Industrial Traceability Generative AI Situation Compilation and Task Graph Generation Method and System

By collecting and encapsulating evidence units, parsing natural language targets, constructing controlled generation contexts, and verifying generated candidate semantic patches, this approach solves the problems of unfounded inference and outdated rules in generative artificial intelligence in manufacturing environments. It achieves traceability and security of generation results and supports task graph generation for MES, APS, WMS, and personnel terminals.

CN123116022APending Publication Date: 2026-11-17XUANJI INTELLIGENT CHAIN (SHANGHAI) TECHNOLOGY SOFTWARE CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611456650.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-18
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In manufacturing environments characterized by diverse sources, inconsistent timeliness, and the inclusion of natural language objectives, generative artificial intelligence may write unfounded inferences or outdated rules into the factory state, leading to production execution risks and a lack of traceability in the generated results.

Method used

By collecting and encapsulating evidence units, parsing natural language targets, constructing a controlled generation context, and verifying the generated candidate semantic patches by an independent deterministic verifier, an executable task graph is formed, ensuring that the generated results are consistent with evidence, rule versions, and permissions, and supplementary collection or approval is performed when necessary.

Benefits of technology

The generated task graph is bound to the evidence unit to prevent unfounded output. An independent verifier blocks unauthorized operations and ensures that the generated results are traceable. It supports the generation of executable task graphs for MES, APS, WMS and personnel terminals.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123116022A_ABST
    Figure CN123116022A_ABST
Patent Text Reader

Abstract

This invention discloses a method and system for industrial traceability generative AI situational awareness compilation and task graph generation. The method encapsulates multi-source data and natural language targets from the manufacturing site into evidence units carrying source, time, version, and confidence level; retrieves matching industrial ontologies and process rules, constructs a controlled generative context, and outputs candidate semantic patches with evidence references by the generative AI. An independent deterministic verifier verifies structure, object references, time, rule conflicts, and evidence coverage; patches that pass verification are merged into a versioned factory situational semantic graph, while non-compliant items are converted into data supplementation or manual confirmation tasks. Further compilation generates an industrial task graph containing preconditions, postconditions, and rollback actions, and performs release verification. The generative model only outputs candidate content; permission approval is controlled by deterministic rules and manual intervention. This invention reduces the production risks caused by AI illusions, achieves verifiable and traceable task results, and supports controlled execution of the manufacturing system.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of intelligent manufacturing, industrial software, knowledge graphs, generative artificial intelligence, and controlled task orchestration, specifically to a method and system for compiling multi-source evidence and natural language targets from the manufacturing site into verifiable factory situation semantic graphs and industrial task graphs. Background Technology

[0002] In manufacturing enterprises, orders, process routes, equipment status, personnel qualifications, material batches, quality releases, and abnormal events are typically distributed across different systems. The same object may use different codes, status data has acquisition delays and expiration dates, and on-site management objectives are often expressed in natural language. Existing systems require planners to manually convert natural language objectives into objects, constraints, and system parameters, a process prone to overlooking applicable scope, time conditions, or access requirements.

[0003] Generative AI can recognize intent from text and generate task steps, but its probabilistic generation mechanism may produce non-existent equipment numbers, reference outdated process rules, confuse planned and actual values, or draw definitive conclusions when material batch and quality evidence is lacking. If the generated results are directly written into MES, APS, or equipment interfaces, semantic errors will be amplified into production execution risks. Simply requiring the model to comply with rules in the prompts does not prove that the facts, rule versions, and permissions referenced in the output are valid.

[0004] Therefore, an industrial situation compilation mechanism is needed to enable generative AI to generate candidate semantics only within a controlled vocabulary and evidence scope. Subsequently, independent deterministic components verify the object, time, rules, and evidence coverage, and then compile the verified results into an executable task graph. For content without evidence, conflict, or high risk, supplementary collection, consultation, or approval tasks should be generated instead of being directly issued. Summary of the Invention

[0005] The technical problem to be solved by this invention is: in a manufacturing environment with multiple heterogeneous sources, inconsistent timeliness, and natural language objectives, how to prevent generative artificial intelligence from writing unfounded inferences or outdated rules into the factory state, how to generate task graphs that can be executed by industrial software and personnel, and how to make each generated result traceable back to the data source, rule version, verification conclusion, and approval record.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a traceable generative artificial intelligence situation compilation and task graph generation method for industrial sites, comprising the following steps: Step S101: Collect orders, processes, equipment, personnel, materials, quality, abnormal events, technical documents, and natural language targets, and register the data source, collection time, and business object.

[0007] Step S102: Complete the encoding alignment according to the object master data mapping table, and encapsulate each fact into an immutable evidence unit; the evidence unit uses the evidence identifier as the smallest unit for subsequent semantic reference.

[0008] Step S103: Analyze the object candidates, target types, time expressions, and constraint expressions in the natural language target, and use them only as retrieval clues, without directly writing them into the factory situation diagram.

[0009] Step S104: Retrieve the industrial ontology, task templates, and valid rule versions by object, process, time window, rule scope, and user permissions, and construct a controlled generation context.

[0010] Step S105: The generative artificial intelligence model generates candidate semantic patches according to a predefined structural pattern; the candidate semantic patches include at least operation type, object, attribute or relationship, candidate value, evidence reference, uncertainty and generation batch.

[0011] In step S106, the deterministic validator sequentially performs structural schema, object reference, time consistency, rule conflict, and evidence coverage checks to generate a machine-processable verification report.

[0012] Step S107: Perform version merging on candidate semantic patches that pass verification; perform rejection, downgrade, or solicitation on candidate semantic patches that fail verification, and do not allow generative artificial intelligence to bypass the validator to modify the already effective graph.

[0013] Step S108: Based on the verified target, factory situation diagram and task template library, candidate task nodes are generated, and the task graph compiler completes the preconditions, postconditions, resource requirements, time windows, completion criteria, risk levels, permission levels and rollback actions.

[0014] Step S109: Detect circular dependencies, lack of producer input, resource mutual exclusion, missing quality gates, and out-of-bounds permissions in the task graph, and output a task graph that can be published or a task that can be repaired.

[0015] Step S110: Based on the risk level, the task is routed to one of three paths: automatic execution, execution after approval, or only a suggestion is provided; equipment control, safety interlocks, and quality release are maintained by a deterministic control system or authorized personnel.

[0016] Step S111: Receive execution receipts and manufacturing events with idempotency flags, form semantic differences, locate affected objects and task nodes, and perform incremental verification.

[0017] Step S112 records the generated context summary, candidate patches, verification report, graph version, task version, approval record, and execution receipt, forming a traceable audit chain.

[0018] Key data structures An evidence unit can be represented as e=<eid, object, attribute, value, source, observed_at, valid_time, version, confidence, hash>. Here, eid is the evidence identifier, object is the business object identifier, valid_time is the valid business timeframe, and hash is used to detect whether the evidence content has changed. A candidate semantic patch can be represented as p=<op, subject, predicate, object_or_value, evidence_refs, uncertainty, batch_id>.

[0019] The weighted evidence coverage ratio for candidate semantic patch p is defined as: Cov(p) = Σwk·support(k) / Σwk, where k is a required field of the candidate patch, support(k) is 1 if valid evidence exists in the field, and 0 otherwise; the weights of safety, quality, eligibility, capability, and material batch fields are higher than those of descriptive fields. The comprehensive gate value Gate(p) is the minimum value among structural verification, object verification, time verification, rule verification, and evidence coverage verification. Candidate patches cannot be merged if Gate(p) does not reach the threshold.

[0020] When the constraint set cannot be satisfied, the verifier outputs the minimum conflict constraint set (MUS) and associates the rule number and evidence identifier of each conflict constraint. For example, if the objective requires order A1001 to complete testing before 16:00, but equipment M02 is under maintenance from 14:00 to 16:00 and the alternative equipment does not have testing capabilities, the system will not directly compile the objective into an execution task. Instead, it will generate a task for equipment replacement confirmation, delivery risk confirmation, or objective time adjustment.

[0021] Compared with the prior art, the beneficial effects of the present invention are: 1. The generated content is bound to the evidence unit, rule version, and scope of application, and can distinguish between on-site facts, model candidates, and manual confirmation results.

[0022] 2. Generative AI does not directly control devices or modify their effective status; independent verifiers and access control systems block outputs without basis or beyond their authority.

[0023] 3. Natural language objectives can be compiled into task graphs with deterministic completion criteria and compensation actions, facilitating execution by MES, APS, WMS, QMS, and personnel terminals.

[0024] 4. Validation failures are transformed into tasks such as supplementary data collection, consultation, or approval, making the missing information part of the closed loop, rather than being filled in by model guesses.

[0025] 5. After an event occurs, only the objects and tasks affected by semantic difference are re-verified, which can reduce the overhead of full recalculation while maintaining the audit chain. Attached Figure Description

[0026] Figure 1 This is a flowchart of the traceable generative situation compilation method of the present invention; Figure 2 This is a block diagram of the system composition of the present invention; Figure 3 This is a schematic diagram illustrating the generation and verification of candidate semantic patches for evidence constraints in this invention. Figure 4 This is a schematic diagram of the task graph compilation and execution boundary of the present invention; Figure 5 This is a schematic diagram of the event-driven incremental update of the present invention; Figure 6 This is an example of a factory situation semantic diagram of the present invention; Figure 7 This is an example table of evidence units in Embodiment 1 of the present invention; Figure 8 This refers to the candidate semantic patch and verification action table in Embodiment 1 of the present invention; Figure 9 This is the industrial task graph node table compiled in Embodiment 1 of the present invention. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] Example 1: Urgent Order Support for Electronic Assembly and Testing Workshop This embodiment illustrates the complete process from natural language target to verified situation map and task map. The workshop is connected to ERP, MES, WMS, QMS, equipment ledger, and planner terminals. User input: Tomorrow, prioritize the shipment of A1001; M02 will be under maintenance from 2 PM to 4 PM; rework and retesting can only be performed by senior technicians. The system does not directly interpret this text as a scheduling command, but first establishes the following evidence units, such as... Figure 7 As shown.

[0029] The target parser identifies A1001, M02, rework and retesting, and senior technicians as candidate objects. The knowledge retrieval unit retrieves process route R-A1001, equipment capability rule CAP-M02, rework qualification rule SK-07, and quality access rule QG-03 by object and effective time. The generative artificial intelligence model outputs candidate semantic patches under structural pattern constraints, as shown in the example. Figure 8 As shown.

[0030] P204 was rejected due to lack of quality release evidence. The system generated a QMS review task T-Q01 and placed the assembly nodes that depend on quality release into a blocked state. P201 was merged after confirming that the current user had the authority to adjust order priorities; if the user did not have this authority, it was converted into a planner approval task, such as... Figure 9 As shown.

[0031] Example 2: Conflicting Evidence and Secondary Inquiry When the equipment log indicates that M02 was restored at 16:00, while the equipment controller event indicates that M02 is still under maintenance lock, the two evidence units conflict within the same valid timeframe. If the system cannot eliminate the conflict based on source level and freshness, the candidate semantic patch is marked as pending confirmation, and tasks using M02 are not allowed to be published. The system automatically generates a query, listing the source of the conflict, the collection time, and the equipment status that needs confirmation; after the equipment position is confirmed, a new evidence unit is formed. The new evidence does not overwrite the old evidence; instead, the old evidence is marked as invalid based on version relationships.

[0032] If the generative AI model outputs device M04 that is not in the object master data, the reference verification will fail immediately. The verifier returns an UNKNOWN_OBJECT error and a list of similar objects, but will not automatically replace M04 with M03. Verification can only be re-entered after the user confirms the object mapping or the master data system adds M04.

[0033] Example 3: Incremental Feedback and Traceable Calibration After the MES receipt shows that T-X01 has started, the system records the execution version V17 and the idempotent key. Subsequently, a delay event occurs on device M02. Semantic difference only involves the availability time of M02, unstarted tasks dependent on M02, and their successors. Started nodes remain frozen. The system re-retrieves the current rule version, generates new candidate patches, and performs the same five-class validation. Manual rejection reasons are used to adjust the next candidate ranking, but the content of hard constraint rules and thresholds are not modified by the generative model.

[0034] This embodiment can use the following test scheme to verify the system: Object reference accuracy is calculated by dividing the number of verified object references by the total number of object references; Evidence coverage is calculated by the weight of required fields; No-evidence candidate blocking rate is calculated by dividing the number of no-evidence candidates blocked by the validator by the total number of no-evidence candidates; Task graph executability rate is calculated by dividing the number of task graphs that pass the deterministic gate by the total number of candidate task graphs; Audit traceability rate is calculated by dividing the number of tasks that can recover and generate context summaries, rule versions, verification reports, and approval records by the number of published tasks. The test results should be compared with the control configurations with evidence gates disabled, time verification disabled, or permission verification disabled.

[0035] Deployment method Candidate generation units can be deployed in enterprise private clouds, edge servers, or approved external model services; evidence encapsulation, deterministic verification, permission approval, and controlled execution units are deployed within the enterprise control domain. When sending context to external models, order numbers, personnel identifiers, and process parameters can be tokenized or anonymized. Model services cannot directly hold write permissions for MES, WMS, QMS, or equipment control. Systems interact through interface messages with tenant, user, task version, and idempotent keys.

[0036] Generative artificial intelligence models can be large language models, multimodal generative models, or specially trained structured generative models. Deterministic validators can consist of one or more of the following: rule engines, graph query engines, constraint solvers, type checkers, and digital signature validators. The above substitutions do not change the technical concept of this invention: evidence-constrained candidate generation, independent validator-controlled merging, and task graph compilation for industrial execution.

[0037] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A traceable generative artificial intelligence situation compilation and task graph generation method for industrial sites, characterized in that, include: Structured data, event streams, technical documents, and natural language targets from the manufacturing site are collected. Each data item is encapsulated into an evidence unit, which includes at least a business object identifier, attribute value, source identifier, collection time, validity period, version identifier, and confidence level. Based on the candidate business objects and target time windows in the natural language targets, valid knowledge fragments are retrieved from the industrial ontology library, process rule library, resource capability library, quality rule library, and permission rule library. The knowledge fragments and the evidence units are then combined to form a controlled generation context. Generative artificial intelligence models are used to output candidate semantic patches according to predefined situational patterns. Each object, attribute, relationship, or constraint in the candidate semantic patch contains evidence references and uncertainty markers. Using a deterministic validator independent of the generative artificial intelligence model, structural pattern verification, object reference verification, time consistency verification, rule conflict verification, and evidence coverage verification are performed on the candidate semantic patches. Candidate semantic patches that pass verification are merged into a versioned factory situation semantic graph, while candidate semantic patches that fail verification or have a confidence level below a threshold are transformed into supplementary collection tasks, secondary solicitation tasks, or manual confirmation tasks. Based on the merged factory situation semantic graph and task template library, an industrial task graph is generated, containing preconditions, postconditions, execution objects, resource requirements, time constraints, completion criteria, risk levels, permission levels, and rollback actions. Before the industrial task graph is released, release verification is performed by invoking process safety, quality release, resource capability, and permission approval rules.

2. The method for traceable generative artificial intelligence situation compilation and task graph generation for industrial sites according to claim 1, characterized in that: The structured data and event streams come from at least two types of data sources, including ERP, MES, APS, WMS, QMS, PLM, equipment controllers, sensors, maintenance systems, or personnel terminals, and are unified through an object master data mapping table to identify business objects such as orders, processes, equipment, personnel, materials, tooling, quality characteristics, and abnormal events. The evidence unit also includes data owner, verification status, confidentiality level, and hash digest; when there are multiple conflicting evidence units for the same attribute of the same business object, the evidence priority is calculated according to source level, time freshness, integrity, and historical reliability, and unused evidence and reasons for conflict are retained.

3. The method for traceable generative artificial intelligence situation compilation and task graph generation for industrial sites according to claim 1, characterized in that: The controlled generation context contains only knowledge fragments that match the target object, target process, target time window, and user permissions. Each knowledge fragment carries a rule number, rule version, effective time, and applicable scope. Knowledge fragments that are outside the applicable scope do not enter the generation context. The predefined situational pattern limits the types of objects, attributes, relationships, constraint operators, and value ranges that can be generated; the generative artificial intelligence model outputs machine-parseable candidate semantic patches, and cannot directly modify the already effective factory situational semantic graph.

4. The method for traceable generative artificial intelligence situation compilation and task graph generation for industrial sites according to claim 1, characterized in that: The evidence coverage verification calculates the weighted evidence coverage rate of candidate semantic patches; and sets indispensable markers for fields such as security constraints, quality release, personnel qualifications, equipment capabilities, and material batches. If any indispensable field lacks valid evidence, the corresponding task is prohibited from entering the automatic release state. The rule conflict verification process compiles candidate constraints into Boolean constraints, interval constraints, priority constraints, resource capacity constraints, or permission constraints, and calls the rule engine, constraint solver, or a combination of both to determine whether the constraint set can be satisfied; if it cannot be satisfied, it outputs the minimum set of conflicting constraints and its evidence references.

5. The method for traceable generative artificial intelligence situation compilation and task graph generation for industrial sites according to claim 1, characterized in that: The industrial task graph is a directed attribute graph. Task nodes include node identifier, task type, input, output, preconditions, postconditions, idempotent keys, execution interface, timeout conditions, and compensation actions. Task edges include sequential dependencies, data dependencies, resource mutual exclusion, quality access control, and manual approval relationships. The task templates in the task template library have applicable object types, triggering conditions, parameter slots, deterministic completion criteria, and allowed call interfaces; Generative AI models are used to select candidate templates and populate candidate parameters, while task graph compilers are used to verify parameter types, object permissions, interface capabilities, and dependency loops.

6. The method for traceable generative artificial intelligence situation compilation and task graph generation for industrial sites according to claim 1, characterized in that: The release path is determined based on the task's risk level: low-risk read-only query tasks that have passed all verifications are allowed to be executed automatically; tasks involving changes in production sequence, resource usage, or business status write-back are placed into the post-approval execution path. Tasks involving safety interlocks, quality release, equipment control, or irreversible operations must be confirmed by personnel with the corresponding authority; After receiving the task execution receipt and the manufacturing event, deduplication and timing alignment are performed based on the event identifier to generate a semantic difference relative to the current situation map version. Only the objects, constraints and task nodes affected by the semantic difference are re-verified, and the map versions before and after the update and the verification report are retained.

7. The method for traceable generative artificial intelligence situation compilation and task graph generation for industrial sites according to claim 1, characterized in that: The reliability of evidence sources or candidate ranking parameters are updated based on the results of manual confirmation, reasons for rejection, and task execution results; the update does not change the hard constraints of process safety, quality release, equipment capability, and authority approval.

8. A traceable generative artificial intelligence situation compilation and task graph generation system for industrial sites, characterized in that, It includes an evidence encapsulation unit, a knowledge retrieval unit, a candidate generation unit, a deterministic verification unit, a situation map management unit, a task graph compilation unit, an authorization approval unit, and a feedback calibration unit, each of which is configured to perform the method described in any one of claims 1-12.

9. An electronic device comprising a processor and a memory, the memory storing a program, characterized in that, When the program is executed by the processor, it causes the electronic device to perform the method described in any one of claims 1-7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method according to any one of claims 1-7.