level performance indicators, including time-varying state parameters during resource operation and design parameters with stable change patterns. The state parameters of the resource can be obtained by directly accessing the state sensors of the resource instance through the information system, while the design parameters come from the manuals, user manuals, etc. provided by the product design unit. The calculation of performance indicators needs to take into account the impact of environmental parameters. The environment specifically refers to the characteristics of climate, wind, temperature, humidity, electromagnetic fields, etc. that affect the execution of task activities. The acquisition of the above environmental parameters depends on specific environmental monitoring capabilities, and the demand for environmental parameters also depends on the sensitivity design of resources to environmental parameters. The calculation of performance indicators needs to consider the characteristics of the target object, including the time-varying state parameters of the target during operation and the design parameters of the target. The acquisition of the target's design parameters mainly relies on the intelligence information system of the mission party, which depends on the long-term tracking of the objective laws and information of the target to establish an understanding of the target. Its information carrier can be natural language text. The acquisition of the target's state parameters mainly relies on the monitoring system of the mission party. Its information may be intermittent and uncertain, but it should be timely. Its information carrier should be data stream. Intelligence and monitoring information will inevitably have deviations from objective information. Therefore, this method emphasizes rapid task planning under ad-hoc conditions, including rapid adjustments to meet the task triggering conditions and rapid regeneration of solutions after the information deviation effect occurs.

[0033] In one embodiment of the present invention, the output task solution is constrained to satisfy feasibility and optimality in at least one evaluation dimension, and is displayed graphically to facilitate human-computer interaction decision-making, including: The graphical solution display includes the spatiotemporal location of resource instances, the scope of their effectiveness, and the direction of their effect. The graphical display helps people intuitively grasp the information of the solution and supports decision-makers in confirming the task plan. The solution presentation also includes its expected timeline process, including the start time, end time, and expected execution effect time of the corresponding activities of the resources. The timeline process will facilitate the monitoring of task execution. When execution deviations occur due to the resources themselves or due to cognitive errors in the target, the next round of task planning and iteration will be launched.

[0034] Specifically, such as Figure 2 and Figure 3 As shown, natural language instructions are preprocessed. Specifically, Natural Language Processing (NLP) or Large Language Modeling (LLM) techniques are used to perform syntactic checks and semantic extraction on user-input instructions. This process aims to extract keywords related to task planning from the instructions and to standardize different expressions of the same intent, such as typos, abbreviations, or omissions.

[0035] After extracting and aligning keywords, preprocessing further utilizes a system modeling language to formally model the operational flow implied by the instructions, generating a preliminary process execution architecture and laying the foundation for subsequent semantic understanding. If necessary, personnel with general business knowledge can be brought in for on-loop intervention to improve the reliability of preprocessing.

[0036] After preprocessing, the system performs deep semantic understanding on the extracted keywords based on a pre-established knowledge model. This knowledge model is a specialized knowledge system for specific task scenarios, with the core purpose of eliminating ambiguity in nouns and verbs in instructions. For nouns, the model traces them layer by layer, from the class layer and model layer to the specific instance layer object. For verbs, the model aligns them through a relational network to understand the activities between the entities they refer to, thus forming an initial task solution space that meets semantic requirements. The knowledge model is stored in the form of a graph database or ontology model database, where nodes represent entities such as organizations, resources, and threats, relationships between nodes represent various activities between entities, and attributes describe the spatiotemporal, state, and capability characteristics of entities.

[0037] The knowledge model includes descriptions of the relationships between entities, including parent-child relationships, entity-entity relationships, activity-activity relationships, entity-activity relationships, and entity-attribute relationships. Parent-child relationships include "is a" relationships; entity-entity relationships include "structural relationships," further including "belonging," "assigned," "belonging to," "equipped," "composed," and "deployed" relationships; activity-activity relationships include "prior to," "subsequent to," "meeting," "being met," "overlapping," "being overlapped," "containing," "being contained," "starting at," "being started at," "ending at," "being ended at," and "equal to"; entity-activity relationships include "participation"; and entity-attribute relationships include "having."

[0038] The specific implementation forms of knowledge models can include graph databases, ontology model databases, and table databases that reflect SPO relationships.

[0039] This method filters the initial task solution space under spatiotemporal constraints. Spatiotemporal constraints ensure the logical feasibility of task execution; their core mechanism lies in the fact that only when multiple resource instances generate activity effects under specific spatiotemporal conditions can these effects superimpose to produce the expected overall task performance. Thirteen activity-activity relationships are used for characterization. The constraint solver filters the solution space based on these constraints, eliminating all solutions with spatiotemporal conflicts, and ultimately outputs a set of feasible task solutions that satisfy all spatiotemporal constraints.

[0040] This method constructs a multi-dimensional performance measurement and evaluation system to quantitatively evaluate each option in the set of feasible task solutions. The system aims to establish a quantitative relationship between the state of resource instances and the expected task performance, with evaluation dimensions covering multiple aspects such as task completion quality, time consumption, and resource consumption. The evaluation process comprehensively considers activity-level performance indicators, real-time state parameters and fixed design parameters of resource instances, environmental characteristics (such as climate and electromagnetic fields), and characteristic parameters of the target object. All data required for calculation is acquired through access to sensors, intelligence information systems, etc. Using this system, the coordinates of each option in the multi-dimensional performance space are calculated, and based on multi-objective optimization theory, a set of options that satisfy Pareto optimality across all evaluation dimensions is selected, providing optimal trade-offs for decision-making.

[0041] This method graphically displays feasible solutions from the Pareto optimal solution set to support intelligent or human-computer interactive decision-making. The graphical display includes the spatiotemporal location of resource instances, their effective range, and the expected task execution timeline. This intuitive presentation facilitates decision-makers' confirmation of solution details. During the task execution phase, dynamic monitoring is performed by comparing the planned timeline with real-time feedback data. Once an execution deviation is detected due to resource failure or intelligence bias, the system will immediately issue an alert and trigger a new round of ad-hoc task planning iterations, thus forming a closed-loop management system of "planning-decision-monitoring-replanning," ensuring that tasks can proceed efficiently and reliably even in dynamic environments.

[0042] Example 2 This invention proposes another instruction-driven ad hoc task planning method based on a knowledge model, comprising: S101 uses natural language processing technology to preprocess the input natural language instructions, extract keywords related to task planning, and standardize the fuzzy expressions in the instructions.

[0043] Specifically, in some implementations, the preprocessing step of this invention uses Natural Language Processing (NLP) technology to structure the input natural language instructions. Its core objective is to extract keywords related to task planning and standardize ambiguous expressions in the instructions, thereby providing clear and computable input for subsequent semantic understanding and task plan generation. This step is technically implemented based on NLP modules such as lexical analysis, syntactic parsing, and semantic role labeling (SRL), combined with the contextual understanding capabilities of Large Language Models (LLMs), to achieve deep semantic extraction and alignment of user instructions.

[0044] The preprocessing process typically includes subtasks such as word segmentation, part-of-speech tagging, named entity recognition (NER), and dependency parsing. The NER module identifies organizational, resource, and threat entities involved in the instructions, such as "drone," "radar system," and "enemy target," requiring an accuracy of over 90% (F1-score ≥ 0.9) to ensure the reliability of subsequent semantic mapping. For fuzzy expressions, such as "complete as soon as possible" or "use first," the system identifies their intent through semantic role labeling and maps it to quantifiable task constraints. For example, "complete as soon as possible" is transformed into a minimum objective function for minimizing task completion time. This could transform "priority use" into a weighting adjustment mechanism for resource selection. ,in Representing resources The selection weight, This is the priority coefficient.

[0045] This step is widely applicable to fields with high requirements for real-time and reliability of mission planning, such as military command, emergency response, and intelligent scheduling. For example, in a battlefield environment, a commander might input "prioritize using Type A UAVs to conduct reconnaissance of the target area." The system needs to recognize keywords such as "prioritize use," "Type A UAV," and "reconnaissance," and convert "prioritize use" into a priority parameter for resource selection, and map "reconnaissance" to a specific activity type in the knowledge model, thereby constructing a semantically consistent mission execution flow.

[0046] Furthermore, the technical effect of this step is to significantly improve the semantic understanding capability and input robustness of the task planning system. By standardizing fuzzy expressions, the system can more accurately identify user intent, reduce semantic ambiguity, and improve the credibility and consistency of task solution generation. Simultaneously, this step provides structured input for subsequent spatiotemporal constraint modeling and multidimensional performance evaluation, serving as a crucial prerequisite for realizing the closed-loop process of "instruction-driven - knowledge reasoning - solution generation".

[0047] S201, based on a predefined multi-layer structured knowledge model, perform semantic parsing on the keywords, map noun words to instance layer objects layer by layer, align verb words to the activity relationships between entities, and generate an initial task scheme space that meets semantic requirements.

[0048] Specifically, in some implementations, semantic parsing of keywords based on a predefined multi-layered structured knowledge model is a key step in the method of this invention to achieve instruction semantic regularization and task scheme generation. The core of this step lies in mapping nouns in natural language layer by layer through the hierarchical structure of the knowledge model (class layer, model layer, instance layer), thereby clarifying the specific instance objects they refer to in the real-world task scenario; simultaneously, semantic alignment is performed on verbs to identify the inter-entity activity relationships they express in the knowledge model, ultimately constructing an initial task scheme space that meets semantic requirements.

[0049] This semantic parsing process is typically implemented based on graph databases or ontology model databases. In these databases, nodes represent entities such as organizations, resources, and threats, while edges represent the relationships between entities. For nouns, the system first performs semantic matching at the class level to identify their category (e.g., "detection and sensing," "control," etc.). Then, it further refines the matching scope based on attributes at the model level (e.g., design parameters, functional characteristics). Finally, it determines the specific resource instance it refers to at the instance level. For example, if the command contains "drone," the system will first identify it as a "platform" resource, and then match it to a specific "MQ-9 Reaper" instance based on attributes such as "endurance" and "payload capacity" at the model level.

[0050] For verbs, the system aligns them using a predefined network of activity relationships within the knowledge model. These relationships include terms like "participate," "deploy," "enable," and "control," and support various semantic combinations. For example, the verb "deploy" is mapped to a "deployment relationship" and further associated with the entities "platform" and "payload," thus clarifying the dependencies and collaboration logic between resources. In implementation, this alignment process can combine Semantic Role Labeling (SRL) and Relation Extraction (RE) techniques to ensure semantic consistency between verbs and entities.

[0051] The accuracy of semantic parsing depends on the completeness and semantic coverage of the knowledge model. In this invention, the entity classification of the knowledge model must meet at least a three-level hierarchical structure, and each entity should have no fewer than five key attributes (such as location, state, ability, scope of action, and temporal characteristics). The accuracy of verb alignment should reach over 90% to ensure the semantic integrity of the task scheme space. Furthermore, the semantic parsing process should support multilingual input and have the ability to automatically identify and normalize synonyms, abbreviations, and ellipsis.

[0052] This step is widely applicable to complex mission scenarios such as military mission planning, emergency response, and IoT resource scheduling. In the military field, for example, with instructions such as "using drones to conduct reconnaissance of the target area and deploy jamming equipment," the system can align verbs such as "reconnaissance" and "deployment" to activity relationships such as "enable" and "control," and map nouns such as "drone" and "jamming equipment" to specific resource instances, thereby generating an initial mission plan that conforms to tactical logic.

[0053] This step effectively addresses the issues of semantic ambiguity and unclear referencing in natural language instructions through the semantic mapping mechanism of a structured knowledge model, significantly improving the semantic understanding accuracy and interpretability of task planning. Simultaneously, by transforming instructions into executable entity-activity relationships, it provides structured input for subsequent spatiotemporal constraint filtering and multidimensional performance evaluation, thereby supporting the system's rapid and reliable task planning capabilities in dynamic environments.

[0054] Furthermore, S201 includes: S21, the entity classification in the knowledge model includes organizational entities, resource entities, and threat entities. Organizational entities are further subdivided into functional organizations and task organizations. Functional organizations manage the ownership of resource entities, while task organizations manage the right to use resource entities.

[0055] Specifically, in the knowledge model, entity classification is the fundamental structured mechanism for achieving semantic understanding and resource scheduling in task planning. Specifically, entities are divided into three categories: organizational entities, resource entities, and threat entities. Organizational entities are further subdivided into functional organizations and task organizations, respectively responsible for managing resource ownership and usage rights. This classification system has clear semantic boundaries and functional divisions in the task planning system, ensuring that the system can accurately identify the task execution entity and its associated authority when parsing natural language instructions.

[0056] Functional organizations are responsible for the ownership management of resource entities. Their permission relationships are typically modeled using the "ownership" attribute. For example, in a graph database, a "owns" relationship edge is established between functional organization nodes and resource entity nodes, with the ownership attribute attached. This is used to identify whether a resource can be scheduled or allocated. The Mission Organization is responsible for managing temporary usage rights of resources, establishing a "uses" relationship with the resource entity and attaching usage right attributes. This is used to indicate the availability status of resources within a specific task cycle. This mechanism supports the dynamic allocation and reclamation of resources during task execution, and is particularly suitable for rapid response requirements in ad-hoc task scenarios.

[0057] The classification of organizational entities must meet certain granularity and hierarchical specifications. For example, functional organizations can be divided into multiple levels based on their management functions, such as strategic, tactical, and operational levels, with each level corresponding to different resource management permissions. Task organizations, on the other hand, are dynamically combined based on task type, execution cycle, and collaboration requirements, and their combination rules can be formally described using system modeling languages ​​such as SysML. The classification of resource entities needs to consider their functional attributes, deployment dependencies, and state parameters; for example, detection and sensing resources must possess certain attributes. , Perception capability indicators, control resources must possess , Control response parameters, etc.

[0058] This step is widely applicable in practical scenarios involving high reliability tasks such as military command, emergency response, and intelligent operation and maintenance. For example, in military missions, functional organizations can represent logistics support units, task organizations can represent temporary combat teams, resource entities can include weapon platforms, communication equipment, reconnaissance payloads, etc., and threat entities are used to model enemy targets or environmental interference factors. By clearly defining the authority boundaries of organizations and resources, the system can achieve accurate identification and scheduling of mission execution entities, thereby improving the credibility and efficiency of mission planning.

[0059] This entity classification mechanism provides structured knowledge support for subsequent semantic alignment, spatiotemporal constraint modeling, and performance evaluation. By separating the functions and responsibilities of organizational entities, the system can more flexibly respond to changes in task requirements, realize dynamic authorization and reallocation of resources, and thus enhance the adaptability and robustness of task planning.

[0060] S22, the activity relationships between entities include parent-child relationship, entity-entity relationship, activity-activity relationship, entity-activity relationship and entity-attribute relationship, among which the activity-activity relationship is specifically represented by 13 temporal relationships of Allen interval algebra.

[0061] Specifically, in this invention, modeling the activity relationships between entities is one of the core steps in knowledge model construction, and its technical implementation is based on the structured representation of the semantic relationships between entities and their activities. Specifically, the activity relationships between entities include five categories: parent-child relationship, entity-entity relationship, activity-activity relationship, entity-activity relationship, and entity-attribute relationship. Among them, the activity-activity relationship is accurately modeled using 13 temporal relationships from Allen's Interval Algebra to support logical constraints and conflict detection in the time dimension during task execution.

[0062] At the technical implementation level, modeling activity-activity relationships first defines each activity as a time interval. , ],in Indicates the start time of the activity. Indicates the end time of the activity. Two activities. and The temporal relationships between them are described by one of the 13 fundamental relations of Allen's interval algebra, including: before, after, meet, met-by, overlaps, overlapped-by, contain, during, starts, started-by, finishes, finished-by, and equals. These relations are formally expressed through logical predicates, such as... express Prior to , express and Meeting, that is .

[0063] Modeling activity-activity relationships relies on the precise definition of time intervals and the logical combination of relationships. The system quantifies activities using attributes such as timestamps, durations, and activity types. For example, the duration of an activity... and the timing offset between activities. These parameters are used to construct constraints, such as The corresponding constraints are ,and Then it is required In practical applications, these relationships are embedded in the graph structure of the knowledge model and stored and reasoned as edge attributes between active nodes.

[0064] This step is widely used in fields with high requirements for temporal logic, such as military mission planning, emergency response scheduling, and industrial process optimization. For example, in military missions, reconnaissance activities must be completed before strike operations, and the time interval between the two must meet the minimum response time requirements for target identification and lock-on. Using 13 relationships from Allen's interval algebra, the system can automatically identify and eliminate schemes that do not conform to temporal logic, thereby improving the reliability and executability of mission planning.

[0065] By introducing the temporal relationships of Allen's interval algebra, the system achieves accurate modeling and reasoning of the sequence of activities during task execution, effectively avoiding conflicts and infeasibility issues caused by ambiguity in temporal logic in traditional task planning. Furthermore, this modeling approach provides structured temporal dimension support for subsequent constraint satisfaction solutions and multi-dimensional performance evaluation, thereby improving the intelligence level and decision support capabilities of task planning.

[0066] S301, the initial task scheme space is filtered according to the preset spatiotemporal constraint template to exclude schemes with spatiotemporal conflicts and retain the set of feasible schemes that meet the activity-activity relationship constraints.

[0067] Specifically, in some implementations, filtering the initial task scheme space according to a preset spatiotemporal constraint template is a key step in ensuring the logical feasibility of task planning in the method of this invention. The core of this step lies in verifying the constraint satisfaction of the initial task schemes generated after semantic parsing through a formal spatiotemporal constraint model, eliminating invalid schemes with spatiotemporal conflicts, and retaining a set of feasible schemes that conform to the activity-activity relationship constraints, thereby providing a high-quality candidate scheme foundation for subsequent multi-dimensional performance evaluation.

[0068] This step relies on 13 predefined Allen Interval Algebra relations in the knowledge model, including "before", "after", "meet", "meet", "overlap", "overlap", "contain", "contained", "begin at", "begin at", "end at", "end at", and "equal to". Each activity is represented as a time interval. , ],in Indicates the start time of the activity. This indicates the end time of the activity. The constraint solver will use these relationships to logically deduce the time sequence between activities, determining whether there are issues such as time overlap, reversed order, or resource conflicts. For example, if two activities... and Each from resources and Execute, and and There are deployment dependencies (such as) Deployed at (above), then The time interval must be consistent with The time interval must satisfy logical relationships such as "contains" or "is contained"; otherwise, it will be judged as infeasible.

[0069] The key parameters involved in this step include the start time of the activity. End time Duration and resource availability window In addition, the deployment relationships of resources, their state attributes (such as whether they are in use or faulty), and the strength of dependencies between activities (such as hard or soft constraints) must also be considered. During the constraint solving process, the system will perform conflict detection based on these parameters to ensure that the temporal logic between activities is consistent with the resource state.

[0070] This step is widely applicable to complex mission scenarios requiring multi-resource collaboration, such as military operations, emergency response and dispatch, and IoT resource management. In these scenarios, the order of mission execution, the spatiotemporal distribution of resources, and the dependencies between activities have a decisive impact on the success or failure of the mission. Through the filtering mechanism in this step, the system can quickly identify logically consistent feasible solutions in a dynamically changing environment, providing decision-makers with actionable planning recommendations.

[0071] This step significantly enhances the logical rigor and feasibility of task planning. By introducing Allen's interval algebra theory, the system can formally express the temporal relationships between activities, avoiding the subjectivity and uncertainty of traditional experience-based judgments. Simultaneously, by combining real-time monitoring of resource status and deployment relationships, the system ensures the dual feasibility of the task plan at both the physical and logical levels, laying a solid foundation for subsequent performance evaluation and Pareto optimal solution generation.

[0072] Furthermore, S301 includes: S31, 13 types of activity-activity relationships include prior to, subsequent to, meeting, being met, overlapping, being overlapped, containing, being contained, starting at, being started, ending at, being ended, and equal. The constraint satisfaction problem solver performs precise matching of activity time intervals.

[0073] Specifically, in the method of this invention, the 13 activity-activity relationships (including prior to, subsequent to, meeting, being met, overlapping, being overlapped, containing, being contained, starting at, being started, ending at, being ended, and equal) are the core elements for constructing the task execution time constraint model. These relationships are based on Allen's Interval Algebra theory and are used to accurately describe the relative temporal relationship between two activity time intervals, thereby providing a strict logical constraint basis for task planning.

[0074] Each activity is modeled as a time interval. , ],in Indicates activity The start time, This indicates its end time. The temporal relationships between activities are defined using these 13 Allen relations, such as "before" indicating... activities In the activity Completely ended before; "overlap" indicates activities With activities They partially overlap in time. These relationships are explicitly modeled through activity-activity relationship edges in the knowledge model and are input as constraints to the Constraint Satisfaction Problem Solver (CSP Solver) during the task solution generation process.

[0075] The activity time interval must meet the minimum duration. With maximum duration The limitation, namely Furthermore, time interval constraints between activities are typically expressed as time offsets. This indicates, for example, in the "after" relationship, the activity start time Should meet ,in This is the minimum interval time, and its value is set according to task logic or resource scheduling rules. In practical applications, these parameters can be dynamically configured based on factors such as task type, resource response time, and environmental interference to adapt to task planning needs in different scenarios.

[0076] This step is widely used in fields with strict requirements for timing logic, such as military mission scheduling, emergency response, and IoT resource coordination. For example, in multi-platform collaborative strike missions, reconnaissance activities must precede strike activities, and a minimum time interval must be met between the two. This ensures the timeliness of the target information. By inputting these relationships into the CSP solver, the system can quickly filter out feasible solutions that satisfy all timing constraints from a vast task solution space, thereby improving the logical rigor and execution reliability of task planning.

[0077] This step, by introducing 13 relations from Allen's interval algebra, achieves precise modeling and matching of the time dimension of task activities, effectively avoiding logical conflicts and time overlaps that may occur in traditional task planning based on fuzzy time descriptions. Simultaneously, combined with the efficient reasoning capabilities of the CSP solver, the system can quickly converge to a feasible solution under complex constraints, significantly improving the automation level and response speed of task planning, and providing a solid foundation for subsequent multi-dimensional performance evaluation and Pareto optimal solution generation.

[0078] S32, the spatiotemporal constraint template contains constraint rules that cause the loads deployed on it to become disabled due to platform failure. That is, when the platform state attribute is disabled, the state attribute of its associated load entity is automatically updated to disabled.

[0079] Specifically, in some implementations, the spatiotemporal constraint template includes constraint rules that cause platform failure to lead to the failure of payloads deployed on it. This technical implementation is based on a dynamic binding mechanism of deployment relationships and state attributes between entities in the knowledge model. Specifically, the platform and payloads are modeled through a "deployment" relationship, represented in the knowledge model as directed edges. The platform acts as the deployment subject, and the payload as the deployed object. Its logical expression is as follows: ,in Indicates the platform entity, Represents the load entity. When the platform status attribute... Marked as "disabled" (i.e.) When this occurs, the system will automatically trigger the adjustment of load status attributes based on the deployment relationship. The update synchronizes it to a "disabled" state (i.e., ...). This update process is implemented by the state propagation engine, and its core logic is: if... Then for all satisfying of ,implement .

[0080] Furthermore, the implementation of this constraint rule relies on a real-time monitoring and updating mechanism for entity state attributes within the knowledge model. Platform state attributes typically include operational status, availability, and fault level, and their updates can originate from sensor data streams, system logs, manual intervention, etc. During system operation, the update frequency of state attributes can be set to periodic polling (e.g., updating once per second) or event-driven updates (e.g., triggered by platform fault alarm events). During state propagation, the system can introduce a propagation delay parameter. It is used to simulate the transmission delay of state updates in a network or system, ensuring the timing consistency of state synchronization.

[0081] Optionally, this constraint rule can be extended by incorporating Allen Interval Algebra to support more complex temporal dependencies. For example, if platform failure occurs within the time interval of an activity, that activity will be marked as unexecutable, thus affecting the availability assessment of its associated payloads. This mechanism is of great significance in mission planning, especially in scenarios with strong resource deployment dependencies, such as when a drone platform failure will cause its onboard reconnaissance payloads, communication payloads, etc., to fail simultaneously, thereby affecting the feasibility and effectiveness assessment of mission execution.

[0082] This step plays a crucial role in constraint filtering within the mission planning system, ensuring that payloads on disabled platforms are not included in the available resource pool when generating mission plans, thereby avoiding logically infeasible planning results. Its technical value lies in improving the reliability and interpretability of mission planning, particularly in high-risk, highly dependent mission scenarios such as military operations and emergency response, demonstrating significant engineering application implications.

[0083] S401, a multi-dimensional performance evaluation system is used to quantitatively analyze the set of feasible solutions, and the Pareto optimal set of task solutions is selected by comprehensively considering task-level and activity-level performance indicators and resource status parameters.

[0084] Specifically, in some implementations, employing a multi-dimensional performance evaluation system to quantitatively analyze the set of feasible solutions is a core step in the method of this invention for optimizing task solutions. This step, by comprehensively considering task-level and activity-level performance indicators and resource status parameters, performs multi-objective performance evaluation on multiple task solutions that meet spatiotemporal constraints, thereby selecting the Pareto-optimal set of task solutions and providing a scientific basis for subsequent human-computer interaction decisions.

[0085] This performance evaluation system constructs a multi-dimensional performance indicator function based on structured entity, activity, and attribute information in the knowledge model. Task-level performance indicators include task completion quality. Task completion time and task resource consumption These are used to measure the degree of achievement of task objectives, execution efficiency, and resource usage costs, respectively. Activity-level performance indicators include the execution quality of each activity. Execution time and resource consumption Its calculation relies on analytical methods based on physical laws or empirical table lookup methods. Resource status parameters This includes resource availability, fault status, load capacity, etc., which are used to reflect the dynamic characteristics of resources when performing activities.

[0086] Task completion quality A normalized scoring mechanism is typically used, with values ​​ranging from [0,1]. This indicates that the mission objective has been fully achieved. Mission completion time. Calculate the total execution time of the plan based on time units (such as seconds or hours). Resource consumption. The metrics used are resource units (such as energy, manpower, and cost) to reflect the economic efficiency of task execution. Activity-level and task-level metrics are correlated through a weighted summation, with weights... It can be set by decision-makers in the human-computer interaction interface to reflect the priorities of different dimensions.

[0087] In application scenarios, this step is widely applicable to complex task environments such as emergency response, military missions, and IoT dispatch. For example, in disaster relief scenarios, the system needs to coordinate various heterogeneous resources (such as drones, rescue teams, and medical equipment) to complete multiple sub-tasks (such as search, rescue, and material transportation) within a limited time. In this case, a multi-dimensional performance evaluation system can help the system weigh time, resources, and quality, and select the optimal combination solution.

[0088] By introducing multi-objective optimization theory, the system can achieve an optimal balance among multiple conflicting objectives, avoiding local optima problems caused by optimizing a single indicator. The generation of the Pareto optimal solution set ensures that the efficiency of other dimensions cannot be further improved without reducing the efficiency of one dimension, thus providing decision-makers with a set of candidate solutions with global optimal potential, significantly improving the scientific nature and interpretability of task planning.

[0089] Furthermore, S401 includes: S41, task-level performance indicators include task completion quality, task completion time, and resource consumption, and their calculation formulas are as follows:

[0090] in For the quality of task completion, For the first The weights of each quality factor For the first Scores for each quality factor; The time taken to complete the task. For the first The duration of activities along the critical path; For resource consumption, For the first The consumption value of each resource instance.

[0091] Specifically, in a multi-dimensional performance measurement and evaluation system, task-level performance indicators are key quantitative parameters for measuring the overall performance of a task planning scheme. Their calculation formulas are task completion quality, task completion time, and resource consumption. In some implementations, task completion quality... A comprehensive evaluation can be conducted by considering both the achievement of task objectives and the tolerance for deviations during execution. This is typically achieved using a weighted average method, i.e.:

[0092] in, Indicates the first The quality of completion of each sub-task or key activity Each factor corresponds to a weighting coefficient, and these weights can be dynamically adjusted based on task priority, resource importance, or user-inputted preferences. The evaluation of task completion quality may also incorporate a task risk index. It is used to measure the robustness of a task in an uncertain environment, and its calculation can be based on the distribution density of threat entities and the anti-interference ability of resource entities.

[0093] Task completion time This refers to the total time span from task initiation to the achievement of the final goal, and its calculation formula is:

[0094] in, and These represent the start and end times of task execution, respectively. The time values ​​are derived from the temporal attributes of the entity-activity relationships in the knowledge model, and the temporal constraints in the activity-activity relationships defined by Allen's interval algebra. In practical applications, task duration also needs to consider the critical path length. The longest non-parallel path during task execution determines the shortest completion time of the task.

[0095] resource consumption It measures the total amount of various resources consumed during task execution, and its calculation formula is:

[0096] in, Indicates the first The resource cost consumed by a resource instance when performing its corresponding activity includes, but is not limited to, energy, manpower, time, and equipment wear and tear. The calculation of resource consumption depends on the real-time matching of resource instance status parameters (such as operating status and load capacity) and design parameters (such as maximum working duration and energy consumption coefficient), as well as the resource usage function based on physical laws or empirical models in the activity-level performance indicators.

[0097] This step is widely applicable in practical scenarios such as military mission planning, emergency response scheduling, and IoT resource collaboration. By introducing multi-dimensional performance indicators, the system can quantitatively compare multiple feasible solutions while meeting spatiotemporal constraints, thereby selecting the Pareto optimal solution set. Furthermore, this evaluation system supports a human-computer interaction interface, allowing decision-makers to dynamically adjust the weights of each dimension based on information such as mission urgency and environmental situation, achieving personalized and contextualized mission planning decisions. The technical value of this step lies in its significant improvement in the scientific rigor and response efficiency of mission planning through a structured and interpretable performance evaluation model, providing reliable support for intelligent decision-making in complex mission scenarios.

[0098] S42, the activity-level performance index is generated through physical law analysis or empirical value lookup table calculation, specifically including the joint calculation of time-varying state parameters of resource instances and design parameters, wherein the time-varying state parameters are acquired in real time by sensors, and the design parameters are obtained from the product manual.

[0099] Specifically, in some implementations, the generation of the activity-level performance indicators is based on physical law analysis or empirical value lookup table calculations. The technical implementation principle integrates the joint calculation of real-time sensor data and static design parameters to achieve dynamic performance evaluation of resource instances in specific activities. Specifically, the activity-level performance indicators include key dimensions such as the quality of resource completion of activities, the time required, and resource consumption. Their calculation relies on the joint modeling of time-varying state parameters of resource instances (such as current load, operating status, and location coordinates) and design parameters (such as maximum output power, theoretical operating distance, and response time).

[0100] Time-varying state parameters are collected in real time by sensors deployed on the resource entity. For example, spatial coordinate information is obtained through GPS, and the operational status of the resource (such as idle, in use, faulty, etc.) is obtained through a status monitoring module. The start and end times of the activity are recorded using timestamps. Design parameters are extracted from documents such as product specifications and technical manuals, and typically include the rated performance, theoretical limit values, and standard response time of the resource. These parameters have high stability and are suitable for long-term modeling and benchmark comparison.

[0101] Furthermore, the calculation model for activity-level effectiveness can be expressed as:

[0102] in, Indicates activity-level performance indicators. This represents the set of real-time status parameters for a resource instance. This represents the set of design parameters for the resource. This is a performance calculation function based on physical laws or table lookup operations. In the analytical method of physical laws, This may involve dynamic models, thermodynamic models, or communication models, for example, in control-related resources, their response time... It can be calculated using the following formula:

[0103] in, The instantaneous distance between resources and targets. The maximum movement speed of resources. This represents the average time delay of resource processing instructions.

[0104] This step is widely applicable to fields with high requirements for real-time performance and reliability, such as military mission planning, emergency response scheduling, and IoT resource management. For example, in UAV collaborative strike missions, activity-level performance indicators can reflect the accuracy and time taken for a single UAV to complete a strike mission under current environmental and load conditions, thereby supporting the system in selecting the Pareto optimal solution from multiple options.

[0105] By combining the dynamic state of resources with static design parameters, this method enables precise quantitative evaluation of task execution performance, enhancing the scientific rigor and interpretability of task planning. Furthermore, it supports rapid adjustments under different environments and target characteristics, improving the system's adaptability and decision-making efficiency in ad-hoc task scenarios.

[0106] S501 displays the Pareto optimal task solution set as a graph or visualization, dynamically updates the solution based on real-time status monitoring data, and triggers local or global replanning iterations when execution deviations are detected.

[0107] Specifically, in some implementations, displaying the Pareto optimal task solution set as a graph or visualization, and dynamically updating the solution based on real-time status monitoring data, is a key step in realizing human-machine collaborative decision-making and closed-loop control in the task planning system. The technical implementation of this step is based on the structured output of a knowledge model and multi-dimensional performance evaluation results. Through graph modeling and a visualization engine, abstract planning solutions are transformed into intuitive graphical representations, facilitating quick understanding and selection by decision-makers.

[0108] The system first maps each task solution in the Pareto optimal solution set into a graph structure, where nodes represent resource instances, task activities, or target objects, and edges represent the activity relationships, spatiotemporal dependencies, or performance paths between them. The graph can be constructed using a graph database (such as Neo4j) or a table-based database based on SPO (Subject-Predicate-Object) relationships to support efficient query and update operations. Visualization graphics are then transformed into multi-dimensional views such as Gantt charts, network topology graphs, resource load graphs, and timeline graphs through a front-end rendering engine (such as D3.js, ECharts, or WebGL). Each view corresponds to a different performance dimension, such as time performance, resource performance, and quality performance.

[0109] The visualization system needs to support dynamic refresh rate configuration, typically set to refresh every 30 seconds to 5 minutes, to adapt to the real-time requirements of different task scenarios. When detecting execution deviations, the system employs a threshold mechanism; for example, if the execution time of a task activity deviates from the expected time by more than [a certain amount], [the system will take certain action]. When a resource status changes from "in use" to "faulty", a replanning mechanism is triggered. It can be configured according to the task type and system responsiveness; a typical value is... Second.

[0110] This process is widely applicable to fields with high requirements for timeliness and reliability in mission execution, such as military command, emergency response, and intelligent logistics. During mission execution, the system continuously receives real-time data streams from sensors, status monitoring modules, and intelligence systems. By comparing these streams with the planned timeline, deviations are identified and their impact is assessed. If the deviation affects local activities, the system triggers local replanning, adjusting only the affected sub-task chains. If the deviation affects the global structure, a global iterative replanning is initiated, regenerating a Pareto-optimal set of solutions that satisfies the current state.

[0111] This step transforms task planning from static generation to dynamic adaptation, significantly improving the system's real-time responsiveness and decision support capabilities. Through graphical visualization, the system enhances the interpretability and interactive nature of the solutions, enabling decision-makers to quickly identify the optimal path in complex task environments and intervene promptly in case of execution anomalies, thereby ensuring efficient and reliable task execution.

[0112] Furthermore, it also includes: S601 accesses real-time status sensor data of resource instances and target status data of intelligence information systems, and determines whether to trigger local or global replanning iterations based on preset deviation thresholds. Local replanning is for single resource failure or target status deviation, while global replanning is for multi-resource collaborative failure or significant changes in target characteristics.

[0113] Specifically, during task execution, the system accesses real-time status sensor data from resource instances and target status data from the intelligence information system, and determines whether to trigger local or global replanning iterations based on a preset deviation threshold. This step is a key link in achieving dynamic adaptability in task planning, and its technical implementation relies on the coordinated operation of multi-source data fusion, status monitoring mechanisms, and threshold determination logic.

[0114] The system accesses real-time status data of resource instances through standardized interfaces (such as OPC UA, MQTT, REST API, etc.), including but not limited to location coordinates, running status (such as idle, in use, faulty), task execution progress, and resource load capacity. Simultaneously, target status data provided by the intelligence information system is typically accessed in the form of a data stream, containing the target's time-varying state (such as movement trajectory, behavior patterns), design parameters (such as target type, structural characteristics), and cognitive bias information (such as target identification confidence level, intelligence update frequency). The system maps this data to entity nodes in the knowledge model, forming a dynamic representation of the task execution status.

[0115] The system sets multiple deviation thresholds to determine whether to trigger replanning. For example, a time deviation threshold. Used to measure the deviation between actual execution time and planned time, if If this occurs, replanning will be triggered; spatial deviation threshold Used to measure the deviation between the actual location and the expected location of a resource, if If so, it is considered abnormal. In addition, the system also sets a resource failure detection threshold. When resource status parameters (such as failure rate and availability) fall below a certain threshold, the resource is considered disabled. These thresholds can be dynamically configured based on task type, resource characteristics, and environmental complexity, and users can make on-the-spot adjustments through a human-computer interaction interface.

[0116] The instruction-driven task planning method based on a knowledge model in this invention can achieve accurate semantic parsing and task planning of natural language instructions. By combining the knowledge model with spatiotemporal constraints, it generates a reliable task plan that satisfies the optimal performance in multiple dimensions, and supports visual human-computer interaction and dynamic adjustment, thereby improving the agility and intelligence of task planning.

[0117] Example 3 To implement the methods of the above embodiments, the present invention also provides an instruction-driven ad hoc task planning system 10 based on a knowledge model, such as... Figure 4 As shown, it includes: The keyword extraction module 100 is used to preprocess natural language instructions and extract keywords related to the task. The task scheme space generation module 200 is used to perform semantic understanding of keywords based on a pre-built knowledge model, resolve fuzzy references into specific instance objects and their activities, map nouns to the instance layer, align verbs to relational constraints, and generate a semantically consistent task scheme space. The task scheme set generation module 300 is used to filter the task scheme space under spatiotemporal constraints to obtain a feasible task scheme set; The optimal subset selection module 400 is used to perform multi-dimensional performance measurement and evaluation on each task scheme in the feasible task scheme set based on the instance, so as to select the Pareto optimal subset. The interactive output module 500 is used to output the Pareto optimal subset in the form of a graph or visualization, supporting intelligent or human-computer interactive decision-making.

[0118] Furthermore, Figure 5 This is a schematic diagram of the composition architecture of the instruction-driven ad hoc task planning system based on a knowledge model provided in an embodiment of the present invention.

[0119] The instruction-driven task planning system based on a knowledge model in this invention can achieve accurate semantic parsing and task planning of natural language instructions. By combining the knowledge model with spatiotemporal constraints, it generates a reliable task plan that satisfies the optimal performance in multiple dimensions, and supports visual human-computer interaction and dynamic adjustment, thereby improving the agility and intelligence of task planning.

[0120] To implement the above embodiments, this application also proposes a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the method described in the foregoing embodiments.

[0121] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0122] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0123] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. A knowledge-model-based instruction-driven ad hoc task planning method, characterized in that, The method includes: Preprocess the natural language instructions to extract task-related keywords; Based on a pre-built knowledge model, semantic understanding of keywords is performed, fuzzy references are parsed into specific instance objects and their activities, nouns are mapped to the instance layer, verbs are aligned to relational constraints, and a semantically consistent task solution space is generated. The task scheme space is filtered under spatiotemporal constraints to obtain a set of feasible task schemes. Based on examples, the performance of each task scheme in the feasible task scheme set is evaluated in multiple dimensions to select the Pareto optimal subset. The Pareto optimal subset is output in the form of a graph or visualization, supporting intelligent or human-computer interactive decision-making.

2. The method according to claim 1, wherein preprocessing of the natural language instruction includes: Handle situations where words in natural language instructions refer to the same denotative intent but cause symbolic differences; Based on general knowledge, keywords are extracted using large language models or natural language processing techniques and aligned to concepts or instances in the knowledge model. After completing the semantic extraction and alignment of keywords, the implicit operational logic of the instructions is transformed into a structured execution flow, and an executable flow architecture is generated using SysML or an equivalent system modeling language.

3. The method according to claim 1, characterized in that, Knowledge models are used to represent the specialized knowledge required for task planning, including: The knowledge model is a system architecture for specific real-world tasks, including the construction of the components and relationships of task planning objects, and the adjustment of the instance architecture based on natural language instructions; The nodes of the knowledge model represent entities, and the relationships between nodes represent the activities between entities. There are multiple types of relationships between entities, and a contextual understanding of entities is established through a complex relationship network. Obtain adjustment instructions for the corresponding knowledge model; The knowledge model includes organizational entities, resource entities, and threat entities. Organizational entities include functional organizations and task organizations. Functional organizations are organizations that manage resources with specific professional functions and have ownership of resource entities. Task organizations are organizations that are created on the spot for specific tasks and have the right to use resource entities. The existence of functional organizations and task organizations is predefined and hierarchical, and they are adjusted on the spot according to instructions, reflecting the decision-making level's need to adjust the form of resource organization.

4. The method according to claim 3, characterized in that, The adjustment instruction is an instruction message for adjusting the task knowledge model; The adjustment instructions are categorized by object, including instructions for adjusting at the knowledge model class level and instructions for adjusting at the knowledge model instance level. Adjustment instructions can be categorized by their intent into adjustments to organizational relationships and adjustments to collaborative relationships.

5. The method according to claim 3, characterized in that, The classification of resource entities is diverse, reflecting the heterogeneity of resources at different levels. According to the system composition, it is divided into system, subsystem, and system components, which are related to each other through their compositional relationships; Based on their mutual deployment dependencies, they can be divided into payloads and platforms. Payloads and platforms are associated through deployment relationships, and these relationships reflect existence; that is, if a platform fails, the payloads deployed on it will also fail. The knowledge model includes the concept of attributes that quantitatively describe entities, and these attributes include those that reflect spatiotemporal characteristics. The knowledge model includes descriptions of the relationships between entities, including parent-child relationships, entity-entity relationships, activity-activity relationships, entity-activity relationships, and entity-attribute relationships. The specific forms of knowledge models include graph databases, ontology model databases, and table databases that reflect SPO relationships.

6. The method according to claim 1, characterized in that, The spatiotemporal constraints are used to characterize the sequential information of task execution, including: Constructing a basic representation framework for spatiotemporal constraints: Defining task execution as an emergent process in which multiple resource instances perform activities and generate activity effects under specific spatiotemporal constraints, where constraints within the time frame are a prerequisite for the superposition of activity effects and the generation of an overall task effect; Define the time interval constraints of the activity process: describe the activity process as a time interval, and use 13 Allen interval algebras to represent the relationship between the time intervals of two activity processes, including prior to, subsequent to, meeting, being met, overlapping, being overlapped, containing, being contained, starting at, being started, ending at, being ended, and equal relationships; Establish an effectiveness mechanism under spatiotemporal constraints: measure the effectiveness of task execution within the spatiotemporal scope, including setting the precondition that the interval between the discovery of the target and the actual effect acting on the target should be less than the interval between the target's escape and the range of effect.

7. The method according to claim 1, characterized in that, A multi-dimensional performance measurement and evaluation system is used to construct the relationship between the status of task execution resource instances and the expected task execution performance, including: Establish a multi-dimensional evaluation system that includes task-level performance indicators, activity-level performance indicators, resource instance status, and environmental and target characteristic parameters; By directly accessing resource instance sensors through information systems to obtain time-varying state parameters, and obtaining stable design parameters from product design specifications; by acquiring environmental parameters that affect task activities through environmental monitoring capabilities, and designing and determining parameter requirements based on the sensitivity of resources to environmental parameters; by acquiring design parameters of target objects through intelligence information systems, and by acquiring time-varying state parameters of targets through monitoring systems. To address the discrepancies between intelligence and surveillance information and objective reality, rapid adjustments are made to the solution based on the current resource instance status, environmental parameters, and target characteristic parameters, while meeting the task triggering conditions. When the information discrepancy effect occurs, a solution is quickly regenerated. By conducting multi-dimensional quantitative performance evaluations of all feasible task solutions, the optimal solution under the current conditions is selected.

8. The method according to claim 7, characterized in that, The task-level performance indicators cover three categories: task completion quality, completion time, and resource consumption, and reserve a human-computer interaction interface. Decision-makers can predefine or adjust them on the spot based on the urgency of the task and environmental situation information. Activity-level performance indicators include the quality, time, and resource consumption of each entity's resources in completing the activity. Their calculation relies on the analysis of physical laws or empirical value lookup table calculations.

9. The method according to claim 1, characterized in that, The output task solution is constrained to satisfy feasibility and optimality in at least one evaluation dimension, and is displayed graphically, including: The graphical interface presents the spatiotemporal location, effectiveness range, and direction of action of resource instances, enabling decision-makers to intuitively grasp the overall picture of the plan and confirm its feasibility through visualization. It presents the start time, end time, and expected time of the execution effect of the corresponding activities for resource execution, forming a complete time process view; When execution deviations occur due to resource limitations or errors in understanding the target, the system automatically initiates the next round of task planning and iteration based on the monitored deviation information, thereby achieving dynamic optimization and continuous improvement of the solution.

10. A knowledge-model-based instruction-driven ad hoc task planning system, characterized in that, include: The keyword extraction module is used to preprocess natural language instructions and extract keywords related to the task. The task solution space generation module is used to perform semantic understanding of keywords based on a pre-built knowledge model, resolve fuzzy references into specific instance objects and their activities, map nouns to the instance layer, align verbs to relational constraints, and generate a semantically consistent task solution space. The task scheme set generation module is used to filter the task scheme space under spatiotemporal constraints to obtain a feasible task scheme set; The optimal subset selection module is used to perform multi-dimensional performance measurement and evaluation on each task scheme in the feasible task scheme set based on the instance, so as to select the Pareto optimal subset. The interactive output module is used to output the Pareto optimal subset in the form of a graph or visualization, supporting intelligent or human-computer interactive decision-making.

Images