Distributed filling scheduling control method and system based on knowledge graph

Abstract

The invention discloses a distributed filling scheduling control method and system based on a knowledge graph, and the system comprises a knowledge graph construction module which is used for mapping a filling production object into an entity and a relation, and writing the entity and the relation into a graph rule; the edge data acquisition and processing module is used for forming a time sequence observation sequence and synchronously updating the knowledge graph fact layer; the change point detection module is used for executing Bayesian online change point detection to obtain a change point event; the map reasoning verification module is used for generating a feasible action set and an executable parameter set; the scheduling generation module is used for constructing an improved GlowNet model and outputting a scheduling scheme; and the instruction issuing and rescheduling module is used for issuing the instruction to an edge node for execution and triggering incremental rescheduling on an unexecuted part. According to the method, the incremental rescheduling closed-loop control of an executable scheduling scheme and an unexecuted part is realized through Bayesian online change point detection and improved GlowNet distributed coordination.

Description

Technical Field

[0001] This invention relates to the field of industrial automation technology, and in particular to a distributed filling scheduling and control method and system based on knowledge graphs. Background Technology

[0002] Current production scheduling for filling operations typically relies on fixed rules, manual experience, or static planning methods based on a limited number of process parameters. Information on orders, batches, formulas, materials, packaging specifications, production lines and workstations, and equipment capabilities is often scattered across different systems or ledgers, resulting in inconsistent data definitions, missing correlations, and difficulty in forming a unified semantic view of production. In production environments with frequent order and formula changes, the triggering conditions for line changeover cleaning tasks, cleaning steps and durations, compatibility relationships between equipment, formulas, and packaging, resource unavailability intervals caused by maintenance windows, and the simultaneous existence of mutual exclusion and time window constraints for multiple equipment and workstations make it difficult for traditional scheduling methods to uniformly express and consistently verify these constraints using calculable rules. This often leads to problems such as the need for repeated manual verification after plan generation, new conflicts arising after partial adjustments, or omissions in resource conflicts and cleaning insertions. Consequently, the executability of scheduling results is unstable, scheduling quality is difficult to guarantee, and it is difficult to output multiple differentiated plans at the solution level.

[0003] Filling workshops typically operate with multiple production lines and control nodes working collaboratively. Data on equipment operating status, malfunctions, cleaning stages, material arrival status, and quality inspection results exhibit strong temporal and volatile characteristics. Disturbances such as temporary equipment shutdowns, cycle time drift, changes in cleaning time, and material arrival delays can cause the original plan to fail within a short period. Existing systems often employ fixed-cycle recalculation or alarm triggering methods based on simple thresholds, lacking online identification and event-based output of abrupt changes in the time-series observation sequence, making it difficult to accurately determine the timing and scope of rescheduling triggers. Under network jitter or node instability, the consistency between the issuance and execution of scheduling instructions is difficult to guarantee. The lack of versioned instruction set management, trigger queue orchestration, and incremental rescheduling processes for unexecuted parts easily leads to asynchronous issuance of plans and on-site execution, duplicate issuance, or disordered execution order, thus affecting production line continuity, production efficiency, and the stability of the on-site control loop.

[0004] Therefore, how to provide a distributed filling scheduling control method and system based on knowledge graphs is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a distributed filling scheduling and control method and system based on knowledge graphs. This invention comprehensively utilizes knowledge graph semantic modeling and rule reasoning, edge-side Bayesian online change point detection, generator flow network scheduling generation, and edge-center collaborative control technology to form a complete process from filling production element modeling, edge data acquisition and processing, change point event generation, graph feasibility verification, scheduling scheme generation to versioned instruction issuance and incremental rescheduling. In terms of algorithm structure, this invention innovatively introduces Bayesian online change point detection with runtime space truncation update and state transition matrix sparsification, as well as an improved GFlowNet model including reverse trajectory modeling branches, state history embedding memory mechanism, distributed reparameterized path modeling structure, and a final state structure diversity discrimination module. This enables the generation of feasible solutions for multi-constraint filling production scheduling and event-triggered incremental rescheduling under on-site disturbances. Compared with existing technologies, this invention has the advantages of high scheduling feasibility, fast response to real-time fluctuations, strong distributed execution stability, and strong engineering feasibility.

[0006] A distributed filling scheduling control method based on knowledge graphs according to an embodiment of the present invention includes the following steps: S1. Construct a knowledge graph for filling production, map filling production objects to entities and relationships, and write production constraints into graph rules; S2. Collect and process filling production data at edge nodes to form a time-series observation sequence, and update the fact layer of the filling production knowledge graph synchronously. S3. Perform Bayesian online change point detection on the time series observation sequence, introduce a run length space truncation update strategy to retain run length candidates and update the posterior distribution, and obtain the change point event by limiting the set of transition edges based on the sparsity modeling of the state transition matrix. S4. Receive change point events and perform semantic reasoning based on the filling production knowledge graph, call the graph rules to perform feasibility verification, and generate a set of actionable actions and a set of executable parameters. S5. Construct an improved GFlowNet model, introduce a reverse trajectory modeling branch to learn the consistency of the reverse trajectory between the action set and the executable parameter set, encode it through a state history embedding memory mechanism and fuse it with the current state, perform differentiable reparameterization modeling based on the distributed reparameterization path modeling structure, use the final state structure diversity discrimination to discriminate the structural difference degree, and output the scheduling scheme. S6. Issue the scheduling plan to the edge nodes as versioned instructions for execution and read back the execution status, update the fact layer of the filling production knowledge graph, and perform incremental rescheduling for the unexecuted parts.

[0007] Optionally, the production constraints include compatibility constraints, switching and cleaning constraints, maintenance window constraints, and resource usage constraints.

[0008] Optionally, the construction of the filling production knowledge graph includes: Establish an ontology layer for the knowledge graph of filling production, defining entity types, relationship types, and attribute types. Entity types include orders, batches, formulas, materials, maintenance plans, and quality rules. Relationship types include membership relationships, compatibility relationships, sequential relationships, and occupancy relationships. According to the preset mapping rules, the filling production objects are mapped to entity nodes and written into the object attributes. The objects are associated and mapped to relationship edges and written into the relationship attributes. The entity nodes and relationship edges are written into the knowledge graph storage unit. Production constraints are written into the knowledge graph rules to form a rule set, which includes compatibility constraint rules, switchover and cleaning constraint rules, maintenance window constraint rules, and resource consumption constraint rules. The rule set is then written into the knowledge graph storage unit.

[0009] Optionally, the fact layer for forming a time-series observation sequence and synchronously updating the filling production knowledge graph includes: At the edge node, filling production data is acquired through an industrial communication interface. The filling production data includes equipment operating status, fault events, cleaning status, material arrival status and quality inspection results. A collection timestamp and data source identifier are written for each data record. Data processing is performed on the filling production data. Data processing includes missing value handling, outlier removal, data denoising, data alignment, and data aggregation. Data alignment includes resampling data from different sampling periods to a preset sampling period according to a unified time base. Data aggregation includes calculating the mean, variance, maximum value, minimum value, and rate of change within a sliding window. The rate of change is the difference between the values ​​of two adjacent sampling times divided by the value of the previous sampling time. The processed filling production data is combined into a time-series observation sequence in chronological order, and each observation vector in the time-series observation sequence is associated with the corresponding entity node and relation edge in the knowledge graph. The equipment availability status, workstation occupancy status, cleaning stage status, material inventory status, and quality judgment status are written into the fact layer of the knowledge graph.

[0010] Optionally, obtaining the change point event includes: At the edge node, a set of Bayesian online change point detection instances are established for each critical device. The time-series observation sequence is received, and the observation likelihood is calculated for the observation value at each sampling time. The observation probability under different running lengths is calculated. A spatial truncation update strategy is implemented for the posterior distribution of running length. At each sampling time, the posterior probability of running length is calculated and sorted from high to low according to the posterior probability. A preset number of running length candidates are retained and the posterior distribution is updated only for the retained running length candidates. The posterior distribution and observation state corresponding to the running length candidates that are not retained are discarded. The state transition matrix is ​​sparsified during the runtime transition process. The runtime transition matrix is ​​constructed and the transition probabilities of all transitions except for two types are set to zero. The first type of transition is that the runtime at the current sampling time is equal to the runtime at the previous sampling time plus one, and the second type of transition is that the runtime at the current sampling time is equal to zero. Based on the candidate runtime and the runtime transition matrix, the posterior probability that the runtime is equal to zero at each sampling time is calculated as the change point probability. When the change point probability is greater than the threshold, a change point event is generated.

[0011] Optionally, the generation of the set of actionable actions and the set of executable parameters includes: The central node receives change events and writes them into the scheduling event table. The change event includes the change time, associated production line identifier, associated equipment identifier, and associated observation indicator identifier. Using the fact layer state of the knowledge graph corresponding to the moment of change as the input for reasoning, read the order batch status, formula and packaging specification status, material arrival status, equipment availability status and maintenance window status; Call the graph rules to perform semantic reasoning and complete compatibility verification, switch cleaning verification, maintenance window verification and resource usage conflict verification. Output the set of actionable actions and the set of executable parameters that pass the verification.

[0012] Optionally, the output scheduling scheme includes: An improved GFlowNet model is constructed, including a reverse trajectory modeling branch, a state history embedding memory module, a distributed reparameterized path modeling structure, and a final state structure diversity discrimination module; A state and action space is constructed based on the set of actionable actions and the set of executable parameters. The state includes order pool, resource occupancy and time window information. The actions include batch selection, equipment allocation, workstation allocation and cleaning insertion. A historical state buffer is established by the state history embedding memory module, a preset number of historical state vectors are stored and encoded to obtain historical embedding vectors. The historical embedding vectors are fused with the current state vector to form an enhanced state vector. The enhanced state vector is input into the forward generation unit, which outputs the action distribution parameters. The distribution reparameterization path modeling structure represents the action sampling as a differentiable sampling process jointly determined by random noise variables and action distribution parameters. The random noise variables are sampled from the probability distribution. Based on the differentiable sampling process, action samples are generated and scheduling trajectories are iteratively formed. When the scheduling trajectory reaches the termination condition, a preliminary scheduling scheme is obtained. The reverse trajectory modeling branch takes the scheduling scheme as input to generate a reverse action sequence and outputs the reverse action probability. It uses the consistency between the probability of the forward action sequence and the probability of the reverse action sequence as the training objective to complete the reverse trajectory consistency learning. The final state structure diversity discrimination module extracts the resource occupancy sequence, line switching and cleaning sequence and workstation allocation sequence of the scheduling scheme to form a structural representation, calculates the structural difference degree between the preliminary scheduling schemes and filters and outputs the scheduling scheme.

[0013] Optionally, the step of issuing the scheduling scheme as versioned instructions to edge nodes for execution and reading back the execution status includes: At the central node, the scheduling scheme is converted into a versioned execution instruction set. The versioned execution instruction set includes version number, start and end time of effectiveness, resource allocation, line switching and cleaning task sequence and maintenance avoidance time window, and writes target production line identifier, target equipment identifier and target workstation identifier for each instruction. Versioned execution instruction sets are sent to edge nodes. Edge nodes verify the version number and effective time window, parse the work sequence and generate a trigger queue in chronological order, map the instructions in the trigger queue into control instructions that the production line control unit can recognize and send them out for execution. Control instructions include equipment start / stop instructions, cycle time setting instructions, workstation switching instructions and cleaning start instructions. During the execution process at the edge node, a change point event is received and the starting point of incremental rescheduling is determined. The starting point of incremental rescheduling is the start time corresponding to the first unexecuted instruction in the trigger queue. A rescheduling request is sent to the central node, carrying the sequence of executed jobs, the current resource occupancy status, and the set of unexecuted jobs. A scheduling scheme for the unexecuted part is generated and converted into a new versioned execution instruction set and sent to the edge node.

[0014] On the other hand, a knowledge graph-based distributed filling scheduling control system according to an embodiment of the present invention includes the following modules: The knowledge graph construction module is used to map filling production objects into entities and relationships, and to write production constraints into graph rules; The edge data acquisition and processing module is used to collect and process filling production data to form a time-series observation sequence and synchronously update the fact layer of the knowledge graph; The change point detection module is used to perform Bayesian online change point detection on time-series observation sequences to obtain change point events; The graph reasoning and verification module is used to receive change point events and perform semantic reasoning and feasibility verification based on graph rules to generate a set of feasible actions and a set of executable parameters. The scheduling generation module is used to build an improved GFlowNet model, learn and sample the trajectory of the set of actionable actions and the set of executable parameters, and output a scheduling scheme. The instruction issuance and rescheduling module is used to convert the scheduling scheme into versioned instructions, issue them to edge nodes for execution, and trigger incremental rescheduling for the unexecuted parts.

[0015] The beneficial effects of this invention are:

[0016] This invention maps filling production objects and constraints into a unified knowledge graph of entities, relationships, and rules. After a change event is triggered, semantic reasoning and feasibility verification are performed based on the graph rules. This enables the generation of a set of actionable actions and executable parameters consistent with the actual situation, even under conditions of frequent order and formula changes, complex line changeover cleaning tasks, diverse equipment compatibility, dynamic changes in maintenance windows, and mutually exclusive resource usage. This reduces scheduling conflicts, omissions, and rework caused by repeated manual checks from the source. Simultaneously, Bayesian online change point detection, employing runtime space truncation updates and state transition matrix sparsity, identifies online mutations and generates change events based on the temporal observations of equipment status, fault events, cleaning stages, material arrival, and quality inspection results. This improves the real-time performance and stability of disturbance detection while reducing the computational and storage overhead of parallel detection by multiple devices on the edge, thereby enhancing the accuracy and timeliness of on-site triggered rescheduling.

[0017] This invention constructs an improved GFlowNet model that includes a reverse trajectory modeling branch, a state history embedding memory mechanism, a distributed reparameterized path modeling structure, and a final state structure diversity discrimination module. This model enables trajectory learning and sampling generation of the action set and the executable parameter set, outputting scheduling schemes and alternative scheme sets with structural differences. It maintains the stability and diversity of scheme generation in a multi-constraint discrete decision space, improving scheme availability and on-site switching options. The scheduling scheme is converted into a versioned execution instruction set containing a version number and effective time window and distributed to edge nodes. Edge nodes parse and output control instructions for equipment start / stop, cycle time setting, workstation switching, and cleaning initiation according to the trigger queue. When a change point event occurs, an incremental rescheduling request is initiated based on the unexecuted portion, achieving closed-loop collaboration between the center and the edge. Compared to existing technologies, this invention improves the consistency between plan issuance and on-site execution, continuous production capacity and recovery efficiency under disturbances, reduces the overall losses caused by downtime and line changeover cleaning, and has stronger engineering implementation and stable operation capabilities. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a structural block diagram of a knowledge graph-based distributed filling scheduling control method proposed in this invention. Figure 2 This is a functional diagram of the improved GFlowNet model for a knowledge graph-based distributed filling scheduling control method proposed in this invention. Figure 3This is a flowchart of a distributed filling scheduling control method and system based on knowledge graphs proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figure 1 and Figure 2 A distributed filling scheduling control method based on knowledge graphs includes the following steps: S1. Construct a knowledge graph for filling production, map filling production objects to entities and relationships, and write production constraints into graph rules; S2. Collect and process filling production data at edge nodes to form a time-series observation sequence, and update the fact layer of the filling production knowledge graph synchronously. S3. Perform Bayesian online change point detection on the time series observation sequence, introduce a run length space truncation update strategy to retain run length candidates and update the posterior distribution, and obtain the change point event by limiting the set of transition edges based on the sparsity modeling of the state transition matrix. S4. Receive change point events and perform semantic reasoning based on the filling production knowledge graph, call the graph rules to perform feasibility verification, and generate a set of actionable actions and a set of executable parameters. S5. Construct an improved GFlowNet model, introduce a reverse trajectory modeling branch to learn the consistency of the reverse trajectory between the action set and the executable parameter set, encode it through a state history embedding memory mechanism and fuse it with the current state, perform differentiable reparameterization modeling based on the distributed reparameterization path modeling structure, use the final state structure diversity discrimination to discriminate the structural difference degree, and output the scheduling scheme. S6. Issue the scheduling plan to the edge nodes as versioned instructions for execution and read back the execution status, update the fact layer of the filling production knowledge graph, and perform incremental rescheduling for the unexecuted parts.

[0021] In this embodiment, the production constraints include compatibility constraints, switching and cleaning constraints, maintenance window constraints, and resource usage constraints.

[0022] In this embodiment, the construction of the filling production knowledge graph includes: Establish an ontology layer for the filling production knowledge graph, defining entity types, relationship types, and attribute types. Entity types include orders, batches, formulas, materials, maintenance plans, and quality rules. Relationship types include membership, compatibility, precedence, and occupancy relationships. Specifically, establishing the ontology layer for the filling production knowledge graph involves: Define the semantic boundaries and naming conventions of the filling scheduling control scenario, provide a set of entity types and define a unique identifier, core attribute fields and primary key constraints for each type of entity, define a set of relationship types, clarify the starting entity type, ending entity type, directionality, repeatability and required attributes for each type of relationship, establish a set of attribute types, specify the data type, value range, unit and update timestamp fields of the attributes, configure inheritance and reference rules for entities and relationships, solidify entities, relationships and attributes into a graph schema layer using ontology description files or schema definition tables, and form version numbers and change records to support expansion and consistency maintenance; According to a preset mapping rule, filling production objects are mapped to entity nodes and written into object attributes; object associations are mapped to relationship edges and written into relationship attributes; entity nodes and relationship edges are written into the knowledge graph storage unit. The preset mapping rule is specifically as follows: Preset mapping rules refer to a set of deterministic rules used to convert filling production objects and their relationships into a knowledge graph data structure. These rules include entity mapping rules and relationship mapping rules. Entity mapping rules specify the entity type, unique identifier generation method, primary key field source, field lookup table from attribute fields to graph attributes, data type and unit conversion method, enumeration value normalization method, and missing field filling method for different business objects. Relationship mapping rules specify the relationship type, starting and ending point matching key, relationship direction, unique key generation method for relationship edges, relationship attribute field lookup table, and valid time interval field for the relationship between objects. Ultimately, this ensures that the same object is mapped to the same entity node under different data sources while maintaining consistent relationship edge connections. Production constraints are written into the knowledge graph rules to form a rule set. The rule set includes compatibility constraint rules, switchover and cleanup constraint rules, maintenance window constraint rules, and resource consumption constraint rules. The rule set is then written into the knowledge graph storage unit. Specifically, the rule set is as follows: A rule set refers to an executable set of constraint rules fixed in a knowledge graph in the form of preconditions to conclusions. It is used to generate feasible resource allocation and job arrangement results during reasoning and verification. Compatibility constraint rules are used to define the matching relationship between formulas, materials and equipment and the prohibition of allocation when they are mismatched. Switching and cleaning constraint rules are used to define that a cleaning task must be inserted when adjacent batches switch formulas or materials and to determine the cleaning step sequence and duration. Maintenance window constraint rules are used to define that equipment cannot be allocated or occupied during the interval from the start time of maintenance to the end time of maintenance. Resource occupation constraint rules are used to define the start time and end time of occupation of equipment by the job, the mutual exclusion relationship of occupation, and the occupation conflict judgment and conflict resolution actions.

[0023] In this embodiment, the fact layer of forming a time-series observation sequence and synchronously updating the filling production knowledge graph includes: At the edge node, filling production data is acquired through an industrial communication interface. The filling production data includes equipment operating status, fault events, cleaning status, material arrival status and quality inspection results. A collection timestamp and data source identifier are written for each data record. Data processing is performed on the filling production data. Data processing includes missing value handling, outlier removal, data denoising, data alignment, and data aggregation. Data alignment includes resampling data from different sampling periods to a preset sampling period according to a unified time base. Data aggregation includes calculating the mean, variance, maximum value, minimum value, and rate of change within a sliding window. The rate of change is the difference between the values ​​of two adjacent sampling times divided by the value of the previous sampling time. The processed filling production data is combined into a time-series observation sequence in chronological order, and each observation vector in the time-series observation sequence is associated with the corresponding entity node and relation edge in the knowledge graph. The equipment availability status, workstation occupancy status, cleaning stage status, material inventory status, and quality judgment status are written into the fact layer of the knowledge graph.

[0024] In this embodiment, obtaining the change point event includes: At the edge nodes, a set of Bayesian online change point detection instances is established for each critical device. These instances receive time-series observation sequences and calculate the observation likelihood for each sampling time, as well as the observation probability for different runtime lengths. The establishment of a set of Bayesian online change point detection instances is specifically as follows: a set of change point detection instances is created at the edge node using the key device identifier as an index. Each instance is bound to a set of observation indicators and sampling period. The posterior distribution of the running length is initialized so that the probability of the running length being equal to zero is one, and the rest are zero. The instance is configured with the observation model parameter structure, prior parameter structure, running length candidate buffer area and transition matrix storage area. When the instance runs, it receives the observation vector according to the sampling time, maintains the correspondence table of running length candidate, posterior probability and observation unit status, and generates or releases the corresponding instance according to the same initialization rule when the device indicator is added or deleted. The method of calculating the observation likelihood for each sampling time and the observation probability under different running lengths is as follows: at each sampling time, the predicted distribution parameters are first obtained for the observation vector based on the observation unit, and then the observation value at the time is substituted into the predicted distribution to calculate the observation likelihood value. The observation likelihood value is defined as the probability density or probability mass of the observation value under the predicted distribution. For each running length candidate, the predicted distribution parameters are generated using the observation model state corresponding to the candidate and the observation likelihood is calculated to obtain the observation probability of the observation value when the running length is equal to the candidate. The observation probability is then multiplied by the posterior probability of the running length at the previous sampling time to form an unnormalized joint probability for the next posterior update. A runtime spatial truncation update strategy is implemented on the runtime posterior distribution. At each sampling time, the posterior probability of the runtime is calculated and sorted from high to low. A preset number of runtime candidates are retained, and the posterior distribution is updated only for the retained runtime candidates. The posterior distributions and observation states corresponding to the unretained runtime candidates are discarded. Specifically, retaining a preset number of runtime candidates and updating the posterior distribution only for the retained runtime candidates involves: At each sampling time, the unnormalized posterior probability is calculated for all candidates for run length. The unnormalized posterior probability is then divided by the sum of the unnormalized posterior probabilities of all candidates to obtain the normalized posterior probability. The candidates are sorted from high to low according to their normalized posterior probabilities. A set of 32 candidates for run length is selected as the retention set. The posterior probabilities of candidates outside the retention set are set to zero and the corresponding observation unit states are released. Posterior recursion is performed only for candidates within the retention set. During the recursion, the unnormalized posterior probability of the next sampling time is calculated based on the observation probability and the run length transition probability. The probability is then normalized again to obtain the updated posterior distribution. At the same time, the updated candidate set, posterior probabilities, and observation unit states are written into the run length candidate buffer. The state transition matrix is ​​sparsified during the runtime transition process. A runtime transition matrix is ​​constructed, and the transition probabilities of all types of transitions except for two types are set to zero. The first type of transition is when the runtime at the current sampling time equals the runtime at the previous sampling time plus one; the second type of transition is when the runtime at the current sampling time equals zero. Specifically, the state transition matrix sparsification modeling is performed as follows: A running length transition matrix is ​​constructed using the set of running lengths at the previous sampling time as the row index and the set of running lengths at the current sampling time as the column index. The default transition probability of the matrix is ​​set to zero. Then, only two types of transitions are allowed to write non-zero transition probabilities: the first type is a continuation transition from any previous running length to a current running length equal to the previous running length plus one, and the second type is a reset transition from any previous running length to a current running length equal to zero. All matrix elements except for these two types of positions are kept at zero, and the non-zero elements of each row are normalized row by row so that the sum of the transition probabilities of each row is equal to one. Finally, the sparse transition matrix is ​​written into the transition matrix storage area of ​​the change point detection instance for posterior recursion. Based on the candidate runtime and the runtime transition matrix, the posterior probability that the runtime is equal to zero at each sampling time is calculated as the change point probability. When the change point probability is greater than a threshold, a change point event is generated, and the threshold is set to 0.9.

[0025] In this embodiment, generating the set of actionable actions and the set of executable parameters includes: The central node receives change events and writes them into the scheduling event table. The change event includes the change time, associated production line identifier, associated equipment identifier, and associated observation indicator identifier. Using the fact layer state of the knowledge graph corresponding to the moment of change as the input for reasoning, read the order batch status, formula and packaging specification status, material arrival status, equipment availability status and maintenance window status; The system invokes graph rules to perform semantic reasoning and completes compatibility checks, switch cleaning checks, window maintenance checks, and resource usage conflict checks. It outputs a set of valid actions and a set of executable parameters. Specifically, the semantic reasoning process involves: During semantic reasoning, the central node uses the fact layer state corresponding to the change point as the reasoning input. It loads the order batch state, formula and material state, equipment availability state, maintenance window state, cleaning unit state, and resource occupancy state into fact triples. The fact triples are matched according to the preconditions in the rule set. After a successful match, the rule is instantiated to generate a conclusion triple and written to the reasoning working memory. The next round of rule matching is triggered for the newly generated conclusion triples until no new conclusion triples are generated. Candidate actions that satisfy the compatibility relationship, cleaning insertion condition, maintenance availability interval, and occupancy mutual exclusion relationship are extracted from the reasoning working memory. The cycle time parameter, operation duration parameter, cleaning duration parameter, and maintenance avoidance time window parameter associated with the candidate actions are organized into an executable parameter set for output.

[0026] In this embodiment, the output scheduling scheme includes: An improved GFlowNet model is constructed, including a reverse trajectory modeling branch, a state history embedding memory module, a distributed reparameterized path modeling structure, and a final state structure diversity discrimination module. Specifically, the construction of the improved GFlowNet model involves: The state history embedding memory module is embedded into the state encoding entry position of the original GFlowNet. The single-step state representation consisting only of the current state is replaced by the historical state buffer and the historical embedding encoding result. The fused enhanced state vector is used as the unified input of the forward generation unit. The distributed reparameterized path modeling structure is connected at the action sampling position of the forward generation unit. The differentiable sampling mapping determined by random noise variables and action distribution parameters is used to replace the traditional discrete sampling. Differentiable action samples are output during the trajectory construction process and the scheduling trajectory is iteratively formed. After the scheduling scheme is obtained when the scheduling trajectory reaches the termination condition, the reverse trajectory modeling branch is bypassed. The reverse starting state is constructed with the scheduling scheme and the reverse action sequence and reverse action probability are generated. The consistency between the forward action sequence probability and the reverse action sequence probability is used as the training target and fed back to the forward and reverse branches to complete the reverse trajectory consistency learning. A final state structure diversity discrimination module is added to the output end of the scheduling scheme. The structural representation is constructed with the resource occupation sequence, line switching cleaning sequence and workstation allocation sequence and the structural difference degree between the schemes is calculated. The difference degree is used to filter and output a set of scheduling schemes with structural differences. A state and action space is constructed based on the set of actionable actions and the set of executable parameters. The state includes order pool, resource occupancy, and time window information. Actions include batch selection, equipment allocation, workstation allocation, and cleaning insertion. A state history embedding memory module establishes a historical state buffer, stores a preset number of historical state vectors, and encodes them into historical embedding vectors. These historical embedding vectors are then fused with the current state vector to form an enhanced state vector, where: The construction state and action space are specifically as follows: The order pool, batch set, available equipment and workstation set, cleaning unit status, maintenance window, and resource occupation time window output by knowledge graph reasoning are organized into a set of status fields and encoded into a current status vector in a unified order. The current status vector includes the number of batches to be scheduled, the formula and material identifier of each batch, the start and end time of occupation of each resource, the start and end time of availability, and the cleaning stage identifier. The set of actionable actions is divided into batch selection actions, equipment allocation actions, workstation allocation actions, and cleaning insertion actions according to action type. Action parameter fields and action output fields are defined for each type of action. The action parameter fields include target batch identifier, target equipment identifier, target workstation identifier, target time window, and cleaning task parameters. The action output fields include resource usage update items and time window update items. Using state vectors and action types and parameters as state transition inputs, a transition mapping from the current state to the next state is established, thereby forming a state space and action space for the progressive expansion of the generation flow network. The state history embedding memory module is specifically composed of a historical state buffer, a historical state vector, a preset number, a historical embedding encoder, and a fusion operator. The historical state buffer is a first-in-first-out storage structure with fixed boundaries, used to store the state history obtained from the most recent state updates in chronological order. The historical state vector is a unified encoding result of the state field at each moment, including order pool encoding, resource occupancy encoding, and time window encoding, and is marked with a moment identifier. The preset number is the number of historical state vectors that the buffer can hold, which is fixed at eight. The historical embedding encoder maps the historical state vectors in the buffer to a single historical embedding vector, used to extract the sequence pattern and change trend in the historical state. The fusion operator concatenates the historical embedding vector with the current state vector and then performs a linear transformation to obtain an enhanced state vector, so that the forward generation unit can use the current state and recent historical state information simultaneously when outputting action distribution parameters. The enhanced state vector is input into the forward generation unit, which outputs action distribution parameters. The distribution-reparameterized path modeling structure represents action sampling as a differentiable sampling process jointly determined by random noise variables and action distribution parameters. The random noise variables are sampled from a probability distribution. Action samples are generated based on the differentiable sampling process, and a scheduling trajectory is iteratively formed. When the scheduling trajectory reaches the termination condition, a preliminary scheduling scheme is obtained, where: The output action distribution parameters are as follows: the forward generation unit takes the enhanced state vector as input, first obtains the state representation vector through feature transformation, and then inputs the state representation vector into the action type head and the action parameter head respectively. The action type head outputs the unnormalized score of each action type and performs normalization on the unnormalized score to obtain the action type probability vector. The action parameter head outputs the corresponding parameter distribution descriptor for the defined action parameter field. The parameter distribution descriptor includes the category probability vector used to characterize discrete parameters and the distribution location parameter and distribution scale parameter used to characterize continuous parameters, forming a set of action distribution parameters for action sampling. The differentiable sampling process is as follows: first, a random noise variable is generated from the preset probability distribution, which is a standard normal distribution; then, the random noise variable and the action distribution parameters are input into the sampling mapping function to obtain action samples. In the forward calculation, the action samples are used as discrete actions or parameter values ​​to participate in the construction of the scheduling trajectory. In the backward calculation, the gradient is passed to the action distribution parameters through the sampling mapping function. The reverse trajectory modeling branch generates a reverse action sequence and outputs the reverse action probability using the scheduling scheme as input. It then uses the consistency between the probabilities of the forward and reverse action sequences as the training objective to complete the reverse trajectory consistency learning. The reverse trajectory modeling branch specifically consists of a reverse starting state construction, a reverse action set, a reverse sampling process, and consistency learning. The reverse starting state construction is used to parse the scheduling scheme into a termination state vector as the starting point for reverse generation. The reverse action set is used to map forward actions to cancel actions, including canceling batch selection, canceling equipment allocation, canceling workstation allocation, and canceling cleaning insertion. The reverse sampling process samples cancel actions step by step according to conditional probability and updates the reverse state until the preset starting state condition is met. The preset starting state condition is that the order pool is not allocated and the resource occupancy is empty. Consistency learning is used to calculate the difference between the probability of the forward action sequence and the probability of the reverse action sequence and update the forward and reverse action parameters. The process of generating a reverse action sequence and outputting the reverse action probability is as follows: a set of reverse actions in the opposite direction to the forward action is constructed. The batch allocation cancellation, equipment allocation cancellation, workstation allocation cancellation, and cleaning insertion cancellation are defined as reverse action types. The termination state vector is input into the reverse generation unit to output the reverse action distribution parameters. At each step, the reverse generation unit samples a reverse action from the reverse action distribution parameters and updates the reverse state. This process is repeated until the initial state is reached, resulting in a reverse action sequence arranged in reverse time. At the same time, the probability value of each reverse action in the reverse action distribution is recorded as the reverse action probability and multiplied together to obtain the reverse action sequence probability. The completion of reverse trajectory consistency learning specifically involves: Using the forward action sequence and reverse action sequence corresponding to the same scheduling scheme as training samples, the probabilities of the forward action sequence and the reverse action sequence are calculated respectively. The probability of the forward action sequence is the product value obtained by multiplying the conditional probabilities of the selected forward action given by the forward generation unit at each step along the trajectory. The probability of the reverse action sequence is the product value obtained by multiplying the conditional probabilities of the selected reverse action given by the reverse trajectory modeling branch at each step along the reverse trajectory. A consistency loss is constructed to ensure that the probabilities of the forward action sequence and the reverse action sequence are consistent with the probability expressions of the same trajectory in the forward and reverse directions. The consistency loss is defined as the square of the difference between the logarithm of the forward action sequence probability and the logarithm of the reverse action sequence probability. Gradient backpropagation is performed on the consistency loss to update the parameters of the forward generation unit and the reverse generation unit until the consistency loss converges. The final-state structural diversity discrimination module extracts resource occupancy sequences, line-changing cleaning sequences, and workstation allocation sequences from the scheduling schemes to form structural representations. It then calculates the structural difference degree between the preliminary scheduling schemes and filters and outputs the final scheduling schemes. The final state structure diversity discrimination module is specifically composed of a sequence extractor, a structure representation unit, a difference calculator, and a filter. The sequence extractor extracts three types of sequences from each preliminary scheduling scheme: the resource occupation sequence records the start and end times and the occupied objects of each resource occupation; the line switching and cleaning sequence records the cleaning insertion position, cleaning stage, and duration; and the workstation allocation sequence records the workstation assignment corresponding to each batch. The structure representation unit encodes the three types of sequences in a unified order and concatenates them into a structure representation vector. The difference calculator calculates the normalized structure difference degree for any two structure representation vectors and forms a difference degree matrix. The filter performs deduplication selection based on the difference degree threshold and the number of outputs. The calculation of the structural differences between preliminary scheduling schemes and the selection of output scheduling schemes are specifically as follows: The resource occupancy sequence, line switching and cleaning sequence, and workstation allocation sequence of each preliminary scheduling scheme are encoded into three types of sequence features, and then concatenated in a unified order to form a structure representation vector. The difference degree is calculated for any two structure representation vectors. The difference degree is defined as the sum of the differences of the three types of sequence features. The resource occupancy difference is defined as the complement of the ratio of the intersection length to the union length of the occupancy time windows of the two schemes on the same resource. The line switching and cleaning difference is defined as the edit distance of the line switching and cleaning event sequences of the two schemes divided by the larger of the two sequence lengths. The workstation allocation difference is defined as the number of inconsistent workstation assignments of the two schemes in the same batch divided by the number of batches. The difference degree is used as the distance metric between the two schemes to construct a difference degree matrix. The schemes with the difference degree greater than the preset threshold of 0.4 are selected and added to the output set in turn until the output set reaches the preset number of five or the candidate scheme traversal is completed, thus obtaining the scheduling scheme and the set of alternative scheduling schemes.

[0027] In this embodiment, the step of issuing the scheduling scheme to the edge nodes for execution as versioned instructions and reading back the execution status includes: At the central node, the scheduling scheme is converted into a versioned execution instruction set. This versioned instruction set includes a version number, effective start and end times, resource allocation, line switching and cleaning task sequence, and maintenance avoidance time window. Each instruction is also written with a target production line identifier, target equipment identifier, and target workstation identifier. Specifically, the conversion of the scheduling scheme into a versioned execution instruction set involves: The central node first sorts the job sequences in the scheduling plan according to the planned start time, determines the version number and writes the effective start time and effective end time. Each job is broken down into several instruction records. The instruction records are generated according to the type of batch execution instruction, equipment allocation instruction, workstation allocation instruction, line change cleaning instruction, and maintenance avoidance instruction. The target production line identifier, target equipment identifier, and target workstation identifier are written in each instruction, and the planned start time, planned end time, and resource occupation time window corresponding to the instruction are also written. Resource allocation is directly mapped to instruction parameters by the equipment and workstation assignment fields in the scheduling scheme. The line change cleaning task sequence is mapped to cleaning start instructions and cleaning phase instructions by the cleaning insertion position and cleaning duration fields in the scheduling scheme. The maintenance avoidance time window is mapped to avoidance start and end times and prohibited allocation flags by the maintenance disabled interval field in the scheduling scheme. An instruction sequence number and a unique instruction identifier are generated for each instruction. All instructions are written into the versioned execution instruction set under the same version number according to the sequence number. Versioned execution instruction sets are sent to edge nodes. Edge nodes verify the version number and effective time window, parse the work sequence and generate a trigger queue in chronological order, map the instructions in the trigger queue into control instructions that the production line control unit can recognize and send them out for execution. Control instructions include equipment start / stop instructions, cycle time setting instructions, workstation switching instructions and cleaning start instructions. During the execution process at the edge node, a change point event is received and the starting point of incremental rescheduling is determined. The starting point of incremental rescheduling is the start time corresponding to the first unexecuted instruction in the trigger queue. A rescheduling request is sent to the central node, carrying the sequence of executed jobs, the current resource occupancy status, and the set of unexecuted jobs. A scheduling scheme for the unexecuted part is generated and converted into a new versioned execution instruction set and sent to the edge node.

[0028] refer to Figure 3 A distributed filling scheduling control method and system based on knowledge graphs, comprising the following modules: The knowledge graph construction module is used to map filling production objects into entities and relationships, and to write production constraints into graph rules; The edge data acquisition and processing module is used to collect and process filling production data to form a time-series observation sequence and synchronously update the fact layer of the knowledge graph; The change point detection module is used to perform Bayesian online change point detection on time-series observation sequences to obtain change point events; The graph reasoning and verification module is used to receive change point events and perform semantic reasoning and feasibility verification based on graph rules to generate a set of feasible actions and a set of executable parameters. The scheduling generation module is used to build an improved GFlowNet model, learn and sample the trajectory of the set of actionable actions and the set of executable parameters, and output a scheduling scheme. The instruction issuance and rescheduling module is used to convert the scheduling scheme into versioned instructions, issue them to edge nodes for execution, and trigger incremental rescheduling for the unexecuted parts. Example

[0029] To verify the feasibility of this invention in practice, it was applied to the bottling workshop of a beverage company. This workshop employs a multi-variety, small-batch, multi-batch production model, with multiple parallel bottling lines and edge control nodes. Frequent order insertions, rescheduling, and formula switching occur daily. The main pain points on-site fall into two categories: First, due to the overlapping constraints of line changeover cleaning, equipment compatibility, maintenance windows, and resource usage mutual exclusion, scheduling personnel need to repeatedly verify across multiple systems. Even after the plan is released, workstation conflicts, cleaning omissions, or maintenance window collisions still occur, leading to repeated rework. Second, when equipment downtime, cleaning time drift, or material arrival delays occur, the original plan quickly becomes invalid. Furthermore, under network fluctuations, multiple nodes are prone to duplicate instruction issuance, disordered sequence, or temporary changes on the execution side, making it difficult to form a stable closed loop.

[0030] After deploying the system of this invention in the workshop, a knowledge graph ontology and mapping rules are first established based on production factors. Four types of constraints—compatibility, switching and cleaning, maintenance window, and resource occupancy mutual exclusion—are solidified into graph rules and written into the graph storage. Edge nodes continuously collect equipment status, fault events, cleaning stages, material arrival, and quality inspection results through industrial communication interfaces. After alignment resampling and sliding window aggregation, a time-series observation sequence is formed and written into the graph fact layer. The edge side performs Bayesian online change point detection on key equipment, outputting change point events with candidate truncation updates based on runtime length and sparse transition matrices. After receiving change point events, the central side performs graph semantic reasoning to obtain a set of actionable actions and a set of executable parameters. Then, multiple scheduling schemes with significant structural differences are generated by the improved GFlowNet, which are converted into execution instruction sets with version numbers and effective time windows and sent to the edge nodes. The edge nodes map the instructions into control instructions such as start / stop, cycle time setting, workstation switching, and cleaning start according to the trigger queue. When a change point event occurs again, incremental rescheduling is triggered only for the unexecuted part, and a new version instruction set is sent to achieve smooth continuation.

[0031] Statistics from two consecutive production cycles of the system show that the first version of the knowledge graph has solidified over 300 compatible relationships, over 40 switching and cleaning rules, over 20 maintenance window rules, and over 60 resource mutual exclusion rules. Edge-side change point detection runs in parallel across multiple nodes, with the average processing latency per node remaining within seconds. It outputs over 20 change point events daily and can reliably trigger incremental rescheduling. Compared to before go-live, the average daily time spent by scheduling personnel to check conflicts and modify plans has decreased from about two hours to about 20 minutes. The number of manual rollbacks after plan release has decreased from over ten times per week to one to two times per week. On-site confirmation communication caused by duplicate instruction issuance has decreased from over 20 times per day to less than five times per day. Even with multiple short-term network jitters, edge nodes can still maintain consistency in the trigger queue based on version number and effective time window, with no duplicate batch execution or missed cleaning tasks, significantly improving production organization stability.

[0032] Table 1 Comparison of Key Indicators of Filling Scheduling Control Scheme

[0033] As shown in Table 1, in terms of planning-level indicators, the system of this invention significantly outperforms the comparative methods in both planning feasibility and on-time delivery rate. The planning feasibility rate reaches 98.9%, an increase of 13.2 percentage points compared to manual scheduling (85.7%), and also higher than MES rules (90.6%), heuristic scheduling (92.8%), reinforcement learning (93.6%), and generative models (94.5%). The on-time delivery rate reaches 96.4%, an increase of 8.5 percentage points compared to manual scheduling (87.9%), and maintains a leading position among similar automated methods. This demonstrates that semantic modeling and reasoning verification based on knowledge graphs can reduce unexecutable plans under multiple constraints and improve the ability to complete plans on schedule after implementation.

[0034] From the perspective of production changeover and efficiency loss indicators, the system of this invention performs better in terms of the number of line changeover cleanings and downtime. The number of line changeover cleanings is 11.8 times / day, which is 8.5 times / day less than manual scheduling (20.3 times / day), 6.3 times / day less than MES rules (18.1 times / day), and also lower than heuristic scheduling (16.4 times / day), reinforcement learning (15.7 times / day), and generative models (14.8 times / day). The downtime is 34 minutes / day, which is 69 minutes / day less than manual scheduling (103 minutes / day), 52 minutes / day less than MES rules (86 minutes / day), and better than the comparison methods. This indicates that the scheduling scheme generated by this invention is more compact in terms of resource allocation, cleaning insertion, and maintenance avoidance, and can reduce time losses caused by waiting for resources and unreasonable changeovers.

[0035] From the perspective of disturbance response and execution metrics, the system of this invention demonstrates stronger real-time closed-loop performance and engineering executability. The disturbance response latency is 2.0s, which is orders of magnitude higher than manual scheduling (990s), MES rules (750s), heuristic scheduling (372s), reinforcement learning (234s), and generative models (186s). Simultaneously, the instruction execution success rate reaches 99.5%, higher than manual scheduling (94.2%) and various algorithm solutions, indicating that versioned instruction issuance and edge execution orchestration are more stable. The OEE improvement is 9.1%, significantly higher than MES rules (2.4%), heuristic scheduling (3.7%), reinforcement learning (5.1%), and generative models (5.6%). Combined with a planned version switching frequency of 3.3 times / day, this shows that the system can maintain execution stability while maintaining a high response frequency, achieving efficiency improvement rather than frequent disturbances.

[0036] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A distributed filling scheduling control method based on knowledge graphs, characterized in that, Includes the following steps: S1. Construct a knowledge graph for filling production, map filling production objects to entities and relationships, and write production constraints into graph rules; S2. Collect and process filling production data at edge nodes to form a time-series observation sequence, and update the fact layer of the filling production knowledge graph synchronously. S3. Perform Bayesian online change point detection on the time series observation sequence, introduce a run length space truncation update strategy to retain run length candidates and update the posterior distribution, and obtain the change point event by limiting the set of transition edges based on the sparsity modeling of the state transition matrix. S4. Receive change point events and perform semantic reasoning based on the filling production knowledge graph, call the graph rules to perform feasibility verification, and generate a set of actionable actions and a set of executable parameters. S5. Construct an improved GFlowNet model, introduce a reverse trajectory modeling branch to learn the consistency of the reverse trajectory between the action set and the executable parameter set, encode it through a state history embedding memory mechanism and fuse it with the current state, perform differentiable reparameterization modeling based on the distributed reparameterization path modeling structure, use the final state structure diversity discrimination to discriminate the structural difference degree, and output the scheduling scheme. S6. Issue the scheduling plan to the edge nodes as versioned instructions for execution and read back the execution status, update the fact layer of the filling production knowledge graph, and perform incremental rescheduling for the unexecuted parts.

2. The distributed filling scheduling control method based on knowledge graphs according to claim 1, characterized in that, The production constraints include compatibility constraints, switching and cleaning constraints, maintenance window constraints, and resource usage constraints.

3. The distributed filling scheduling control method based on knowledge graphs according to claim 1, characterized in that, The construction of the knowledge graph for filling production includes: Establish an ontology layer for the knowledge graph of filling production, defining entity types, relationship types, and attribute types. Entity types include orders, batches, formulas, materials, maintenance plans, and quality rules. Relationship types include membership relationships, compatibility relationships, sequential relationships, and occupancy relationships. According to the preset mapping rules, the filling production objects are mapped to entity nodes and written into the object attributes. The objects are associated and mapped to relationship edges and written into the relationship attributes. The entity nodes and relationship edges are written into the knowledge graph storage unit. Production constraints are written into the knowledge graph rules to form a rule set, which includes compatibility constraint rules, switchover and cleaning constraint rules, maintenance window constraint rules, and resource consumption constraint rules. The rule set is then written into the knowledge graph storage unit.

4. The distributed filling scheduling control method based on knowledge graphs according to claim 1, characterized in that, The fact layer that forms the time-series observation sequence and synchronously updates the filling production knowledge graph includes: At the edge node, filling production data is acquired through an industrial communication interface. The filling production data includes equipment operating status, fault events, cleaning status, material arrival status and quality inspection results. A collection timestamp and data source identifier are written for each data record. Data processing is performed on the filling production data. Data processing includes missing value handling, outlier removal, data denoising, data alignment, and data aggregation. Data alignment includes resampling data from different sampling periods to a preset sampling period according to a unified time base. Data aggregation includes calculating the mean, variance, maximum value, minimum value, and rate of change within a sliding window. The rate of change is the difference between the values ​​of two adjacent sampling times divided by the value of the previous sampling time. The processed filling production data is combined into a time-series observation sequence in chronological order, and each observation vector in the time-series observation sequence is associated with the corresponding entity node and relation edge in the knowledge graph. The equipment availability status, workstation occupancy status, cleaning stage status, material inventory status, and quality judgment status are written into the fact layer of the knowledge graph.

5. The distributed filling scheduling control method based on knowledge graphs according to claim 1, characterized in that, The obtained change point event includes: At the edge node, a set of Bayesian online change point detection instances are established for each critical device. The time-series observation sequence is received, and the observation likelihood is calculated for the observation value at each sampling time. The observation probability under different running lengths is calculated. A spatial truncation update strategy is implemented for the posterior distribution of running length. At each sampling time, the posterior probability of running length is calculated and sorted from high to low according to the posterior probability. A preset number of running length candidates are retained and the posterior distribution is updated only for the retained running length candidates. The posterior distribution and observation state corresponding to the running length candidates that are not retained are discarded. The state transition matrix is ​​sparsified during the runtime transition process. The runtime transition matrix is ​​constructed and the transition probabilities of all transitions except for two types are set to zero. The first type of transition is that the runtime at the current sampling time is equal to the runtime at the previous sampling time plus one, and the second type of transition is that the runtime at the current sampling time is equal to zero. Based on the candidate runtime and the runtime transition matrix, the posterior probability that the runtime is equal to zero at each sampling time is calculated as the change point probability. When the change point probability is greater than the threshold, a change point event is generated.

6. The distributed filling scheduling control method based on knowledge graphs according to claim 1, characterized in that, The generation of the set of actionable actions and the set of executable parameters includes: The central node receives change events and writes them into the scheduling event table. The change event includes the change time, associated production line identifier, associated equipment identifier, and associated observation indicator identifier. Using the fact layer state of the knowledge graph corresponding to the moment of change as the input for reasoning, read the order batch status, formula and packaging specification status, material arrival status, equipment availability status and maintenance window status; Call the graph rules to perform semantic reasoning and complete compatibility verification, switch cleaning verification, maintenance window verification and resource usage conflict verification. Output the set of actionable actions and the set of executable parameters that pass the verification.

7. The distributed filling scheduling control method based on knowledge graphs according to claim 1, characterized in that, The output scheduling scheme includes: An improved GFlowNet model is constructed, including a reverse trajectory modeling branch, a state history embedding memory module, a distributed reparameterized path modeling structure, and a final state structure diversity discrimination module; A state and action space is constructed based on the set of actionable actions and the set of executable parameters. The state includes order pool, resource occupancy and time window information. The actions include batch selection, equipment allocation, workstation allocation and cleaning insertion. A historical state buffer is established by the state history embedding memory module, a preset number of historical state vectors are stored and encoded to obtain historical embedding vectors. The historical embedding vectors are fused with the current state vector to form an enhanced state vector. The enhanced state vector is input into the forward generation unit, which outputs the action distribution parameters. The distribution reparameterization path modeling structure represents the action sampling as a differentiable sampling process jointly determined by random noise variables and action distribution parameters. The random noise variables are sampled from the probability distribution. Based on the differentiable sampling process, action samples are generated and scheduling trajectories are iteratively formed. When the scheduling trajectory reaches the termination condition, a preliminary scheduling scheme is obtained. The reverse trajectory modeling branch takes the scheduling scheme as input to generate a reverse action sequence and outputs the reverse action probability. It uses the consistency between the probability of the forward action sequence and the probability of the reverse action sequence as the training objective to complete the reverse trajectory consistency learning. The final state structure diversity discrimination module extracts the resource occupancy sequence, line switching and cleaning sequence and workstation allocation sequence of the scheduling scheme to form a structural representation, calculates the structural difference degree between the preliminary scheduling schemes and filters and outputs the scheduling scheme.

8. The distributed filling scheduling control method based on knowledge graphs according to claim 1, characterized in that, The step of issuing the scheduling scheme to edge nodes as versioned instructions for execution and reading back the execution status includes: At the central node, the scheduling scheme is converted into a versioned execution instruction set. The versioned execution instruction set includes version number, start and end time of effectiveness, resource allocation, line switching and cleaning task sequence and maintenance avoidance time window, and writes target production line identifier, target equipment identifier and target workstation identifier for each instruction. Versioned execution instruction sets are sent to edge nodes. Edge nodes verify the version number and effective time window, parse the work sequence and generate a trigger queue in chronological order, map the instructions in the trigger queue into control instructions that the production line control unit can recognize and send them out for execution. Control instructions include equipment start / stop instructions, cycle time setting instructions, workstation switching instructions and cleaning start instructions. During the execution process at the edge node, a change point event is received and the starting point of incremental rescheduling is determined. The starting point of incremental rescheduling is the start time corresponding to the first unexecuted instruction in the trigger queue. A rescheduling request is sent to the central node, carrying the sequence of executed jobs, the current resource occupancy status, and the set of unexecuted jobs. A scheduling scheme for the unexecuted part is generated and converted into a new versioned execution instruction set and sent to the edge node.

9. A knowledge graph-based distributed filling scheduling control system, executing the knowledge graph-based distributed filling scheduling control method according to any one of claims 1 to 8, characterized in that, Includes the following modules: The knowledge graph construction module is used to map filling production objects into entities and relationships, and to write production constraints into graph rules; The edge data acquisition and processing module is used to collect and process filling production data to form a time-series observation sequence and synchronously update the fact layer of the knowledge graph; The change point detection module is used to perform Bayesian online change point detection on the time series observation sequence to obtain change point events; The graph reasoning and verification module is used to receive change point events and perform semantic reasoning and feasibility verification based on graph rules to generate a set of feasible actions and a set of executable parameters. The scheduling generation module is used to build an improved GFlowNet model, perform trajectory learning and sampling on the set of actionable actions and the set of executable parameters, and output a scheduling scheme. The instruction issuance and rescheduling module is used to convert the scheduling scheme into versioned instructions, issue them to edge nodes for execution, and trigger incremental rescheduling for the unexecuted parts.

Images