An AI intelligent human resource management system for manufacturing industry

Abstract

The application discloses a manufacturing industry AI intelligent human resource management system and relates to the technical field of human resource management.The system comprises a business logic multi-source collection and semantic analysis module, a rule formalization modeling and knowledge graph construction module, a dynamic rule engine and conflict resolution module and a manufacturing scene intelligent scheduling execution module.The semantic analysis module extracts a four-tuple feature vector of a business rule from multi-source heterogeneous data by using a large language model;the rule compiling module classifies and compiles the vector and injects the vector into a knowledge graph, thereby realizing version management;the dynamic engine listens to changes, scans conflicts, makes multi-target decisions and performs gray release.The optimal scheduling scheme vector of a flexible target is outputted, an execution instruction stream is generated and is distributed to subordinate terminals, the technical problem of how to convert dynamically changing business logic into structured rules executable by a software system in the prior art is solved, and the effect of full-process automation from automatic rule extraction to intelligent scheduling is realized.

Description

Technical Field

[0001] This invention relates to the field of human resource management technology, and in particular to an AI-powered intelligent human resource management system for the manufacturing industry. Background Technology

[0002] Global manufacturing is rapidly transitioning towards smart factories and lighthouse factories. Smart manufacturing is also evolving from manual labor to human-machine collaboration. Traditional human resource management in manufacturing often only handles administrative functions such as recruitment, payroll, and employee departures, becoming disconnected from the actual needs of the production floor. Meanwhile, manufacturing companies have increasingly higher demands for human efficiency, requiring precise monitoring of employee skills, time utilization, and safety production status. Against this backdrop, AI-powered intelligent human resource management systems for manufacturing, incorporating artificial intelligence technology, have emerged.

[0003] Existing AI-powered human resource management systems in the manufacturing industry operate within a closed-loop logic, from data collection to decision-making and execution. This involves first collecting real-time production, working hours, and attendance data via IoT hardware such as attendance machines and access control systems. Then, an intelligent scheduling engine combines order forecasts, employee skill maps, and labor laws to generate work schedules. Subsequently, payroll calculations are automatically completed based on actual output and attendance records. Finally, the results are pushed to employees' mobile devices for confirmation, and any anomalies are fed back to the system, forming a closed management loop. However, in terms of scheduling, the manufacturing industry involves dozens of mutually constraining rules, including multi-shift rotations, continuous work across days, time off in lieu of leave, skill-based job matching, and legal limits on overtime. Temporary adjustments made by team leaders based on experience, as well as various unforeseen circumstances, are difficult to accurately translate into structured code. Forcing fixed schedules leads to rigid scheduling and employee resistance, while relying entirely on manual intervention results in a significant increase in labor costs. For payroll calculation, manufacturing companies employ multiple pay models, including piece-rate, hourly, position allowance, quality-based rewards and penalties, and overtime pay. Furthermore, the formula combinations vary significantly across workshops and job types. Manual calculation not only requires several days to compile production details from the Manufacturing Execution System (MES), time records from the attendance system, and performance evaluation data, but more importantly, any formula configuration error or decimal point precision deviation can lead to wage calculation errors for hundreds or even thousands of workers, potentially triggering labor disputes or even legal proceedings. This presents a technical challenge: how to transform the dynamically changing business logic of manufacturing human resource management into structured rules executable by a software system.

[0004] In view of this, the present invention proposes an AI-powered intelligent human resource management system for the manufacturing industry to solve the above problems. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art and to achieve the above objectives, the present invention provides the following technical solution: an AI-powered intelligent human resource management system for manufacturing, comprising:

[0006] The business logic multi-source acquisition and semantic parsing module is used to acquire multi-source heterogeneous input stream data, preprocess the multi-source heterogeneous input stream data, construct a semantic parsing input matrix based on the multi-source heterogeneous input stream data, input the semantic parsing input matrix into the pre-trained manufacturing human resources big language model, and extract the four-tuple feature vector of each business logic.

[0007] The rule formal modeling and knowledge graph construction module is used to receive the semantic quadruple feature vectors output by the business logic multi-source collection and semantic parsing module, perform routing and distribution according to the rule type classifier, formally compile the distributed rules, inject the rule nodes into the manufacturing human resources knowledge graph, build a reasonable relationship network, assign a version number and effective fingerprint to each rule, and write it into the versioned rule library.

[0008] The dynamic rule engine and conflict resolution module are used to listen for rule base version change events, trigger the incremental compilation pipeline, use knowledge graph reasoning to scan whether there are combinations of overlapping conditions but contradictory conclusions between new rules and existing rules, mark conflict pairs, solve the priority order of conflict pairs based on the preset multi-objective adjudication strategy function, automatically retain high-priority rules, put suppressed rules into a dormant state, generate conflict adjudication logs, and distribute the new rule set to each execution sub-engine through atomic hot switching instructions. A canary release strategy is adopted, and after monitoring for no abnormal indicators, the full release is pushed out, and a set of valid rule snapshots is output.

[0009] The intelligent scheduling and execution module for manufacturing scenarios is used to acquire the real-time operation state flow matrix, input the state matrix and rule constraints into the multi-objective scheduling optimization network, output the optimal scheduling scheme vector of flexible objectives, pre-simulate the output of the scheduling scheme in the virtual simulation environment, generate the execution instruction flow and distribute it to subordinate terminals.

[0010] Furthermore, acquire multi-source heterogeneous input stream data and preprocess the multi-source heterogeneous input stream data, including:

[0011] Multi-source heterogeneous input stream data includes: work order information data, equipment operating status and equipment operating hours obtained from the production execution system; unique identification employee number and attendance information data obtained from the attendance system; and production plan data, material requirements data and order scheduling data obtained from the enterprise resource planning system.

[0012] The specific steps for preprocessing multi-source heterogeneous input stream data are as follows: collect multi-source heterogeneous input stream data from the production execution system, attendance system and enterprise resource planning system at regular intervals; clean the collected multi-source heterogeneous input stream data, including deduplication, missing value imputation and outlier deletion; standardize the cleaned multi-source heterogeneous input stream data and associate them with each other using a unique identifier employee number.

[0013] Furthermore, a semantic parsing input matrix is ​​constructed based on multi-source heterogeneous input stream data, including:

[0014] Entity recognition is performed on multi-source heterogeneous input stream data to identify employee entities, job entities, time entities, location entities, and business rule entities. The heterogeneous original representations of each entity are mapped to the system's standard naming to generate a standardized entity set.

[0015] Based on standardized entity sets, and using a preset time granularity as the slicing unit, time series alignment is performed on data from different sources. The historical data backtracking window and the future plan data forward window corresponding to each slice are determined. All event data within the window are aggregated according to the time axis to generate a time-aligned structured dataset.

[0016] Semantic fusion processing is performed on the time-aligned structured dataset to convert numerical data into natural language descriptions with context, rule-based data and applicable conditions are concatenated to generate rule description text, and employee attributes, job requirements and production plan data are combined to generate scenario description text, thus constructing a semantic unit pool composed of several semantic units.

[0017] Using each decision unit as the row index and employee characteristics, job requirements, rule constraints, and scenario context as column dimensions, semantic units corresponding to each dimension are retrieved from the semantic unit pool and populated. Multiple structured prompt texts are then concatenated and generated. These multiple structured prompt texts are arranged in rows to construct a semantic parsing input matrix. This matrix is ​​used to input the pre-trained manufacturing human resources big language model for rule feature extraction.

[0018] Furthermore, the semantic parsing input matrix is ​​input into the pre-trained large language model for manufacturing human resources to extract the four-tuple feature vector for each business logic, including:

[0019] The semantic parsing input matrix is ​​divided into several independent prompt units by row. The prompt units are lexicalized and a preset task instruction prefix and output format constraint suffix are added. The output format constraint suffix specifies that the large language model is output in a four-tuple structure of conditional entity, relation type, target entity and effective time.

[0020] All prompting units are input into a pre-trained big language model for manufacturing human resources in parallel. The big language model encodes the semantic information in the prompting units based on the attention mechanism. Through context-aware reasoning, it extracts the four-tuple information of human resources domain semantics from the prompting units and summarizes the four-tuple information into a four-tuple information set.

[0021] The conditional entity, relation type, target entity, and effective time in the quadruple information set are vectorized and encoded respectively. The four component vectors of the same quadruple are concatenated into a quadruple feature vector with fixed dimensions.

[0022] Using the row index of the semantic parsing input matrix as the association key, a mapping relationship is established between each extracted quadruple feature vector and the corresponding business logic entry, and a set of quadruple feature vectors is output.

[0023] Furthermore, the semantic quadruple feature vectors output by the business logic multi-source acquisition and semantic parsing module are received, routed according to the rule type classifier, and the routed rules are formally compiled, including:

[0024] The source system identifier and the original text position index are added to each quadruple feature vector to form a rule vector to be processed with a source label. The rule vector to be processed is then input into a pre-trained rule type classifier.

[0025] The classifier performs type discrimination on each rule based on the semantic embedding of the relation type component and condition entity component of the quadruple feature vector, and outputs the rule type label.

[0026] Rule type tags must include at least one of the following: working hours rules, payroll calculation rules, scheduling constraints, skills matching rules, and compliance review rules.

[0027] Based on the rule type label, each quadruple feature vector is routed to the corresponding rule compilation channel. Each rule compilation channel is pre-set with a compilation template and syntax parser for different rule types. The compilation template defines the mapping relationship between the rule and the structured language.

[0028] In each rule compilation channel, the component values ​​of the quadruple feature vector are extracted. Based on the corresponding compilation template, the condition entity is mapped to the condition predicate, the relation type is mapped to the logical operator, the target entity is mapped to the target predicate, and the effective time is mapped to the time window limit condition. The components are combined according to the preset syntax rules to generate a structured rule description text that conforms to the formal language specification.

[0029] The structured rule description text output by each compilation channel is subjected to syntax verification and semantic integrity check. After the verification is passed, the formally compiled set of structured rules is output, and a mapping index table between each structured rule and the corresponding original quadruple feature vector is generated.

[0030] Furthermore, rule nodes are injected into the manufacturing human resources knowledge graph to construct a reasonable relationship network. Each rule is assigned a version number and an effective fingerprint, and written into a versioned rule base, including:

[0031] Each structured rule in the mapping index table is parsed to extract the conditional entity, target entity, relation type, and effective time window information. The extracted conditional entity and target entity are used as entity nodes of the knowledge graph, the relation type is used as the directed relation edge between entity nodes, and the effective time window is used as the attribute label of the time dimension of the relation edge. According to the preset ontology model, each structured rule is converted into a knowledge graph triple.

[0032] Knowledge graph triples are injected into the manufacturing human resources knowledge graph. Entity nodes are deduplicated and merged. When the entity to be injected and an existing entity in the graph point to the same business object, the entity to be injected is merged into the existing entity node and a synonym mapping chain is established. When the relationship edge to be injected overlaps with the existing relationship edge, the relationship edge is fused. The time attributes of multiple rules are stored on the same relationship edge in the form of a time interval set, forming a reasonable relationship network. The reasonable relationship network supports relationship chain query based on graph traversal and reasoning calculation based on rule logic.

[0033] For each rule injected into the knowledge graph, the original quadruple feature vector corresponding to the rule and the generated structured rule description text are obtained and concatenated to form a rule content summary. The rule content summary is calculated using a hash algorithm to calculate the rule content hash value. The rule content hash value is combined with the current timestamp and the operator identifier to generate the effective fingerprint of the rule. The effective fingerprint is used to uniquely identify the content status of the rule at a specific time.

[0034] Each rule injected into the knowledge graph is assigned a version number. The version number adopts the semantic version number format. When the rule content changes, the corresponding major version number, minor version number or revision version number is incremented. The current version number is associated with the effective fingerprint, stored and written into the versioned rule base.

[0035] Furthermore, it monitors rule base version change events and triggers the incremental compilation pipeline, including:

[0036] Deploy a rule base version change event listener. The listener continuously monitors rule addition events, rule modification events, and rule deletion events in the versioned rule base. It captures the rule identifier, change type, version number before and after the change, and change timestamp corresponding to the event, generates a version change event message, and pushes it to the message queue.

[0037] The change rule identifier and change type are obtained from the version change event message parsing. The changed rule content is read from the versioned rule base according to the change rule identifier. The incremental compilation scope is determined according to the change type. The incremental compilation scope includes at least the change rule itself, the associated rules with the target entity of the change rule as the head entity, and the same group of rules with the same rule type as the change rule.

[0038] Based on the incremental compilation scope, each rule node and its relationship within the incremental compilation scope is extracted from the knowledge graph. The extracted rule nodes are arranged in a preset compilation order, and the rule compilation channel corresponding to the compilation order is called to perform incremental compilation on the changed rules and related rules, generating an executable rule instruction set after incremental compilation, and recording the dependency resolution log during the compilation process.

[0039] The compilation results of the incrementally compiled executable rule instruction set are verified. The effective fingerprint of the compiled instruction is compared with the effective fingerprint of the corresponding rule in the versioned rule base. If the comparison is consistent, the compilation is marked as successful. The successful compilation status and executable instruction set are output to the waiting queue of the dynamic rule engine. At the same time, an incremental compilation report is generated to record the number of rules involved in this compilation and the compilation time.

[0040] Furthermore, using knowledge graph reasoning to scan for combinations of overlapping conditions but contradictory conclusions between new rules and existing rules includes:

[0041] Take the newly injected rule nodes and their relationships in the knowledge graph. Starting from the condition entity of the new rule node, perform a graph traversal query in the knowledge graph. The graph traversal query includes at least a one-hop relationship query and a multi-hop relationship query. Retrieve all existing rule nodes in the existing rule library that have the same entity as the new rule node and are reachable through relationship edges, and construct a candidate conflict detection set.

[0042] Extract the condition entity set of each existing rule node and the condition entity set of the new rule node from the candidate conflict detection set and calculate the overlap. When the overlap is greater than or equal to a preset threshold, mark the existing rule node as a condition overlap candidate node.

[0043] The rule conclusions corresponding to the overlapping candidate nodes and the rule conclusions of the new rule nodes are checked for contradiction. The contradiction detection includes judging the conflict of conclusion entities, judging the opposite or mutually exclusive relation types, and judging the non-intersection of quantitative constraint intervals. If any detection result is true, the overlapping candidate node and the new rule node are determined to be a conflict pair.

[0044] Record the rule identifiers, overlapping condition descriptions, and contradictory conclusion descriptions involved in the conflict pairs, generate a conflict detection log, and output a set of conflict pairs.

[0045] Furthermore, based on a preset multi-objective adjudication strategy function, the priority order of conflict pairs is solved, high-priority rules are automatically retained, suppressed rules are put into a dormant state, and a conflict adjudication log is generated, including:

[0046] Extract the rule attribute vectors of the two rules involved in the conflict. The rule attribute vectors shall include at least the rule source system weight, rule effective timestamp, rule applicable scope level, rule type priority coefficient, and rule historical usage frequency.

[0047] The rule attribute vector is input into a preset multi-objective adjudication strategy function. The multi-objective adjudication strategy function calculates a comprehensive priority score for each rule by combining multiple adjudication strategy dimensions. The adjudication strategy dimensions include at least the scores of the time priority strategy dimension, the source authority priority strategy dimension, the specialization priority strategy dimension, and the stability priority strategy dimension. The comprehensive priority score is obtained by weighted summation of the scores of each strategy dimension.

[0048] Within the same conflict pair, compare the overall priority scores of the two rules, retain the rule with the higher overall priority score as the active rule, mark the rule with the lower overall priority score as the suppressed rule, and attach a suppression mark and a description of the suppression reason to the suppressed rule. The description of the suppression reason should include at least the identifier of the conflicting rule and a description of the conflict dimension.

[0049] Suppressed rules are transferred to a dormant state. The dormant state is a state in which the rule does not participate in subsequent scheduling decisions but is still retained in the versioned rule base and can be queried and restored. The dormant state record is recorded for the rule that is transferred to the dormant state, including the dormant effective timestamp and the conflict pair identifier of the adjudication source.

[0050] Summarize the adjudication results of all conflict pairs in this batch and generate a conflict adjudication log. The conflict adjudication log should include at least the adjudication time of each conflict pair, the rule identifiers involved in the adjudication, the scores of each strategy dimension on which the adjudication is based, the adjudication result, and the dormant status record of the suppressed rule. Write the conflict adjudication log to the log storage area of ​​the versioned rule base.

[0051] Furthermore, the real-time operational state flow matrix is ​​obtained, and the state matrix, along with the rule constraints, is input into the multi-objective scheduling optimization network to output the optimal scheduling scheme vector with flexible objectives, including:

[0052] Real-time operational status data is obtained from the production execution system, attendance system and equipment IoT platform. The operational status data includes at least the quantity of work-in-process and process cycle time of each production line, the operating status and utilization rate of each piece of equipment, and the real-time on-duty status and remaining available working hours of each position. The operational status data is matrix-coded according to a unified time granularity to construct a real-time operational status flow matrix.

[0053] Read a set of valid rule snapshots from the versioned rule base. The set of valid rule snapshots includes at least work hour rules, skill matching rules, compliance constraint rules, and cost control rules.

[0054] Each rule in the effective rule snapshot set is parsed into a constraint tensor. The constraint tensor is then concatenated and fused with the real-time operation state flow matrix along the feature dimension to construct the scheduling optimization input tensor.

[0055] The input tensor of the scheduling optimization is input into the pre-trained multi-objective scheduling optimization network. The multi-objective scheduling optimization network is based on an encoder and decoder architecture. The encoder performs attention encoding on the input tensor, and the decoder generates multiple sets of candidate scheduling schemes in parallel based on the flexible objective weight vector. The flexible objectives include at least minimizing labor costs, maximizing skill matching degree, and maximizing order delivery timeliness.

[0056] Multiple candidate scheduling schemes are sorted by multi-objective non-dominated ordering to select Pareto optimal solution set. From the Pareto optimal solution set, the scheme with the highest comprehensive score is selected as the optimal scheduling scheme vector according to the preset flexible objective preference weight, and the optimal scheduling scheme vector is output.

[0057] The technical effects and advantages of the AI-powered intelligent human resource management system for manufacturing industry of this invention are as follows:

[0058] By leveraging knowledge graph reasoning, the system proactively scans for overlapping conditions and contradictory conclusions between newly added and existing rules. A multi-objective adjudication strategy function automatically prioritizes rules, transferring suppressed rules to a recoverable dormant state. This effectively resolves the logical conflicts arising from the superposition of multiple rules, avoids scheduling errors caused by rule contradictions, and significantly improves the system's stability and decision reliability.

[0059] Incremental compilation is triggered by listening for version change events, compiling only the changed rules and their associated rules, thus avoiding the resource waste of full compilation. Simultaneously, atomic hot-swapping instructions and a canary release strategy are employed to monitor abnormal indicators as the new rule set gradually takes effect, and a full rollout is only pushed out after confirming no anomalies. This ensures uninterrupted execution of rule changes in the production environment, minimizing the business risks associated with rule changes.

[0060] By integrating a real-time operational state flow matrix with multi-dimensional rule constraints into a multi-objective scheduling optimization network, an optimal solution is generated using Pareto non-dominated sorting to balance labor costs, skill matching, and on-time order delivery. The feasibility of the solution is verified through digital twin simulation before implementation. This approach upgrades scheduling from static planning based on historical experience to dynamic optimization based on real-time operating conditions, significantly improving the scientific rigor, adaptability, and executability of scheduling solutions. Attached Figure Description

[0061] Figure 1 A schematic diagram of the structure of an AI-powered intelligent human resource management system for manufacturing industry provided in this application embodiment;

[0062] Figure 2This is a flowchart illustrating the process of receiving semantic quadruple feature vectors output by the multi-source acquisition and semantic parsing module of the business logic provided in this application embodiment, routing and splitting the traffic according to the rule type classifier, and formally compiling the split rules.

[0063] Figure 3 This is a schematic diagram illustrating the process of injecting rule nodes into a manufacturing human resources knowledge graph, constructing a reasonable relationship network, assigning a version number and an effective fingerprint to each rule, and writing it into a versioned rule base, as provided in the embodiments of this application. Detailed Implementation

[0064] 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.

[0065] This application provides an AI-powered intelligent human resource management system for the manufacturing industry. The system comprises a multi-source heterogeneous data acquisition and semantic parsing module, which preprocesses the data to construct a semantic parsing input matrix. This matrix is ​​then input into a pre-trained manufacturing human resource big language model to extract a four-tuple feature vector for each business logic, including conditional entity, relation type, target entity, and effective time. Next, a rule formal modeling and knowledge graph construction module receives the four-tuple feature vectors, routes them according to rule type classifiers, performs formal compilation, injects the rule nodes into the manufacturing human resource knowledge graph to construct a reasonable relational network, and assigns a version number and effective fingerprint to each rule, writing them into a versioned rule base. Finally, a dynamic rule engine and conflict resolution module monitors rule base version change events to trigger incremental updates. The compilation process utilizes knowledge graph reasoning to scan conflicting pairs with overlapping conditions but contradictory conclusions. Priority order is determined based on a multi-objective adjudication strategy function, retaining high-priority rules and transferring suppressed rules to a dormant state. The new rule set is securely distributed to each execution sub-engine via atomized hot-switching instructions and a canary release strategy, outputting a set of valid rule snapshots. Finally, the intelligent scheduling execution module acquires the real-time operational state flow matrix, which, along with rule constraints, is input into the multi-objective scheduling optimization network to generate the optimal scheduling scheme vector with flexible objectives. After pre-verification in a virtual simulation environment, an execution instruction stream is generated and distributed to subordinate terminals, completing a full closed loop from rule extraction, compilation and injection, conflict resolution to intelligent scheduling.

[0066] Please see Figure 1 The image below is a schematic diagram of the structure of an AI-powered intelligent human resource management system for manufacturing, provided in an embodiment of this application. The system includes:

[0067] The business logic multi-source acquisition and semantic parsing module is used to acquire multi-source heterogeneous input stream data, preprocess the multi-source heterogeneous input stream data, construct a semantic parsing input matrix based on the multi-source heterogeneous input stream data, input the semantic parsing input matrix into the pre-trained manufacturing human resources big language model, and extract the four-tuple feature vector of each business logic.

[0068] The rule formal modeling and knowledge graph construction module is used to receive the semantic quadruple feature vectors output by the business logic multi-source collection and semantic parsing module, perform routing and distribution according to the rule type classifier, formally compile the distributed rules, inject the rule nodes into the manufacturing human resources knowledge graph, build a reasonable relationship network, assign a version number and effective fingerprint to each rule, and write it into the versioned rule library.

[0069] The dynamic rule engine and conflict resolution module are used to listen for rule base version change events, trigger the incremental compilation pipeline, use knowledge graph reasoning to scan whether there are combinations of overlapping conditions but contradictory conclusions between new rules and existing rules, mark conflict pairs, solve the priority order of conflict pairs based on the preset multi-objective adjudication strategy function, automatically retain high-priority rules, put suppressed rules into a dormant state, generate conflict adjudication logs, and distribute the new rule set to each execution sub-engine through atomic hot switching instructions. A canary release strategy is adopted, and after monitoring for no abnormal indicators, the full release is pushed out, and a set of valid rule snapshots is output.

[0070] The intelligent scheduling and execution module for manufacturing scenarios is used to acquire the real-time operation state flow matrix, input the state matrix and rule constraints into the multi-objective scheduling optimization network, output the optimal scheduling scheme vector of flexible objectives, pre-simulate the output of the scheduling scheme in the virtual simulation environment, generate the execution instruction flow and distribute it to subordinate terminals.

[0071] Furthermore, acquire multi-source heterogeneous input stream data and preprocess the multi-source heterogeneous input stream data, including:

[0072] Multi-source heterogeneous input stream data includes: work order information data, equipment operating status and equipment operating hours obtained from the production execution system; unique identification employee number and attendance information data obtained from the attendance system; and production plan data, material requirements data and order scheduling data obtained from the enterprise resource planning system.

[0073] The specific steps for preprocessing multi-source heterogeneous input stream data are as follows: collect multi-source heterogeneous input stream data from the production execution system, attendance system and enterprise resource planning system at regular intervals; clean the collected multi-source heterogeneous input stream data, including deduplication, missing value imputation and outlier deletion; standardize the cleaned multi-source heterogeneous input stream data and associate them with each other using a unique identifier employee number.

[0074] In this embodiment, the system collects work order information, equipment operating status, and equipment operating hours from the Production Execution System (MES) at a set time every morning, collects the employee's unique identification number and clock-in records from the attendance system, and collects the production plan, material availability rate, and order priority data for the next three days from the Enterprise Resource Planning System (ERP). After collection, the system automatically deduplicates duplicate clock-in entries for the same employee within five minutes in the attendance records and retains the first record. For missing process cycle time fields in the work order data, the system interpolates using the historical median value of the same model of product on the same production line over the past week. Abnormal records in the equipment operating hours that exceed 1.5 times the theoretical maximum operating time of the equipment (such as a value of 9999 hours due to sensor failure) are deleted.

[0075] The system standardizes order numbers in the Enterprise Resource Planning (ERP) system to the format "PO-YYYYMMDD-XXX", maps equipment operating status to four standard enumeration values ​​"RUN / IDLE / STOP / FAULT", converts attendance clock-in time to the format "YYYY-MM-DD HH:MM:SS", and standardizes the mapping of classification fields such as gender, department, and shift in various systems using a unified coding table.

[0076] The system uses the employee's unique ID number as the primary key to horizontally link the same employee's daily clock-in records, work order load of the production line, and equipment operating status, thus constructing a unified structured dataset containing all elements.

[0077] By using timed automated data collection and cleaning, time window bias and duplicate records between multi-source data are eliminated, avoiding transcription errors caused by manual data entry. Median imputation and standard deviation anomaly detection are employed to preserve the statistical validity of the data and avoid subjective assumptions, thereby improving the reliability of subsequent analysis. Unified format and encoding mapping enable data from heterogeneous systems to be comparable and fusionable, effectively supporting the construction of a cross-system holographic view based on employee ID. The association method based on employee ID as the primary key can accurately match labor resources and production load, providing a high-quality "human-machine-material" association data foundation for subsequent semantic parsing and rule extraction, fundamentally ensuring the input quality of upper-level AI models.

[0078] Furthermore, a semantic parsing input matrix is ​​constructed based on multi-source heterogeneous input stream data, including:

[0079] Entity recognition is performed on multi-source heterogeneous input stream data to identify employee entities, job entities, time entities, location entities, and business rule entities. The heterogeneous original representations of each entity are mapped to the system's standard naming to generate a standardized entity set.

[0080] Based on standardized entity sets, and using a preset time granularity as the slicing unit, time series alignment is performed on data from different sources. The historical data backtracking window and the future plan data forward window corresponding to each slice are determined. All event data within the window are aggregated according to the time axis to generate a time-aligned structured dataset.

[0081] Semantic fusion processing is performed on the time-aligned structured dataset to convert numerical data into natural language descriptions with context, rule-based data and applicable conditions are concatenated to generate rule description text, and employee attributes, job requirements and production plan data are combined to generate scenario description text, thus constructing a semantic unit pool composed of several semantic units.

[0082] Using each decision unit as the row index and employee characteristics, job requirements, rule constraints, and scenario context as column dimensions, semantic units corresponding to each dimension are retrieved from the semantic unit pool and populated. Multiple structured prompt texts are then concatenated and generated. These multiple structured prompt texts are arranged in rows to construct a semantic parsing input matrix. This matrix is ​​used to input the pre-trained manufacturing human resources big language model for rule feature extraction.

[0083] In this embodiment, the system takes the night shift scheduling task of senior welders in Workshop A of a manufacturing enterprise as the decision unit. First, it performs entity recognition on the preprocessed multi-source data: it maps "EMP_100234" in the attendance system and "Zhang Wei" in the HR system to the employee entity "EMP_ZW_001", maps "WELD-ADV" in the MES system and "senior welding" in the skill tag library to the job entity "SKILL_WELD_ADV", maps "2026-09-01" in the ERP order and "NIGHT_SHIFT" in the shift table to the time entity "2026-09-01_NIGHT", maps "Workshop A" and "Assembly_Line_3" in various systems to the location entity "LOC_A_WELD", and identifies the provisions of the Labor Law that "continuous work shall not exceed 6 hours" and the clause of the enterprise system that "night shift allowance shall be paid at 1.2 times the basic salary" as business rule entities. After completing the standardized mapping, a standardized entity set is generated.

[0084] The system uses "shift" as the time slice unit, takes the night shift on September 1, 2026 as the target slice, looks back 30 days to obtain the workshop's historical production capacity data and employees' historical attendance performance, and looks forward 7 days to obtain the order delivery plan for the next week. All event data within the window are aggregated according to the time axis to form an aligned structured dataset.

[0085] The system performs semantic fusion processing, transforming the numerical "average daily output of 850 pieces in the past 30 days" into a natural language description of "the average daily output of the welding process in this workshop in the past month is about 850 pieces, and the output of the night shift is usually 85% of that of the day shift." The system concatenates the rule clauses and applicable conditions into "the night shift allowance is paid at 1.2 times the basic salary, applicable to employees who work from 22:00 to 06:00 the next day." The system combines employee attributes and job requirements into "Zhang Wei, senior welding qualification, attendance rate of 98% in the past 30 days, specializing in thin plate welding, and able to accept night shift arrangements." This constructs a semantic unit pool containing more than 100 semantic units.

[0086] The system uses the night shift scheduling task as the row index, retrieves and fills the corresponding semantic units in the semantic unit pool with four dimensions: employee characteristics, job requirements, rule constraints, and scenario context. It then concatenates these semantic units to generate a structured prompt text: "Please arrange 5 senior welders for the night shift of Workshop A on September 1, 2026. Current candidates include Zhang Wei (senior welder, attendance rate 98%, available for night shifts), Li Qiang (senior welder, attendance rate 95%, available for night shifts)... They must comply with the rules that the night shift allowance is paid at 1.2 times the basic salary and that continuous work cannot exceed 6 hours. They must complete the welding task of 1000 products for order PO-20260901-015 on the same day." Multiple similar prompt texts arranged in rows form the semantic parsing input matrix.

[0087] Entity recognition and standardization eliminate different representations of the same business object across multiple systems, enabling the large language model to understand business logic based on a unified semantic space and avoiding feature extraction errors caused by name ambiguity. Time series alignment and sliding window design allow the model to perceive historical trends and predict future needs when parsing each rule, enhancing the temporal context awareness of rule extraction. Semantic fusion transforms dry numerical values ​​and codes into natural language descriptions rich in business meaning, significantly reducing the understanding difficulty of the large language model and improving the accuracy and recall of quadruple feature vector extraction. The matrix structure with decision units as row indices allows the system to process hundreds or thousands of scheduling tasks in parallel, transforming management systems that originally relied on manual interpretation into structured inputs that the model can directly consume, providing a high-quality semantic starting point for subsequent rule formalization and knowledge graph injection.

[0088] Furthermore, the semantic parsing input matrix is ​​input into the pre-trained large language model for manufacturing human resources to extract the four-tuple feature vector for each business logic, including:

[0089] The semantic parsing input matrix is ​​divided into several independent prompt units by row. The prompt units are lexicalized and a preset task instruction prefix and output format constraint suffix are added. The output format constraint suffix specifies that the large language model is output in a four-tuple structure of conditional entity, relation type, target entity and effective time.

[0090] All prompting units are input into a pre-trained big language model for manufacturing human resources in parallel. The big language model encodes the semantic information in the prompting units based on the attention mechanism. Through context-aware reasoning, it extracts the four-tuple information of human resources domain semantics from the prompting units and summarizes the four-tuple information into a four-tuple information set.

[0091] The conditional entity, relation type, target entity, and effective time in the quadruple information set are vectorized and encoded respectively. The four component vectors of the same quadruple are concatenated into a quadruple feature vector with fixed dimensions.

[0092] Using the row index of the semantic parsing input matrix as the association key, a mapping relationship is established between each extracted quadruple feature vector and the corresponding business logic entry, and a set of quadruple feature vectors is output.

[0093] In this embodiment, the system sequentially extracts each line of structured prompt text from the semantic parsing input matrix constructed in the previous stage. For the prompt unit "Please arrange 5 senior welders for the night shift of Workshop A on September 1, 2026. Candidates include Zhang Wei (senior welder, attendance rate 98%), Li Qiang (senior welder, attendance rate 95%), etc., and must comply with the rules such as night shift allowance being paid at 1.2 times the basic salary and continuous work not exceeding 6 hours. The welding task of 1000 products for order PO-20260901-015 must be completed on the same day," the system first performs lexicalization processing to decompose it into the smallest semantic unit sequence that the model can recognize. Then, "Please extract all business rules from the following production schedule description and output them according to the four dimensions of condition entity, relation type, target entity, and effective time" is appended to the beginning of the sequence as a task instruction prefix. At the same time, "Output format: [condition entity] | [relation type] | [target entity] | [effective time]" is appended to the end of the sequence as a format constraint suffix to ensure that the output of the large language model strictly follows the four-tuple structure.

[0094] The system inputs this prompt unit along with hundreds of other prompt units in the matrix into the pre-trained large language model for manufacturing human resources in a parallel batch processing manner. The model uses its multi-head self-attention mechanism to encode the semantic information in each prompt unit, and identifies the four-tuple information of business logic such as "night shift allowance - paid at 1.2 times - basic salary - daily from 22:00 to 06:00 the next day" and "continuous work - not exceeding - 6 hours - per shift" from the description through context-aware reasoning, and summarizes them into a four-tuple information set.

[0095] The system performs vectorization encoding on the condition entity "night shift allowance", relation type "paid at 1.2 times", target entity "basic salary", and effective time "22:00 to 06:00 the next day" for each quadruple in the set. (For example, a pre-trained word embedding model is used to map each component into a 768-dimensional dense vector, and then the four 768-dimensional components are concatenated end to end into a 3072-dimensional fixed-dimensional vector, converting all quadruples into quadruple feature vectors of the same dimension.)

[0096] The system uses the row index "ROW_20260901_NIGHT_WELD" of the semantic parsing input matrix as the association key to establish a mapping relationship between all the extracted quadruple feature vectors under the scheduling task and the original business logic entries corresponding to the row, and outputs a complete set of quadruple feature vectors for downstream rule formal modeling modules to call.

[0097] Through a dual-guidance mechanism of task instruction prefixes and output format constraint suffixes, the free generation of large language models is restricted within a pre-defined four-dimensional structured framework, significantly reducing the probability of model output deviating from the expected structure and ensuring accurate parsing by subsequent compilation modules. The parallel batch processing input method fully utilizes GPU computing resources, compressing the extraction time of hundreds of business logic lines from tens of minutes in serial processing to minutes, significantly improving the system's throughput. The context-aware capability of the multi-head self-attention mechanism can accurately understand the semantic differences of the same entity in different contexts within a scheduling scenario (e.g., "night shift" can be used as both a time entity and a...). The prefix of job type avoids semantic misjudgment caused by one-size-fits-all rule matching; vectorized encoding transforms discrete symbolic quadruples into numerical representations in a continuous dense vector space, which not only facilitates computer storage and retrieval, but also enables downstream rule type classifiers to directly route traffic based on the cosine similarity of vectors, providing a directly computable mathematical foundation for rule formalization compilation and knowledge graph construction; the design of establishing mapping relationships with row indexes preserves the complete link from the quadruple feature vector back to the original prompt text and corresponding business logic entries, giving the entire rule extraction process full traceability, which is convenient for subsequent auditing and error correction.

[0098] Furthermore, the semantic quadruple feature vectors output by the business logic multi-source acquisition and semantic parsing module are received, routed according to the rule type classifier, and the routed rules are formally compiled, including:

[0099] The source system identifier and the original text position index are added to each quadruple feature vector to form a rule vector to be processed with a source label. The rule vector to be processed is then input into a pre-trained rule type classifier.

[0100] The classifier performs type discrimination on each rule based on the semantic embedding of the relation type component and condition entity component of the quadruple feature vector, and outputs the rule type label.

[0101] Rule type tags must include at least one of the following: working hours rules, payroll calculation rules, scheduling constraints, skills matching rules, and compliance review rules.

[0102] Based on the rule type label, each quadruple feature vector is routed to the corresponding rule compilation channel. Each rule compilation channel is pre-set with a compilation template and syntax parser for different rule types. The compilation template defines the mapping relationship between the rule and the structured language.

[0103] In each rule compilation channel, the component values ​​of the quadruple feature vector are extracted. Based on the corresponding compilation template, the condition entity is mapped to the condition predicate, the relation type is mapped to the logical operator, the target entity is mapped to the target predicate, and the effective time is mapped to the time window limit condition. The components are combined according to the preset syntax rules to generate a structured rule description text that conforms to the formal language specification.

[0104] The structured rule description text output by each compilation channel is subjected to syntax verification and semantic integrity check. After the verification is passed, the formally compiled set of structured rules is output, and a mapping index table between each structured rule and the corresponding original quadruple feature vector is generated.

[0105] In this embodiment, please refer to Figure 2 This is a flowchart illustrating the process of receiving semantic quadruple feature vectors output by the multi-source acquisition and semantic parsing module of the business logic provided in this application embodiment, routing and splitting the traffic according to the rule type classifier, and formally compiling the split rules. The system receives the set of quadruple feature vectors output from the upstream, and one feature vector representing the rule "continuous work shall not exceed 6 hours" is processed first: First, the source system identifier "MES_20260901" and the original text position index "Line15" are added to the vector to form a vector to be processed with a source tracing label, which is then input into the pre-trained rule type classifier; The classifier, based on the semantic embedding of the relation type component in the feature vector (such as "shall not exceed" being encoded as a constraint class vector in the semantic space) and the semantic representation of the condition entity "continuous work", identifies it as a "compliance review class" rule and outputs the corresponding label.

[0106] The system routes the quadruple feature vector to a dedicated compilation channel for compliance review based on the tag. This channel has a pre-set compilation template for handling labor laws and working hour limits (which defines "continuous work" as the predicate "WorkDuration", "not exceeding" as the logical operator "≤", "6 hours" as the target predicate "MaxHours=6", and "per shift" as the time window limit "PerShift"). The compilation channel maps and combines the values ​​of each component of the quadruple according to the template to generate the structured rule description text "WorkDuration ≤ MaxHours(6) WITHINPerShift". After compilation, the system performs syntax verification (checks the matching of brackets and the legality of predicates) and semantic integrity check (confirms that the MaxHours parameter is not empty and the units are consistent) on the text. After the verification is passed, the rule is included in the structured rule set, and a mapping index record from the structured rule to the original quadruple feature vector and the original text position "Line15" is generated so that the source of the rule can be traced in subsequent audits.

[0107] The rule type classifier performs type discrimination by semantic embedding of relation type components and condition entity components. Compared with traditional rule classification methods based on keyword matching, it has stronger semantic generalization ability and can accurately identify rules with the same semantics but different expressions (e.g., "continuous work must not exceed 6 hours" and "the maximum continuous working time per shift is 6 hours" can both be correctly classified), significantly reducing the rate of missed and false positives. The design of splitting rule types to dedicated compilation channels allows various rules to be compiled in parallel, and each channel can independently maintain and optimize its own compilation template, avoiding the problem of bloated compilation logic and mutual interference caused by large differences in rule types in a single general-purpose compiler. The quadruplet component mapping... A formal compilation method for conditional predicates, logical operators, target predicates, and time window constraints transforms natural language-level business descriptions into precise and unambiguous formal language, eliminating rule execution deviations caused by differences in understanding in traditional manual interpretation. Syntax validation and semantic integrity checks intercept incomplete or formatted rule descriptions during the compilation stage, preventing such issues from flowing into downstream knowledge graphs and scheduling execution stages and causing system anomalies. The mapping index table establishes a full-link traceability path from structured rules to original quadruple feature vectors, enabling any rule with compilation anomalies to quickly locate its source text and extraction stage, significantly improving the system's debuggability and auditability.

[0108] Furthermore, rule nodes are injected into the manufacturing human resources knowledge graph to construct a reasonable relationship network. Each rule is assigned a version number and an effective fingerprint, and written into a versioned rule base, including:

[0109] Each structured rule in the mapping index table is parsed to extract the conditional entity, target entity, relation type, and effective time window information. The extracted conditional entity and target entity are used as entity nodes of the knowledge graph, the relation type is used as the directed relation edge between entity nodes, and the effective time window is used as the attribute label of the time dimension of the relation edge. According to the preset ontology model, each structured rule is converted into a knowledge graph triple.

[0110] Knowledge graph triples are injected into the manufacturing human resources knowledge graph. Entity nodes are deduplicated and merged. When the entity to be injected and an existing entity in the graph point to the same business object, the entity to be injected is merged into the existing entity node and a synonym mapping chain is established. When the relationship edge to be injected overlaps with the existing relationship edge, the relationship edge is fused. The time attributes of multiple rules are stored on the same relationship edge in the form of a time interval set, forming a reasonable relationship network. The reasonable relationship network supports relationship chain query based on graph traversal and reasoning calculation based on rule logic.

[0111] For each rule injected into the knowledge graph, the original quadruple feature vector corresponding to the rule and the generated structured rule description text are obtained and concatenated to form a rule content summary. The rule content summary is calculated using a hash algorithm to calculate the rule content hash value. The rule content hash value is combined with the current timestamp and the operator identifier to generate the effective fingerprint of the rule. The effective fingerprint is used to uniquely identify the content status of the rule at a specific time.

[0112] Each rule injected into the knowledge graph is assigned a version number. The version number adopts the semantic version number format. When the rule content changes, the corresponding major version number, minor version number or revision version number is incremented. The current version number is associated with the effective fingerprint, stored and written into the versioned rule base.

[0113] In this embodiment, please refer to Figure 3This is a schematic diagram illustrating the process of injecting rule nodes into the manufacturing human resources knowledge graph, constructing a reasonable relationship network, assigning a version number and an effective fingerprint to each rule, and writing it into a versioned rule base. The system inherits the structured rule set output from the upstream, with "WorkDuration ≤ MaxHours(6) WITHIN" as the rule. Taking "PerShift" as an example: First, the structured rule is parsed to extract the condition entity "WorkDuration", the target entity "MaxHours(6)", the relation type "≤" and the effective time window "PerShift". According to the preset ontology model, it is converted into a knowledge graph triple ("continuous work" → "limited by" → "6 hours / shift"). When it is injected into the manufacturing human resources knowledge graph, the system detects that the "continuous work" entity already exists in the graph. The node to be injected is merged into the existing node and a synonym mapping chain is established. At the same time, the system detects that there is already a relation edge "continuous work shall not exceed 8 hours". The time attribute "PerShift" of the new relation edge is merged and stored in the same relation edge in the form of a time interval set. This makes the relation edge carry both the "6 hours / shift" and "8 hours / day" rules and each is marked with a different effective time interval. This forms a reasonable relation network, which supports subsequent graph traversal query of "all time limit constraints related to continuous work" and rule matching based on time reasoning.

[0114] The system concatenates the original quadruple feature vector of the rule with the structured description text to form a summary, calculates the hash value of the rule content "a3f8c9d2…" using the SHA-256 hash algorithm, and concatenates it with the current timestamp "2026-09-01T10:30:00Z" and the operator identifier "ADMIN_HR_001" to generate a unique effective fingerprint "f7e2a9c3…". At the same time, the system assigns a semantic version number "v1.0.0" to the rule, associates it with the effective fingerprint, and writes it into the versioned rule base.

[0115] The entity node deduplication and merging mechanism, along with the synonym mapping chain mechanism, avoids duplicate storage and semantic splitting of the same business object in the knowledge graph, maintaining the simplicity and semantic consistency of the graph structure and providing a clear and non-redundant query path for subsequent graph traversal and reasoning. Relationship edge attribute fusion, using a time interval set to overlay multiple rules on the same relationship edge, enables the knowledge graph to express differentiated constraints of the same business relationship under different time windows. This avoids the graph expansion and increased query complexity caused by creating redundant relationship edges to express multiple versions of rules, significantly improving the graph's storage efficiency and reasoning performance. Based on rule content hash values ​​and timestamps... The operator identifier is used to generate an effective fingerprint, ensuring that each rule has a unique and immutable identifier for its content status at a specific moment. Any minor change to the rule content (such as changing "6 hours" to "7 hours") will result in a complete change in the hash value and effective fingerprint, providing an accurate comparison basis for subsequent rule change tracking and conflict detection. The association storage of semantic version numbers and effective fingerprints constructs a complete rule version evolution chain, supporting the backtracking of the rule status of any historical version by time point. When the scheduling system encounters an anomaly, it can quickly locate the version and content of the most recent rule change, significantly improving the system's rollback capability and troubleshooting efficiency.

[0116] Furthermore, it monitors rule base version change events and triggers the incremental compilation pipeline, including:

[0117] Deploy a rule base version change event listener. The listener continuously monitors rule addition events, rule modification events, and rule deletion events in the versioned rule base. It captures the rule identifier, change type, version number before and after the change, and change timestamp corresponding to the event, generates a version change event message, and pushes it to the message queue.

[0118] The change rule identifier and change type are obtained from the version change event message parsing. The changed rule content is read from the versioned rule base according to the change rule identifier. The incremental compilation scope is determined according to the change type. The incremental compilation scope includes at least the change rule itself, the associated rules with the target entity of the change rule as the head entity, and the same group of rules with the same rule type as the change rule.

[0119] Based on the incremental compilation scope, each rule node and its relationship within the incremental compilation scope is extracted from the knowledge graph. The extracted rule nodes are arranged in a preset compilation order, and the rule compilation channel corresponding to the compilation order is called to perform incremental compilation on the changed rules and related rules, generating an executable rule instruction set after incremental compilation, and recording the dependency resolution log during the compilation process.

[0120] The compilation results of the incrementally compiled executable rule instruction set are verified. The effective fingerprint of the compiled instruction is compared with the effective fingerprint of the corresponding rule in the versioned rule base. If the comparison is consistent, the compilation is marked as successful. The successful compilation status and executable instruction set are output to the waiting queue of the dynamic rule engine. At the same time, an incremental compilation report is generated to record the number of rules involved in this compilation and the compilation time.

[0121] In this embodiment, the rule base version change event listener deployed in the system continuously monitors the versioned rule base. When the HR administrator modifies a compliance review rule "Continuous work hours shall not exceed 6 hours" from v1.0.0 to v1.0.1 (the content is changed to "Continuous work hours shall not exceed 7 hours") and submits the change, the listener immediately captures the event, obtains the rule identifier "RULE_HR_0001", the change type "modification", the version number before the change "v1.0.0", the version number after the change "v1.0.1", and the change timestamp "2026-09-01T14:30:00Z", assembles this information into a version change event message, and pushes it to the Kafka message queue.

[0122] The incremental compilation consumer in the system backend pulls the message from the queue and parses it to obtain the change rule identifier and change type. Then, based on "RULE_HR_0001", it reads the modified rule content "continuous work shall not exceed 7 hours" from the versioned rule base. At the same time, based on the change type "modified", it determines the scope of incremental compilation. In addition to the change rule itself, it also includes all related rules with "continuous work" as the head entity (such as "continuous work requires a break of no less than 15 minutes"), as well as other rules in the same group belonging to the "compliance review category" into the scope of incremental compilation.

[0123] The system extracts relevant rule nodes and their relationships from the knowledge graph based on the compilation scope, sorts these nodes from shallow to deep according to their dependency depth, and calls the corresponding rule compilation channels in sequence to compile and generate executable rule instruction sets (such as compiling "continuous work must not exceed 7 hours" into "WorkDuration ≤ MaxHours(7) WITHINPerShift"). At the same time, it records the dependency resolution logs between rules involved in the compilation process.

[0124] After compilation, the system compares the effective fingerprint of the new compilation instructions with the effective fingerprint of version "RULE_HR_0001" v1.0.1 in the versioned rule base. If the comparison is consistent, the compilation is marked as successful. The executable instruction set is output to the waiting queue of the dynamic rule engine, and an incremental compilation report is generated to record that this incremental compilation involves 3 rules and takes 47 milliseconds.

[0125] By decoupling event listeners and message queues, rule change events can be captured and processed asynchronously and reliably. Even if a temporary failure occurs during the incremental compilation process, messages can still be retained in the queue and consumed again after the system recovers, avoiding the risk of loss of changes. The incremental compilation scope is intelligently determined based on the change type instead of a full recompile. Only the necessary scope—the changed rules, the associated rules with the target entity of the changed rule as the head entity, and the rules in the same group—is compiled, significantly reducing the amount of computation and time consumed during compilation. This keeps the response time from rule change submission to effectiveness within milliseconds. Dependency resolution logs fully record the references and dependencies between rules during the compilation process. When a rule change causes downstream compilation failure, the cause can be quickly located based on the logs (e.g., a constraint conflict between "continuous work should not exceed 7 hours" and "rest time should not be less than 15 minutes"), significantly improving the system's debuggability. The bidirectional comparison mechanism between the compiled effective fingerprint and the versioned rule base ensures that the executable instructions generated by compilation are completely consistent with the rule versions stored in the base, eliminating the risk of semantic deviation between instructions and expected rules due to compilation caching or intermediate processing errors. The incremental compilation report provides operations and maintenance personnel with quantitative indicators of the compilation efficiency and impact of each rule change, facilitating system performance monitoring and continuous optimization.

[0126] Furthermore, using knowledge graph reasoning to scan for combinations of overlapping conditions but contradictory conclusions between new rules and existing rules includes:

[0127] Take the newly injected rule nodes and their relationships in the knowledge graph. Starting from the condition entity of the new rule node, perform a graph traversal query in the knowledge graph. The graph traversal query includes at least a one-hop relationship query and a multi-hop relationship query. Retrieve all existing rule nodes in the existing rule library that have the same entity as the new rule node and are reachable through relationship edges, and construct a candidate conflict detection set.

[0128] Extract the condition entity set of each existing rule node and the condition entity set of the new rule node from the candidate conflict detection set and calculate the overlap. When the overlap is greater than or equal to a preset threshold, mark the existing rule node as a condition overlap candidate node.

[0129] The rule conclusions corresponding to the overlapping candidate nodes and the rule conclusions of the new rule nodes are checked for contradiction. The contradiction detection includes judging the conflict of conclusion entities, judging the opposite or mutually exclusive relation types, and judging the non-intersection of quantitative constraint intervals. If any detection result is true, the overlapping candidate node and the new rule node are determined to be a conflict pair.

[0130] Record the rule identifiers, overlapping condition descriptions, and contradictory conclusion descriptions involved in the conflict pairs, generate a conflict detection log, and output a set of conflict pairs.

[0131] In this embodiment, when the new rule node "Continuous work shall not exceed 6 hours (night shift only)" is injected into the knowledge graph, the system performs a graph traversal query in the knowledge graph starting from the condition entity "continuous work" of the node: first, a one-hop query is performed to directly retrieve all rule nodes directly connected to the "continuous work" entity, and then a two-hop query is performed to retrieve rule nodes indirectly associated through the "night shift" entity. In this way, existing rule nodes such as "Continuous work on the night shift shall not exceed 8 hours" and "Night shift requires a 15-minute break every 4 hours" are retrieved, and a candidate conflict detection set is constructed.

[0132] The system extracts the condition entity set {continuous work, night shift} of the new rule and the condition entity set {continuous work, night shift} of the candidate node “continuous work on night shift shall not exceed 8 hours”. The Jaccard overlap is calculated to be 1.0, which exceeds the preset threshold of 0.7. Therefore, it is marked as a candidate node with conditional overlap.

[0133] The system checks for contradictions between the candidate node's conclusion "must not exceed 8 hours" and the new rule's conclusion "must not exceed 6 hours". Both are numerical upper limit constraints, but the quantitative constraint intervals of 6 hours and 8 hours are inclusive rather than mutually exclusive (the 6-hour interval completely falls within the 8-hour interval, which is actually a stricter constraint rather than a contradiction), so they do not constitute a conflict pair.

[0134] When the system simultaneously checks the conclusion of another candidate node, "Night shifts need to arrange 15-minute breaks every 4 hours," against the conclusion of the new rule, it finds that the conclusion entity "break arrangement" and the conclusion entity "working hours limit" of the new rule are not of the same type. Therefore, there is no need to perform quantitative interval judgment, nor do they constitute a conflict pair.

[0135] Record that no conflicts were found in this scan, output an empty set of conflict pairs, and generate a conflict detection log containing the scan range and the judgment criteria.

[0136] Leveraging the graph traversal capabilities of knowledge graphs (one-hop and multi-hop queries), this method retrieves association rules radiating outwards from conditional entities. Compared to traditional keyword-matching-based retrieval methods, it can discover candidate conflicting rules indirectly related through relationship chains (such as those connected via the "night shift" entity), significantly improving the recall rate of conflict detection and avoiding missed conflicts due to indirect associations between rules. The method quantitatively calculates the degree of overlap of conditional entities using Jaccard overlap, and uses a configurable threshold as the judgment criterion, replacing the subjective approach of relying on human experience to judge conditional similarity. This makes the judgment process of conditional overlap objective and quantifiable, while the adjustable threshold allows the system to flexibly adapt to the sensitivity requirements of different manufacturing scenarios. From the conclusion entity... Comprehensive conflict detection is performed across three dimensions: conflict, opposite or mutually exclusive relationship types, and no intersection of quantitative constraint intervals. In particular, it provides refined interval relationship judgment for numerical interval conflicts that are most common and prone to causing problems in manufacturing scheduling (such as the simultaneous existence of two rules, "6-hour limit" and "8-hour limit", where 6 hours is a strict subset of 8 hours and is not actually contradictory). This avoids the problem of misjudging stricter constraints as contradictions. The conflict detection log fully records the candidate range, overlap calculation value, and contradiction judgment basis for each scan, providing a transparent and traceable decision basis for subsequent manual review or priority adjudication. This allows the system to quickly trace whether unexpected behavior is caused by missed conflict judgment or inappropriate adjudication strategy.

[0137] Furthermore, based on a preset multi-objective adjudication strategy function, the priority order of conflict pairs is solved, high-priority rules are automatically retained, suppressed rules are put into a dormant state, and a conflict adjudication log is generated, including:

[0138] Extract the rule attribute vectors of the two rules involved in the conflict. The rule attribute vectors shall include at least the rule source system weight, rule effective timestamp, rule applicable scope level, rule type priority coefficient, and rule historical usage frequency.

[0139] The rule attribute vector is input into a preset multi-objective adjudication strategy function. The multi-objective adjudication strategy function calculates a comprehensive priority score for each rule by combining multiple adjudication strategy dimensions. The adjudication strategy dimensions include at least the scores of the time priority strategy dimension, the source authority priority strategy dimension, the specialization priority strategy dimension, and the stability priority strategy dimension. The comprehensive priority score is obtained by weighted summation of the scores of each strategy dimension.

[0140] Within the same conflict pair, compare the overall priority scores of the two rules, retain the rule with the higher overall priority score as the active rule, mark the rule with the lower overall priority score as the suppressed rule, and attach a suppression mark and a description of the suppression reason to the suppressed rule. The description of the suppression reason should include at least the identifier of the conflicting rule and a description of the conflict dimension.

[0141] Suppressed rules are transferred to a dormant state. The dormant state is a state in which the rule does not participate in subsequent scheduling decisions but is still retained in the versioned rule base and can be queried and restored. The dormant state record is recorded for the rule that is transferred to the dormant state, including the dormant effective timestamp and the conflict pair identifier of the adjudication source.

[0142] Summarize the adjudication results of all conflict pairs in this batch and generate a conflict adjudication log. The conflict adjudication log should include at least the adjudication time of each conflict pair, the rule identifiers involved in the adjudication, the scores of each strategy dimension on which the adjudication is based, the adjudication result, and the dormant status record of the suppressed rule. Write the conflict adjudication log to the log storage area of ​​the versioned rule base.

[0143] In this embodiment, when the conflict detection module outputs a set of conflict pairs (for example, the newly injected compliance review rule "night shift continuous work shall not exceed 6 hours" conflicts with the existing rule "night shift continuous work shall not exceed 8 hours"), the adjudication module first extracts the attribute vectors of the two rules respectively: the new rule comes from "relevant internal regulations of the enterprise" (source system weight 0.95), the effective timestamp is "2026-09-01", the applicable scope level is "national level", the rule type priority coefficient is "compliance review type - 0.9", and the historical usage frequency is "0 times" (first injection); the existing rule comes from "internal enterprise system" (source system weight 0.70), the effective timestamp is "2024-01-01", the applicable scope level is "workshop level", the rule type priority coefficient is "shift constraint type - 0.6", and the historical usage frequency is "180 times".

[0144] The system inputs each component of the two vectors into a preset multi-objective adjudication strategy function. This function calculates the scores for the time priority dimension (new rules score higher due to time updates), the source authority priority dimension (new rules are significantly superior because their source is relevant internal company regulations), the specificity priority dimension (existing rules are more specific because their application scope is limited to the workshop), and the stability priority dimension (existing rules are more stable due to 180 historical uses). After weighting and summing the scores of each dimension with preset weights (time weight 0.15, source authority weight 0.40, specificity weight 0.25, and stability weight 0.20), the overall priority score of the new rule is 0.82, and the overall priority score of the existing rule is 0.61.

[0145] Based on this, the system determined that the new rule had a higher priority, retained the new rule as an active rule, and marked the existing rule as a suppressed rule with an additional description of the suppression reason: "It conflicts with rule RULE_HR_20260901 (internal release of the enterprise) in the dimension of continuous working time limit, and is suppressed because it scores poorly in the dimension of source authority."

[0146] The system puts the existing rule into a dormant state (i.e., the rule no longer participates in subsequent scheduling decisions, but the records and effective fingerprints in its versioned rule base are still fully preserved and can be queried and recovered), and records the dormant effective timestamp as "2026-09-01T15:00:00Z" and the conflict pair identifier of the adjudication source as "CONFLICT_20260901_001".

[0147] The system summarizes the adjudication results of all conflict pairs in this batch, generates a conflict adjudication log that includes the adjudication time, rule identifiers involved in the adjudication, detailed scores for each strategy dimension, adjudication results, and records of the dormant status of suppressed rules, and writes it to the log storage area of ​​the versioned rule base.

[0148] By employing a multi-dimensional weighted scoring mechanism, the determination of rule priority is transformed from a subjective qualitative decision relying on expert experience into a repeatable and traceable objective quantitative calculation, eliminating the inconsistency of precedents caused by individual biases in manual adjudication. Four adjudication strategy dimensions (time, source authority, specificity, and stability) comprehensively evaluate the importance of each rule from different perspectives. Time priority ensures timely response to legislative updates; source authority guarantees the absolute priority of higher-level regulations; specificity maintains the flexibility needs of grassroots management; and stability protects long-term, proven production experience. The weighting of these four factors ensures that the adjudication results balance compliance and flexibility. The system addresses the triple management requirements of activity and stability. Low-priority rules are handled by remaining dormant rather than being deleted. This allows dormant rules to quickly become active again if a new rule is later withdrawn or its scope of application is adjusted, without needing to be re-extracted and recompiled from the original text, significantly reducing the cost of rule rollback. A complete conflict resolution log records all quantitative dimension scores and reasons for suppression for each resolution, providing transparent and self-consistent decision-making chain evidence for subsequent manual audits. When production line managers question the resolution results, the entire resolution logic can be traced back for review, significantly enhancing the system's explainability and management's trust.

[0149] Furthermore, the real-time operational state flow matrix is ​​obtained, and the state matrix, along with the rule constraints, is input into the multi-objective scheduling optimization network to output the optimal scheduling scheme vector with flexible objectives, including:

[0150] Real-time operational status data is obtained from the production execution system, attendance system and equipment IoT platform. The operational status data includes at least the quantity of work-in-process and process cycle time of each production line, the operating status and utilization rate of each piece of equipment, and the real-time on-duty status and remaining available working hours of each position. The operational status data is matrix-coded according to a unified time granularity to construct a real-time operational status flow matrix.

[0151] Read a set of valid rule snapshots from the versioned rule base. The set of valid rule snapshots includes at least work hour rules, skill matching rules, compliance constraint rules, and cost control rules.

[0152] Each rule in the effective rule snapshot set is parsed into a constraint tensor. The constraint tensor is then concatenated and fused with the real-time operation state flow matrix along the feature dimension to construct the scheduling optimization input tensor.

[0153] The input tensor of the scheduling optimization is input into the pre-trained multi-objective scheduling optimization network. The multi-objective scheduling optimization network is based on an encoder and decoder architecture. The encoder performs attention encoding on the input tensor, and the decoder generates multiple sets of candidate scheduling schemes in parallel based on the flexible objective weight vector. The flexible objectives include at least minimizing labor costs, maximizing skill matching degree, and maximizing order delivery timeliness.

[0154] Multiple candidate scheduling schemes are sorted by multi-objective non-dominated ordering to select Pareto optimal solution set. From the Pareto optimal solution set, the scheme with the highest comprehensive score is selected as the optimal scheduling scheme vector according to the preset flexible objective preference weight, and the optimal scheduling scheme vector is output.

[0155] In this embodiment, a manufacturing company's Workshop A encountered an emergency at 14:30 on September 1, 2026—a critical piece of equipment on the welding production line malfunctioned, causing a decrease in production capacity. The system obtained real-time data from the MES showing that the work-in-process quantity on the production line was 850 pieces, and the process cycle time dropped from the standard 120 seconds / piece to 180 seconds / piece. The equipment IoT platform showed that the equipment's operating status was "faulty" and its utilization rate was 42%. The attendance system showed that there were currently 3 senior welders on duty (Zhang Wei with 4.5 hours of remaining working time and Li Qiang with 3.2 hours of remaining working time) and those who could be urgently dispatched. Two backup personnel (Wang Fang has 5 hours of remaining working time, and Zhao Lei has 4.8 hours of remaining working time) are assigned to the system. The system encodes this data in a matrix format with a 15-minute time granularity to form a real-time operation status flow matrix. Then, it reads a set of valid rule snapshots from the versioned rule library, including the working time rule "each person shall not work continuously for more than 6 hours per shift", the skill matching rule "welding positions require at least intermediate welding qualifications", the compliance constraint rule "night shift allowance shall be paid at 1.2 times the basic salary" and the cost control rule "overtime work costs shall not exceed 1.5 times the normal working hours".

[0156] The system parses these rules into constraint tensors and concatenates them with the real-time operational state flow matrix along the feature dimension to construct a scheduling optimization input tensor, which is then input into a pre-trained multi-objective scheduling optimization network. This network is based on an encoder-decoder architecture. The encoder uses a multi-head self-attention mechanism to interactively encode the state information and rule constraints in the input tensor. The decoder generates 50 candidate scheduling schemes in parallel based on a flexible objective weight vector (labor cost weight 0.4, skill matching degree weight 0.3, and order delivery timeliness weight 0.3). These schemes include: Scheme A, which arranges Zhang Wei, Li Qiang, and Wang Fang to be on duty and extends their working hours to 6 hours; Scheme B, which arranges Zhang Wei, Li Qiang, and Zhao Lei to be on duty and uses backup equipment to temporarily restore some production capacity; and Scheme C, which schedules all four welders to work in two shifts.

[0157] The system performs multi-objective non-dominated ranking on 50 sets of solutions, selects Pareto optimal solution set (where solution A has the lowest labor cost but only 75% on-time delivery rate, solution C has a 95% on-time delivery rate but the highest labor cost, and solution B achieves a balance between the two objectives), calculates the comprehensive score according to the preset flexible objective preference weights, and selects solution B as the optimal scheduling solution vector output.

[0158] By acquiring real-time operational status data such as work-in-process quantity, process cycle time, equipment utilization, and personnel on-duty status from MES, IoT devices, and attendance systems, and constructing a state flow matrix at a unified time granularity, scheduling decisions can accurately respond to sudden fluctuations in production lines (such as reduced capacity due to equipment failure). This upgrades scheduling from "static scheduling based on historical experience" to "dynamic adaptive scheduling based on real-time operating conditions," significantly improving the matching degree between scheduling plans and actual production needs. Effective rule snapshots in the versioned rule base are parsed into constraint tensors and integrated with the state matrix, ensuring that scheduling plans are constrained by multiple rules such as working hour limits, skill thresholds, compliance requirements, and cost budgets during the generation stage. This avoids the inefficient cycle of repeated adjustments due to constraint violations in the traditional "scheduling first, verification later" model. The multi-objective scheduling optimization network with a device-decoder architecture utilizes an attention mechanism to jointly encode state and rule tensors, enabling the network to capture the potential correlation between "equipment failure" and "an employee holding the qualification to operate the equipment," generating candidate solutions with business interpretability rather than simple combination enumeration. By filtering Pareto optimal solutions through multi-objective non-dominated sorting, the system can simultaneously optimize three conflicting objectives: labor cost, skill matching, and order delivery timeliness. This avoids the local optimum trap of simply weighting multiple objectives, where optimizing a single objective leads to the severe deterioration of other objectives. The flexible objective preference weights can be dynamically adjusted according to the enterprise's management strategy, ensuring that the scheduling optimization direction is consistent with the enterprise's phased management priorities, enhancing the system's business adaptability and flexibility.

[0159] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0160] This invention is described with reference to flowchart illustrations and / or block diagrams of systems, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0161] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0162] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0163] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0164] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A manufacturing AI-powered intelligent human resource management system, characterized in that, include: The business logic multi-source acquisition and semantic parsing module is used to acquire multi-source heterogeneous input stream data, preprocess the multi-source heterogeneous input stream data, construct a semantic parsing input matrix based on the multi-source heterogeneous input stream data, input the semantic parsing input matrix into the pre-trained manufacturing human resources big language model, and extract the four-tuple feature vector of each business logic. The rule formal modeling and knowledge graph construction module is used to receive the semantic quadruple feature vectors output by the business logic multi-source collection and semantic parsing module, perform routing and distribution according to the rule type classifier, formally compile the distributed rules, inject the rule nodes into the manufacturing human resources knowledge graph, build a reasonable relationship network, assign a version number and effective fingerprint to each rule, and write it into the versioned rule library. The dynamic rule engine and conflict resolution module are used to listen for rule base version change events, trigger the incremental compilation pipeline, use knowledge graph reasoning to scan whether there are combinations of overlapping conditions but contradictory conclusions between new rules and existing rules, mark conflict pairs, solve the priority order of conflict pairs based on the preset multi-objective adjudication strategy function, automatically retain high-priority rules, put suppressed rules into a dormant state, generate conflict adjudication logs, and distribute the new rule set to each execution sub-engine through atomic hot switching instructions. A canary release strategy is adopted, and after monitoring for no abnormal indicators, the full release is pushed out, and a set of valid rule snapshots is output. The intelligent scheduling and execution module for manufacturing scenarios is used to acquire the real-time operation state flow matrix, input the state matrix and rule constraints into the multi-objective scheduling optimization network, output the optimal scheduling scheme vector of flexible objectives, pre-simulate the output of the scheduling scheme in the virtual simulation environment, generate the execution instruction flow and distribute it to subordinate terminals.

2. The AI-powered intelligent human resource management system for manufacturing as described in claim 1, characterized in that, The acquisition of multi-source heterogeneous input stream data and the preprocessing of the multi-source heterogeneous input stream data include: Multi-source heterogeneous input stream data includes: work order information data, equipment operating status and equipment operating hours obtained from the production execution system; unique identification employee number and attendance information data obtained from the attendance system; and production plan data, material requirements data and order scheduling data obtained from the enterprise resource planning system. The specific steps for preprocessing multi-source heterogeneous input stream data are as follows: collect multi-source heterogeneous input stream data from the production execution system, attendance system and enterprise resource planning system at regular intervals; clean the collected multi-source heterogeneous input stream data, including deduplication, missing value imputation and outlier deletion; standardize the cleaned multi-source heterogeneous input stream data and associate them with each other using a unique identifier employee number.

3. The AI-powered intelligent human resource management system for manufacturing as described in claim 1, characterized in that, The construction of the semantic parsing input matrix based on multi-source heterogeneous input stream data includes: Entity recognition is performed on multi-source heterogeneous input stream data to identify employee entities, job entities, time entities, location entities, and business rule entities. The heterogeneous original representations of each entity are mapped to the system's standard naming to generate a standardized entity set. Based on the standardized entity set, data from different sources are time-series aligned using a preset time granularity as the slicing unit. The historical data backtracking window and the future plan data forward window corresponding to each slice are determined. All event data within the window are aggregated according to the time axis to generate a time-aligned structured dataset. Semantic fusion processing is performed on the time-aligned structured dataset to convert numerical data into natural language descriptions with context, to concatenate rule data with applicable conditions to generate rule description text, and to combine employee attributes, job requirements and production plan data to generate scenario description text, thus constructing a semantic unit pool composed of several semantic units. Using each decision unit as a row index and employee characteristics, job requirements, rule constraints, and scenario context as column dimensions, semantic units corresponding to each dimension are retrieved and populated from the semantic unit pool, and multiple structured prompt texts are generated by concatenating them. The multiple structured prompt texts are arranged in rows to construct a semantic parsing input matrix. The matrix is ​​used to input the pre-trained manufacturing human resources big language model for rule feature extraction.

4. The AI-powered intelligent human resource management system for manufacturing as described in claim 1, characterized in that, The semantic parsing input matrix is ​​input into a pre-trained large language model for manufacturing human resources, and the four-tuple feature vector of each business logic is extracted, including: The semantic parsing input matrix is ​​divided into several independent prompting units by row. The prompting units are lexicalized and a preset task instruction prefix and output format constraint suffix are added. The output format constraint suffix specifies that the large language model is output in a four-tuple structure of conditional entity, relation type, target entity and effective time. All prompting units are input in parallel into a pre-trained large language model for manufacturing human resources. The large language model encodes the semantic information in the prompting units based on an attention mechanism. Through context-aware reasoning, it extracts the four-tuple information of human resources semantics from the prompting units and summarizes the four-tuple information into a four-tuple information set. The conditional entity, relation type, target entity, and effective time in the quadruple information set are vectorized and encoded respectively. The four component vectors of the same quadruple are concatenated into a quadruple feature vector with fixed dimensions. Using the row index of the semantic parsing input matrix as the association key, a mapping relationship is established between each extracted quadruple feature vector and the corresponding business logic entry, and a set of quadruple feature vectors is output.

5. The AI-powered intelligent human resource management system for manufacturing as described in claim 5, characterized in that, The semantic quadruple feature vector output by the receiving business logic multi-source acquisition and semantic parsing module is routed according to the rule type classifier, and the routed rules are formally compiled, including: The source system identifier and the original text position index are added to each quadruple feature vector to form a rule vector to be processed with a source tracing label. The rule vector to be processed is then input into a pre-trained rule type classifier. The classifier performs type discrimination on each rule based on the semantic embedding of the relation type component and condition entity component of the quadruple feature vector, and outputs the rule type label. The rule type labels include at least one of the following: working hour rules, payroll calculation, scheduling constraints, skills matching, and compliance review. According to the rule type label, each quadruple feature vector is routed to the corresponding rule compilation channel. Each rule compilation channel is pre-set with a compilation template and a syntax parser for different rule types. The compilation template defines the mapping relationship between the rule and the structured language. In each rule compilation channel, the component values ​​of the quadruple feature vector are extracted. Based on the corresponding compilation template, the condition entity is mapped to the condition predicate, the relation type is mapped to the logical operator, the target entity is mapped to the target predicate, and the effective time is mapped to the time window limit condition. The components are combined according to the preset syntax rules to generate a structured rule description text that conforms to the formal language specification. The structured rule description text output by each compilation channel is subjected to syntax verification and semantic integrity check. After the verification is passed, the formally compiled set of structured rules is output, and a mapping index table between each structured rule and the corresponding original quadruple feature vector is generated.

6. The AI-powered intelligent human resource management system for manufacturing as described in claim 1, characterized in that, The process of injecting rule nodes into the manufacturing human resources knowledge graph, constructing a reasonable relationship network, assigning a version number and an effective fingerprint to each rule, and writing them into a versioned rule base includes: Each structured rule in the mapping index table is parsed to extract the conditional entity, target entity, relation type, and effective time window information. The extracted conditional entity and target entity are used as entity nodes of the knowledge graph, the relation type is used as directed relation edges between entity nodes, and the effective time window is used as the attribute label of the time dimension of the relation edge. Each structured rule is converted into a knowledge graph triple according to the preset ontology model. Knowledge graph triples are injected into the manufacturing human resources knowledge graph. Entity nodes are deduplicated and merged. When the entity to be injected and an existing entity in the graph point to the same business object, the entity to be injected is merged into the existing entity node and a synonym mapping chain is established. When the relationship edge to be injected overlaps with an existing relationship edge, the relationship edge is fused with attributes. The time attributes of multiple rules are stored on the same relationship edge in the form of a time interval set, forming a reasonable relationship network. The reasonable relationship network supports relationship chain query based on graph traversal and reasoning calculation based on rule logic. For each rule injected into the knowledge graph, the original quadruple feature vector corresponding to the rule and the generated structured rule description text are obtained and concatenated to form a rule content summary. The rule content summary is calculated using a hash algorithm to calculate the rule content hash value. The rule content hash value is combined with the current timestamp and the operator identifier to generate the effective fingerprint of the rule. The effective fingerprint is used to uniquely identify the content status of the rule at a specific time. Each rule injected into the knowledge graph is assigned a version number, which adopts the format of semantic version number. When the rule content changes, the corresponding major version number, minor version number or revision version number is incremented, and the current version number is associated with the effective fingerprint and stored and written into the versioned rule base.

7. The AI-powered intelligent human resource management system for manufacturing as described in claim 1, characterized in that, The monitoring rule base version change event triggers the incremental compilation pipeline, including: Deploy a rule base version change event listener. The listener continuously monitors rule addition events, rule modification events, and rule deletion events in the versioned rule base, captures the rule identifier, change type, version number before and after the change, and change timestamp corresponding to the event, generates a version change event message, and pushes it to the message queue. The change rule identifier and change type are obtained from the version change event message parsing. The changed rule content is read from the versioned rule base according to the change rule identifier. The incremental compilation scope is determined according to the change type. The incremental compilation scope includes at least the change rule itself, the associated rules with the target entity of the change rule as the head entity, and the same group of rules with the same rule type as the change rule. Based on the incremental compilation scope, each rule node and its relationship within the incremental compilation scope is extracted from the knowledge graph. The extracted rule nodes are arranged in a preset compilation order, and the rule compilation channel corresponding to the compilation order is called to perform incremental compilation on the changed rules and related rules, generating an executable rule instruction set after incremental compilation, and recording the dependency resolution log during the compilation process. The compilation results of the incrementally compiled executable rule instruction set are verified. The effective fingerprint of the compiled instruction is compared with the effective fingerprint of the corresponding rule in the versioned rule base. If the comparison is consistent, the compilation is marked as successful. The successful compilation status and executable instruction set are output to the waiting queue of the dynamic rule engine. At the same time, an incremental compilation report is generated to record the number of rules involved in this compilation and the compilation time.

8. The AI-powered intelligent human resource management system for manufacturing as described in claim 1, characterized in that, The method of using knowledge graph reasoning to scan for combinations of overlapping conditions but contradictory conclusions between new rules and existing rules includes: Take the newly injected rule nodes and their relationships in the knowledge graph, and perform a graph traversal query in the knowledge graph starting from the condition entity of the new rule node. The graph traversal query includes at least a one-hop relationship query and a multi-hop relationship query. Retrieve all existing rule nodes in the existing rule library that have the same entity as the new rule node and are reachable via relationship edges, and construct a candidate conflict detection set. Extract the condition entity set of each existing rule node and the condition entity set of the new rule node from the candidate conflict detection set and calculate the overlap. When the overlap is greater than or equal to a preset threshold, mark the existing rule node as a condition overlap candidate node. The rule conclusions corresponding to the overlapping candidate nodes and the rule conclusions of the new rule nodes are checked for contradictions. The contradiction detection includes judging the conflict of conclusion entities, judging the opposite or mutually exclusive relation types, and judging the non-intersection of quantitative constraint intervals. If any detection result is true, the overlapping candidate node and the new rule node are determined to be a conflict pair. Record the rule identifiers, overlapping condition descriptions, and contradictory conclusion descriptions involved in the conflict pairs, generate a conflict detection log, and output a set of conflict pairs.

9. The AI-powered intelligent human resource management system for manufacturing as described in claim 1, characterized in that, The priority order of conflict pairs is determined based on a preset multi-objective adjudication strategy function. High-priority rules are automatically retained, suppressed rules are put into a dormant state, and a conflict adjudication log is generated, including: Extract the rule attribute vectors of the two rules involved in the conflict. The rule attribute vectors include at least the rule source system weight, rule effective timestamp, rule applicable scope level, rule type priority coefficient, and rule historical usage frequency. The rule attribute vector is input into a preset multi-objective adjudication strategy function. The multi-objective adjudication strategy function calculates a comprehensive priority score for each rule by integrating multiple adjudication strategy dimensions. The adjudication strategy dimensions include at least the scores of the time priority strategy dimension, the source authority priority strategy dimension, the specialization priority strategy dimension, and the stability priority strategy dimension. The comprehensive priority score is obtained by weighted summation of the scores of each strategy dimension. Within the same conflict pair, compare the overall priority scores of the two rules, retain the rule with the higher overall priority score as the active rule, mark the rule with the lower overall priority score as the suppressed rule, and attach a suppression mark and a description of the suppression reason to the suppressed rule. The description of the suppression reason includes at least the identifier of the conflicting rule and a description of the conflict dimension. Suppressed rules are transferred to a dormant state, which is a state in which the rules do not participate in subsequent scheduling decisions but are still retained in the versioned rule base and can be queried and restored. The dormant state records the dormant effective timestamp and the conflict pair identifier of the adjudication source for the rules transferred to the dormant state. Summarize the adjudication results of all conflict pairs in this batch and generate a conflict adjudication log. The conflict adjudication log includes at least the adjudication time of each conflict pair, the rule identifiers involved in the adjudication, the scores of each strategy dimension on which the adjudication is based, the adjudication results, and the dormant status records of the suppressed rules. Write the conflict adjudication log into the log storage area of ​​the versioned rule base.

10. The AI-powered intelligent human resource management system for manufacturing as described in claim 1, characterized in that, The process of obtaining the real-time operational state flow matrix, inputting the state matrix along with rule constraints into a multi-objective scheduling optimization network, and outputting an optimal scheduling scheme vector with flexible objectives includes: Real-time operational status data is obtained from the production execution system, attendance system and equipment IoT platform. The operational status data includes at least the quantity of work-in-process and process cycle time of each production line, the operating status and utilization rate of each piece of equipment, and the real-time on-duty status and remaining available working hours of each position. The operational status data is matrix-encoded according to a unified time granularity to construct a real-time operational status flow matrix. Read a set of valid rule snapshots from a versioned rule base. The set of valid rule snapshots includes at least working hour rules, skill matching rules, compliance constraint rules, and cost control rules. Each rule in the effective rule snapshot set is parsed into a constraint tensor. The constraint tensor is then concatenated and fused with the real-time operation state flow matrix along the feature dimension to construct the scheduling optimization input tensor. The input tensor of the scheduling optimization is input into a pre-trained multi-objective scheduling optimization network. The multi-objective scheduling optimization network is based on an encoder and decoder architecture, wherein the encoder performs attention encoding on the input tensor, and the decoder generates multiple sets of candidate scheduling schemes in parallel based on flexible objective weight vectors. The flexible objectives include at least minimizing labor costs, maximizing skill matching degree, and maximizing order delivery timeliness. Multiple candidate scheduling schemes are sorted using a multi-objective non-dominated ranking method to select the Pareto optimal solution set. From the Pareto optimal solution set, the scheme with the highest comprehensive score is selected as the optimal scheduling scheme vector according to the preset flexible objective preference weight, and the optimal scheduling scheme vector is output.

Images