A complex task decision system and method based on distributed cross-domain knowledge service

Abstract

The application discloses a complex task decision system and method based on distributed cross-domain knowledge service. The application constructs a five-dimensional collaborative framework of knowledge representation dimension, relationship structure dimension, reasoning calculation dimension, situation interaction dimension and evolution control dimension. Under a unified global unique identifier mapping system, loose coupling communication and independent deployment are realized through a standardized data interface. Multi-source heterogeneous field knowledge is converted into parameterized and calculable micro-service components. Based on a group behavior closed-loop mechanism with stability constraints, incremental evolution of the knowledge network is realized. Thus, problems such as data island, search dimension disaster and poor system dynamic self-healing ability in complex high-concurrency decision scenarios are effectively solved.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence, big data, distributed systems, knowledge graphs, and comprehensive decision-making technologies, specifically to a complex task decision-making system and method based on distributed cross-domain knowledge services. Background Technology

[0002] With the continuous development of IoT, distributed computing, and AI technologies, the operating environments of complex systems (such as the Industrial Internet, aviation support, and unmanned swarm collaboration) are exhibiting high dynamism and complexity. Against this backdrop, their decision-making models have gradually moved beyond reliance on single-domain experience, shifting towards a comprehensive decision-making paradigm driven by multi-source data and cross-domain knowledge collaboration. In this context, the Knowledge-as-a-Service (KaWa) model, which encapsulates structured domain knowledge into services and provides support in the form of callable components, has become a crucial technological path for supporting complex decision-making tasks.

[0003] Despite some advancements in related technologies, existing solutions still face several limitations when dealing with highly dynamic and uncertain real-world application scenarios. These limitations are mainly reflected in the following aspects:

[0004] In terms of cross-domain knowledge fusion and unified representation, knowledge systems from different domains are usually distributed in heterogeneous structures across different system nodes, lacking a unified expression and alignment mechanism. When decision-making tasks involve multi-domain collaboration, existing methods struggle to achieve effective mapping and consistent modeling between multi-source knowledge, leading to semantic biases or distorted relationships in cross-domain information fusion, thus affecting the reliability of subsequent reasoning and decision-making.

[0005] Regarding the organization and evolution capabilities of distributed knowledge systems, existing knowledge graphs mostly adopt centralized or weakly distributed architectures, with their computation and update mechanisms primarily based on static construction and querying. As the scale of nodes and the frequency of interactions increase, the system faces significant pressure in terms of computational efficiency and response time. Simultaneously, existing methods generally lack effective utilization of user or agent interaction behavior, making it difficult to promptly transform feedback information from actual use into incremental optimizations of the knowledge structure, resulting in knowledge system updates lagging behind practical application needs.

[0006] Furthermore, in multi-agent decision support processes, existing systems typically output results based on predefined rules or fixed strategies, failing to adequately consider the differences among agents in terms of cognitive state, risk preference, and task environment. Related modeling processes are often independent of each other, lacking unified coordination between knowledge relationship modeling, path search mechanisms, and strategy adjustment processes, making it difficult to form a continuous, closed-loop dynamic adjustment mechanism. Summary of the Invention

[0007] In view of this, the present invention provides a complex task decision-making system and method based on distributed cross-domain knowledge services. By constructing a five-dimensional collaborative framework of knowledge representation dimension, relation structure dimension, reasoning and computation dimension, contextual interaction dimension, and evolution control dimension, each dimension achieves loosely coupled communication and independent deployment through standardized data interfaces under a unified globally unique identifier mapping system. This transforms multi-source heterogeneous domain knowledge into parameterized and computable microservice components, and realizes incremental evolution of the knowledge network based on a group behavior closed-loop mechanism with stability constraints. This effectively solves problems such as data silos, the curse of search dimensionality, and poor dynamic self-healing ability of the system in complex high-concurrency decision-making scenarios.

[0008] For ease of understanding, the basic concepts involved in this invention are defined as follows: The decision entity refers to an object, resource, event, state, or constraint extracted from business data, environmental data, or interactive data in various fields that can influence the decision-making process of complex tasks. For example, in an emergency dispatch scenario, the decision entity may include roads, vehicles, hospitals, rescue equipment, disaster events, road blockage status, resource availability status, and safety constraints.

[0009] The atomic knowledge unit refers to the smallest computable knowledge object formed after semantic vectorization and metadata binding of the decision entity. Each atomic knowledge unit includes at least a globally unique identifier, entity type, semantic embedding vector, priority weight, timeliness label, and availability status identifier.

[0010] The knowledge node refers to the graph structure representation of the atomized knowledge unit in a cross-domain knowledge relationship network.

[0011] The metadata refers to additional attributes associated with the decision-making entity and used to assist in subsequent filtering, sorting, routing, and evolutionary updates. It includes at least priority weights, timeliness tags, and availability status identifiers. The priority weights characterize the importance of the entity in business decisions, the timeliness tags characterize the valid time range of the entity's data, and the availability status identifiers characterize whether the entity is currently able to participate in subsequent inference or scheduling.

[0012] The semantic embedding vector refers to a continuous vector obtained by encoding decision entities through a representation learning model, which is used to calculate the semantic similarity, spatial distance, or association strength between different entities in a unified global semantic space.

[0013] The complex task decision-making method based on distributed cross-domain knowledge services of the present invention includes the following steps: S1, extract decision entities that can influence the decision-making process from multi-source heterogeneous domain data, wherein the decision entities include objects, resources, events, states or constraints; encode the decision entities into semantic embedding vectors of a unified dimension; Construct atomic knowledge units for each decision entity; each atomic knowledge unit includes the semantic embedding vector and metadata of the decision entity; the metadata includes a globally unique identifier, entity type, and status identifier, and the status identifier includes priority weight, timeliness label, and availability; S2, for each domain, uses the atomized knowledge units of the domain as knowledge nodes, the business logic relationships between knowledge nodes in the domain as edges, and the association strength or passage cost as the weight of the edges to construct a locally weighted directed graph for the domain. Extract cross-domain alignment anchor pairs that have strong business relationships between different domains; For each domain, a mapping matrix for that domain is constructed to map the local semantic space of that domain to a shared global reference semantic space; A cross-domain knowledge relationship network is constructed, where network nodes are mapping points of knowledge nodes from all domains in the global reference semantic space. The edges of the cross-domain knowledge relationship network include local domain edges and cross-domain connection edges. The local domain edges correspond to the edges of the local weighted directed graph of that domain and retain the corresponding weights. The cross-domain connection edges are generated based on cross-domain alignment anchors, and edge weights are assigned according to the semantic distance, alignment confidence, and business association strength between the cross-domain alignment anchors. The greater the semantic distance, the greater the weight; the higher the alignment confidence and business association strength, the smaller the weight. S3 encodes the decision task request, which includes the task objective, current state information, and constraints, to obtain the query vector. ; For each domain, firstly, knowledge nodes representing the current state information in that domain are extracted from the decision task request context as the initial triggering starting set. Then, the query vector is projected using a graph constraint projection operator. Mapping this to a local weighted directed graph within the domain yields a localized intent vector for that domain. Based on the semantic embedding vector of knowledge nodes in this field, calculate the spatial geometric distance between the intent vector in this field and the knowledge nodes in this field, and select knowledge nodes whose distance is less than a preset matching threshold as candidate response nodes in this field. The candidate response nodes in all domains are merged, deduplicated, and sorted to obtain the target response anchor point set; Starting from the initial set of triggering nodes and using the set of target response anchors as the candidate target range, topological routing is performed in the cross-domain knowledge relationship network. During the path expansion process, the path expansion cost is calculated based on the connection cost between the current node and the next node, as well as the semantic deviation of the next node relative to its domain localized intent vector. The path is then expanded in order of increasing path expansion cost until the node in the target response anchor set is reached, generating a cross-domain decision micro-orchestration chain. S4, the cross-domain decision micro-orchestration chain is distributed to the decision-making entity terminals participating in the current decision-making task according to task roles, domain responsibilities or execution permissions, and the interaction status data of each decision-making entity terminal is collected in real time; the interaction status data is compared with the preset security threshold matrix, and when any data exceeds the corresponding threshold, a decision execution blockage interruption signal is generated, and intervention is performed to update the decision scheme; S5, each decision-making entity terminal and the corresponding distributed sharding node in each field write the status log, path log and feedback log related to task execution according to the unified log structure, and logically aggregate them based on task identifier, field identifier, knowledge node identifier and globally unique identifier to form a distributed global log storage structure; For the interactive data in the distributed global log library, aggregate statistics and smoothing are performed within a sliding time window to obtain smoothed statistical features; based on the smoothed statistical features, the cross-domain knowledge relationship network topology is updated.

[0014] Preferably, in step S1, the decision entity is encoded into a semantic embedding vector of a uniform dimension through a representation learning model.

[0015] Preferably, the representation learning model is a cross-domain semantic encoding model based on contrastive learning, or a graph attention network that incorporates local graph structure constraints.

[0016] Preferably, in S2, the business logic relationship includes the dependency relationship, constraint relationship, causal relationship or temporal relationship between knowledge nodes.

[0017] Preferably, in step S2, the mapping matrix is ​​obtained by solving the following optimization problem:

[0018]

[0019]

[0020] in, As a balancing coefficient, Ω represents the domain set, Ω = {transportation, water, healthcare, emergency}. Ω indicates that the current process is in progress. k Each field Within the k-th neighborhood, sum each edge in the local graph of that neighborhood. Indicates the first k In each field, nodes i With nodes j There is a directed correlation edge between them. Indicates the first k In each field, nodes i With nodesj The weight corresponding to this edge, Indicates the first k In the first field i The original semantic embedding vector of each knowledge node in the local semantic space Indicates the first k In the first field i Each node is mapped to a vector in a unified global semantic space. Indicates the first k Knowledge nodes in each field i With the l Knowledge nodes in each field j This forms a set of cross-domain alignment anchor point pairs. This is the preset distance tolerance threshold.

[0021] By constraining the spatial distance between related knowledge nodes in a unified global semantic space, we ensure that logically related cross-domain entities remain clustered in the unified global semantic space without disrupting the original topological rigidity of each domain subgraph.

[0022] Preferably, in step S3, the query vector is projected using a graph constraint operator. Mapping this to a local weighted directed graph within the domain yields a localized intent vector for that domain. Specifically: Construct a topological mask function constrained by the local weighted directed graph connectivity relationships in this domain. ; Filter knowledge nodes that have a logical connection with the initial triggering point set, specifically: when a knowledge node With the initial triggering start set When logical connectivity exists, ;otherwise All satisfy The nodes constitute a local knowledge subgraph ; Calculate the query vector using a graph-constrained attention mechanism. semantic embedding vectors of each knowledge node within the local knowledge subgraph Association weight :

[0023] in, This is a vector similarity function; Then the localized intent vector in this domain for .

[0024] Preferably, in step S3, when performing topology routing in a cross-domain knowledge relationship network, the objective function is: to minimize the weighted sum of the cumulative topology routing cost and semantic deviation. Specifically, when performing single-step node jump expansion, a set of next-hop candidate nodes for the current node is obtained. The set of next-hop candidate nodes includes knowledge nodes in the same domain connected by local domain edges and knowledge nodes in different domains connected by cross-domain edges. For each candidate node, the topology path cost between the current node and the candidate node is read. The topology path cost between the current node and the candidate knowledge node in the same domain is the weight of the local domain edge, and the topology path cost between the current node and the candidate knowledge node in different domains is the weight of the cross-domain edge. And determine the corresponding localized intent vector based on the domain to which the candidate node belongs. Calculate the semantic deviation of the candidate node relative to the corresponding localized intent vector. :

[0025] in, The vector of the current node; The vector of candidate nodes for the next hop; The path expansion cost is calculated based on the topology routing cost and semantic deviation, and the path is expanded in order of increasing path expansion cost until a node in the target response anchor set is reached or a preset termination condition is met. When the generated path contains at least one cross-domain connection edge, or when the knowledge nodes in the path belong to two or more domains, the path is identified as a cross-domain decision path. The system organizes the domain-specific knowledge nodes and cross-domain knowledge nodes into a cross-domain decision micro-arrangement chain according to the sequential dependencies between the knowledge nodes in the path.

[0026] Preferably, the distributed topology routing employs a parallel computing mechanism, and simultaneously, a small non-negative real number is assumed. To preset the pruning threshold, if This terminates the subsequent expansion of the branch, thereby suppressing branch paths with high semantic deviation.

[0027] Preferably, in step S4, the interaction status data includes task execution response time, operation error rate, and node retry or rejection frequency. Set global abstraction level parameters Local abstraction level parameters Explain the granularity parameters The global abstraction level parameter controls the overall display level of the cross-domain decision path in the decision-making entity terminal. The larger the global abstraction level parameter, the smaller the graph display depth threshold, the more abstract the displayed content, and the more hidden underlying execution details. The local abstraction level parameter controls the degree of local expansion within the neighborhood of abnormal or blocked nodes. The smaller the local abstraction level parameter, the more finely the child nodes, dependencies, and constraints around abnormal or blocked nodes are expanded. The explanation presentation granularity parameter controls the scope and level of detail of the explanation content output by the text generation and processing module. The larger the explanation presentation granularity parameter, the greater the path backtracking depth, the higher the degree of information expansion, and the more detailed the output explanation content. The decision execution stall and interruption signals are classified and processed to trigger different parameterized intervention mechanisms: (1) If the task execution response time exceeds the preset time threshold and no effective execution action is detected, the corresponding decision-making entity terminal is determined to be in a state of cognitive overload, triggering the parameterized topology folding and interpretation enhancement mechanism: Increase global abstraction level parameters Based on the depth threshold A bottom-up node aggregation process is performed on the current cross-domain decision path subgraph to generate a simplified decision path representation, reducing the instantaneous cognitive load on the decision-making terminal; at the same time, the granularity parameters of the interpretation presentation are adjusted upwards. Based on the reverse analysis of the execution path of the current blocked node, key dependencies and constraints are extracted. Based on the key dependencies and constraints, corresponding auxiliary explanations are generated to show the key logical relationships of the decision path. The current cross-domain decision path subgraph is a subgraph extracted from the cross-domain knowledge relationship network based on the cross-domain decision micro-arrangement chain generated by S3, which contains the knowledge nodes and their connecting edges traversed by the cross-domain decision micro-arrangement chain. (2) If the error rate or retry / rejection frequency exceeds the preset frequency threshold, the corresponding decision-making entity terminal is determined to be in an execution conflict state, and the conflict intervention mechanism is triggered: Preserve global abstraction level parameters Remain unchanged, but lower the local abstraction level parameter corresponding to the abnormal node. It performs fine-grained expansion on abnormal nodes, displaying their child nodes and dependencies, so that the decision-making terminal can locate the source of the conflict; at the same time, it increases the granularity of the interpretation presentation parameters. Extract the specific dependency constraints that caused the error or path rejection and encapsulate them into an explanation context, driving the text generation and processing module to output a micro-constraint tracing report; (3) Scheme restructuring and feedback recording mechanism: The decision path generated after intervention is integrated with the corresponding explanatory information to form an updated auxiliary decision-making scheme, which is then redistributed to the relevant decision-making terminals. The response time and number of compensations during this interaction are recorded, and a cost bias term is generated based on the type of obstruction / interruption signal. Based on this, the joint traversal cost, including a time-based weight term and a compensation weight term, is calculated. The joint traversal cost is written as scalar data along with the corresponding state flag into a distributed global log library. Specifically,

[0028] Among them Joint traversal cost; t is the response time during this interaction process; The preset standard response time; This refers to the number of times compensation, retries, or reissues were performed after this intervention. This is the time-consuming weighting coefficient; To compensate for the weighting coefficient; It is a type of blockage; This is the cost bias term corresponding to the type of blocking / interruption signal.

[0029] Preferably, in step S5, the statistical features include call frequency, execution success rate, average response time, average traversal cost, and number of rejections; the update of the cross-domain knowledge relationship network topology specifically involves: (1) Calculate the comprehensive performance index of each preset recommended path based on the statistical characteristics; if a path is blocked multiple times or manually rejected within the statistical window, and its average crossing cost is higher than the preset penalty threshold, increase the passage cost weight of the corresponding edge of the path; conversely, if the execution success rate and comprehensive performance index of a path are higher than the preset threshold, decrease the passage cost weight of the corresponding edge of the path; the comprehensive performance index of the preset recommended path refers to the comprehensive evaluation parameter used to evaluate the actual execution effect of the cross-domain decision path recommended by the system within the statistical window; the comprehensive performance index is determined based on the execution success rate, call frequency, average response time, average crossing cost and rejection number of the path within the statistical window, wherein the higher the execution success rate, the higher the call frequency, the lower the average response time, the lower the average crossing cost and the fewer the rejections, the higher the comprehensive performance index of the preset recommended path; conversely, the lower the comprehensive performance index of the preset recommended path.

[0030] (2) Analyze the actual execution sequence in the distributed global log data to identify atypical flow paths; the atypical flow path refers to a cross-node operation sequence that does not explicitly define direct connection edges in the initial cross-domain knowledge relationship network, but repeatedly appears in the actual task execution process and has a stable success rate; based on the atypical flow path, instantiate implicit relationship connection edges between the starting knowledge node and the ending knowledge node in the unified global semantic space and assign them initial weights; (3) The local relaxation optimization algorithm is adopted to perform local mapping matrix parameter and topological connection relationship update operations on the affected knowledge nodes and their neighboring nodes, so as to complete the autonomous anti-aging incremental evolution of the cross-domain knowledge relationship network.

[0031] Preferably, in S5, the implicit relationship instantiation is as follows: if there is a cross-node flow in the distributed global log library that is not defined in the initial logic, and its success rate within the time window is higher than a preset threshold, then a corresponding directed topological edge is automatically created in the cross-domain knowledge relationship network to solidify the cross-domain collaborative shortcut generated by decision-making practice.

[0032] This invention provides a distributed architecture system employing the above method, comprising: The atomic entity representation module is used to realize the semantic vectorization and dynamic metadata binding of heterogeneous entities; The cross-domain isomorphic fusion module is used to construct a global cross-domain relationship network through affine transformation and manifold alignment; The distributed topology routing engine is used to perform target anchor locking and distributed topology routing based on sharding computation. The adaptive perception intervention module is used to monitor decision-making obstacles and reconstruct decision schemes based on parameterized adjustment. The controlled incremental evolution module is used for incremental updates and implicit logic discovery of the group feedback driven network based on smoothing processing.

[0033] Beneficial effects: (1) The distributed five-dimensional cross-domain knowledge service collaborative evolution architecture and method described in this invention realizes the isomorphic mapping of heterogeneous knowledge in a unified global semantic space through affine mapping and cross-domain alignment and fusion mechanism, which completely breaks the cross-domain data silos between traditional systems and effectively supports the rapid synthesis of cross-domain complex decision chains.

[0034] (2) The distributed five-dimensional cross-domain knowledge service collaborative evolution architecture and method described in this invention adopts distributed sharding computation and geometric heuristic pathfinding, which greatly compresses the computational complexity of large-scale graph search from the global full scale to the local subgraph scale, significantly improving the performance robustness of the system, thereby effectively supporting real-time decision response in massive concurrent scenarios.

[0035] (3) The distributed five-dimensional cross-domain knowledge service collaborative evolution architecture and method described in this invention introduces a dynamic adjustment mechanism for the abstract hierarchy of decision instructions and the granularity of interpretation presentation parameters, realizes situational adaptive intervention against overload, effectively alleviates the cognitive overload and analysis paralysis of decision-makers under extreme high pressure environment, and significantly improves the success rate of human-machine collaborative decision-making.

[0036] (4) The distributed five-dimensional cross-domain knowledge service collaborative evolution architecture and method described in this invention establishes an evolution mechanism based on time window smoothing, which can automatically solidify the "implicit shortcut" generated in actual combat based on group interaction feedback and automatically avoid high resistance paths, giving the system a strong self-repair capability, ensuring that the underlying logic of the system can continuously approach the real physical operation law, and has extremely high evolution stability. Attached Figure Description

[0037] Figure 1 This is an overall diagram of the complex task decision-making system architecture based on distributed cross-domain knowledge services of the present invention.

[0038] Figure 2 This is the main flowchart of the complex task decision-making method based on distributed cross-domain knowledge services of the present invention.

[0039] Figure 3 This is a schematic diagram of the cross-domain manifold mapping and isomorphic space generation of the present invention.

[0040] Figure 4 This is a schematic diagram of the heuristic topology pathfinding based on semantic deviation gradient of the present invention.

[0041] Figure 5 This is a schematic diagram of the parameterized topology folding mechanism under cognitive overload conditions of the present invention.

[0042] Figure 6 This is a schematic diagram illustrating the cooperative evolution and implicit relationship generation of the present invention. Detailed Implementation

[0043] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0044] This invention provides a complex task decision-making system and method based on distributed cross-domain knowledge services, deployed in a distributed cluster environment. The distributed cluster consists of multiple physically isolated and logically orthogonal computing nodes. Each computing node is configured with a multi-core processor (CPU) and a graphics processing unit (GPU) or tensor processing unit (TPU / NPU) as the underlying computing power support.

[0045] Each node independently maintains its own in-memory graph database shards. Nodes communicate via Remote Procedure Call (RPC) protocols, preferably gRPC, and use protocol buffers for structured data transmission. At the system level, a unified Globally Unique Identifier (GUID) mapping system is established to enable cross-node, cross-dimensional data association and service governance.

[0046] Furthermore, the functional modules are decoupled through standardized interfaces. When some modules become unavailable, the system can perform degraded operation based on existing data, thereby ensuring the continuity of overall service.

[0047] Based on the aforementioned system operating environment and considering the scenario of cross-domain urban extreme disaster emergency command and dispatch, the specific implementation process of the distributed five-dimensional cross-domain knowledge service collaborative evolution method described in this invention is explained. In sudden events such as urban flooding caused by torrential rain, systems such as transportation, water management, healthcare, and emergency response are typically operating in a decentralized manner, corresponding to different domain subspaces, lacking a unified dispatch mechanism. This invention achieves dynamic collaborative decision-making among multiple systems through unified representation, cross-domain modeling, and collaborative evolution. The specific steps are as follows: Step 1: Construct an atomized knowledge unit layer In this embodiment, this step is performed by the atomized entity representation module. The system extracts decision entities that can influence the current decision-making process from data sources in fields such as transportation, water resources, healthcare, and emergency response. These decision entities include objects, resources, events, states, or constraints such as road flooding points, road closures, chemical defense vehicles, water pumping trucks, hospital emergency resources, power outages in flooded areas, and on-site safety constraints. Subsequently, the system encodes each decision entity into a unified dimension using a preset representation learning model. 3D continuous semantic embedding vector ,in Preferably, the value is 256 or 512. The representation learning model is preferably a cross-domain semantic encoding model based on contrastive learning, or a graph attention network that incorporates local graph structure constraints.

[0048] For each decision entity, the system constructs an atomic knowledge unit for it. This atomic knowledge unit includes a globally unique identifier, entity type, and semantic embedding vector. And including priority weights Time-sensitive labels and availability status indicators Metadata, including.

[0049] in, Continuous semantic features used to represent the decision entity; Used to characterize the prior importance of the decision-making entity in the current business logic; Used to indicate the valid time range of the decision entity's data; This is used to indicate whether the decision-making entity is currently able to participate in subsequent reasoning or scheduling, and .when When the decision entity's corresponding knowledge node participates in subsequent calculations; when During the routing phase, the system automatically filters the node. When the emergency resources of a hospital are already at full capacity, the availability status of its corresponding entity can be set to unavailable to prevent the knowledge node from being included in subsequent scheduling routes.

[0050] Compared to processing methods that only summarize data, this step uses the joint representation of semantic embedding vectors and metadata to enable different types of decision entities to have a unified measurement basis. The resulting atomized knowledge units further serve as knowledge nodes in cross-domain knowledge relationship networks, providing a unified computable input for subsequent cross-domain relationship modeling, distributed reasoning, and collaborative evolution.

[0051] Step 2: Construct a cross-domain manifold relationship layer In this embodiment, this step is performed by the cross-domain isomorphic fusion module. Based on the knowledge nodes formed in step 1, the system constructs locally weighted directed graphs in subspaces of various fields such as transportation, water affairs, medical care, and emergency response.

[0052] against Local subgraphs of each domain ,in This represents the set of knowledge nodes in the local subgraph. This represents the set of edges in the local subgraph. The edge weight matrix (with dimensions ) represents the strength of the topological association between knowledge nodes. ,in (This represents the total number of knowledge nodes in the corresponding local subgraph). The edges in the local weighted directed graph are used to represent the dependencies, constraints, causal relationships, or temporal relationships between knowledge nodes in the same domain, and the edge weights are used to represent the association strength or passage cost of the corresponding relationship.

[0053] In cross-domain urban extreme disaster emergency command and dispatch scenarios, the local subgraph in the transportation domain consists of knowledge nodes such as road blockage nodes, road closure nodes, and road capacity reduction nodes; the local subgraph in the water domain consists of knowledge nodes such as water-affected area nodes and water pumping vehicle nodes; the local subgraph in the medical domain consists of knowledge nodes such as hospital emergency resource nodes and patient reception capacity nodes; and the local subgraph in the emergency response domain consists of knowledge nodes such as chemical protection vehicle dispatch nodes and on-site disposal nodes. Thus, each domain first forms a knowledge structure reflecting its local business logic within its own domain.

[0054] The system extracts cross-domain alignment anchor pairs that have strong business relationships between different domains, forming a set of cross-domain alignment anchor pairs. In this embodiment, the cross-domain alignment anchor pairs are determined manually by experts or learned from historical data. A road blockage node in the transportation domain and an emergency medical delay node in the medical domain form a set of cross-domain alignment anchor pairs. These cross-domain alignment anchor pairs do not require that the different domains have completely identical shared entities, but rather that the corresponding knowledge nodes have a stable and clear cross-domain association in business logic.

[0055] Based on the set of anchor points, the system solves for the first... k Mapping matrix from local semantic space of each domain to a unified global semantic space Considering the high-order nonlinear manifold characteristics of the multi-domain semantic space, the mapping matrix employs a local affine approximation transformation to perform piecewise alignment of the cross-domain subspace under manifold constraints. It should be specifically noted that the mapping matrix... Essentially, dimension is The semantic evolution matrix, whose object is the semantic evolution matrix obtained in step 1. Dimensional semantic embedding vectors are used to map knowledge nodes from different domains to a unified global semantic space.

[0056] The transformation matrix is ​​solved through iterative optimization, and its joint objective function is defined as:

[0057]

[0058]

[0059] in, As a balancing coefficient, Ω represents the domain set, Ω = {transportation, water, healthcare, emergency}. Ω indicates that the current process is in progress. k Each field Within the k-th neighborhood, sum each edge in the local graph of that neighborhood. Indicates the first k In each field, nodes i With nodes j There is a directed correlation edge between them. Indicates the first k In each field, nodes i With nodes j The weight corresponding to this edge, Indicates the first k In the first field i The original semantic embedding vector of each knowledge node in the local semantic space Indicates the first k In the first field i Each node is mapped to a vector in a unified global semantic space. Indicates the first k Knowledge nodes in each field i With the l Knowledge nodes in each field j This forms a set of cross-domain alignment anchor point pairs. A preset distance tolerance threshold is used to limit the maximum permissible deviation of cross-domain related knowledge nodes in a unified global semantic space. When the distance between a pair of cross-domain related knowledge nodes after mapping exceeds the distance tolerance threshold, the deviation is considered to be greater than or equal to the distance tolerance threshold. At that time, a squared penalty is applied to the portion exceeding the threshold, driving the mapping matrix. Continue to optimize in subsequent iterations.

[0060] For the local topology preservation term, based on the edge set and edge weight matrix of the local weighted directed graph in each domain. By connecting the matrix through the graph The constraint is constructed using the mean square error of the geometric distance between adjacent nodes before and after projection, and is used to constrain the relative structural stability of adjacent knowledge nodes in the same domain before and after mapping. For cross-domain alignment deviation terms, based on the set of cross-domain alignment anchor pairs The cross-domain alignment anchor point pairs are constructed using Euclidean distance in a unified global semantic space. This constrains the distance deviation between knowledge nodes with business relationships between different domains in the unified global semantic space, ensuring that cross-domain related entities maintain semantic proximity in the unified global semantic space.

[0061] By optimizing the joint objective function, the system obtains the optimized mapping matrix. Based on this, knowledge nodes in locally weighted directed graphs from various fields such as transportation, water affairs, healthcare, and emergency response are mapped to a unified global semantic space. Cross-domain connection edges are generated based on the cross-domain alignment anchor pairs. These edges are assigned weights based on the semantic distance, alignment confidence, and business relevance strength between the cross-domain alignment anchors; a larger semantic distance results in a larger weight, while higher alignment confidence and business relevance strength result in a smaller weight. This forms a cross-domain knowledge relationship network containing local domain edges and cross-domain connection edges. The originally physically isolated and logically dispersed multi-domain knowledge structures are unified into a unified global semantic space, providing a unified relational and geometric metric foundation for subsequent distributed reasoning and cross-domain decision-making micro-orchestration chain generation.

[0062] Compared to cross-department integration methods that rely on predefined rules or simple data table cascading, this step unifies knowledge nodes from different domains into a unified global semantic space through semantic space mapping and cross-domain alignment under manifold constraints. This breaks through the isolation limitations between discrete network topologies and provides a unified spatial geometric measurement basis for cross-domain physical constraints and business decision-making needs.

[0063] Step 3: Construct a distributed inference service layer In this embodiment, this step is performed by the distributed topology routing engine.

[0064] Before performing reasoning, the semantic embedding vectors of knowledge nodes in each domain subspace are first mapped to a unified global semantic space through the affine transformation mapping matrix described in step 2, so as to establish a unified geometric metric benchmark required for cross-domain reasoning.

[0065] When a cross-regional urban extreme disaster emergency command and dispatch mission is triggered, the system receives a current decision-making task request. The decision-making task request includes the task objective, current status information, and constraints. The task objective may be to complete personnel evacuation and chemical rescue; the current status information may include severe flooding on some roads and ongoing power outage repairs in several areas; and the constraints may include prioritizing the avoidance of high-risk areas and ensuring the accessibility of critical medical resources.

[0066] After receiving the decision task request, the microservice gateway extracts knowledge nodes representing the current event state from the request context as the initial triggering starting point set. Simultaneously, a multimodal parser is used to extract the original modal features from the decision task request and encode them into a query vector. Subsequently, the system uses graph-constrained projection operators. The query vector is mapped to a local weighted directed graph for each domain to obtain the localized intent vector for that domain. .

[0067] Specifically, the system introduces a topological mask function that constrains the connectivity relationships of a locally weighted directed graph in this domain. The selection of knowledge nodes that have a logical connection with the initial triggering point set is defined as follows: when a node... With the triggering start set When logical connectivity exists, ;otherwise All satisfied The nodes constitute a local knowledge subgraph .

[0068] Based on this, the system utilizes a graph-constrained attention mechanism to compute the query vector. semantic embedding vectors of knowledge nodes within a local knowledge subgraph Association weight :

[0069] in, A correlation scoring function for measuring the similarity between vectors.

[0070] Through topological filtering using the masking function, the weights of disconnected knowledge nodes are suppressed or set to zero, ensuring that the domain-localized intent vector is jointly generated by nodes within the local knowledge subgraph relevant to the current task. The final domain-localized intent vector... Only from local knowledge subgraphs The nodes within the range are generated, satisfying the following convex combinatorial mathematical form:

[0071] The constraints are as follows: and .

[0072] To ensure consistency of the metrics, the model parameters of the attention mechanism and the representation learning model in step 1 are obtained through offline joint contrastive learning pre-training, and are continuously calibrated online during system operation using incremental interaction logs to ensure the mapped intent vector. Strictly anchored within the convex hull spanned by prior knowledge nodes, it can serve as a standard metric coordinate in the unified global semantic space to participate in subsequent geometric distance calculations.

[0073] After generating localized intent vectors for each domain, the microservice gateway distributes them to distributed sharding nodes corresponding to the domains of transportation, water, healthcare, and emergency response. Each distributed sharding node stores and manages local knowledge nodes and local graph data for its corresponding domain. Based on the semantic embedding vectors of the local knowledge nodes, it calculates the spatial geometric distance between the localized intent vectors of that domain and the local knowledge nodes in parallel, and filters knowledge nodes with a distance less than a preset matching threshold as candidate response nodes. Each distributed sharding node returns its local filtering results to the microservice gateway, which merges, deduplicates, and sorts the candidate response nodes from different domains to generate a target response anchor set. The target response anchor point set includes key resource nodes, dependency state nodes, and target result nodes involved in executing the current decision task, which are used to limit the range of candidate targets for subsequent path search.

[0074] In this embodiment, traffic segment nodes can filter out passable road nodes or perimeter control nodes; water segment nodes can filter out water pump truck nodes or flood-affected area treatment nodes; medical segment nodes can filter out hospital receiving nodes; and emergency segment nodes can filter out chemical defense vehicle dispatch nodes. After being uniformly aggregated by the microservice gateway, the system obtains the target response anchor point set corresponding to the current personnel evacuation and chemical defense rescue mission. .

[0075] Based on the initial triggering start set and target response anchor set The system performs topology routing in the cross-domain knowledge relationship network formed in step 2: using the initial triggering starting set... Using the target response anchor set as the starting point for path search Define the candidate target range and utilize the localized intent vector of the relevant domain. The path expansion is guided and performed within a cross-domain knowledge relationship network. The essence of the path search is to solve a joint optimization objective: minimizing the cumulative topological pathfinding cost (given by the edge weight matrix). (determined by) and semantic deviation (by gradient) The weighted sum of the factors (determined by the system) is used as the objective function. During path expansion, the system extracts semantic deviation as a heuristic pruning factor. .

[0076] Specifically, during single-step node jump expansion, a set of next-hop candidate nodes for the current node is obtained. This set includes candidate knowledge nodes in the same domain connected by local domain edges and candidate knowledge nodes in different domains connected by cross-domain edges. For each candidate node, the topology path cost between the current node and that candidate node is read. The topology path cost between the current node and the candidate knowledge node in the same domain is the weight of the local domain edge, and the topology path cost between the current node and the candidate knowledge node in different domains is the weight of the cross-domain edge. Based on the domain to which the candidate node belongs, a corresponding localized intent vector is determined, and the semantic deviation of the candidate node relative to the corresponding localized intent vector is calculated. Let the vector of the current node be... The next hop candidate node vector is The calculation rule for the semantic deviation is as follows:

[0077] Furthermore, preset For small non-negative real numbers ( This is used to define the system's tolerance for local deviations during semantic exploration. When At this time, it indicates that the next hop node is gradually moving away from the target intention in geometric space; if If the branch exceeds the preset geometric divergence pruning threshold, it is determined that the extension branch deviates semantically from the current core task objective, and the system directly terminates the subsequent extension of the branch.

[0078] In this embodiment, if a candidate scheduling path needs to traverse a severely flooded area or pass through a high-risk area undergoing power outage repairs, the semantic deviation of the candidate knowledge node corresponding to that path from the intent of personnel evacuation and chemical rescue missions will significantly increase. (Preset pruning threshold) The system prunes the branch accordingly; while for candidate branches that simultaneously meet the constraints of accessibility, availability of chemical defense resources and hospital acceptance, the system will prioritize their retention and continue to expand.

[0079] The path expansion cost is calculated based on the topology routing cost and semantic deviation, and the path is expanded in ascending order of path expansion cost until a node in the target response anchor set is reached or a preset termination condition is met. When the generated path contains at least one cross-domain connection edge, or when the knowledge nodes in the path belong to two or more domains, the path is determined as a cross-domain decision path. The system organizes intra-domain knowledge nodes and cross-domain knowledge nodes into a cross-domain decision micro-orchestration chain according to the sequential dependencies between knowledge nodes in the path.

[0080] In this embodiment, the system can generate a cross-domain decision-making micro-arrangement chain of "peripheral traffic control - power outage confirmation in water-affected areas - chemical protection vehicle entry - hospital reception and treatment", providing a direct execution basis for subsequent multi-subject interactive intervention and collaborative evolution.

[0081] Step 4: Construct a multi-agent context interaction layer In this embodiment, this step is performed by the adaptive perception intervention module.

[0082] The system distributes the cross-domain decision-making micro-orchestration chain generated in step 3 to the decision-making entity terminals participating in the current decision-making task, according to task roles, domain responsibilities, or execution permissions. These decision-making entity terminals include command terminals, traffic operation terminals, medical dispatch terminals, and related intelligent agent execution terminals. The system collects real-time interaction status data from each decision-making entity terminal. This interaction status data includes task execution response time, operation error rate, and node retry or rejection frequency. The system compares this interaction status data with a preset security threshold matrix. When any indicator exceeds the corresponding threshold, a decision execution obstruction / interruption signal is generated.

[0083] Furthermore, the system constructs a state feature vector based on the limit-crossing index and classifies the decision execution stall and interruption signals to trigger corresponding parameterized intervention mechanisms. It should be noted that the abstract-level parameters involved in the mechanisms described later differ from the granularity parameters presented in the explanation. All parameters are constrained by a preset normalization range. Their dynamic values ​​are calculated in real time by inputting the state feature vector into a preset nonlinear mapping function (such as the sigmoid function) to ensure the stability of the human-machine system's control boundary. The abstract level parameter controls the display level of the cross-domain decision micro-orchestration chain on the terminal, determining whether to display high-level task steps or low-level execution details. This embodiment includes a global abstract level parameter. and local abstraction level parameters The global abstraction level parameter controls the overall display level of the cross-domain decision path in the decision-making entity's terminal. A larger global abstraction level parameter results in a smaller graph display depth threshold, more abstract displayed content, and more hidden underlying execution details. The local abstraction level parameter controls the degree of local expansion within the neighborhood of abnormal or blocked nodes. A smaller local abstraction level parameter results in finer expansion of child nodes, dependencies, and constraints around abnormal or blocked nodes. The explanation presentation granularity parameter controls the degree of expansion of dependencies, constraints, and causal explanations. A larger explanation presentation granularity parameter results in greater path backtracking depth, higher information expansion, and more detailed output explanations.

[0084] (1) Cognitive overload intervention mechanism When it is detected that the task execution response time exceeds the preset time threshold and there is no effective execution action, the corresponding decision-making subject is determined to be in a state of cognitive overload, and the parameterized topology folding and interpretation enhancement mechanism is triggered.

[0085] During topology folding, the system adjusts the global abstraction level parameters upwards. Read the maximum topology depth of the current cross-domain knowledge relationship network. And based on parameters Calculated map displays depth threshold The threshold With parameters The relationship exhibits a monotonically decreasing trend, and the preferred option satisfies the relationship. .

[0086] Based on the depth threshold The system performs bottom-up node aggregation processing on the current cross-domain decision path subgraph. It traverses the node topology depth. For satisfying For each node, its execution constraint information is extracted, including execution time, resource consumption, and related state attributes. This constraint is then propagated upwards and fused to a depth of [depth value missing] using feature vector concatenation or scalar accumulation. In the parent node.

[0087] After constraint fusion is completed, the system performs a hiding process on the underlying nodes at the display layer and replaces the original node structure with the fused parent node, forming a decision node with a higher semantic level. Through the above processing, while retaining the underlying constraint information, the system performs structural compression on the multi-level nested cross-domain decision micro-arrangement chain, generating a simplified decision path representation. In this embodiment, when massive alarm information in the early stage of a disaster causes the command terminal to time out, the system can only present high-level task steps such as activating flood control and evacuation plans and dispatching chemical and biological rescue, while hiding the underlying resource scheduling details, reducing the instantaneous cognitive load on the decision-making entity.

[0088] While performing topology folding, the system upscales the granularity parameters of the interpretation presentation. Based on the current blocked node, the system execution path is reverse-analyzed to extract key dependencies and constraints. The explanation presents granular parameters. This is used to control the output range and granularity of the text generation and processing module, including the path backtracking depth and the degree of information expansion. Based on the dependent paths and constraint information, the text generation and processing module generates corresponding auxiliary explanatory content to demonstrate the key logical relationships of the decision path.

[0089] (2) Implement conflict intervention mechanisms If the micro-operation error rate or the frequency of retry / rejection exceeds the preset frequency threshold, the corresponding decision-making entity is determined to be in an execution conflict state.

[0090] In this state, the system keeps the global abstraction level parameters unchanged and lowers the local abstraction level parameters corresponding to the abnormal nodes. Fine-grained expansion is performed on abnormal nodes to display their child nodes and dependencies. That is, when the system determines that the current conflict is concentrated in a local execution stage, it does not compress the overall task structure again, but expands the low-level details of the local abnormal nodes so that the decision-maker can locate the source of the conflict.

[0091] At the same time, the granularity parameter of the explanation was increased. The system extracts the specific dependency constraints that caused the error or path rejection and encapsulates them into an explanatory context, driving the text generation and processing module to output a micro-constraint tracing report. Simultaneously, it provides optional alternative path suggestions in the interface. In this embodiment, if a system-recommended scheduling path is rejected by the command terminal because it needs to traverse a high-risk area undergoing power outage repairs, the system can expand the underlying dependencies of that path and output a constraint explanation stating that the high-voltage lines along the path are undergoing power outage repairs, posing secondary risks. This helps the decision-maker understand the difference between the system's recommendation and their experience-based judgment.

[0092] (3) Scheme restructuring and feedback recording mechanism The system integrates the decision path generated after intervention with the corresponding explanatory information to form an updated auxiliary decision-making scheme, which is then redistributed to the relevant decision-making terminals. Upon receiving an execution confirmation signal from the terminal, the system records the response time and number of compensation attempts during this interaction, and generates a cost bias term based on the obstruction type. Based on this, it calculates the joint traversal cost, including both time-based and compensation-based weights. :

[0093] Among them Joint traversal cost; t is the response time during this interaction process; The preset standard response time; This refers to the number of times compensation, retries, or reissues were performed after this intervention. This is the time-consuming weighting coefficient; To compensate for the weighting coefficient; It is a type of blockage; This is the cost bias term corresponding to the type of stagnation.

[0094] The joint traversal cost, as scalar data, is written together with the corresponding status flag into a distributed global log library jointly written by each decision-making entity terminal and each distributed shard node. This log is used for subsequent steps such as topology edge weight optimization, implicit relationship discovery, and continuous updating and evolution of the underlying model. The distributed global log library does not directly mix and store the original business data from various domains. Instead, each decision-making entity terminal and each distributed shard node writes status logs, path logs, and feedback logs related to task execution according to a unified log structure. These logs are logically aggregated based on task identifiers, domain identifiers, knowledge node identifiers, and globally unique identifiers to form a distributed global log storage structure.

[0095] Step 5: Construct a global-local co-evolution layer In this embodiment, this step is performed by the controlled incremental evolution module.

[0096] The system performs aggregation statistics and smoothing processing on the interaction data in the distributed global log library within a sliding time window to reduce the impact of short-term fluctuations on statistical results under high-concurrency scenarios. Specifically, the aggregation statistics refer to summarizing and statistically analyzing task execution logs, path logs, and feedback logs within the sliding time window; the smoothing processing involves applying a moving average, weighted average, or exponential smoothing to the aggregation statistics in chronological order to weaken the instantaneous impact of single abnormal events or short-term high-frequency fluctuations on statistical features. Based on the smoothed statistical features, an update operation on the cross-domain knowledge relationship network topology is triggered. These statistical features include call frequency, execution success rate, average response time, average traversal cost, and number of rejections, used to characterize the execution effect of paths, nodes, or edges within the preset time window.

[0097] (1) Edge weight dynamic modulation mechanism The system extracts and calculates the comprehensive performance index of each preset recommended path. The comprehensive performance index of the preset recommended path refers to a comprehensive evaluation parameter used to assess the actual execution effect of the cross-domain decision-making path recommended by the system within a statistical window. The comprehensive performance index is determined based on the path's execution success rate, call frequency, average response time, average traversal cost, and number of rejections within the statistical window. Higher execution success rate, higher call frequency, lower average response time, lower average traversal cost, and fewer rejections result in a higher comprehensive performance index for the preset recommended path; conversely, a lower comprehensive performance index results in a lower comprehensive performance index. If a path experiences multiple blockages or is manually rejected within the statistical window, and its average traversal cost exceeds a preset penalty threshold, the passage cost weight of the corresponding edge of that path is increased; conversely, if the execution success rate and comprehensive performance index of a path exceed the preset threshold, the passage cost weight of the corresponding edge of that path is decreased. Through this mechanism, bidirectional modulation and adaptive evolution of the topological priority of the locally weighted directed graph in each domain are achieved.

[0098] In this embodiment, for cross-domain collaborative paths that statistically show high success rates and low average traversal costs, the system reduces the search cost of their associated edges in subsequent topology pathfinding; for paths with frequent blockages, high rejection rates, or high average traversal costs, the system increases the search cost of their associated edges. Simultaneously, the system uses a threshold to limit the variation of edge weights, suppressing excessive fluctuations in a short period to avoid drastic oscillations in the knowledge network topology.

[0099] (2) Implicit relation generation mechanism The system analyzes actual execution sequences in distributed global log data to identify frequently occurring atypical flow paths with stable execution results. These atypical flow paths refer to cross-node operation sequences that are not explicitly defined as direct connections in the initial cross-domain knowledge relationship network, but repeatedly occur during actual task execution and have a stable success rate. To accurately capture contextual dependencies in the dynamic decision-making process, the system introduces a sequence modeling mechanism based on Long Short-Term Memory (LSTM) networks. This mechanism tracks the evolution of continuous cross-domain micro-operation sequences of the decision-making terminal subject based on Markov hidden states, thereby more accurately capturing the deep contextual dependencies of multiple subjects in complex cognitive flows.

[0100] When the sequence modeling network detects that the implicit state transition probability of a certain atypical flow path meets a preset convergence threshold, and its actual success rate within the current time window reaches the admission criteria, the atypical flow path is identified as a valid empirical shortcut. Based on the atypical flow path, the system instantiates implicit relational connection edges between the starting knowledge node and the ending knowledge node in a unified global semantic space, assigns them initial weights, and sets an upper limit on the number of newly added edges and constraints on the magnitude of edge weight changes to avoid excessive expansion or drastic disturbance of the graph structure in a short period of time.

[0101] In this embodiment, if the system observes a high frequency of pre-operation confirmation of power outage in the water-affected area before mobilizing the water pump truck, and the operation sequence continuously meets the success rate requirement within a preset time window, the system can add implicit relationship connection edges between the corresponding knowledge nodes to solidify cross-domain collaborative experience.

[0102] (3) Local coordinate update mechanism Because the aforementioned edge weight adjustments or the addition / deletion of implicit relational edges alter the connectivity at the discrete graph level, thereby breaking local geometric isomorphism constraints, the system employs a local relaxation optimization algorithm to perform local coordinate updates. While maintaining the coordinates of most knowledge nodes in the unified global semantic space, only the continuous semantic features of the affected knowledge nodes and their neighboring nodes are recalculated. Specifically, the system will... Update links of edge weight matrices in locally weighted directed graphs of dimension The update link of the continuous semantic evolution matrix is ​​processed separately. The discrete graph-level connection matrix represents the topological connections between knowledge nodes, while the continuous semantic evolution matrix represents the continuous semantic mapping relationships of knowledge nodes in a unified global semantic space. The system first fine-tunes the affected nodes in... Unify continuous semantic vectors in the global semantic space to generate calibrated ones. dimensional vector Simultaneously, based on local evolution errors, the corresponding updates are performed synchronously. Semantic evolution matrix The local mapping parameters. To prevent distortion of the unified global semantic space structure caused by long-term local updates, the system asynchronously introduces a low-frequency global recalibration and verification mechanism in the background to periodically verify and correct the local update results.

[0103] Ultimately, the system incrementally synchronizes the updated continuous semantic coordinates and discrete topological relationships to the corresponding distributed sharding nodes in each domain through a globally unique identifier (GUID) mechanism, thus completing the autonomous anti-aging incremental evolution of the cross-domain knowledge relationship network.

[0104] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. 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.

Claims

1. A complex task decision-making method based on distributed cross-domain knowledge services, characterized in that, include: S1, extract decision entities that can influence the decision-making process from multi-source heterogeneous domain data, wherein the decision entities include objects, resources, events, states or constraints; encode the decision entities into semantic embedding vectors of a unified dimension; Construct atomic knowledge units for each decision-making entity; The atomized knowledge unit includes the semantic embedding vector of the decision entity and metadata; the metadata includes a globally unique identifier, entity type, and status identifier, and the status identifier includes priority weight, timeliness label, and availability; S2, for each domain, uses the atomized knowledge units of the domain as knowledge nodes, the business logic relationships between knowledge nodes in the domain as edges, and the association strength or passage cost as the weight of the edges to construct a locally weighted directed graph for the domain. Extract cross-domain alignment anchor pairs that have strong business relationships between different domains; For each domain, a mapping matrix for that domain is constructed to map the local semantic space of that domain to a shared global reference semantic space; A cross-domain knowledge relationship network is constructed, where network nodes are mapping points of knowledge nodes from all domains in the global reference semantic space. The edges of the cross-domain knowledge relationship network include local domain edges and cross-domain connection edges. The local domain edges correspond to the edges of the local weighted directed graph of that domain and retain the corresponding weights. The cross-domain connection edges are generated based on cross-domain alignment anchors and are assigned edge weights based on the semantic distance, alignment confidence, and business association strength between the cross-domain alignment anchors. S3 encodes the decision task request, which includes the task objective, current state information, and constraints, to obtain the query vector. ; For each domain, firstly, knowledge nodes representing the current state information in that domain are extracted from the decision task request context as the initial triggering starting set. Then, the query vector is projected using a graph constraint projection operator. Mapping this to a local weighted directed graph within the domain yields a localized intent vector for that domain. ; Based on the semantic embedding vector of knowledge nodes in this field, the spatial geometric distance between the intent vector and the knowledge nodes in this field is calculated, and knowledge nodes whose distance is less than a preset matching threshold are selected as candidate response nodes in this field. The candidate response nodes in all domains are merged, deduplicated, and sorted to obtain the target response anchor point set; Starting from the initial set of triggering nodes and using the set of target response anchors as the candidate target range, topological routing is performed in the cross-domain knowledge relationship network. During the path expansion process, the path expansion cost is calculated based on the connection cost between the current node and the next node, as well as the semantic deviation of the next node relative to its domain localized intent vector. The path is then expanded in order of increasing path expansion cost until the node in the target response anchor set is reached, generating a cross-domain decision micro-orchestration chain. S4, the cross-domain decision micro-orchestration chain is distributed to the decision-making entity terminals participating in the current decision-making task according to task roles, domain responsibilities or execution permissions, and the interaction status data of each decision-making entity terminal is collected in real time; the interaction status data is compared with the preset security threshold matrix, and when any data exceeds the corresponding threshold, a decision execution blockage interruption signal is generated, and intervention is performed to update the decision scheme; S5, each decision-making entity terminal and the corresponding distributed sharding node in each field write the status log, path log and feedback log related to task execution according to the unified log structure, and logically aggregate them based on task identifier, field identifier, knowledge node identifier and globally unique identifier to form a distributed global log storage structure; For the interactive data in the distributed global log library, aggregate statistics and smoothing are performed within a sliding time window to obtain smoothed statistical features; based on the smoothed statistical features, the cross-domain knowledge relationship network topology is updated.

2. The method as described in claim 1, characterized in that, In step S1, the decision entity is encoded into a semantic embedding vector of a unified dimension through a representation learning model.

3. The method as described in claim 2, characterized in that, The representation learning model is a cross-domain semantic encoding model based on contrastive learning, or a graph attention network that integrates local graph structure constraints.

4. The method as described in claim 1, characterized in that, In S2, the business logic relationship includes the dependency relationship, constraint relationship, causal relationship or temporal relationship between knowledge nodes.

5. The method as described in claim 1, characterized in that, In S2, the mapping matrix is ​​obtained by solving the following optimization problem: in, As a balancing coefficient, Ω represents the domain set, Ω = {transportation, water, healthcare, emergency}. Ω indicates that the current process is in progress. k Each field Within the k-th neighborhood, sum each edge in the local graph of that neighborhood. Indicates the first k In each field, nodes i With nodes j There is a directed correlation edge between them. Indicates the first k In each field, nodes i With nodes j The weight corresponding to this edge, Indicates the first k In the first field i The original semantic embedding vector of each knowledge node in the local semantic space Indicates the first k In the first field i Each node is mapped to a vector in a unified global semantic space. Indicates the first k Knowledge nodes in each field i With the l Knowledge nodes in each field j This forms a set of cross-domain alignment anchor point pairs. This is the preset distance tolerance threshold.

6. The method as described in claim 1, characterized in that, In step S3, the query vector is projected using the graph constraint operator. Mapping this to a local weighted directed graph within the domain yields a localized intent vector for that domain. Specifically: Construct a topological mask function constrained by the local weighted directed graph connectivity relationships in this domain. ; Filter knowledge nodes that have a logical connection with the initial triggering point set, specifically: when a knowledge node With the initial triggering start set When logical connectivity exists, ;otherwise All satisfy The nodes constitute a local knowledge subgraph ; Calculate the query vector using a graph-constrained attention mechanism. semantic embedding vectors of each knowledge node within the local knowledge subgraph Association weight : in, This is a vector similarity function; Then the localized intent vector in this domain for 。 7. The method as described in claim 1, characterized in that, In S3, when performing topology routing in a cross-domain knowledge relationship network, the objective function is to minimize the weighted sum of the cumulative topology routing cost and semantic deviation. Specifically, when performing single-step node jump expansion, a set of next-hop candidate nodes for the current node is obtained. The set of next-hop candidate nodes includes knowledge nodes in the same domain connected by local domain edges and knowledge nodes in different domains connected by cross-domain edges. For each candidate node, the topology path cost between the current node and the candidate node is read. The topology path cost between the current node and the candidate knowledge node in the same domain is the weight of the local domain edge, and the topology path cost between the current node and the candidate knowledge node in different domains is the weight of the cross-domain edge. And determine the corresponding localized intent vector based on the domain to which the candidate node belongs. Calculate the semantic deviation of the candidate node relative to the corresponding localized intent vector. : in, The vector of the current node; The vector of candidate nodes for the next hop; At the same time, let a small nonnegative real number be denoted as . To preset the pruning threshold, if This terminates any further expansion of the branch. The path expansion cost is calculated based on the topology routing cost and semantic deviation, and the path is expanded in order of increasing path expansion cost until a node in the target response anchor set is reached or a preset termination condition is met. When the generated path contains at least one cross-domain connection edge, or when the knowledge nodes in the path belong to two or more domains, the path is identified as a cross-domain decision path. The system organizes the domain-specific knowledge nodes and cross-domain knowledge nodes into a cross-domain decision micro-arrangement chain according to the sequential dependencies between the knowledge nodes in the path.

8. The method as described in claim 1, characterized in that, In S4, the interaction status data includes task execution response time, operation error rate, and node retry or rejection frequency. Set global abstraction level parameters Local abstraction level parameters Explain the granularity parameters ; The decision execution stall and interruption signals are classified and processed to trigger different parameterized intervention mechanisms: (1) If the task execution response time exceeds the preset time threshold and there is no effective execution action, the corresponding decision-making entity terminal is determined to be in a state of cognitive overload, triggering the parameterized topology folding and interpretation enhancement mechanism: Increase global abstraction level parameters Based on the depth threshold A bottom-up node aggregation process is performed on the current cross-domain decision path subgraph to generate a simplified decision path representation, reducing the instantaneous cognitive load on the decision-making terminal; at the same time, the granularity parameters of the interpretation presentation are adjusted upwards. Based on the reverse analysis of the execution path of the current blocked node, key dependencies and constraints are extracted. Based on the key dependencies and constraints, corresponding auxiliary explanations are generated to show the key logical relationships of the decision path. (2) If the error rate or retry / rejection frequency exceeds the preset frequency threshold, the corresponding decision-making entity terminal is determined to be in an execution conflict state, and the conflict intervention mechanism is triggered: Preserve global abstraction level parameters Remain unchanged, but lower the local abstraction level parameter corresponding to the abnormal node. It performs fine-grained expansion on abnormal nodes, displaying their child nodes and dependencies, so that the decision-making terminal can locate the source of the conflict; at the same time, it increases the granularity of the interpretation presentation parameters. Extract the specific dependency constraints that caused the error or path rejection and encapsulate them into an explanation context, driving the text generation and processing module to output a micro-constraint tracing report; (3) Scheme restructuring and feedback recording mechanism: The decision path generated after intervention is integrated with the corresponding explanatory information to form an updated auxiliary decision-making scheme, which is then redistributed to the relevant decision-making terminals. The response time and number of compensations during this interaction are recorded, and a cost bias term is generated based on the type of obstruction / interruption signal. Based on this, the joint traversal cost, including a time-based weight term and a compensation weight term, is calculated. The joint traversal cost is written as scalar data along with the corresponding state flag into a distributed global log library. Specifically, Among them Joint traversal cost; t is the response time during this interaction process; The preset standard response time; This refers to the number of times compensation, retries, or reissues were performed after this intervention. This is the time consumption weighting coefficient; To compensate for the weighting coefficient; It is a type of blockage; This is the cost bias term corresponding to the type of blocking / interruption signal.

9. The method as described in claim 1, characterized in that, In S5, the statistical features include call frequency, execution success rate, average response time, average traversal cost, and number of rejections; the update of the cross-domain knowledge relationship network topology specifically involves: (1) Calculate the comprehensive performance index of each preset recommended path based on the statistical characteristics; if a path is blocked multiple times or manually rejected within the statistical window, and its average crossing cost is higher than the preset penalty threshold, increase the passage cost weight of the corresponding edge of the path; conversely, if the execution success rate and comprehensive performance index of a path are higher than the preset threshold, decrease the passage cost weight of the corresponding edge of the path. (2) Analyze the actual execution sequence in the distributed global log data to identify atypical flow paths; the atypical flow path refers to a cross-node operation sequence that does not explicitly define direct connection edges in the initial cross-domain knowledge relationship network, but repeatedly appears in the actual task execution process and has a stable success rate; based on the atypical flow path, instantiate implicit relationship connection edges between the starting knowledge node and the ending knowledge node in the unified global semantic space and assign them initial weights; (3) The local relaxation optimization algorithm is adopted to perform local mapping matrix parameter and topological connection relationship update operations on the affected knowledge nodes and their neighboring nodes, so as to complete the autonomous anti-aging incremental evolution of the cross-domain knowledge relationship network.

10. A decision-making system employing the method described in any one of claims 1 to 9, characterized in that, include: The atomic entity representation module is used to realize the semantic vectorization and dynamic metadata binding of decision entities in multi-source heterogeneous domains; The cross-domain isomorphic fusion module is used to construct local weighted directed graphs in various domains, and to construct a global cross-domain relationship network through affine transformation and manifold alignment; The distributed topology routing engine is used to map decision task requests to each shard domain and drive each shard domain to calculate the candidate response nodes in its own domain. It performs topology routing in the cross-domain knowledge relationship network and generates a cross-domain decision micro-orchestration chain. The adaptive perception intervention module is used to monitor the interaction status of the terminals of each decision-making entity, intervene in decision-making obstacles, and realize the reconstruction of decision-making schemes. The controlled incremental evolution module is used to update the topology of cross-domain knowledge relationship network based on the statistical characteristics of interaction data in a distributed global log library.

Images